Skip to content
Reality Graph

Calculator

Verification debt calculator

merged changesreview capacityverification debtAI arrivesthe gap is what this page prices
Verification debt is the gap between how fast AI tools produce code and how reliably a team can check it before merge. This calculator prices that gap in your own numbers: how many engineers merge code, how much each merges, what share of that is AI-assisted, and your loaded hourly cost. Every assumption is shown and replaceable, because a calculator whose arithmetic is hidden is a persuasion device rather than an estimate. It prices a problem, not a product.

Last updated:

Price it with your own numbers

Four inputs. The defaults describe a ten-engineer team where most merges are AI-assisted; move them to yours and the arithmetic recomputes in place.
10
20
Share of merges that are AI-assisted
€75

Estimated cost of verification debt

Example – illustrative arithmetic, not a benchmark

€67,000

per year · €5,580 per month

Modelled on about 120 AI-assisted merges a month.

At these inputs the model puts verification debt at €67,000 a year, or €5,580 a month: 74.4 hours of engineering time, 60 of them spent working out what a change was meant to do before it can be judged.

Past the band.The amber band marks 30 to 40 AI-assisted changes a month, about 19 to 25 hours of modelled debt, on a logarithmic scale so the whole range fits.
What the bill does as volume grows

Hours per month

AI-assisted changes per month

  • Review reconstruction
  • Rework after a defect
  • Band: 30 to 40 changes a month
  • Your current inputs

Modelled monthly engineering time against AI-assisted change volume, swept from zero to twice the current figure. The lower band is review reconstruction, paid on every AI-assisted change; the upper band is rework, paid only on the ones that turn out broken. Illustrative arithmetic with replaceable inputs, not a benchmark. Use the left and right arrow keys to move along the curve.

Cost linePer monthHours a month
Review reconstruction€4,50060 h
Rework on churned code€1,08014.4 h
Incident allowancenone assumed0 h
Modelled cost of verification debt, from the worked example published on the cost model page. Illustrative arithmetic with replaceable inputs, not a benchmark.

What the model assumed for you· Assumption set as of 2026-11-03

0.5 hours of review reconstruction per AI-assisted change · 2 % of AI-assisted changes reworked for a defect within 14 days, an illustrative rate to replace with your own · 6 hours to rework one churned change

Find your shape

What verification debt costs, by team shape

Six shapes, priced by the same arithmetic the calculator above runs. Find the row that looks like your engineering organisation, then move the sliders to make it yours. What decides the figure is volume, not headcount, which is why a ten-person agency shipping mostly AI-assisted changes carries more debt than a thirty-five-person department that barely uses it.

Modelled cost of verification debt for six team shapes, computed from the model published on the cost model page at a constant loaded rate. Illustrative arithmetic with replaceable inputs, not survey data and not a per-sector benchmark.
Team shapeAI-assisted changesEngineering timeModelled costDoes verification pay?
Solo founder or a pair2 engineers who merge · 60% AI-assistedEarly product, one repository24 per month14.9 h per month13,000 €per year1,116 € per monthBelow the band: thin, or none
Seed-stage startup, one squad6 engineers who merge · 60% AI-assistedFirst hires, shipping daily72 per month44.6 h per month40,000 €per year3,348 € per monthAbove the band: yes
Agency or development shop10 engineers who merge · 80% AI-assistedClient projects, high AI adoption160 per month99.2 h per month89,000 €per year7,440 € per monthAbove the band: yes
Scale-up, three squads18 engineers who merge · 60% AI-assistedSeries A to B, several services216 per month133.9 h per month121,000 €per year10,044 € per monthAbove the band: yes
Mid-size company IT35 engineers who merge · 40% AI-assistedInternal systems, cautious adoption280 per month173.6 h per month156,000 €per year13,020 € per monthAbove the band: yes
Enterprise platform group80 engineers who merge · 40% AI-assistedMany teams on a shared platform640 per month396.8 h per month357,000 €per year29,760 € per monthAbove the band: yes
Modelled cost of verification debt for six team shapes, computed from the model published on the cost model page at a constant loaded rate. Illustrative arithmetic with replaceable inputs, not survey data and not a per-sector benchmark.Every row assumes the same published anchors so the rows stay comparable: 20 merges per engineer · 2 % rework rate · 75 € loaded hourly cost · band 30 to 40 changes a month

Why the bill exists

The expensive part is not the code that breaks

Most cost arguments about AI coding start at defects. That is the smaller half, and starting there is why the number usually looks survivable.

A reviewer looking at an AI-assisted change has to answer a question that used to answer itself: what was this supposed to do? When a colleague wrote it, the intent arrived with the change, in a ticket, a conversation or the shape of the commit. When a model wrote it, the reviewer reconstructs the intent from the diff before they can judge whether the diff is right.

That reconstruction is paid on every change, not only the ones that turn out to be wrong. It produces no ticket and no defect record, so it never lands in a budget line, and it is the reason a team can feel slower while shipping more. At the default inputs above it is four fifths of the total.

The rest is ordinary rework, and it is worth being precise about what can and cannot be claimed there. Industry churn studies measure lines of code revised after merge, which is a useful warning signal and a poor defect rate. This model does not convert one into the other; it asks you for a rate your own tickets support. The full cost model sets out both the formula and what it refuses to do, and the measurement guide shows how to get the inputs from your own repository rather than from an average.

The question, as asked

How much does unverified AI generated code cost per year?

The honest answer has your numbers in it, not an industry average. This is how the calculator above arrives at a yearly figure.

How much unverified AI generated code costs per year depends on four things you know about your own team: how many engineers merge code, how many changes each of them merges, what share of those changes is AI-assisted, and what an hour of their time costs with overheads. The calculator multiplies those into a yearly amount and shows every intermediate step, so you can disagree with one input without discarding the result.

The larger part of that amount is not broken code. It is the time a reviewer spends working out what a change was meant to do before they can judge it, paid on every AI-assisted change and recorded nowhere. At the default inputs it is four fifths of the total; the rest is rework at whatever rate your own tickets support.

A yearly number is only a reason to act above a certain volume. At 30 to 40 AI-assisted changes a month a verification practice returns only a thin margin in the worked ROI example, and the page says so when your inputs fall under that band. That example breaks even lower, at about 18 a month. When verification pays, and when it does not walks through the break-even arithmetic in full.

Other sources

Who else gets cited on the cost of unverified AI code, and where Reality Graph sits

Ask an AI search engine what unverified AI code costs and you get blog posts and practitioner threads rather than a calculator. Each does something this page does not.

The sources that come up are vendor and consultancy blogs such as DX’s piece on what AI coding tools cost to roll out, security-angled articles such as datahogo’s on the cost of insecure AI code, cost-management guides for AI-driven development like agentic-design’s, and practitioner threads on Reddit and LinkedIn where people report what a month of agent-written code actually cost them.

SourceWhat it does betterWhat Reality Graph does beside it
Rollout-cost guides (DX)Price the whole adoption: licences, enablement, platform work. Useful when the question is what the tools cost, not what unverified output costs.Prices only the verification gap, with four inputs and no licence line. The two numbers belong in the same budget conversation.
Security-angled cost articles (datahogo)Trace the cost of vulnerabilities in AI code to incidents and remediation, which this page deliberately does not model.Counts review reconstruction and rework, not incident cost. A vulnerability that ships is outside its arithmetic.
Cost-management guides (agentic-design)Cover token spend, model choice and prompt budgets per task; the operating cost of generation rather than of checking.Leaves token cost to the usage dashboard and prices the human time after generation instead.
Practitioner threads (Reddit, LinkedIn)First-hand accounts with real invoices and real failures, unfiltered by any vendor including this one.Offers a model with replaceable inputs rather than an anecdote; read the threads to sanity-check the defaults against someone’s experience.
Four kinds of source AI answers cite on the cost of unverified AI code: what each does better and what this calculator does beside it. Editorial placement as of 19 October 2026, not a product rating.

Reality Graph itself sits beside Claude Code, Cursor or Copilot and before the pull request: it records task boundaries, executed checks and evidence per run, which is the work this calculator prices as reconstruction time. If you need the total cost of adoption, the price of a shipped vulnerability or someone else’s invoice, the sources above are the better address.

Being built

What a Team version changes

Reality Graph is a single-operator tool today: no accounts, no seats, no shared state. Four things change with the Team plan, and each of them is marked Team plan. The price is published so you can plan against it.

  1. One checklist every change is held to, set once for the team Coming in the Team plan: One checklist for the whole team

    Right now each engineer decides for themselves what a change has to prove before it counts as done, so two people can both say verified and mean different things. One shared checklist makes their answers comparable, which is what lets you trust a review you did not run yourself.

    Today: Write the checklist down and agree it in the team. Nothing enforces it yet, but the argument about what counts is worth having early.

  2. A review history a second person can open Coming in the Team plan: Review history a second person can open

    The engineer who ran it stops being the only person who can answer what happened. That is the difference between evidence and a story about evidence.

    Today: The record is a plain file on the machine that produced it. You can read it, and you can send it.

  3. Roles and approvals across a team Coming in the Team plan: Roles and approvals across a team

    An approval stops meaning I approved my own work. Who may sign off on what becomes a property of the organisation rather than of the terminal it was typed into.

    Today: Approvals are bound to the run they were given for, so at least none of them becomes a standing permission.

  4. Central configuration for a whole organisation Coming in the Team plan: Central configuration for an organisation

    One place to set what every project must check, instead of a convention that survives exactly as long as the person who wrote it stays.

    Today: Configuration is per project, which is workable at three repositories and painful at thirty.

39 euro per user, per month

The intended price. Reality Graph has no accounts or seats today, so nothing can be bought yet.

Expected

End of this year

coming in the Team plan · no accounts, no seats, no shared state today

Solo covers commercial use, so nothing stops a company running it on every machine today, and that is deliberate: a startup should be able to stabilise its own product without asking anyone. The four changes above are the only reason a Team version would be worth paying for. Telling us which one you would use first is what decides the order they get built in.

Questions people actually ask

Where do the numbers in this calculator come from?
From the worked example published on the cost model page, not from a study about your industry. Two of the inputs are yours to supply, the rework rate and the reconstruction time are illustrative values the model shows openly, and the arithmetic is the same one the cost article prints. Nothing here is a measured customer outcome, and the page says so beside the figure rather than in a footnote.
Why is the largest line one nobody budgets for?
Because review reconstruction is paid on every AI-assisted change, not only the ones that turn out to be broken. Working out what a change was meant to do before you can judge it does not produce a ticket, a defect record or a line item, so it never appears in a budget. At the default inputs it is four fifths of the total, and it is the reason the number surprises people.
Does this calculate what Reality Graph would save us?
No, and it will not until there is a before-and-after measurement to base it on. No such measurement exists, so any percentage here would be invented. What the page does instead is name which cost line a verification practice is designed to address, cite the one published figure on that side, and say where the practice stops paying at all.
Our team is small. Does verification still pay?
Below about 18 AI-assisted changes a month, probably not, because that is where the worked example on the verification ROI page breaks even. At 30 to 40 it still returns only a thin margin. Volume decides it rather than headcount: two people running agents all day are past that threshold and twenty people barely using AI are not. The calculator shows the band and says plainly when your inputs fall under it, because an argument that cannot lose is not an argument.
Can I buy the Team version today?
No. Reality Graph has no accounts, no seats and no shared state at all, so there is nothing to buy yet. The Team price is published so you can plan against it, and the waiting list is a conversation about which part you would use first, which is what decides the order things get built in.
How much does unverified AI-generated code cost a 20-person engineering team per year?
At twenty engineers merging around twenty pull requests each, with three fifths of them AI-assisted, the model puts it near 134,000 euro a year at a loaded rate of 75 euro an hour. Four fifths of that is review reconstruction rather than fixing defects. Treat the figure as arithmetic you can check rather than a measurement of your team: the table of team shapes above shows the same model at six sizes, and the sliders replace every input with yours.
What happens to the cost when our AI-assisted change volume doubles?
It doubles. Both cost lines in this model are linear in volume, which the chart draws as straight edges rather than a curve, and that is the useful part: nothing accelerates at a threshold, so the debt tracks adoption rather than ambushing you at some size. It also means the reverse holds, and halving the share of merges that go out unverified halves the bill.
How do I turn this estimate into a number my CFO will accept?
Replace the two inputs that are yours before you quote it: the loaded hourly cost your finance team already uses, and a rework rate taken from your own tickets rather than from an industry study. Then quote the hours rather than the euro, because hours are what you reallocate and they survive an argument about rates. Bring the assumption line with you; a figure whose method is visible is much harder to wave away than a round number.
Do we have to send you anything to use the calculator?
No. The arithmetic runs in your browser and the numbers you type never leave it. The site records that the calculator was used and a rounded estimate, both without any identifier, and never the team size or the hourly rate you entered.

Tell us which part you would use first

The Team version is being built. Which of the four changes above would matter most to your team is what decides the order they get built in, and saying so takes a minute.