What ROI argument works for an internal AI tool with no revenue line?

Latchpoint watches badge and door logs across Vaultridge Pharma's eleven warehouses, two of them holding controlled medicine behind a caged vault, and flags anything that breaks a person's normal pattern. Ruven Straathof owns it. It has never had a revenue line, and every fall Guthrie Pryor's finance team cuts any cost centre that can't show real numbers by default. This year, with four more warehouses opening, Ruven had to walk into that review with an ROI case that would actually hold.

How to answer this, stage by stage

Nobody is grading whether you can name four things Latchpoint might be worth. They're grading whether you know which of those four falls apart the moment someone with a spreadsheet pushes on it.

Let's learn

Latchpoint watches every badge swipe and door sensor across Vaultridge Pharma's eleven warehouses, and flags anything that looks off, before a person has to go looking for it.

Five steps. The fourth one, the confidence check, is the step every one of the four ROI arguments quietly depends on.

Before Latchpoint, nobody watched the logs in real time. Two of the eleven warehouses hold controlled medicine behind a caged vault, and a full count of that vault is required twice a year. That count was the only time anyone opened the access logs at all. If something odd happened in between, on average it sat there forty-seven days before anyone found it.

Latchpoint changed that. It watches all eleven sites, all the time, and flags anything that breaks a person's normal pattern: an unfamiliar door sequence, a badge active after someone's last day, a door held open too long. When it launched fourteen months ago, one setting decided what counted as worth a look, the same setting for every site, cages and coat closets alike.

That one setting flagged about a hundred and thirty things a day, company-wide. Nine out of every ten turned out to be nothing: a real employee taking an unfamiliar door because of construction, a supply run at an odd hour. Checking one flag took about two minutes. Multiply that across every day of the week and the five-person security team spent about thirty hours a week just clearing the queue, more than six times what the twice-a-year count used to cost them.

Here's the turn. Those extra hours were never the real problem. The real problem is what a team does with a tool that cries wolf ninety-one times out of a hundred: it stops trusting it. Anzor Osoro, who covers most of the overnight desk at the Ashfeld site, started clearing flags without reading them closely. Some nights he muted the queue and caught up in the morning.

What that costs at its worst: a team that stops trusting its own tool is no better than the twice-a-year count it replaced, except now it also costs thirty hours a week.

What I would leave alone: the nine warehouses that hold nothing controlled. A missed flag there is a shrug, not a loss, so the looser threshold that saves the team review time is the right call, even though it lets a few more small things through.

Latchpoint itself costs Vaultridge four hundred ten thousand dollars a year: three engineers, cloud compute, and the escalation link to each site's cameras. Against that, the case has two real numbers behind it. Two hundred eleven thousand dollars a year in analyst hires the company never had to make as its warehouse count grew from seven to eleven. And an average of three hundred forty thousand dollars, the cost of each of the two bad badge incidents in the three years before Latchpoint, the kind of event the Ashfeld save stopped from becoming a third.

The lesson: a tool with no revenue line still has to earn its keep in dollars somebody else can check. How useful it feels doesn't survive a budget meeting. Records do.

Now here is the same thing as a story

Read the short version above when you're in the room. Read this one when you want to feel why the number Ruven almost led with was the one that could have sunk him.

Ruven Straathof has run security engineering at Vaultridge Pharma for three years, and he still remembers pitching Latchpoint off eleven slides and one number: forty-seven days, the average time it took anyone to notice a bad badge swipe.

Latchpoint went live across all eleven warehouses on a Monday in March. For the first few weeks, Anzor Osoro, who has covered the overnight security desk at Ashfeld for six years, loved it. He used to check the vault log by memory and a paper sign-out sheet. Now a flag showed up the second something looked wrong, and Anzor was proud of how fast he could clear one: read it, check the camera, close it out, on to the next, usually inside two minutes.

It thinned in three beats, and none of them looked like a mistake at the time. Beat one: the model shipped with one setting for every site, cages and coat closets alike, because nobody had eleven warehouses' worth of history yet to tell them apart. Beat two: as more sites came online, the daily flag count climbed past a hundred, then past a hundred and twenty, and nine out of ten of them turned out to be nothing. Beat three: Anzor stopped reading each flag closely. He started scanning for a few keywords and clearing the rest, the way you skim a form you've filled out forty times.

No single bad night caused this. A tool that flagged too much taught a good analyst to stop reading it, one shift at a time.

The trigger wasn't an incident. It was smaller than that. A new hire on the day shift asked him, half joking, why he even bothered opening the queue if he was going to close everything anyway. Anzor didn't have a good answer. That night, he muted the alert sound on his terminal for the first time.

Six weeks after that, at 2:14 in the morning, a badge opened the vault door at Ashfeld. It belonged to a technician who had left the company eleven days earlier; a sync bug on the badge vendor's side meant nobody had actually deactivated it. Latchpoint flagged the door sequence in forty seconds, a mismatch between a terminated badge and a live vault entry, exactly the kind of thing it was built to catch. Anzor's terminal buzzed. Because he'd muted the sound, he didn't see it for six minutes.

He caught it at minute seven, deactivated the badge himself, and dispatched the on-site guard. Nothing was taken. Nobody outside the security team ever heard about it. But if he'd been asleep at the console, or if the flag had landed on one of the nights he cleared the queue without reading it, this would have been Vaultridge's third bad badge incident in three years, on top of the two that had already cost the company an average of three hundred forty thousand dollars each, in lost product, mandatory reporting, and cleanup.

The decision that opened the door traced back to launch week, a short meeting about the dashboard. Someone asked whether a dismissed flag needed a reason code, true anomaly, false alarm, or unclear, before it could be closed. The answer was no. At the time, the team was three people who all knew each other's judgement by feel, and a reason code felt like paperwork nobody needed.

Run the six weeks again, with one change: the reason code exists from day one, and the threshold isn't one setting for every site. The two vault sites get a tighter bar, catching more, costing the team more minutes there. The other nine get a looser one, catching less, but freeing up most of the week. Under that design, daily flags across all eleven sites drop from about a hundred and thirty to about forty. Anzor's queue at Ashfeld shrinks to something he can actually read every one of. And because every dismissed flag now carries a reason code, Ruven walks into the budget meeting fourteen months later with a real, countable number: two hundred eleven thousand dollars in analyst hires Vaultridge never had to make as the warehouse count grew from seven to eleven, checked against a hiring ledger Guthrie can read himself, without asking Ruven to vouch for a single figure.

One design asked a team to trust a tool that hadn't earned it. The other gave that same team a record clean enough that the tool didn't need anyone's trust at all, just its own numbers.

What Ruven would tell himself, back in that short meeting about the reason code: skipping it wasn't lazy. It was the sensible call for a three-person team that trusted each other completely. It just wasn't a call that was going to survive being asked to prove anything, fourteen months and four more warehouses later.

ORDER, for an argument with nothing to sell

Not a way to dress up "trust me, it's worth it" in five letters. ORDER is what forces you to say, out loud, which of your four best arguments would actually survive a CFO leaning on it.

One of these can be re-checked in front of the CFO without falling apart. The other only survives if nobody asks a follow-up question.

The order a spreadsheet alone won't give you: check the model against real cases before any number built on top of it gets said out loud.

The whole answer, in one picture. A budget case that sounds right and a budget case that's checked against a record are not the same argument wearing different clothes.

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

Same five letters, an expense ledger instead of a badge log, and this time the number that almost got the tool cancelled wasn't the loudest one, it was the one built on a threshold nobody had rechecked.

Ledgerwatch is Dunmarsh Freight's internal tool: it reads every driver's expense claim, per diem, fuel card swipe, and lodging receipt, and flags the ones that don't match a driver's own normal pattern. Havard Kesteven runs internal audit there. Ledgerwatch has never had a revenue line either; it just watches money the company was already going to spend.

Dunmarsh's fleet grew from sixteen hundred drivers to twenty-four hundred over three years, without adding a single new auditor to its six-person team. Under the old hiring ratio, about one auditor per five hundred new drivers, that growth should have cost roughly two hundred eight thousand dollars a year in new hires. Instead, Ledgerwatch's backtest against eighteen months of closed audit cases found eleven confirmed duplicate-claim cases it had caught, including one driver submitting the same fuel receipt to two card systems after a company merger, worth about sixty-one thousand dollars on its own.

Different company, different ledger, same corner of the chart wins. The argument built on a record nobody has to take on faith is the one that survives a challenge.

Same rank, different lever: validating the model against real, closed cases still comes first here too, but the reversibility risk points somewhere new. For Vaultridge, the fragile argument was self-reported hours. For Dunmarsh, it's the flat threshold itself: any claim about "fraud caught" is worthless until someone checks whether the cross-border contract quietly inflated the false-positive count on legitimate tolls, the way one global setting once did at every Vaultridge warehouse.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to the order: lead with the ledger-checked headcount number, back it with a real backtest, never open with self-reported hours.
Cost: there's no budget this quarter to run a full backtest. Pull the smallest cheap slice first, the last ninety days of closed cases, since that's already sitting in the case-management system.
The model got better, for real: say Ledgerwatch's fraud-detection accuracy doubles overnight. The rank barely moves. A better model changes how much the "fraud caught" argument is worth, not whether the headcount number still needs the hiring ledger behind it.

Where people run it wrong.
They lead with whichever number already exists on a dashboard, instead of the one that survives a challenge.
They let a threshold ship once and never recheck it as the business the tool watches actually changes shape.
They treat one dramatic near miss as proof the case is solid, instead of proof it's worth digging for the boring, checkable number underneath it.

How to use it live. Ask the real question before naming an order: "which of these arguments can I re-check in front of you right now, and which one only works if nobody asks a follow-up." That buys the room to rank by what survives, instead of by what sounds best on a slide.

Three things worth stating directly, since the real judgement sits here. The alternative Vaultridge considered, and rejected, was hiring a third-party overnight monitoring vendor to watch live camera and badge feeds at the two vault sites instead of building Latchpoint. It lost, because a vendor's cost scales with sites and hours, exactly the headcount growth the case exists to avoid, and a remote command centre has no way to learn what counts as a normal deviation at any one specific warehouse, a distribution-shift problem for a person watching a screen, not just for a model. The AI-specific failure worth naming by name is silent drift after a facility change: any time a site swaps its badge system or changes its floor plan, the model's learned sense of "normal" breaks, and false positives spike right when the team can least afford to distrust the tool. The guardrail is a standing rule, not a one-off fix: any site with a recent facility change gets frozen out of live escalation for fourteen days, every flag there routes to manual review only, while Latchpoint quietly relearns the new baseline. And the trade-off being accepted plainly: the two vault sites run a tighter threshold, more real minutes spent by the security team, to catch more of what actually matters there, while the other nine run looser, accepting a slightly higher chance a trivial anomaly slips through, in exchange for a queue Anzor can actually keep up with.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0