Shiplog

AI FinOps

Know what your
AI spend is buying.

Import a provider export and this page names how much of your AI spend is recoverable, what to do first, and how far to trust the number.

Bundled synthetic example

Did what we committed to work?

No track record on file: this browser is keeping no period. Importing one provider export keeps the month it covers and starts the record.

Where should we cut AI spend first, and how much monthly spend is recoverable?

Modelled recoverable AI spend · USD per month $51,254

Summed over the 5 of 5 departments carrying a completed FinOps score for 2026-06-01 to 2026-07-01. Unscored departments contribute zero and are never extrapolated to. Modelled potential, not realized savings.

Move Atlas Platform's short, low-context requests to the standard model
How we know this
Provenance
The bundled synthetic example: invented usage records for an invented company, shipped with this page. Not your spend, not customer data, and not a realized saving. No file of yours is read, uploaded, or stored to produce it.
Basis
The monthly figure is modelled recoverable spend: each scored department's recoverableUsd for the reporting month, summed and rounded once to whole USD. Unscored departments contribute zero. No annual figure is presented.
Limits
Supporting context only: Illustrative — Pricing provenance: Limited confidence (67/100) on the prices, Confidence: High on spend coverage. Rates come from the published-list reference card, not your contract. A ceiling at list prices until 2 destination models state contracted rates and committed-use. State contracted input and output rates for the premium text tier and the standard text tier — that moves this figure to Declared · Medium. Until then it stays illustrative: a list-price ceiling, not a saving. Measured throughput is left out of the arithmetic too. It is stated as of the bundled synthetic example and moves the moment a provider export of your own is analyzed. Confidence in this figure is graded High by a published rule: High from 80% of spend scored, Moderate from 50%, and High also requires at least 2 scored departments with every department read carrying spend. Nothing this rubric reads holds it lower. Pricing provenance scores 66.8 of 100 — Limited confidence. The pricing basis is directional and should be checked before committing spend. What would raise confidence: Add a citable price source, then fill the uncovered destinations and required fields. Audit inputs: Price source — Published list; weight 45%; input 0.75; contribution 33.75. Rates come from the published-list reference card, not your contract. Destination coverage — Full coverage; weight 25%; input 1; contribution 25.00. The declaration covers all 2 destinations this analysis prices. Field completeness — Partial; weight 20%; input 0.4; contribution 8.00. 6 of the 12 required declaration fields are present and well-formed across 2 destinations. Effective-date age — Undated; weight 10%; input 0; contribution 0.00. The declaration states no effective date, so how old these prices are is unknown. Total arithmetic: 33.75 + 25.00 + 8.00 + 0.00 = 66.75; rounded = 66.8
Coverage detail is computed locally when the analysis loads.

Attested finops-recoverable-attestation/1.0.0 — the headline above is the one figure this region states, on the monthly basis of record · confidence medium · provenance 5 of 7 operands declared by the export, 2 derived here · coverage 5 of 5 departments scored. Each carries its assumption in tests/fixtures/finops-consolidated-answer-attestation.json, and a drift in any of the four fails by naming it.

Analysis readiness

Can I get a trustworthy savings recommendation now?

Why this is the answer

Reading Resolving one annual figure, the benchmark behind it, and the action it implies.

Resolving the annual figure from the analyzed scenario…

Bundled synthetic example Every figure in this briefing is modelled in this browser from invented provider-export records — not your spend, not a realized saving.

Analyzing the bundled provider export selected in the chooser below.

Checking that figure against the analysis benchmark…

The answer is resolved locally, from the analysis this page already loaded.

Provenance is stated with the figure, from the analysis's own signal names.

Yes for an illustrative scenario recommendation; no for an organization-specific savings commitment.

Readiness is calculated from four required evidence categories.

Current evidence is limited to bundled synthetic scenarios.

No conclusion about your organization is supported.

The prioritized illustrative action is read from the bundled scenario's own next-step record.

Ranking the recommended actions…

Modelled recoverable line
Read from the bundled scenario's own next-step record.
Action confidence
Read from the quoted fixture action.
Why this one
Read from the page's one derived next-step record.
Provenance
Bundled synthetic fixture.
Evidence confidence
Computed from the evidence categories the bundled scenario carries.
What later evidence would enable

    Bundled synthetic example · nothing of yours needed

    Evidence behind the recoverable AI spend answer

    Bundled synthetic example Illustrative — invented data for an invented company, not your spend, customer data, or realized savings.

    Illustrative only, from hand-authored synthetic cohort boundaries rather than your own export: Your AI spend is in the most expensive quarter of organizations like yours, at $38.63 per successful task for June 2026. $51,254 of that is modelled as recoverable. Atlas Platform is driving the increase.

    Recoverable spend

    $51,254 · 33% of analyzed spend

    $51,254 of $154,500 analyzed. A modelled ceiling on what re-routing this work could save — not money already saved. Modelled, not graded: the rubric has scored $143,500 of the $154,500 in scope, and this figure is taken over all of it.

    Graded floor · scored departments only

    Rank against similar organizations

    Peer comparison unavailable

    Benchmark fit has not been evaluated. Synthetic cohorts are privacy-preserving and do not imply access to customer, provider, or HRIS data.

    How we know this
    Provenance
    Every figure in this answer — here and inside each layer below — was computed from data invented for a company that does not exist. It is not your spend, customer data, or realized savings, and no file is needed to read it. Import a provider export and the same slots are recomputed in this browser tab from your own file, which is never uploaded.
    Basis
    Recoverable spend is scored per department against this page's published recoverability rule and summed; the rank is this organization's cost per successful task placed on the cohort boundaries the comparison ships with. Which scoring rules and which peer data produced each figure is named, with versions, in the disclosures below this one.
    Limits
    The cohort boundaries are hand-authored: they are in no export file, yours or the example's, so the rank is a position on this page's own scale rather than a measurement of a market. Coverage, the prompt grade and the unclassified remainder are reported only once something has been read; until then no figure here claims a position.

    Department driving the increase

    Atlas Platform

    Atlas Platform contributed +$34,500 of the +$39,200 change (88% of it).

    Open the recommendation evidence

    Do this first, before any spend cap is set. What was actually measured, and which input was not checked? Promoted to rank 1 by clause unchecked_basis under destination-priority/1.0.0: the finding's confidence carries one limit, so the basis is checked before a spend cap is committed.

    Is the trend a one-off? See the period-over-period comparison

    Is our AI spend classification trustworthy enough to act on?

    Spend we can stand behind92.9% of spend in scopeThis is the share of spend classified reliably enough to act on; a higher share is better.as of Bundled synthetic example · nothing of yours needed · June 2026

    Coverage: $143,500 of $154,500 of spend in scope sits in departments the rubric scored. Grade: high coverage tier — At least 80% of imported spend sits in departments the rubric scored. Residue: $11,000 of that spend has no scored query, the largest single block of it in Ember Studio. All of it as of Bundled synthetic example · nothing of yours needed · June 2026.

    Record and verify a savings commitment — go to Act and verify

    Scope: the share of in-scope spend the rubric scored. Whether this analysis is ready to circulate is a separate question, answered below.

    Circulation decision · bundled analysis only

    Is this bundled analysis ready to circulate?

    Ready to circulate

    Scope: the finding, comparison, and first action below. It is not a claim about how much spend the rubric scored.

    1. Finding

      Atlas Platform is the first intervention priority because over-provisioning is its largest recoverable cost line.

    2. Comparison

      Atlas Platform scores 71, 10 points above the synthetic enterprise SaaS cohort median of 61 under rubric literacy-mix/1.0.0.

    3. First action

      First, in Atlas Platform: enable automated down-routing for short, low-context prompts. The bundled example ranks it first on $10,650 of recoverable_spend_usd over 25 Jun–25 Jul 2026 — a modelled figure from invented records, not realized savings.

    High confidence — 760 sampled queries, current through the source-period end; synthetic evidence only

    Source period 2026-06-25 to 2026-07-25 — hand-authored bundled provider aggregates and org mapping; not live provider or HRIS data

    Prompt text is excluded from this circulation decision. Provider-level rows are excluded; only bundled synthetic aggregates are in scope.

    Executive briefing payload

    Briefing claims generated locally

    Open or share this exact printable briefing

    Inspect client-side payload and audit appendix 
    
            
    
          
    
        
    
        
        
        
        
        
        
    
          Hand-authored synthetic cohort boundaries
          
          Moderate confidence
        
    
        
        
        
        
        
    
          
    
            
            
    
          
    
          
    
            
            
    
          
    
          
    
            
            
    
          
    
          
          
    
            
            
    
          
    
          
    
            
            
    
          
    
          
    
            
            
    
          
    
          
    
            
            
    
          
    

          
          
    
            Evidence layer If I only do one thing this month, is this the right one?
            
            
    
              
    
                
    This month · one next step
    
                
    Where do I start this month?
    
              
    
              
              
    Bundled synthetic example Invented review, action, and verification records for an invented company. Not your spend and not realized savings.
    
              
    
                
                
    This month's one step is composed from what this browser holds, and from the bundled synthetic example when it holds nothing.
    
              
    
            
    
          
    

          
          
    
            Evidence layer If I take that step, what will tell me it actually worked?
            
    
              
    
                
    This month · evidence, action, and the checkpoint
    
                
    What do I do next, and what will tell me it worked?
    
              
    
              
              
    Bundled synthetic example Invented review, action, and verification records for an invented company. Not your spend and not realized savings.
    
              
    
                
                
    This step and its checkpoint are composed from what this browser holds, and from the bundled synthetic example when it holds nothing.
    
              
    
              
            
    
          
    
        
    
        
      
    

      
      
      
      
      
    
        
    
          
          
     Example result
    
          
          
    How much of our analyzed AI spend is recoverable?
    
        
    
        
    Bundled synthetic example Illustrative — invented data for an invented company, not your spend, customer data, or realized savings.
    
        
        
    33% of analyzed AI spend is recoverable AI spend
    
        
    $51,254 in recoverable AI spend for the reporting period
    
        
        
    
           How we know this
          
    
            
    Provenance
    Every figure below comes from six months of invented records for a company that does not exist, bundled with this page. It is not your spend, customer data, or realized savings. No file is needed. The prompt sample beside it is the same invented company's conversation corpus, read through the same parser a provider export of yours would go through.
    
            
    Basis
    Each figure carries its own working on the line it is read: the marker beside it says whether this page derived the value or could not find it in the files, and opens to the arithmetic. Unmarked values — department names, the reporting period, the record counts — are read straight out of the example's export files. The prompt grade is the published rubric literacy-mix/1.0.0, weighted by each department's spend.
    
            
    Limits
    Illustrative only. The quartile is read off hand-authored synthetic cohort boundaries: those boundaries are in no export file, not the example's and not yours, so the rank is a position on this page's own scale. The routing figure is a modelled ceiling rather than an invoice line, the accountable role is assigned by rule and named in no column, and 4 of 5 departments carry scored prompts — Ember Studio carries none, so it sits outside both the grade and its coverage figure.
    
          
    
        
    
        
        
    
          
          
    
            
    Where this came from
    A ratio this page computes: $51,254 of recoverable spend over $154,500 of analyzed spend, both summed from the 15 invented usage records, then rounded once to a whole percent. No column in the example's files carries a recoverable share.
    
            
    What it moves
    It sizes the prize before anyone commits time to it. A share this large is worth a pilot; the same share of a tenth of the spend is not.
    
          
    
        
    
        
    Where each figure came from: unmarked values — the department names, the reporting period, and the record counts — are read straight out of the example's export files. Every marked figure below says whether this page derived it or could not find it, and opens to show the working.
    
        
        
    
          
    Correct a derived name or figure
    
          
    The answer: 33% of analyzed AI spend is recoverable AI spend. 14 of the 16 names and figures below were derived on this page rather than read from the example's export files.
    
          
          
    Confidence in this brief is bounded by how much of it was derived: 88% of it still is (14 of 16). Nothing below has been corrected by a reader yet.
    
        
    
        
        
    
          
    
            
    Potentially recoverable share
    
            
    33% of analyzed AI spend
    
            
    $51,254 of $154,500 analyzed · 5 invented departments · 2026-06-01 to 2026-07-01 (end exclusive).
    
            
    
              
              
    
                
    Where this came from
    The same ratio as the answer above, shown against the figures it divides. The $154,500 denominator is the sum of the cost column across all five invented departments for 2026-06-01 to 2026-07-01; the $51,254 numerator is scored per department by the published recoverability rule, not read from any column.
    
                
    What it moves
    It is the number to hold your own export's share against. If yours comes back far lower, the ceiling below is smaller too.
    
              
    
            
    
          
    
          
    
            
    Potential impact · routing scenario
    
            
    $51,254 in the reporting period
    
            
    A modelled routing scenario over invented departments — a ceiling on what re-routing could recover, not realized, invoiced, or promised customer savings.
    
            
    
              
              
    
                
    Where this came from
    A modelled routing scenario, not an invoice line. Each department's baseline spend is multiplied by its normalized waste share and that category's recoverable share, then rounded once — the arithmetic the example data publishes beside every department.
    
                
    What it moves
    It is the ceiling on what re-routing could return, so it caps what the pilot is allowed to cost before it stops being worth running.
    
              
    
            
    
          
    
        
    
        
        
    
          
    
            
    Rank against similar organizations
    
            
    Bottom quartile
    
            
    Bottom quartile · $38.63 per successful task
    
            
    Synthetic comparison against an invented peer cohort; lower cost per successful task is better. This is not customer performance or realized savings.
    
            
    
              
              
    
                
    Where this came from
    $38.63 is analyzed spend divided by the successful-task count in the same invented records. The quartile it lands in is read off hand-authored synthetic cohort boundaries — those boundaries are not in the example's export files and no reader's export contains them either.
    
                
    What it moves
    A bottom-quartile cost per task says the problem is efficiency rather than volume, which is what makes routing the first move instead of a headcount conversation.
    
              
    
            
    
          
    
          
          
    
            
    Internal drill-down · widest department gap
    
            
    A full band behind
    
            
    Atlas Platform is a full band behind Boreal Support on cost per successful task.
    
            
    Synthetic cost-per-successful-task comparison across the invented company's own departments — Boreal Support, Cinder Research, Quartz Analytics, Ember Studio, Atlas Platform — not customer performance or realized savings.
    
            
    
              
              
    
                
    Where this came from
    Both department names are read straight from the files. The gap between them is computed: cost per successful task per department, placed on the same band boundaries as the peer position above, and the widest surviving pair is the one named.
    
                
    What it moves
    It names the department to pilot in. A gap inside your own organization is one you can act on this month without waiting for a cohort to agree with you.
    
              
    
            
    
          
    
          
          
    
            
    AI literacy · graded prompt sample
    
            
    B · 85 of 100 · literacy-mix/1.0.0
    
            
    $143,500 of $154,500 in-scope invented spend was scored — high coverage, 4 of 5 invented departments, 141 of 144 synthetic prompts classified. Not graded: Ember Studio. At least 80% of imported spend sits in departments the rubric scored. Synthetic prompts from an invented company, scored by a published rubric — not customer behaviour and not realized savings.
    
            
    
              
              
    
                
    Where this came from
    141 of 144 invented prompts were classified and scored by the published literacy-mix rubric, then weighted by each department's spend. The letter is the rubric's output; the $143,500 of $154,500 coverage figure beside it is the spend those scored departments account for. Ember Studio carries no prompts in the corpus, so it is outside both figures.
    
                
    What it moves
    It says whether the spend is being wasted on how people ask rather than on which model answers — a B over 93% of the budget means routing, not training, is the cheaper first move.
    
              
    
            
    
          
    
        
    
        
    
          
    Recommended action · rank 1
    
          
    Pilot lower-cost routing in Atlas Platform, the top-spend invented department. Cap the pilot at $51,254, then compare it with a similar period.
    
          
          
    Accountable role: Platform Engineering Lead
    
          
          
    
            
            
    
              
    Where this came from
    The example's export files carry no owner, team-lead, or cost-centre column, so nobody accountable for Atlas Platform's spend could be read out of them. "Platform Engineering Lead" is the role this page assigns by rule to a routing pilot — it is not a person, a title, or a mapping named anywhere in the data.
    
              
    What it moves
    Nothing gets piloted until someone owns it. Check this role against your own org chart before the meeting, because your export will not fill it in either unless it carries an owner column.
    
            
    
          
    
        
    
        
        
    
          
    Confidence in this answer
    
          
    0.85 of 1.00 · moderate
    
          
    Coverage 1.00 — all 15 bundled example records were analyzed and all four required aggregate inputs were present — less 0.15 because the rank-1 routing candidate's call shape could not be verified. Coverage asks how much of the data was read; this score also asks how much of the recommendation was checked, which is why it sits a band below coverage.
    
          
    
            
            
    
              
    Where this came from
    Scored by this page, not read from anything: all 15 invented records were analyzed, which is what holds the figure up, and the score is then reduced because the routing call shape the recommendation depends on could not be verified in the records.
    
              
    What it moves
    It is the gap between quoting this number and standing behind it. At 0.85 the direction is safe to circulate; the exact dollar ceiling is not, until the call shape is confirmed.
    
            
    
          
    
        
    
        
        
    
          
    Estimated position from declared facts
    
          
    Estimated · $39.81 per successful task · Bottom quartile
    
          
    Modelled recoverable range $20,772–$49,854 a month against Enterprise · 2,000+ employees · Software-as-a-service (most expensive quarter of the cohort). Modelled · every declared fact used as given. Estimated from declared facts and published assumptions — not measured from your usage, not verified against an invoice, and not a realized saving.
    
          
    
            
            
    
              
    Inputs used
    $154,500 declared monthly AI spend · 100 declared engineers · provider mix frontier 45% · standard 40% · economy 15% · cohort Enterprise · 2,000+ employees · Software-as-a-service.
    
              
    Tasks attempted
    100 engineers at 49 attempted tasks each per month = 4,900 attempted tasks.
    
              
    Tasks that succeed
    The declared mix weights the published tier success rates (frontier 86%, standard 78%, economy 62%) to 79.2%. 4,900 attempted at 79.2% = 3,881 successful tasks, rounded once to whole tasks.
    
              
    Cost per successful task
    $154,500 ÷ 3,881 successful tasks = $39.81.
    
              
    Cohort basis
    Published cohort cost-enterprise-saas (Enterprise · 2,000+ employees · Software-as-a-service), snapshot 2026-06-30: p25 $18.40, p75 $31.50. Lower cost per successful task is better.
    
              
    Recoverable range
    Headroom to the cohort's cheap boundary is $154,500 less the $18.40 p25 rate over 3,881 successful tasks = $83,090. At 25%–60% of that headroom: $20,772–$49,854 a month, modelled.
    
            
    
          
    
        
    
        
        
    
          
          
    
            
    Inputs
    Invented provider and org exports generated in this tab. No company, account, provider, or person in them is real, and no customer or telemetry data is read.
    
            
    Limits
    A routing scenario is a modelled ceiling, not a realized saving. Your own export will produce different numbers.
    
          
    
        
    
        
        
        
    
          
          
    
            
            
            
    Fills every panel below with six invented months. No file is needed.
    
          
    
          
    
            
            
    Choose your export files. They stay in this browser and are not uploaded.
    
          
    
        
    
        
        
    
          
    Read this example as an Executive FinOps briefing
    
          Open the Executive FinOps briefing for this example
          
    Opens the printable one-page sheet on this same Bundled synthetic example, with the same recoverable share and scenario. Invented figures, not your spend; nothing of yours is read, uploaded, or stored.
    
          
          
    Forward this address to open the figure where it is stated: /evolution.html#workspace-answer
    
        
    
        
        
    
          
    
            Estimate from five facts you already know
            
    
            
    
            
    
            
            
    
            
    Expected columns, exactly two: team_role (the role or department bucket, counted when it reads "engineering") and period (the month, as YYYY-MM). Column order does not matter.
    
            
    Refused columns: every other column, including name, first_name, last_name, email, employee_id, salary, compensation, manager, location, birth_date. An unlisted column is refused on sight, never dropped, so re-export with those two columns only. The file is read in this tab — nothing is uploaded, no account is needed, and no copy is kept.
    
            
    No roster chosen. The headcount above is the one you typed.
    
            
    
            
    
          
    
          
    
        
    
        
    
          
    Estimated: we spend $39.81 per successful AI task — Bottom quartile, the most expensive quarter of the cohort. Lower cost per successful task is better.
    
          
    Estimated recoverable: $20,772 to $49,854 a month, modelled from these five declared facts — not a measured, invoiced, or realized saving.
    
          
    Next: move one month of standard-tier traffic that does not need the tier to a cheaper one, then check the invoice against this estimate before booking anything.
    
        
    
        
        
    
          
    How far this brief has been checked
    
          
    Rung 1 of 3 · Declared · you are hereRung 2 of 3 · EstimatedRung 3 of 3 · Verified
    
          
    A declared fact supports a shared starting point — what your org says is true of itself, written down where everyone can see it. It does not support a figure of yours, because nothing on this screen has been modelled from your answers yet: the five facts above are still the bundled example's.
    
          
    One step up: answer the five facts above and press "Estimate from these five facts" to reach Estimated.
    
        
    
        
        
      
    

      
      
    
        
    What checks support the recoverable AI spend answer?
    
        
    No readiness benchmark has been taken for this analysis yet.
    
        
    Nothing to guide yet: no analysis has been read.
    
        
    The claim above and its inputs are quoted here once the analysis has been read.
    
        
        
    
           Every step, and what the last one moved
          
    
            
      
            
        
          
        
        
        
        
      
        

      
      
      
      
        
        
        Where to go next · 3 destinations
        
        
        Recoverable spend this quarter $16k 
        
        
        
          
        Impact
        Monthly saving from the highest-ranked lever: $5,200 / month. That lever is what the figure above is recoverable through.
        
          
        Confidence
        medium — at least one contributing line item is attributed by allocation rule rather than read off an invoice line.
        
          
        Provenance
        bundled synthetic example, as of 2026-07-01. All figures are synthetic demo data.
        
        
        
        
        
        
          
          
        
            
        How $16k was computed
        Recoverable spend this quarter sums the quarter-to-date cost of every workload flagged as addressable by at least one shipped lever. One qualifies, at $5,200 a month: 5,200 x 3 = 15,600, rounded once after the arithmetic and shown as $16k. It is a subset of spend, never a projection, and never spend already recovered.
        
          
        
        
        
      
        

      
      
        
        
        
        Evidence for the headline figureConfidence and source are not available until the Bundled synthetic example is prepared.
        
        
        
          
        Next step · one of three doors
        
          
        Where do I go next to act on this?
        
        
        
        
        Ranked from the Bundled synthetic example The order was computed from six months of invented data, not from your spend.
        
        
        
          
        No destination is ranked from the Bundled synthetic example.
        
        
        
      
        

      
      

      
      
        
        
        Working area
        
        
        One destination is open at a time. The answer above stays available while you work.
        
        
        
        
          
        Now showing
        
          
        The answer
        
          
        Is our AI spend classification trustworthy enough to act on?
        
        
        
        
        
      
        

      
      
      
      
      

      
      
        
        
        Which bundled provider export has the most recoverable AI spend, and who should act on it first?
        
        
        Local synthetic demonstrations. No real data is entered, uploaded or transmitted — every figure is computed locally in this browser from invented records.
        
        
        
          
        
            
        Analysis pendingWaiting for the page result
        
            
        The finding and action will appear here.
        
          
        
        
        
        
      
        
      
        
        
        Evidence · chosen scenario
        
        
        
      
        
      
        
        
        Department detail · chosen scenario
        
        
        The guided department detail moved: open this department’s own screen
        
        
        
      
        

      
      

      

      
      

      
      
        
        
        AI literacy
        
        
        
        
        
        
        
        Score not available yet
        
        
        
        
          
        
            
        Coverage · share of spend graded
        
            
        Sampled-spend coverage not available yet
        
          
        
          
        
            
        Confidence in this letter
        
            
        Not established yet · no scored sample read
        
          
        
        
        
        
        
        Under review · no letter published
        
        
        
        Why this letter: no scored sample has been read, so no driver can be named yet.
        
        
        
        
        Action not available yet
        
        
        
        
          Per-department scores and the rubric · no departments read yet
          
          
            
        • No department has been read yet.
          • 
          
        
          
        
        
        
        Fills in once a scored query sample is read.
        
      
        

      
      

      

      
      

      
        
        
        
          
        
            
        Bring your own exports · browser only
        
            
        Analyze provider spend without uploading it.
        
            
        One provider export is enough to get a FinOps briefing. Add a second, equal-length period to compare spend across time, and an org mapping to put your own department names on it. Three steps: choose files, check the mapping, read the briefing.
        
            
            
          
          
          
          
            
            Bundled synthetic example
          
        
          
        
        
          
          
          Start here · one question
          
          
          Which provider do you pay?
          
          
          
            
            
            
          
          
          
          
            
          Choose the provider you pay above. This panel then names the one report to pull, hands you a starting file for it, and gives you one next step. Nothing is uploaded either way.
          
          
          
          
        
          
        
        
          
          
          Contracted rates · synthetic, this tab only
          
          
          What do you actually pay per model?
          
          
          Declared rates are synthetic and held locally only: nothing is uploaded, no credential or contact detail is accepted, and no live provider connection is made. One record per line — model, unit, rate, effective date, source label — or paste a JSON array of the same five fields. Accepted units: usd-per-million-input and usd-per-million-output. A line that cannot be read is handed back with every other problem in the paste, and none of them are applied.
          
          
          
            
          
            
          
          
          
          
          No contracted rates declared. The figure above stays priced at the published list.
          
          
          No contracted rate is declared for a destination this analysis prices, so every one of them is priced at the published list — a ceiling, not your contract.
          
          
          
          
          
          Nothing is retained in this browser.
          
        
          
        
          
          
          
            
          Start with one file
          
            
          Use your provider export for this decision. One provider period export — CSV, TSV, or a v1 JSON envelope — is validated and analyzed in this tab. You do not name the provider: which console the file came from is read from its own columns, and if it cannot be read the panel below says which importer it is nearest to and what was missing.
          
            
          The Bundled synthetic example is active. Choose your provider export to answer this question with your own spend.
          
          
          
          
          
            
            
          
          
        
          
        
        
          
          Your files do not leave this tab.
          They are read and analyzed here, in this browser. No upload · no credentials · no network transfer · no browser storage · results disappear on refresh. Column headers, one sample value per column while you check the mapping, and totals are shown; no other cell value is ever rendered, announced, or exported.
          
          Prompt text is never one of them. A conversation or audit export is parsed in memory only to count its characters and estimate its tokens. The prompt itself is never rendered on this page, never exported, never stored in this browser, and never uploaded — see the conversation-export contract.
        
          
        
          
          Which native provider exports are compatible?2 versioned adapters
          
          
            
          Compatibility contract is available when this client-side demo starts.
          
          
          
          
          Full static contract, normalization rules, fixtures, and failure behavior: native provider export contract v1.
          
        
          
        
        
          
          
          Bedrock, Vertex AI, Azure OpenAI · browser only
          
          
          Will your hyperscaler export work here?
          
          
          Check a bundled example against the published contract without a credential, a connector, or a single network request. Your own export does not belong here: drop it into the one import above and it is recognized from its own content.
          
          
          
            What each provider contract requires3 versioned contracts
            
          Contracts are available when this client-side demo starts.
          
            
            
          
            
          
            
            
            
            
            
            
          
            
          
            
            
            No export has been checked yet. Choose a bundled example or one of your own export files, then run the check.
            
          
            
        
            
        
        
            
          
            Recognition score · reproducible, local, browser only
            
          
            How confident is this workspace that it recognized the export?
            
          
            Every bundled example below is a labelled fixture with one expected score. Choosing one recomputes that score in this tab from the published provider contracts: the same example always produces the same number, and every point of the number is listed under it.
            
          
            
          
            
            
            The recognition score for the selected bundled example appears here when this client-side demo starts.
            
          
            
        
            
        
        
            
          
            Your own export · read in this tab
            
          
            Which model in your export costs the most per thousand billed units?
            
          
            Read a bundled example below and the answer, the confidence behind it, and the one thing to do about it appear together. Your own export goes through the one import at the top of this panel, which recognizes the provider from the file itself.
            
          
            
          
          
            
          
            
            
            No export has been read yet. Choose the provider you exported from, then drop a file here, pick one with the file control, or read a bundled example.
            
          
            
          
              
          
              
            How that figure was reached
            
              The rate table and the recognition evidence appear here once an export has been read.
              
          
              
        
              
        
        
        
              
          Evidence layer Can I prove to a security reviewer that my export never left this browser?
          
              
            
              What happens to the file you choose
              
            
              
              
              Processed locally
              
              
              Read with the browser's own File API, then parsed, mapped, joined, and scored by scripts in this tab. The analysis needs no account, no key, and no connection.
              
              
              Stored
              
              
              Nothing, unless you save a review. The rows are held in memory and are gone on refresh. Saving writes two localStorage keys on this origin: shiplog.finops.workspace.v1 (consent, periods, commitments) and shiplog.finops.journey-snapshot.v1 (derived figures and references — never raw rows or prompt text). Clear them with Discard all files and results above, or by clearing this site's data.
              
              
              Never transmitted
              
              
              Your file, any cell in it, any prompt text, any figure derived from it. The import path opens no connection of any kind — no upload, no fetch, no beacon, no socket, no form post — and the page is served with connect-src 'self', so the browser refuses cross-origin requests from it.
              
            
              
            
              Enforced, not promised: a test drives this import with every browser transport replaced by a recorder and fails the build if one is called.
              
          
              
        
              
        
        
        
        
                
        
                
          
          
                
            
                 Import your provider export
                
            
                Drop the file anywhere on this page, or browse for it below. Any supported provider export is recognized from its own content — you are never asked which console it came from.
                
            
            
                
            
                
            
            
                This import reads two files, and each one answers a different half.
                
            
                  
              
                • 
                
                  Spend export · Not imported yet
                  
                
                  
              
                  • 
              
                • 
                
                  Conversation export · Not imported yet
                  
                
                  
              
                  • 
            
                
            
            
                
              
                Which file to pull, and which half it answers
                
              
                
              
                  
            
                  
            
            
            
                  
              
                  Ask your billing owner for one of these five exports.
                  
              
                  
              
                    
            
                    
          
                    
            
            
            
                    Required: at least one provider period export, as CSV, TSV, or a v1 JSON envelope. Optional: an org mapping for your own department names, a query sample, and a Shiplog delivery-history JSON export if the releases you want spend compared against live in another install. Choose them together or add them in batches; every provider period must come from the same source.
                    
            
            
                    
            
            
            
          
                    
            
                    
          
                    
          
          
          
          
          
          
                    
            Where to get a provider export — what to ask for, and what to take out first
            
                    
            
                    
            
                    The full package contract, including the unsupported-package path and what happens to partial, stale, malformed, and reordered data: provider export package contract.
                    
          
                    
          
          
                    
            Which query sources can be read here — and which cannot
            
                    
            
                    
              
              
            
                    
            
                    
            
                    
            
                    Not read here, and what to bring instead:
                    
            
                      
            
                      The full source contract, including attribution, bucketing, sampling bounds, and partial, stale, malformed, and reordered behaviour: organizational query-source contract.
                      
          
                      
          
                      
            Other ways to start — templates, example exports, and a saved FinOps briefing
          
          
          
          
          
          
                      
            
            
            
            
                      Invented sample prose only. Downloads from this tab; nothing is uploaded.
                      
          
                      
          
          
                      
            
            
            
                      This is not the import. It reopens a briefing this page already wrote: choose a FinOps briefing JSON file this page exported. It is read in this browser only — nothing is uploaded, and nothing is saved after you close this tab. It opens read-only below the current briefing and never replaces it.
                      
            
            
          
                      
          
                      
          
          
        
                      
        
        
                      
          
                      Your track record
                      
          
                      The whole record lives in this browser, on this device, and nothing about it is uploaded. Exporting writes a file this tab builds; importing reads a file you choose, in this tab. Neither one sends anything anywhere.
                      
          
                      No period is kept in this browser yet.
                      
          
                      
          
                       
                      
          
                      
        
                      
        
                        

        
        
        
        

        
        

        
        

        
        

        
        
        
        
        
        
        
        
        
                        
          Ready for local files.
          The Bundled synthetic example stays on screen until your first provider export is ready. Its invented figures are not your spend.
        
                        
        
        
        
        

        
        
        
                        
          
                        Want to go through this briefing — or your own numbers — with a person?
                        
          
          
                        Only the work email address you type is submitted. No imported file, figure, column value, department name, or prompt text is attached.
                        
        
                        
        
        

        
        
      
                        

      
      
                        
        
                        
          
                        Coach one prompt · browser only
                        
          
                        Would a model answer one prompt well?
                        
          
                        Grading a single prompt has its own page. Paste a prompt, or the few turns around it, and grade it against the same rubric this page grades an imported corpus with. The bundled synthetic example is graded there on arrival, so a real result is readable before anything is pasted, and the text stays in that tab: nothing is uploaded, stored, or attached to your organization’s grade.
                        
        
                        
        Open the prompt coach
      
                        

      
      
                        
        
                        
          
                        Next step
                        
          
                        Questions about this result?
                        
          
                        Everything above is computed in this tab. If you would rather go through these numbers with a person, request a Shiplog follow-up: submitting sends one thing, the work email address you type.
                        
        
                        
        
        
      
                        

      
                        
        
        
                        
          Illustrative figures · invented sample
          These figures use invented example data. They are not your spend or realized savings.
        
                        
        
                        
          
                        Example recommendation
                        
          
                        Route short, low-context requests to the standard model.
                        
          
                        Use a bounded gateway rule while leaving high-value prompts unchanged.
                        
        
                        
        
                        
          
                        Monthly baseline
                        $7,430
                        
          
                        Projected savings
                        $5,200 / month
                        
          
                        Accountable role
                        Core Services platform director
                        
          
                        Confidence
                        High · 760-query scored sample
                        
        
                        
        
          Inspect rubric and evidence 
        
      
                        

      
      
                        
        
                        Your grade
                        
        
        
                        Is this grade yours?
                        
        
        
                        
          Prompts of yours behind this grade
          0 imported · 25 needed in one department
        
                        
        
                         Import a prompt export to grade your own departments. Until then every panel shows the Bundled synthetic example.
                        
        
        
                         Bundled synthetic example
                        
        
        
        
                        
        
                        
        
        
      
                        

      
      
                        
        
                        Classifier agreement
                        
        
                        How often does the classifier agree with a human reviewer?
                        
        
                        Scoring the labelled sample…
                        
        
                        
        
                        
          Where it disagrees, class by class
          
                        
        
                        
        
                         Recompute it: the labelled queries are in the published corpus, which states who labelled them and how, and they are scored by src/finops-classifier-agreement.js. 
                        
      
                        

      
                        
        Supporting panel How does this grade compare — against a cohort, and against the team that needs coaching?
      
                        
        
                        
          
                        Cohort comparison
                        
          
                        A local import carries no peer organizations, so the comparison stays the bundled cohort's.
                        
          
                         Bundled synthetic example
                        
          
          
          
                        
            Comparable-peer method and eligibility
            
                        
              
                        The full comparable-peer method — the published contract finops-peer-cohort/1.0.0 and its synthetic reference snapshot, the eligibility rule, the three ranked metrics, the percentile and quartile arithmetic, and the trust labels behind the single prioritized action — is loaded when this panel is opened. It contains no customer data and is never changed by an import. If it does not appear, read it directly: the comparable-peer method record.
                        
            
                        
          
                        
        
                        
        
                        
          
                        Which team needs coaching
                        
          
                        The Bundled synthetic example's lowest-scoring team, drawn from invented prompts.
                        
          
                         Bundled synthetic example
                        
          
        
                        
      
                        
      
                        

      
      

      
      

      
      

      
      

      
      

      
      

      
      
                        
        
                        Your plan
                        
        
                        Which of these moves are you committing to, and at what scope?
                        
        
                        $0 planned: no move has been committed at a stated scope.
                        
        
                        
      
                        

      
      
                        
        
                        Routing slate
                        
        
                        Which model routing changes should we ship this month?
                        
        
                        No analysis has been read yet, so there is no routing change to rank.
                        
        
                        
        
        
                        
        
        
                        
      
                        

      
      
                        
        
                        Routing score
                        
        
                        Did last period's routing rules return what they said they would?
                        
        
                        Nothing has been committed in this browser, so there is no prior period to score a routing policy from and no follow-up period it is answerable for.
                        
        
                        
      
                        

      
                        
        Supporting panel What did we spend this period, and how much of it is recoverable?
      
                        
        
        
                         Bundled synthetic example
                        
        
        
        
        
                        
          
                        AI spend · period
                        
          
          
                        –
                        
          
                        Across all providers
                        
        
                        
        
                        
          
                        Recoverable spend
                        
          
          
          
          
                        –
                        
          
                        Down-routing, training, and leakage
                        
        
                        
        
                        
          
                        High-value share
                        
          
          
                        –
                        
          
                        Of scored spend
                        
        
                        
        
                        
          
                        Peer position
                        
          
          
                        –
                        
          
                        Published synthetic peer cohort
                        
          
          
          
          
          
          
          
          
          
          
          
          
        
                        
      
                        
      
                        

      
                        
        Supporting panel Which department needs help first, and how is it trending?
      
                        The department priority drill-down moved: open this department’s own screen
                        
      
      
                        
        
        
                        
          
                        Decision 1 of 3 · intervention
                        
          
                        Which department needs help?
                        
          
                        Lowest eligible performance score first. Unavailable samples are not scored or ranked as poor performance.
                        
        
                        
        
                        Sample provenance not available yet
                        
        
        
                        Open this department’s own screen
                        
        
                        
          
                          
            
                        1. Department ranking not available yet
                          2. 
          
                        

          
                        
            
                        
              
                        
                
                        Selected department
                        
                
                
                        Department result not available yet
                        
                
                        Department results will be available after the Bundled synthetic example is prepared.
                        
              
                        
              
                        –
                        
            
                        

            
                        
              
                        
                
                        
                  
                        Priority 01 · recommended intervention
                        
                  
                        Recommended intervention not available yet
                        
                
                        
                Recommendation not available yet
              
                        
              
                        A recommended intervention will be available after the Bundled synthetic example is prepared.
                        
              
                        
                
                        Expected impact
                        Not estimated
                        
                
                        Confidence
                        Not assessed
                        
                
                        Accountable role
                        Unassigned
                        
                
                        Provenance
                        Bundled static fixture
                        
              
                        
              
                        
                
                        Baseline—
                        
                
                        Target—
                        
                
                        Estimated savings—
                        
                
                        Simulated realized—
                        
              
                        
              
                        Why this action: Diagnosis not available yet
                        
            
                        

            
                        
              
                        Decision 2 of 3 · trajectory
                        
              
                        Is cost/performance worsening?
                        
              
                        Comparison not available yet
                        
              
                        
            
                        

            
                        
              
                        Decision 3 of 3 · comparator & evidence
                        
              
                        How does it compare with the peer benchmark?
                        
              
                        Peer comparison not available yet
                        
              
                        
              
                        
                Inspect supporting scored evidence
                
                
                          
              
                          
            
                          
          
                          
        
                          
      
                          
      
                          

      
                          
        Supporting panel Where does the money go — what was the spend actually asked to do?
      
                          The department spend mix moved: open this department’s own screen
                          
      
                          
        
                          
          
                          Quality engine
                          
          
                          Where the money goes
                          
          
                          Spend split by what the prompt was actually asking for, not by token count alone.
                          
        
                          
        
        
        
        
        
        
                            
      
                            
      
                            

      
                            
        Supporting panel Are projected savings turning into verified savings?
      
                            
        
                            
          
                            
            
                            Finance leader action portfolio
                            
            
                            Start with the highest-priority savings opportunity.
                            
            
                            The ranked recommendation leads; related findings stay available as supporting evidence.
                            
            
                            Open monthly Savings Action Center
                            
            
            
                            Read the one-page Executive FinOps briefing
                            
          
                            
          
                            
            
                            
              
              
              
            
                            
            
                            
              
              
            
                            
          
                            
        
                            
        
                            
          
                            Projected savings—
                            
          
                            Completed savings (includes verified)—
                            
          
                            Verified savings—
                            
        
                            
        
        
                            No actions counted yet
                            
        
                              
          
                            1. 
            
                              No action portfolio yet
                              
            
                              Fills in once the action lifecycle is read.
                              
          
                              2. 
        
                            
      
                            
      
                            

      
                            
        Supporting panel How did a recommendation earn the score it was given?
      
      
                            
        
                            
          
                            Defensible evaluation · static fixture
                            
          
                            Inspect how a recommendation earns its score.
                            
          
                            Every point maps to a labelled criterion, evidence statement, and documented weight assumption. Unsafe content cannot be averaged into approval.
                            
        
                            
        
                            Bundled synthetic fixture · deterministic rubric · no stored prompts, credentials, customer data, or live provider calls.
                            
        
        
                            
          
                            No fixture scores read yet
                            
        
                            
      
                            
      
                            

      
                            
        
                            
          
                            Privacy & compliance
                            
          
                            Scored on shape, never on secrets
                            
          
                            Prompts pass through redaction before any judge model or score record sees them. Structure survives; identity and credentials do not.
                            
        
                            
        
                              
        
                              Sample data only This tab renders hand-authored sample data. No production gateway, HRIS connection, customer prompt, or telemetry store is read by this page.
                              
        
                              
          Gateway unavailable until the bundled sample starts.