Explain the cost argument against AI for a low-volume internal workflow.

Grisdale Structural Engineers is a ninety five person firm. Its field engineers travel to job sites, so its finance team audits about a dozen expense reports a month for policy violations. Arlo Vantreese runs internal platforms there. Demelza Trewin, the finance director, just came back from a vendor conference wanting an AI feature that audits every report automatically. Rooney Angwin, three weeks into the finance job, asked the question nobody else in the room had: is that worth it for twelve reports a month? Arlo has to find out if Rooney is right.

How to answer this, stage by stage

Nobody is grading whether you can recite "break it down, own the numbers, use a range." They are grading whether you can turn "it's only twelve a month" into a number that would actually change what a finance team builds next quarter.

Let's learn

What happens when a low-volume task gets an AI feature built for it anyway, because a vendor made it look easy?

Say a company builds an AI feature that reads an expense report, the receipts, the travel policy, the project code, and decides whether it's clean or needs a second look. Before anything like that exists, a real person audits it by hand: checking each receipt against the policy, checking for a duplicate submission, checking that the project code matches a real job. That is the entire manual cost, nothing hidden in it.

Most reports never reach a human at all. The twelve a month that do are the only twelve this whole argument is about.

At Grisdale, about twelve reports a month cross the hundred fifty dollar threshold that sends them to a real audit. Thirty five minutes each, at Demelza's loaded rate, comes to about four hundred seventy six dollars a month. That's how it goes today.

Here's how it would go if the AI feature got built. The model itself is cheap to run, about eight dollars a month for twelve reports. But a model that judges expense policy for a living needs three more things a plain rules check never needs: a current eval set to prove it's still getting the hard cases right, a monthly check for drift, because policy changes and new job sites change what "reasonable mileage" even looks like, and a person reviewing whatever the model flags, because a wrong "approved, compliant" call doesn't announce itself the way a person's own doubt would.

Add those three up, plus the build cost spread across the first year, and the total climbs to about ten seventy four dollars a month.

Four of these five parts exist only because a model is involved. A plain rules check would need none of them.

Here's the turn. It isn't that the AI feature would do a bad job. It's that the parts of its cost that never go away, the eval set, the drift check, don't get any cheaper just because there are only twelve reports to check. A rules-based gate that flags a missing receipt costs the same whether it runs on twelve reports or twelve hundred. An AI feature's upkeep cost is the same either way too, it's just that twelve reports can't spread that upkeep thin enough to make it worth it.

What it costs at its worst: Grisdale builds the feature, volume never grows past fifteen a month because the firm isn't hiring that fast, and the company pays around a thousand dollars a month, forever, to replace a task that costs under five hundred to just do by hand. Worse, if the underlying model gets retired or changed in eighteen months, which happens, the whole pipeline needs rework, another few engineer weeks, on a feature that was already losing money.

What I would leave alone: if Grisdale's audit volume already sat above thirty reports a month, or the firm were about to merge and volume would triple within the year, this whole hesitation would evaporate. The same build cost would pay for itself inside year one, and I'd greenlight it without a second thought.

The lesson: a feature that's technically easy to build and a feature that's worth building are answering two different questions. Check the real arithmetic before either one wins by default.

Now here is the same thing as a story

The short version is above, for saying out loud. Read this one for the actual Thursday a new hire's small question turned into the real math.

Arlo Vantreese can tell, usually before anyone else in the room finishes the pitch, whether a proposed tool's fixed cost will actually pay for itself. He'd run internal platforms at Grisdale Structural Engineers for four years, and that habit had killed more than a few projects nobody wanted to admit were half thought through.

Demelza Trewin runs finance, and she is good at a different thing entirely: a room. She'd just gotten back from an industry conference where a vendor demoed an AI feature that reads every receipt in real time and flags anything against policy before it ever lands on a person's desk. It looked seamless up on the screen. She booked time with Arlo the next morning to scope it as Grisdale's next internal build.

Nobody in the room disagreed with the pitch. The question that mattered came from the newest person in it.

Rooney Angwin sat at the back. Three weeks into the finance job, still learning where the supply closet was, Rooney raised a hand about halfway through the pitch. "Quick question, sorry. How many of these do we actually do a month?"

The room went quiet for a second. Demelza checked her notes. "Around a dozen," she said, and moved on to the next slide. Nobody else picked the thread back up. But it sat with Arlo the rest of the meeting.

He started scoping it the way he always did, by pricing the alternative first. Manual: twelve reports, thirty five minutes each, Demelza's own loaded rate. Four hundred seventy six dollars a month, five thousand seven hundred twelve a year. That part took him ten minutes.

The AI feature took longer, because it wasn't one number, it was five. Three engineer weeks to build, nine thousand six hundred dollars. Then the parts that never stop: model calls, about eight dollars a month at this volume. An eval set built from Grisdale's own past audits, and kept current every quarter as policy changed, about a hundred thirteen a month. A monthly drift check, comparing the model's calls against a small sample Demelza re-audits herself, a hundred two a month. And a person still reviewing the roughly one in four reports the model would flag rather than clear outright, fifty one a month.

Spread across the first year, the AI feature came to about ten seventy four dollars a month. More than double the four seventy six it was supposed to replace.

Same twelve reports, checked two different ways. One of them costs more than twice the other, before a single receipt gets read.

The nine thousand six hundred dollars was not a build cost. It was twenty months of Demelza's own time, paid up front, before the model read a single report.

Arlo brought both numbers back to Demelza that Thursday, no framing, just the arithmetic laid out next to each other. She looked at it longer than she'd looked at anything in the vendor's own slide deck. She hadn't been wrong that the manual audit was tedious. She'd just never priced what replacing it would actually cost, month after month, at Grisdale's own volume.

"We're not saying never," Arlo told her. "We're saying not yet, and here's the number that would change it." He walked her through the range: at twenty nine reports a month, the build pays for itself inside a year. At Grisdale's growth rate, that was two, maybe three years out, if it happened at all.

Demelza remembered the meeting six months earlier where the finance team had quietly agreed that any tool a good enough demo could sell was probably worth building. Nobody in that room had been careless. The demo really had worked. It just never had to answer for its own eval set or its own drift check, because a vendor's slide deck doesn't carry a monthly bill.

Grisdale extended its existing rules-based gate that quarter instead: two new checks, a missing receipt over seventy five dollars and a duplicate expense ID, built in an afternoon, no model, nothing to keep feeding. The AI feature stayed on a list, with one condition next to it in Arlo's own handwriting: revisit only if audit volume clears twenty a month for two straight quarters.

What Arlo would tell his Thursday-morning self: he almost let a good demo skip the step he always insisted on for everything else, pricing the alternative first. Rooney's question wasn't a challenge. It was the first real requirement anyone had stated all meeting: how many.

BOUND, for pricing whether Grisdale's audit feature is worth building

Not a way to make a reasonable pitch sound suspicious. BOUND turns "we could build this" into a number Arlo, or anyone else at Grisdale, could actually defend in the room.

Today's volume sits between the best case and the worst case. Which one is closer depends on how patient the payback plan is.

One alternative considered and rejected: build the AI feature now anyway, so it's "ready" the day volume grows. It lost, because a model sitting mostly idle still needs its eval set kept current every quarter and its drift checked every month, real ongoing cost, whether or not anyone is using it much. The AI-specific risk underneath all of this is quiet: a model's policy judgment can drift as new job sites or new policy edits show up, and a wrong "approved, compliant" call doesn't announce itself the way a person's own hesitation would. The guardrail is mandatory human review on anything the model flags, plus the monthly drift check comparing its calls against a small hand re-audit. And the trade-off is real, not free: keeping that review in place costs $51 a month and real turnaround time, instead of a fully hands-off system, trading total automation for a wrong "clear" that never slips through unseen.

And if you want to be sure it really works, try it somewhere else

Same five letters, a small college's continuing-education office instead of an engineering firm, and this time the tempting build is an AI tool that resolves tuition-refund disputes instead of expense reports.

Sallowmere Continuing Studies runs evening and weekend courses for working adults. Idony Fellowes runs its operations, and every month a handful of students dispute a tuition refund, arguing they withdrew earlier than the record shows, or that a policy exception should apply. A vendor pitched Idony an AI tool that reads the enrollment record and the withdrawal policy and drafts a recommended ruling.

Run BOUND on it. Break it down: manual cost is cases a month, times review minutes, times the registrar's loaded rate. AI cost is the build, spread over however many months, plus inference, eval-set upkeep, drift checking, and human review of what it flags. Own the numbers: Sallowmere sees about 8 disputed refund cases a month, 40 minutes each, at a $52 hourly rate, about $277 a month manually. The AI tool would cost about $6,400 to build, 2 engineer weeks, plus about $143 a month ongoing, inference, eval upkeep, drift checking, and review of the roughly 30 percent it flags. Spread the build over the first year and the total runs about $676 a month, well over double the manual cost.

A different office, a different policy book, the same shape of gap between what the build costs and what checking it by hand costs.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Name the two totals, manual and AI, say which is bigger today, and name the volume that would flip it.
Cost: no time to build the full monthly build-up before answering. Shrink it to the one number that decides it: how many months of the manual cost the build alone would consume.
The model got better, for real: a new release cuts the per-report inference cost in half. Recheck the total anyway, since inference was already only about 1 percent of the AI feature's monthly cost. The build and the upkeep costs are what actually move this.

Where people run it wrong.
They price the AI feature's build cost and stop there, without adding what it costs to keep running.
They assume a low-volume workflow means a low-stakes decision, when the real question is whether volume will ever grow enough to pay the build back.
They treat "a vendor showed it working" as proof it's worth building in-house, instead of pricing the manual alternative first.

How to use it live. Before answering, ask yourself one plain question out loud: how many times a month does this actually happen. If the honest number is small, say so, and let that be the reason to check the math before agreeing to build, not a detail you skip past to sound decisive.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0