An audit of one Fortune 500 division found more than 150 fixable inefficiencies. Most were not AI problems.

An audit of one Fortune 500 division found more than 150 fixable inefficiencies. Most were not AI problems.

A systematic opportunity audit inside one division shows where AI operationalization actually starts, and it is not with a model.

A division of a Fortune 500 manufacturer set out to map its operational inefficiencies and match them to AI fixes, treating the exercise as its first real test of enterprise AI operationalization. The mandate given to the review team was ordinary enough: catalogue every recurring inefficiency across the division's operations, then decide what AI could do about each one. The team worked through purchasing, planning, fulfillment, and customer service line by line, logging more than 150 fixable inefficiencies before the review closed. When the list was sorted by what would actually respond to an AI system, most of it did not move. The gap between what executives expected and what the spreadsheet showed became the real finding.

What the audit actually found

The review team logged recurring themes across the division: duplicate approval steps in procurement, manual reconciliation between two systems that had never been properly integrated, a scheduling process that ran on a spreadsheet nobody owned. Most of these themes had no AI candidate attached to them at all. They were process problems, ownership problems, and integration debt built up over years, the kind of mess no model can reach.

Of the smaller set that did look like AI candidates, fewer still had usable data behind them. Only a narrow slice of the division's systems held information clean, current, and rights-cleared enough for an AI system to draw on safely today. That distinction, between an inefficiency a model could theoretically help with and one it could actually be pointed at this quarter, is where most enterprise AI transformation programs quietly stall.

Four buckets, only one of them AI-native

The team sorted the full list into four buckets. Some items just needed to be fixed the ordinary way: a policy rewritten, a step removed, an owner assigned, no software involved. A second group needed the process redesigned first, with AI layered on only once the new process existed, because automating a broken sequence just makes the mistake happen faster. A third group was genuinely AI-native, problems that did not exist in this form before AI made a real-time, high-volume judgment call practical.

The fourth bucket is the one budgets never plan for: things that should simply stop happening. A report nobody reads. A reconciliation step that exists because two systems were never properly connected years ago. Eliminating those items produces no software to demo and no vendor to credit, which may be exactly why they survive audit after audit.

Sequencing enterprise AI operationalization around governance first

Sequencing matters as much as identifying the individual fixes. Stop the bleeding comes first: the conventional-fix bucket and the easiest of the eliminate items, work that returns value in weeks and buys the credibility to keep going. Build the substrate comes second, the integration and data cleanup the redesign-then-AI bucket depends on, unglamorous work that rarely gets funded on its own. Re-platform and intelligize comes last, once a clean foundation exists underneath it: the AI-native bucket, plus whichever redesigned processes are now safe enough for a model to sit inside.

As Future Works founder Matt Leta has put it, "Eliminate is the hardest bucket to fund because the savings land in someone else's P&L. Nobody owns the budget for work that stops existing, so it waits until a CFO decides it's theirs." Running all three waves inside one static, multi-year program plan is exactly how value gets promised early and delivered late. A 12-week cycle, with fees staked to the outcome rather than fixed to the scope and the client's own finance team validating a value target at the end of each one, keeps every wave honest before the next one starts. The pairing behind that discipline, AI agents plus named experts rather than a rotating bench of consultants, is what an AI-native transformation engine brings to AI operationalization done in this order.

Why enterprise AI operationalization stalls the same way everywhere else

None of this is unique to one manufacturer. Roughly 88% of AI pilots never reach production, according to IDC research, and the audit above helps explain why: most of what gets logged as an "AI opportunity" in a workshop is not actually AI-shaped once someone checks the data underneath it. The pilots that do reach production tend to sit on that same narrow slice of clean, current, rights-cleared data, smaller than most transformation budgets assume.

The uncomfortable part for anyone running a similar review inside their own division rarely involves AI capability at all. Three of the four buckets, fix conventionally, redesign, eliminate, come down to discipline: an owner assigned, an integration cleaned up, a report nobody reads killed off. A CIO who wants an enterprise AI transformation program to survive a board's budget cycle should expect the AI headline items to be the smallest bucket on the page, not the largest.

None of the individual fixes here required deep technical novelty. What the audit forced was discipline about sequence and ownership, the least exciting part of any transformation program and the part most often skipped when a board wants an AI slide by the next review. The list itself made clear that most of the division's opportunities were never going to need AI at all, and funding a pilot for a problem a policy change would have solved for free is how transformation budgets quietly leak.

If a similar review is on this year's roadmap, the bucket a fix lands in tells you more than the AI label attached to it. Start with a pilot.

Matt Leta, Founder and CEO, Future Works.