Where AI Actually Helps With Lease Accounting Under ASC 842

The FASB looked at its own leases standard and concluded that the cost of implementing and applying it ran significantly higher than the Board expected when it issued the thing, which is a remarkable admission from a standard setter and also the least surprising sentence any controller will read this year. Anyone who has actually done an ASC 842 adoption knows exactly where that cost went - it went into hunting down embedded leases in service contracts, into abstracting terms out of PDFs somebody scanned sideways in 2019, into arguing about incremental borrowing rates with an auditor who wanted documentation you did not have. That grind is where AI belongs, and the roles we see hiring for technical accounting across our database of 80,000-plus accounting and finance postings keep listing lease accounting as a named responsibility, which tells you the work never actually went away after adoption.

What does ASC 842 actually demand that AI can help with?

The standard's core requirement is simple to state and brutal to operationalize: under ASU 2016-02 a lessee puts a right-of-use asset and a lease liability on the balance sheet for essentially every lease longer than twelve months. Stating it takes one sentence. Doing it means you need a complete population of contracts that contain a lease, a defensible term for each one including which renewal options are reasonably certain, a discount rate per lease, and a disclosure package that ties out and rolls forward every period. Three of those four things are document work and arithmetic, and document work and arithmetic is precisely what the current generation of AI tooling is good at. The fourth is judgment, and that is where you keep your hands on the wheel.

Where AI earns its keep first: finding the leases you do not know you have

Embedded leases are the single best place to start, because the problem is a search problem dressed up as an accounting problem. Somewhere in your contract folder there is a logistics agreement that gives you the exclusive use of specific trailers, a data centre agreement that names specific cages, an equipment service contract where you actually control an identified asset - and none of them say the word lease anywhere in the document. Historically you found these by having a senior person read hundreds of contracts, which is expensive and unbelievably boring and therefore done badly. Point a document-extraction model at the full contract population instead, have it flag every agreement containing an identified asset with a right to control its use, and have it produce a citation to the specific clause it relied on. You are not asking the model to conclude anything, you are asking it to shrink a population of 900 contracts down to the 40 a human should actually read. That is a real hour-for-hour saving and the failure mode is benign - a false positive costs you five minutes of reading.

What about the ongoing grind after adoption?

Lease abstraction and remeasurement monitoring are the two that quietly eat a technical accountant's month. Abstraction is pulling commencement date, payment schedule, escalation clauses, renewal and termination options and residual value guarantees out of a document and into structured fields, which is a straight extraction task with a verifiable answer, so run it with the model and have a human review the extracted fields against the source clause rather than retyping them. Remeasurement is subtler and more valuable - a modification, a change in the assessment of a renewal option, an index-based payment reset all trigger a remeasurement, and the reason teams miss them is that nobody is reading the contract amendments as they arrive. Set up an agent that watches for new or amended lease documents, compares the abstracted terms to what is in the lease subledger, and raises a flag when they diverge. That is a control improvement and a time saving in the same move, which is rare enough to be worth building.

Where would I keep AI out of it entirely?

The discount rate, the lease term conclusion, the classification call and the disclosure sign-off. All four are judgments where the answer depends on facts and intentions that live outside the document, and where a confident wrong answer is far worse than no answer. The incremental borrowing rate depends on your actual credit profile, the collateral, the term, and what your lender would genuinely charge you - a model will happily generate a plausible-sounding rate with no basis in any of that, and your auditor will take it apart. Whether a renewal option is reasonably certain of exercise depends on economic incentives, leasehold improvements you have made, and what management actually intends, which is a conversation with the business, not a document parse. Finance and operating classification hangs on those same term and rate inputs. So use AI to assemble the evidence for those calls and to draft the memo once you have made them, and make the call yourself. This is the same line we drew in how AI interacts with SOX controls in the close, and it holds here for the same reason.

How do you run this without blowing up your controls?

Treat every AI output as a prepared-by, never as a reviewed-by, and document the review the same way you would for a junior's work. Practically that means three things. Keep the model's citation to the source clause attached to every extracted field so a reviewer can verify without reopening the contract, because the thing that makes AI defensible in an audit is traceability rather than accuracy claims. Log the prompt, the model version and the date alongside the workpaper, since your auditor will eventually ask how the output was generated and hand-waving is a bad look. And run a sample test before you trust the pipeline - take 30 leases you have already abstracted by hand, run them through, and measure the error rate honestly, because if it is 4% you know exactly how much review to apply and if it is 20% you have learned something cheap. That sampling discipline is what separates teams who get real leverage from teams who quietly create a mess, and it is the same approach that makes AI-assisted audit prep hold up under scrutiny.

What I'd do

Pick your contract population and run the embedded-lease sweep this quarter, before anything else. It is the highest-value, lowest-risk application on the list, it produces a completeness result you can hand to your auditor, and it will almost certainly surface at least one agreement nobody had identified, which is the kind of finding that pays for the whole exercise. Then abstract the leases you already have into structured fields with the model and reconcile against your subledger, which doubles as a data quality audit you have probably been avoiding. If your lease software cannot ingest structured output or expose an API, that is a real signal about the tool, and it is worth seeing what else is out there in our accounting and finance software directory before you renew. The teams pulling ahead on this are running one workflow at a time and measuring, which is exactly the pattern we described in rolling reconciliations.

Frequently asked questions

Can AI do ASC 842 lease accounting?

It can do a large share of the work around the judgments - identifying contracts that contain a lease, extracting terms into structured fields, monitoring for remeasurement triggers and drafting disclosure narratives. It should not make the discount rate, lease term or classification determinations, because those depend on facts and management intent that live outside the document.

What is the best use of AI for lease accounting?

Finding embedded leases. It converts a task where a senior person reads hundreds of contracts into a task where they read the forty the model flagged, with a citation to the clause that triggered each flag. The saving is large and a false positive costs almost nothing.

Why was ASC 842 so expensive to implement?

The FASB's own post-implementation review found initial and ongoing costs ran significantly higher than expected at issuance, driven by identifying embedded leases, extracting data from contracts, determining discount rates and assessing renewal options. Private companies felt it hardest because of limited staff and system capability.

Does using AI in lease accounting create an audit problem?

Only if you cannot show your work. Auditors care about traceability and review, so keep the citation to the source clause with every extracted field, log the model and prompt alongside the workpaper, and document human review of every output. Treat AI as a preparer whose work gets reviewed.

How do I know if my lease accounting software is worth keeping?

Ask whether it can ingest structured data from an external extraction step and whether it exposes an API. Tools that force manual re-keying cap how much leverage you can get from any AI workflow, and that limitation compounds every reporting period.

Lease accounting is a good test case for the whole question of AI in finance, because the judgment and the drudgery are cleanly separable and the drudgery is enormous. Take the boring 80%, automate it deliberately, measure what you saved, and keep the calls that require an accountant. Then go do it again on the next workflow.