Cost per Case: The Number Missing From Most AI Business Cases

The business case for the AI pilot is twelve slides long. It promises thousands of hours saved a year. It lists the licence fee, the integration estimate and a contingency line. The steering committee likes it.

Then the finance director asks one question. What does it cost us to handle one claim today, and what will it cost once this is running?

Nobody in the room knows. The hours are an estimate of time freed, not money spent. The licence fee is fixed. The model calls, the reviews, the rework and the complaints that follow a wrong answer are nowhere on the page.

The paradox: The business case is full of numbers, and none of them is the one finance will check every month.

What We Found

Before we recommend anything about an AI investment, we study what the public record says about how such investments are judged. For this article we went through an official audit report, analyst predictions about AI projects, and a research preprint on the running cost of AI workflows, from 2024 to 2026.

Three findings stand out.

The gains are often estimated before the cost of delivering them. In March 2024 the UK’s National Audit Office reported on the use of AI in government. It found that the centre of government “has identified the potential for large-scale productivity gains from AI use in the public sector but has not yet assessed the feasibility or cost of delivering these improvements”. An indicative analysis had found that almost a third of tasks in the civil service could be automated; it did not examine feasibility or assess cost. The report also records that further scrutiny and evidence collection were recognised as needed, alongside substantial investment.

Cost and unclear value are named as reasons projects stop. In mid-2024 one analyst firm predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, “due to poor data quality, inadequate risk controls, escalating costs or unclear business value”. It added that productivity benefits can be “difficult to directly translate into financial benefit”. A year later the same firm predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls.

The cost of each case depends on the design, not only on the price list. An April 2026 research preprint compared generating workflow code once with calling a model on every transaction. It reported break-even after about 17 transactions, and 57 times fewer tokens at 1,000 transactions. That comparison counts tokens only, not the engineering time to build, test and maintain the code, so a full-cost break-even sits higher. It still shows how far the cost of one case can move with one design choice.

What our review does not show: The audit report is about UK central government, not about firms. The analyst figures are predictions, not measured outcomes; the first one’s date has passed, and we did not examine whether it came true. The preprint measures tokens on its own workflows. We found no public record of a firm’s cost per case before and after an AI workflow went live. We use these as signs of how AI investments are judged, not as evidence of what any firm spends.

Cost per Case, and What It Unlocks

Our reading of the record is that most AI business cases count the benefit in one unit and the cost in another, and never bring them together.

The benefit is counted in hours: time freed, tasks automated, productivity gained. The cost is counted in contracts: licences, seats, an integration estimate. Neither is the unit finance manages the operation by, which is the cost of one completed case: one claim settled, one application decided, one invoice processed.

Cost per case is everything one completed case costs. For an AI workflow that means the model calls (including retries and the longer prompts that grow over time), the people who review or correct the output, the cases that are escalated or redone, the errors that reach customers and come back as refunds or complaints, and a fair share of the platform and its upkeep. Measured today, before the AI, it is the baseline. Projected for the new design, it is the case. Measured every month after go-live, it is the evidence.

Once that number exists, three decisions get easier. A step whose cost per case does not fall should not be automated. A step that runs at high volume with stable rules may be cheaper compiled into code than run on a model each time. And a pilot that reports in finance’s own unit can be funded on evidence, not on a forecast of hours.

The gap we see: many firms approve AI on productivity estimates and judge it later on cost. Firms that put cost per case, today against projected, into the business case from the start can tell early which workflows pay and which do not, and can stop the ones that do not before the bill grows with volume.

This is our reading of the evidence. It is not a measured market figure.

Four Ways to Build the Business Case, Weighed

A firm deciding whether to invest in an AI workflow has four real options for how to make the case. Each one is right for somebody.

Option	When it fits	What it risks
Hours saved, turned into a productivity figure	Early exploration with a small budget	Hours freed rarely show up in the accounts; running costs that grow with volume are missed
Licence and seat cost against the expected benefit	A tool staff use for drafting, where a person reviews everything	Leaves out per-call costs, review time and error costs once volume grows
A token-only comparison of design options	Choosing between running a model on every case and generating code once	Leaves out the engineering time to build, test and maintain the code
Full cost per case, today against projected, workflow by workflow	The workflow will run at volume and its answers are acted on	Needs a short period measuring today’s process, and some costs (errors) are estimates at first

Our recommendation: for any workflow that will run at volume, build the case on full cost per case, today against projected, including review, rework and the cost of errors. Use it to decide, step by step, what to compile, what to keep on a model and what not to automate, and measure it every month after go-live.

Where You Are Now, and Where You Should Be

A recommendation is only useful if you can tell later whether it worked. So before any work begins we set a benchmark: a few measures taken from your own workflow, each with a target.

For this topic, the benchmark would look like this.

Measure	Typical starting point	Where it should be	How it is checked
The unit of the business case	Hours saved, or seats	Cost per completed case, today and projected, for each workflow	Finance’s own cost data, workflow by workflow
What the cost includes	Licence or model fees	Model calls with retries, review time, rework and escalations, errors that reach customers, a share of the platform	A cost sheet for each workflow
How cost moves with volume	Not modelled	Cost per case at today’s volume and at the expected volume	The volume plan
Review after go-live	Rare, or a one-off	Monthly, against the projection	A monthly cost-per-case report
What the number decides	Whether to buy the tool	Which steps to compile, keep on a model, or not automate	The step-by-step triage

The starting points above are what we expect to find. In an engagement they are replaced by your own figures, taken in the first review. The targets are what we would work towards with you. They are not results we have already achieved for a client.

How We Work With You

1. We listen: a free Reliability Review. A first conversation about the workflow, what the business case assumes, and what one case costs today as far as anyone knows. You leave with a short written note of what we found.
2. We research. We research your sector, the regulator’s position and what the public record says about firms in your situation. This article is an example of that work.
3. We diagnose: the Workflow Diagnostic. It sorts each step of the workflow into compile, keep on an AI model, or don’t automate, and sets out the cost per case today against projected, alongside a 100-rerun consistency test of your current pilot, a policy-to-rules map, the evidence gaps and a change plan. It may recommend another firm or tool, or not automating at all. We take no referral fees on model or platform recommendations.
4. If you approve, we prove it: a Proof Sprint. One workflow, on your own cases, measured for consistency and for accuracy against your own experts. You keep the code.
5. We review against the benchmark. Before work starts we set the benchmark. At each stage we show you where each measure stands, and the filled-in benchmark is yours at the end.

The Review is free. Every paid step is fixed price, agreed at scoping. Stop after any step. The findings are yours. The Diagnostic fee is credited against a Proof Sprint within 90 days.

What changes for the business:

Role	What they gain
CFO	A business case in a unit finance can check every month: the cost of one completed case, today against projected.
CEO	AI spend approved workflow by workflow, on a number the board can follow, not on a forecast of hours.
COO	The cost of reviews, rework and escalations is on the table, so staffing decisions follow the real workload.
Head of AI or Data	A pilot that reports in finance’s own unit, so the next workflow is funded on evidence.

The cases in this article, and what IOanyT does about that kind of problem:

The case	What IOanyT does
The National Audit Office found potential gains identified, with feasibility and cost not yet assessed.	The Workflow Diagnostic sets out the cost per case today against projected for each workflow it covers.
An analyst firm named escalating costs and unclear business value as reasons AI projects stop.	We recommend not automating a step when its cost per case does not justify it, and say so in writing.
A preprint showed the cost of one case moving sharply with one design choice.	Each step is triaged into compile, keep on an AI model, or don’t automate, with the cost of each option set out.

IOanyT is a consultancy with its own engineering team: we study the problem, recommend, and prove it by building only what you approve. Sometimes the cheapest option we find is not to automate at all.

Book a free Reliability Review →

Get a fixed quote →

See the Workflow Diagnostic →