You’ve probably sat through a version of this review.

An operations team runs an AI pilot for two quarters. It reads carrier invoices, flags the ones that don’t match the contracted rate, and queues them for dispute before payment goes out. At the annual technology review, the pilot gets scored the way every tool gets scored: hours saved per week, invoices processed per analyst, adoption across the team. The number is fine. Not spectacular. Fine.

It gets shelved.

What the scorecard never showed was the line that mattered. The pilot had stopped a recurring overbilling leak worth far more than the hours it saved, money the company had been paying out quarter after quarter without knowing it. The pilot didn’t fail. The scorecard did.

This is one of the most expensive mistakes in enterprise AI right now. The model usually works. The business case is built on the wrong number.

The borrowed lens

For decades, operations leaders learned to evaluate technology the way you evaluate a software rollout. How many seats. How much adoption. How many hours are saved. Those questions made sense for a CRM or a planning suite, where the value was the tool itself and the cost was the license.

Operations AI doesn’t live there. It lives on the cost line. It sits on top of a process that already leaks money, a missed delivery window, a quality escape, an overcharge nobody caught, and its job is to stop the leak. Measuring that with “hours saved” is like judging a roof repair by how fast the contractor worked.

The data shows how badly this is going. MIT’s NANDA initiative, in its 2025 study of enterprise AI, found that despite an estimated $30 to $40 billion in spend, roughly 95% of organizations are seeing no measurable impact on their P&L. Gartner expects at least 30% of generative AI projects to be abandoned after the proof-of-concept stage, and one of the named reasons is unclear business value. McKinsey’s 2025 survey is blunter still: more than 80% of companies report no tangible effect on enterprise-level earnings from their use of the technology.

The usual read on these numbers is that AI is overhyped. That’s the wrong read. The technology is doing work. The problem is that the work isn’t being measured where the money actually moves.

Productivity is not P&L

Here is the quiet admission buried in Gartner’s own research. Justifying AI investment on productivity grounds, it notes, is “difficult to directly translate into financial benefit.” Read that again. The most-cited reason AI pilots stall isn’t that they fail to deliver. It’s that the metric they deliver against doesn’t convert into a number finance can bank.

A productivity gain is an input. “We saved each analyst four hours a week” is true, and it is also unbankable. Nobody in finance can put four hours a week on next year’s plan. They can’t see it in the margin. So when the budget tightens, the pilot with the soft number loses to the project with the hard one, every time.

McKinsey found that the small group of companies actually capturing earnings from AI do something different. They set cost-reduction targets alongside growth, and they rewire the workflow rather than bolt the tool onto it. The winners aren’t running better models. They’re measuring against a better number.

The right number is margin recovered

Not hours saved. Not seats filled. Margin recovered: the cost of inaction you avoided, the leakage you stopped, the penalty you prevented, expressed in dollars and payback months. It is the one number that survives a budget review, because it is denominated in the same currency as the budget.

Operations AI is almost uniquely suited to be measured this way, and that’s the part most companies miss. Marketing AI has to argue about attribution. Creative AI has to argue about quality. Operations AI argues about neither. It sits on cost lines you already track, against baselines you already know. You know what a chargeback costs. You know what a quality escape costs to contain. You know what you paid in overbilling last year, or you can find out in an afternoon. The dollar case isn’t a projection. It’s already in your books. You’re just not crediting the recovery to the tool that earned it.

And the margin at stake is not small. U.S. business logistics costs reached $2.6 trillion in 2024, about 8.7% of GDP, according to the Council of Supply Chain Management Professionals’ annual report. That isn’t a market size. It’s the size of the cost base operations AI is pointed at. A single point of recovery against it is real money.

Look at where it leaks.

Quality.
The cost of poor quality, scrap, rework, returns, warranty, and containment, is commonly estimated at 15 to 20% of sales for manufacturers, and higher for complex products. Most of it never appears on the standard dashboard. AI that catches a defect class before it ships isn’t saving inspection hours. It’s recovering the failure cost that defect would have become.

Penalties.
Retailer compliance programs turn a late or short shipment into a direct deduction. Walmart’s On-Time In-Full program charges 3% of the cost of goods on non-compliant cases, assessed quarterly and automatically against the supplier’s invoice. AI that protects the delivery window isn’t improving an ops metric. It’s stopping a 3% bleed at the exact point it would have hit the P&L.

Leakage.
Carrier invoices are wrong often enough that auditing them is its own discipline. Industry estimates put the share of freight invoices carrying billing errors in the high single to low double digits, and those errors rarely favor the shipper. AI that checks every line against the contract before payment recovers money that was walking out the door.

None of these is a productivity story. Every one of them is a margin story. And every one was sitting on the cost line the whole time, waiting for someone to measure it.

Change the metric before you change the tooling

So here is our position, plainly. Change the metric before you change the tooling.

Most AI engagements start with a use case. “Let’s do something with our invoices.” “Let’s try AI on quality.” We start somewhere else. We start at the cost line. Before the question of which model or which build, the question is this: where is this operation leaking margin, how much, and what is stopping it worth in dollars and payback months? Get that number right and the engagement has a spine. Skip it for a productivity proxy and you’ve built another pilot that works and dies anyway.

It’s also why the build-versus-buy evidence lands the way it does. NANDA found that partnering with specialists succeeds far more often than building from scratch internally, not because internal teams lack the skill, but because the discipline of starting from the outcome, and being measured on it, is hard to hold while you’re also running the operation. The margin number keeps everyone honest, including the partner.

The fair objection is that not all AI value is a hard dollar. Some of it is better decisions, faster cycles, fewer fire drills, a team that isn’t drowning. That’s real, and we’re not pretending otherwise. But it’s a second-order story, not the business case. Lead with the soft number and you lose the room. Lead with the recovered margin, the penalty you stopped, the overbilling you caught, the failure cost you avoided, and the productivity gains ride in behind it for free. The order matters. Put the bankable number first.

Operations AI rarely fails because the model was wrong. It fails because someone judged a roof repair by how fast the contractor worked. Change the number and most of the “AI doesn’t pay off” problem disappears, because it was never really an AI problem. It was a measurement problem.

At Amazatic, that’s where every engagement starts: the cost line, before the tooling.