AI Lease Accounting: What the Vendors Are Actually Adding

When a lease accounting platform announces AI today, it is nearly always the same move: a capability added to an existing product, and the product then described as AI-powered.

That is not a criticism in itself. Several of these additions are useful, and a vendor improving what it ships is what you want from a vendor. But it creates a problem for the controller sitting through the demo, because “we have added AI” covers at least three different things — and only one of them has any bearing on the number that will land in your financial statements.

Worth getting the distinction straight before your next renewal conversation. Here is the one our practice works from.

Three things the word is covering

AI that reads. Lease abstraction — pulling commencement dates, payment ladders, renewal options and termination clauses out of a scanned contract. This is the most mature application, it works well, and it removes real drudgery. But note what it produces: inputs. An abstraction engine hands you a populated field, and every downstream measurement question — classification, discount rate, term judgement — remains entirely yours. Useful, and not accounting.

AI that explains. A conversational layer over data you already hold: summarizing a portfolio, drafting a disclosure narrative, answering “what changed on this lease.” This is the category most often demonstrated, because it shows well. It is also the category most likely to be mistaken for the one below it. An explanation layer describes the accounting; it does not perform it.

AI that builds. Directing agentic AI to construct the calculation and control machinery itself — the engine, the schedules, the posting logic, the tie-outs, the role enforcement. This is the least discussed of the three and, in our view, the most consequential, because it changes what a finance function is capable of producing on its own rather than what it can buy.

The first two are features. They sit on top of an existing calculation engine, they can be added to a platform already running in production, and adding them is precisely what the market is doing right now.

The third is not a feature. It is how the system underneath was built — which is why it does not arrive in a release note, and why no amount of bolting on produces it.

The line that decides everything: computed, not narrated

Here is the principle we hold to without exception.

A language model must never be the thing that produces a balance.

Large language models are probabilistic. They are extraordinary at reading, drafting, explaining, and reasoning about accounting. They are the wrong instrument for arithmetic that has to tie to the cent and reproduce identically next quarter, because the same question asked twice need not return the same answer — and a lease liability that moves between runs is not a lease liability.

In the subledger our practice built, the numbers come from a deterministic engine. Cells hold formulas evaluated by a formula engine and cached by cell-content hash. The same inputs return the same output, every time, by construction. AI directed the building of that engine; AI does not compute the balance it produces.

That is the whole distinction, and it is the first thing worth establishing about any platform announcing an AI capability. Ask where the number comes from. If the answer is a model, the conversation is over. If the answer is a deterministic calculation the model helped build or helps you interrogate, you are in a different conversation entirely.

Three questions that separate the two

Once a vendor tells you the number is computed rather than generated, the follow-up is how they know it is right. These are the three we ask.

1. Can it be reproduced? Not “is it tested” — can a change be proven not to have moved anything it should not have moved? In our build, every change is gated to zero cell differences against roughly one hundred anchor workbooks before it ships. Correctness is a gate the change has to pass, not a hope it is held to.

2. Is it proven twice, independently? Wherever a disclosure can be computed without reference to the trial balance, it should be, with a “difference must equal zero” line proving agreement. The discipline that matters here is negative: tie-by-construction is forbidden. A total that agrees with itself because it was derived from itself has proven nothing. Every balance in our subledger is proven three independent ways — the cached extract, a fresh re-summation of the journal lines, and each workbook’s own trial balance — and where a road is unavailable it renders as unavailable rather than as a fabricated zero.

3. Does it replicate the profession’s own published examples? This is the test that changes the character of an audit conversation. Our engine reproduces the specific figures published in the FASB codification illustrations, the KPMG handbooks, the Deloitte modification examples, the BDO practice aid and the EY FRD — 51 anchors across six families — to the cent, with every source-rounding divergence documented as a tolerance rather than absorbed. To be precise about the boundary: those sources publish an initial measurement and selected milestone balances, not complete lifecycle results, so the full lifecycle output is the engine’s own standard-compliant computation rather than something validated against a published figure.

A platform that cannot answer all three is not necessarily wrong. But you now know exactly which parts of its output you will have to prove yourself.

Where the human stays in charge

“Built with AI” raises a fair governance question, and the answer belongs in the architecture rather than in a policy document.

In our subledger, five server-enforced roles carry the segregation of duties. The AI agent can read and capture — book leases, pull the trial balance, run the tie-outs — through the same governed door an ERP integration uses. Period close is human-gated: a machine credential attempting to close a period receives the same 403 refusal a junior accountant would, and the denial is logged as audit evidence. The agent surface exposes no close tool at all, so the control point is not merely blocked to the machine; it is absent from the machine’s action space.

Every action, human or machine, lands in a row-level change log — who, which field, the value before and after, and when — with machine actors attributed to their real principal rather than to “system.”

We hold a firm view here: agentic AI belongs inside the control framework, not around it. Any vendor whose AI story does not begin with what the machine is forbidden to do has not thought about the part your auditor will ask about.

What this actually means for your close

The reason the third category matters — AI that builds — is that it moves a capability that used to sit with software vendors into reach of a finance function directing it well.

Our practice built a working dual-standard subledger this way: ASC 842 and IFRS 16 in parallel from one lease population, capture through close and ERP posting, with agentic AI under CPA direction and no line of code written by hand. That was not achievable on this timeline, at this cost, even two years ago. We wrote up what it is and how it is proven in a companion piece, and you can watch the engine run live.

It is not a product we sell. It is proof that the discipline making an AI-built system trustworthy is accounting discipline — tie-outs, segregation of duties, a regression gate, a change log — rather than software discipline. Which is precisely why the work needs a CPA directing it.

So the next time a platform walks you through the AI it has added, three questions will tell you what you are being shown. Where does the number come from. How is it proven. What is the machine forbidden to do.

If those are the questions you have been asking and not getting answers to, start a conversation — or see how we work with finance teams.

The FASB and the accounting firms named above are identified only as the source of published illustrative fact patterns. They do not sponsor, endorse, or review this application of them. Provided for information, not as accounting advice.