Last updated: October 16, 2026
AI pilot ownership means one named person signs for the spend, accepts a measured baseline before launch, and answers on a fixed date whether the number moved. A pilot without that person is a recurring cost waiting for its renewal date.
Account 6140.
That was the answer I got when I asked a CFO at a direct lending manager who owned the firm’s three AI pilots. It was a general ledger line, “Software and subscriptions,” where a covenant extraction trial, a memo drafting assistant, and a borrowing base reader had been booked since the spring at a combined $31,500 a month. The variance report flagged the line as over plan. Nobody was flagged.
She laughed, a little.
Then she pulled the approval emails. Each pilot had been signed by a different person, each under the threshold that would have sent it to the investment committee, and none of the three emails named a number the pilot was supposed to move.
That is what a budget line looks like from the inside. It has a vendor, a monthly cost, and a renewal date. A project has those too, plus a person who will be asked, on a date everyone agreed to, whether it worked.
My engagements mostly come in as finance and accounting consulting work at direct lenders, ABL shops, and insurers, which is why most examples below involve a loan book. If your pilots are already running and you are deciding which ones survive, I covered the portfolio review for stopping AI pilots separately. This piece is the earlier conversation. How a pilot gets an owner before it starts.
What AI Pilot Ownership Actually Requires
AI pilot ownership is a commitment by one named person to fund a pilot from a budget they answer for, to accept a baseline measured before launch, and to report on a fixed date whether the target number moved. Sponsorship means liking the idea. Ownership means being the one asked.
Three conditions. Funded, baselined, dated. Miss any one and you are back to account 6140.
The market data says most firms miss at least one. In its 2025 CEO study, the IBM Institute for Business Value surveyed 2,000 CEOs across 33 countries and found that only 25% of AI initiatives had delivered the expected ROI over the previous few years, and only 16% had scaled enterprise wide. The same survey had 64% of those CEOs admitting that the risk of falling behind drives investment in some technologies.
Read those two numbers together. Fear signs the check. Nobody signs for the result.
How a Pilot Becomes a Budget Line
It almost never starts as negligence. A vendor offers a ninety-day trial priced just under the approval limit.
Someone in operations likes the demo and forwards it. IT gets asked to set up access, so the invoice goes to IT. The analyst who would actually use the thing is invited to the kickoff and then to nothing after it.
Three months later the trial converts to an annual term, because converting needs no meeting and canceling needs a reason.
I have watched this happen at firms with excellent controls everywhere else. Their borrowing base process has four sign-offs. Their AI spend has none, because it arrived in pieces small enough to avoid the process that would have asked the obvious questions.
The accounting rules, oddly, already ask them. Under FASB’s ASU 2025-06, which rewrote the internal-use software guidance in Subtopic 350-40, a company starts capitalizing software costs only once management, “with the relevant authority,” has authorized and committed to funding the project, and it is probable the project will be completed and the software used for its intended function. Until then, costs are expensed as they come in. The standard’s own examples of commitment are things like approving a budget and allocating a team.
I am not your auditor, and most pilot spend should be expensed anyway. My point is narrower.
Your controller already has to decide whether someone with authority committed to funding each software project and whether it will probably be finished, for the financial statements. Your pilot register usually does not ask the same question at all. Ask the controller which side of that line each pilot sits on, and listen to how long the answer takes.
The Owner Sheet
Every pilot I start now begins with one page. Nothing gets provisioned and no vendor gets a data extract until a person has filled it in and put their name on it.
Here is the page, with the failure I see most often on each line.
| Line | What goes in it | What usually goes in it instead |
|---|---|---|
| Owner | One person’s name | “Ops,” “IC,” or a working group |
| Cost center | The owner’s own budget | IT, innovation, or general software |
| The number | One metric the owner already reports upward | “Efficiency” |
| Baseline | That metric, measured on the last two or three periods, before launch | An estimate from the vendor’s ROI calculator |
| Measured by | Someone who does not benefit from the answer | The vendor’s customer success team |
| The work today | The person who would do the work, by name, and their hours on it per month | Blank |
| Review date | A day on the calendar, after one full reporting cycle | “End of pilot” |
| Stop condition | The result below which the owner cancels, written now | Nothing, because nobody wants to write it |
The cost center line does most of the work. When the pilot is charged to the owner’s budget, the owner starts asking the vendor hard questions within a week. When it is charged to IT, the owner is a guest.
The baseline line is where good intentions die, because measuring last quarter’s covenant testing hours honestly means someone has to reconstruct them from calendars, email threads, and the analysts’ own memory of a busy month, and that takes a few days nobody budgeted. Do it anyway. A pilot measured against a guess will always look like it worked.
And the stop condition? Write it while everyone is still optimistic. Owners who agree in advance that “under a 20% cut in review hours by the June close, we cancel” actually cancel. Owners asked to decide after six months of sunk cost almost never do.
If you want to know whether your firm can fill in this sheet at all, the AI readiness scorecard for mid-market operations asks a version of the owner question in its first section.
Six Functions, Six Signatures
I learned most of this running a phased rollout of Claude, GPT, and Copilot at a specialty finance investor, function by function rather than firm-wide. Before each phase, the head of that function filled in a sheet much like the one above. The phase did not start until it was signed.
One function waited almost a quarter. Its head liked the tools and would not put them in her budget, and I respected that more than I expected to. When she did sign, she had already decided which two workflows mattered and which four did not.
Engineering went first because its owner had a baseline nobody had to reconstruct. The team already tracked delivery against the same backlog every sprint, so we froze two quarters of that history before anyone got a license, and agreed with the head of engineering that the vendors would not be the ones counting. The measured lift, after the governance rules and the role-based training were in place, was a 15% gain in engineering productivity.
Fifteen percent will never make a vendor case study. It held, though, and the person who reported it to the board was the person whose budget had paid for it.
That last part mattered more than the figure. When the firm later raised institutional capital, the operational due diligence questions about AI had an answer with a name on it for every function. The investors’ questions were about who answered for each tool and how we knew it worked. Six sheets covered both.
When Nobody Will Sign
Sometimes the sheet stays blank. That is information, and it is cheaper to get it before launch than after.
A blank owner line usually means one of three things. The pilot solves a problem nobody in the firm actually has, which a demo can hide for months. The problem is real but spans two functions, so each head assumes the other one owns it. Or the would-be owner has no hours, which is the most common and the most fixable.
The two-function case needs a decision from above, once, in writing. Borrowing base review is the classic example in credit, sitting between portfolio operations and finance, and a pilot aimed at it will drift until the COO says whose it is.
For the no-hours case, the honest move is to fund the hours before the tool. An owner who signs for a pilot and then has no one to run the baseline, sit with the vendor, and score the review date is an owner on paper. Plenty of the managers I work with bring in a data engineer through contract staffing for the build period, or a fractional chief data officer to run the sheets across functions, and decide on permanent seats after two reporting cycles. KORE1’s average time to fill an IT role is 17 days, shorter than most pilot trials.
If none of the three explains the blank line, do not launch. Put the vendor on a dated expense with a hard cap, call it a trial, and let it end. That is still a budget line. At least it is one with an expiry.
What the Owner Owes on the Review Date
One page again, shorter. The number against the baseline. The hours the person who would do the work actually spent, including the hours spent checking the output. A one-sentence verdict (keep, change, or stop) with the owner’s name under it.
That sentence is the thing your board, your LPs, and any operational due diligence team will eventually ask for, usually phrased as “how do you govern AI.” Firms that can hand over a stack of signed review pages have a short meeting. The pilot-to-production checklist for AI vendor contracts covers the contract side of the same discipline, including the clause that lets an owner actually walk away on the review date.
It is also the moment the pilot either becomes a project or gets honest about being a cost. Both are fine. The only bad outcome is a third one, where the review date passes and nobody notices because nobody owned it.
Readiness, in the end, is an operations property. I made the longer version of that argument in why AI readiness is an operations question, and the owner sheet is the smallest unit of it.
What Owners Ask Me Before They Sign
Can IT hold the budget while the business owns the result?
It can, but in practice the result drifts toward whoever holds the money.
Split ownership works only if the business owner can cancel the IT spend unilaterally on the review date. If canceling needs IT’s agreement, IT is the owner and the business head is an advisor. Better to move the cost and keep IT as the operator.
The owner left halfway through. Now what?
Treat it as a new pilot with a new sheet, signed by the successor within thirty days.
If nobody will sign it, the pilot ends on its next review date. Inheriting someone else’s stop condition is fine. Inheriting a pilot with no stop condition at all is how account 6140 happens.
Does the vendor’s customer success manager count?
No.
They are paid on renewal. Useful people, often genuinely helpful in the first month. Just not owners.
How do you keep a productivity number honest?
Freeze the baseline before anyone gets access, and have someone outside the pilot count.
Productivity claims fail in two predictable ways. The baseline gets estimated after the fact, generously, or the people using the tool are the ones measuring it, and they like the tool. Use a metric the team already reports for other reasons, such as close days, covenant tests completed per analyst, or tickets delivered per sprint, so nobody has to invent it for the pilot. Then say the result once and move on.
Our pilot costs $2,000 a month. Is a sheet really necessary?
Small pilots are exactly where the sheet pays for itself, because nothing else will ever review them.
The license fee is rarely the real cost. The real cost is analyst hours spent checking output, plus the data extract someone built for the vendor and now maintains. Those never show up on the invoice.
The board asked who owns AI. What do we say?
Name a person for each use case, and show the board their signed sheets.
A single “AI owner” usually ends up owning the vendor list and none of the results. What a board or an LP wants to hear is that every live pilot has an owner, a baseline, and a date. Bring the pages to the meeting.
Put a Name on Line One
Pull your three most expensive AI pilots from the general ledger, not from the project list. For each one, write the owner’s name on a blank sheet and send it to that person to fill in the other seven lines by the end of the month.
Whichever sheet comes back blank is your answer. If what’s missing is hours rather than will, talk to KORE1 about staffing the work behind the pilot. And if you want to compare sheets with someone who has filled in a few dozen, I’m on LinkedIn.
