What applying AI to one operational process actually costs, and what it actually returns

Every business case for this kind of work fails in the same place: the baseline is asserted rather than measured. "It takes about a day" becomes a number in a spreadsheet, and no one can say afterwards whether the project worked.

The figures below are the shape of a typical mid-market operational process — invoice handling, order exceptions, customer onboarding, claims triage. They are illustrative. The assessment replaces them with observed values.

Note that this is a cost of effort, not a cost of failure. It excludes errors, rework, delayed cash collection and customer impact — all of which are real and none of which I will estimate for you without evidence. If those matter in your process, the case is better than this page shows.

The intervention is not "AI." It is a specific system that removes named steps from a named process: reading and extracting from documents, classifying and routing incoming work, assembling information a person currently gathers by hand, and drafting output for a person to check rather than compose.

Model and infrastructure costs are billed by your providers directly to you, not resold through me. Maintenance assumes your team owns day-to-day operation with occasional support; a full retainer costs more and is only worth it if your process changes often.

Benefit is discounted twice before it reaches this table: once for the share of work that is realistically automatable (42%), and again for adoption (85%), because some people will keep doing it the old way for a while and some cases will always route to a human.

Everything else in the model is roughly knowable. These two are estimates until a pilot measures them, so the case should be judged across the range rather than at the point estimate.

The full range is 11 to 19 months. Even the pessimistic corner pays back inside two years, which is the useful finding here — the decision is not especially sensitive to getting these estimates exactly right, and that is unusual enough to be worth stating.

Be suspicious of any vendor quoting an automatable share above 60% before measuring your process. In my experience the honest number sits between 30% and 50% for most operational work, and the gap between a demo and a deployment is almost entirely made of the cases nobody showed you.

If you disagree with one of these, change it and the case changes. That is the point of writing them down rather than embedding them in a conclusion.

A7 and A8 pull in opposite directions and roughly cancel in most cases I have seen. A7 makes the case conservative; A8 makes it more so. If your finance function values recovered capacity at zero, this project does not pay back at all and you should read §7 before going further.

It depends entirely on what happens next, and this is where most AI business cases quietly cheat. Recovering 1,488 hours a year does not produce $68,283 of cash unless one of the following is true:

• You avoid a hire you would otherwise have made. The cleanest version. Growth is absorbed by the existing team, and the saving is the salary you did not add.
• You reduce headcount. Real, and I would rather you say so at the start than discover it in month three — it changes what gets built and how success is measured.
• The time goes to work that earns. Sales conversations, collections, backlog that was costing you elsewhere. Requires someone to actually direct it, or it dissipates.
• You accept it as capacity rather than cash. A legitimate answer, but then this is a resilience and service-quality investment, and it should be judged against that bar rather than a payback period.

Prices are fixed and published. The sequencing exists so that the largest commitment is made with measured figures rather than estimated ones.

Downside is capped at $44,000 and in practice at $6,000, because the assessment is designed to be a stopping point. Roughly speaking, if the measured automatable share comes in below 25% I will recommend against the build in writing.

Everything is built in your infrastructure, in your repositories, with documentation. There is no licence, no hosted dependency and nothing that stops working if I do. That matters for the risk register as much as for the budget.

Bahman Shadmehr. Eight years building production backend systems where quiet failure was expensive — automated trading platforms placing orders on the Texas wholesale power market and US equities, a national payment gateway decomposed from a monolith while it was live, and the data infrastructure and monitoring underneath all of it.

That work is the reason this page is a spreadsheet rather than a demonstration. Systems that run unattended are judged on measured behaviour over time, not on how they look when someone is watching. The same standard should apply before you spend $44,000.

Remote, UTC+3 — European hours and US mornings. Two clients at a time. 76 technical posts · LinkedIn