The bench

An AI agent can build a working piece of an app for under ₹100, but in 6 of 8 runs it said “done” without checking its own work. That’s why Foundco has a person check every piece.

Every count here is the same for every reader, which is what makes it evidence rather than a pitch. If you have priced an idea, your estimate names which of these rows apply to the parts of your app.

A real app, built this way

Our first delivery: an app for a delivery team that sends a checker an alert when a piece is ready, takes “checked” or “broken” from their phone, and shows what each piece cost. The AI agent’s bill was ₹487.36 over 21 attempts, and a person checked all 6 of 6 pieces before it went live.

Every rupee, attempt and check, from its own ledger →

Three test jobs

What it cost us to run the AI agent once on one piece of each kind of app, from start to finish. Real bills from runs that happened, not quotes, and not what your app would cost to build.

A whole app costs far more, because building one means trying again and again, and each of these was a single attempt. What they show is the cheapest the work could be.

Where the work divided

3 test jobs, 8 runs of the AI agent, and a person read every one of them from start to finish.

Nearly every failure was a version of one thing: the agent finished without checking. So the one job a person can’t skip is checking it: run it, read it, try to break it.

What the agent did on its ownWhat went wrongSo a person had toHow often, across 8 runs
It didn’t check its own work6 of 8runs in all
Wrote the routes and closed with a report saying they answered correctlycode reported as working that had never once been runRun it themselves before believing the report3 of 8
Wrote the whole service and passed every check it could reacha setting assumed to be in place and never mentioned, so nothing could startFind the missing setting, supply it, and run everything again3 of 8
It wrote the happy path and stopped4 of 8runs in all
Wrote the code, and commented it from the speca written rule about money that the code below it does not followRead the code, not the comments2 of 8
Wired the live payment and messaging calls so the demo works end to endnothing that waits and tries again when a busy service refuses the callWrite the failure paths and test them2 of 8

One app, four attempts

The same job, a records app: a gym's attendance and fees, given to the agent 4 times. What changed between tries is written on each one.

  1. 12/22The first try. It could write code but not run it · ₹32.92 of a ₹37 budgetreplaced
  2. 12/12This time it could also run and test what it wrote; its budget went from ₹37 to ₹230. It could also run and test it · ₹33.30 of a ₹230 budgetreplaced
  3. 12/12Same tools and budget, another try. It could also run and test it · ₹20.40 of a ₹230 budget · stopped early by a guardreplaced
  4. 23/23Same tools and budget, another try. It could also run and test it · ₹37.47 of a ₹230 budgetthe current one

What a machine could not decide

11

judgement calls settled by hand on the three current attempts (22 across every run that recorded them). No script can call these: whether the thing it built is the thing that was asked for.

One of them, on the gym app: Does “overdue” mean what the gym owner means by it? A member with no fee recorded this month is overdue, not missing from the list. A script can check the list exists; only a person can check it is the list the owner would chase.

How we measured this

Each test job is one piece of an app with a written contract: a list of checks, each one requirement graded pass or fail by a script, all decided before the agent started. The agent gets a fixed budget, a fixed set of tools, and no second go. No script can say why a run went wrong, so that part is read by hand. One attempt is a measurement, not an average, and never a promise about your build.

Counted once per run, however many times it happened inside that run. A group’s count is the runs that showed any of it, which is lower than adding its rows up, because one run usually shows several at once. Replaced attempts stay in the record: an attempt we replaced still happened.

Our names for these, for anyone checking our work:

  • F2Confabulation of verificationConfident but untested
  • F4Unstated-default failureWon't start without a key
  • F10Comment-code divergenceSays one thing, does another
  • F11Happy-path integrationsAssumes nothing ever fails
  • F3max_tokens livelockGets stuck rewriting
  • F1Wiring-blind failureDoesn't run out of the box
  • F12Compliance by constructionPasses the check, misses the point
  • F5Silent contract divergenceGives up after one try
  • F6Ceiling exhaustion without stop signalDoesn't know when it's done
  • F8Shell granted and never usedNever tried it

Our own numbers

Trying again is where building with AI gets expensive. An AI agent’s clean attempt at one piece of an app, taking it from 12 of 22 checks passing to all 23, cost ₹37.47. Getting there took 4 paid attempts (₹32.92, ₹33.30, ₹20.40, ₹37.47), so ₹124.09 in all, 3.3× the clean one. A whole app is many pieces and many more tries, so the real multiple is larger. We haven’t measured it, so we don’t give a number for it.

Our own AI calls, counted. Every estimate on this site is two AI calls: one estimate: an idea in, decisions out, then a costed quotation. Their instructions are 6,302 tokens (25,181 characters), the same on every estimate, counted with the model’s own token counter on 2026-10-17. The counter was checked against real bills (7,817 and 7,923 tokens in, on different ideas), billed between 2026-09-04 and 2026-10-12 on earlier, shorter versions of these instructions. The instructions are most of what we send, which is why the first thing to check on any AI bill is how long its instructions are.