There's a story about AI that's quietly writing its own headline, and it isn't the one in the press releases. It doesn't matter which industry you're in or which department you sit in. Talk to a controller in health care, an IT director at a data center, a tax manager at a manufacturer, a facilities lead at a university or a product team at a software company, and you'll hear the same thing. The tools have been bought. The results haven't arrived. And the team is checking everything.

I've been hearing versions of it for the last few years, in conversations with clients, across the software and technology world I've worked in, and in much of what I've been reading and listening to. The Greenbuild AI Summit in New York this month was one more confirmation. I learned a few things there, but mostly I heard the same story again, all day, from people who had never met and work in fields with little in common.

The details change from room to room. The three scenarios underneath them don't, and together they explain why so many finance teams don't trust AI yet and end up checking every answer it gives them:

  • The data isn't ready, so AI doesn't return what people expected.
  • People aren't sure how to work with it: what to ask, or what it could do for a process.
  • There's a real question of trust: is our data safe, and who is actually in charge of the decision?

None of them are technology problems, and all three can be fixed.

Fixed asset data that isn't ready for AI

This is the concern I hear most, and it makes complete sense. AI reads a fixed asset register faithfully. If the register is wrong, the answers will be wrong too, delivered with confidence and good grammar.

Most registers carry the same kinds of things:

  • Servers pulled out in a data center refresh that are still on the books and still on the insurance schedule.
  • CIP lines with no project behind them.
  • Laptops and network gear that were expensed in one department and capitalized in another.
  • Assets in the wrong class with the wrong life, because they were keyed in by hand at the moment they were least verifiable.

None of it is anyone's fault in particular. It's what happens when IT, operations, projects, accounting, insurance and tax each hold a piece of the truth and nobody owes anybody a message. (I wrote about one version of that hand-off in The Disposal Form That Never Made It to Accounting.)

The research says this is widespread. A Datarails survey of 270 CFOs at large U.S. companies, released in October, found that only 4% have a single source of truth for their finance and operating data, and 65% said their most common AI frustration is confident answers built on the wrong data. Before AI can help with the register, someone has to look at whether the register deserves to be believed. That is the first question a Fixed Asset Health Assessment answers.

Knowing what to ask: AI prompts for fixed asset work

For years, much of the value an experienced person brought to fixed assets was knowing which buttons to press: which report, which field, which setting nobody had touched since go-live. That still matters, but it isn't the whole job anymore. With AI, the value is in knowing what to ask, what to give it, and how the process around it needs to change.

Most people start with a one-line question, get a vague answer, and decide AI isn't ready for fixed assets. The difference between that and something useful is usually the setup. Compare “Is this a capital expense?” with something closer to this:

A capital vs. repair prompt

Here is our capitalization policy and the invoice for the rebuild of packaging line 3 at our Ohio plant. Using only the policy, tell me whether this work extends the line's useful life, adds capacity, changes its classification, or only restores it to its prior condition. Quote the section of the policy you relied on. If the invoice doesn't give you enough to decide, say “can't tell” and list what you'd need.

The second version gives the AI the rules, asks it to show its reasoning, and gives it permission not to know. It also leaves the decision where it belongs, with the person reading the answer. The same approach works for CIP aging, closeout questions for project managers, and checking disposals against the insurance schedule and the property tax rendition, which are places where most fixed asset systems have no automation at all.

Trust, data security and who makes the decision

The same survey found finance teams spend about 26% of their working week verifying or correcting AI output. Only 5% of CFOs would trust AI to produce a board-ready report without review, and 4% would trust it with the month-end close.

I don't read that as a failure. I read it as finance doing its job. The goal was never to trust AI blindly. The goal is to design the work so a named person makes the important decisions, the AI's part is visible and traceable, and nobody has to recheck everything because nobody knows what was already checked. A capital authorization releases funds; it isn't permission to act. An AI finding is a finding, not a decision.

There's a practical reason to sort this out now. PCAOB amendments on technology-assisted analysis apply to audits of fiscal years beginning on or after December 15, 2025, so auditors will be asking how AI touched the numbers at this year-end. “Someone ran it through Copilot” is not going to be a satisfying answer.

Why this matters more than it used to

For as long as I've done this work, the same problem has outlived every new system. The information about an asset lives in software, in spreadsheets and in people's heads, and those places almost never meet. The project system, the register, the insurance schedule, the property tax file and the facilities manager each know something true, and for most organizations, bringing it together has simply been too expensive.

That has changed. With the right questions and a clear idea of where it fits, AI can connect those pieces inside an organization, using tools it already pays for. It's also becoming more urgent, because accounting teams are turning over faster, and every departure takes some of the understanding with it. The person who knew why a CIP project was never closed, or which assets were really retired, is often the only record there was.

This is where my own work has been heading. Rather than handing over a report, I'd rather sit down with a team, look honestly at where things stand, organize the data and the process together, and build the tools alongside them in their own systems, so they own them, understand them and can update them. Some teams want help keeping things going afterward; others are glad to take it from there. Either way, the knowledge stays inside the organization. There's more about how we work on the site.

Three problems, one piece of work

It's tempting to treat these as three projects: a data cleanup, a training program and an AI policy. In practice they're one piece of work:

  1. Clean the data where the AI will be used.
  2. Write the questions around your own policy.
  3. Decide in advance which calls a person has to make.

Do that for one process, such as CIP closeout or disposals, and the team can stop checking everything, because they finally know what has been checked. It's the same reason it pays not to automate a broken process.

Most companies have already paid for the AI. The licenses are sitting there. What's usually missing is the fixed asset judgment around them, and that part has always taken people and time.

If you'd like to know where your own register stands before year-end, there's a free 12-question AI readiness checklist on the site. It takes about ten minutes, and it tends to start a useful conversation with the team.

Angela Bolton is the founder of Fixed Asset Consultant. She has spent close to three decades in fixed asset and capital project work, including more than a hundred construction-in-progress implementations.

References