How to cost an AI automation before you build it
Four lines of arithmetic that decide whether a project is worth funding — and that most proposals skip.
The short answer
To cost an AI automation, work out what the current process costs before you look at what the system costs. Multiply volume by minutes per unit by the fully loaded hourly cost, then add rework: error rate times the time to fix. That is your annual baseline. Against it, put the build cost plus twelve months of running cost, where running cost includes model and infrastructure spend, monitoring, and the engineering time to keep it accurate. If the baseline does not comfortably exceed the total, the honest answer is not to build it.
Start with what it costs you today, not with what AI costs
Almost every AI proposal we see starts from the solution and works backwards to a justification. That ordering hides the only question that matters, which is whether the process is expensive enough to be worth automating at all.
The baseline is four numbers and you probably already have three of them.
- Volume — units per month. Invoices, enquiries, calls, documents, tickets.
- Minutes per unit — how long one takes a person, honestly, including the interruptions.
- Fully loaded hourly cost — salary plus everything else, not the salary line.
- Rework — the share that come back wrong, and the minutes to fix each one.
Annual baseline = volume × minutes ÷ 60 × hourly cost × 12, plus the same calculation applied to rework.
Then cost the system honestly, including the year after
A build quote is not the cost of the system. The cost of the system is the build plus what it takes to keep it working, and the second number is the one that gets left out.
- Build — scoped and fixed before work starts, or it is not a number.
- Model and infrastructure — cost per transaction at your volume, not at demo volume.
- Monitoring and maintenance — accuracy checks, eval runs, model migrations, integration drift.
- Human exception handling — the share that still needs a person, times what that costs.
The exception rate is where most models go wrong
A system that handles 80% of cases cleanly does not remove 80% of the cost. Someone still has to identify which 20% needs attention, pick it up cold, and finish it — and handling an exception usually costs more per unit than handling a normal case did.
Model that explicitly. Automation rate, exception rate, and the cost of an exception. A proposal that quotes an accuracy figure without an exception cost is quoting half the arithmetic.
Cost per transaction is not fixed, and it grows in the wrong direction
Token cost at launch is not token cost at month six. Conversation history accumulates, retrieval returns more context, prompts get extended by whoever is iterating.
Model the cost at your real volume and at three times it. If the economics only work at pilot scale, they do not work.
When the answer is not to build it
Roughly a third of the candidates we look at fail this test, and the reasons repeat: volume too low to amortise the build, a process so inconsistent that no system could be scoped against it, or a much cheaper fix hiding underneath — a changed form, a deleted approval step, a report that already exists.
Saying so early is the cheapest value a firm can deliver, and it is the part most vendors have no commercial reason to say.
If a proposal has no list of things the vendor recommends against, they have not done this arithmetic — or they have and they are not showing you it.
A worked shape
Take supplier invoices. Four thousand a month, six minutes each to key and reconcile, a fully loaded cost of ₹600 an hour, and 8% needing rework at fifteen minutes apiece.
The baseline is 4,000 × 6 ÷ 60 × ₹600 × 12, plus 4,000 × 8% × 15 ÷ 60 × ₹600 × 12. That arithmetic gives you a real annual figure to hold a build cost against — and it is arithmetic you can run in a meeting, using your own numbers rather than an industry average somebody put in a slide.
Questions this raises
What is a reasonable payback period for an AI automation?
Most organisations we work with want the build recovered inside twelve months against the current-process baseline, with running cost included in the calculation rather than treated separately. If it only pays back over three years, the technology will have moved underneath you before the return arrives.
How do I estimate running cost before anything is built?
Model cost per transaction at your real volume using the actual context sizes the workflow needs, then add monitoring and maintenance. Anyone who has run a system in production can give you a defensible range in an afternoon. A vendor who cannot has not done it.
What if our process is too inconsistent to measure?
That is a finding, not a blocker — and usually the most valuable one. A process nobody can describe consistently cannot be automated reliably either, so the first return often comes from standardising it. Sometimes that is the whole project.
This is what we do about it
More like this
What actually breaks an AI agent in production
The failure modes, in the order we have met them — and almost none of them are the model being wrong.
Why your RAG system should refuse more often
A retrieval system that never says “I don’t know” is not accurate. It is unmeasured.
What changes when buyers ask ChatGPT instead of Google
AI answers name one or two companies and send no click. Your analytics cannot see any of it.