IFRS 16 and ASC 842 moved most leases onto the balance sheet, and in doing so they quietly created a data extraction problem. A right-of-use asset and a lease liability are computed numbers, and every input to that computation lives in an unstructured document: commencement and end dates, payment amounts and timing, renewal and termination options, escalation clauses, and the evidence relevant to a discount rate. Multiply by a portfolio of hundreds or thousands of contracts in heterogeneous formats, add amendments negotiated over years, and the accounting standard becomes, operationally, a document processing standard.
Most organisations answered it in 2019 with a one-time abstraction project: teams of people reading contracts into spreadsheets, and the spreadsheets feeding the transition numbers. The trouble is that lease accounting is not a one-time event. Contracts get modified, index-linked payments change, options get reassessed, and each event triggers remeasurement. A spreadsheet built during transition decays with every amendment nobody keyed in, and the audit trail from a computed liability back to a contract clause was often never built at all.
The current generation of lease tooling treats the problem as a pipeline: ingestion, abstraction, accounting judgment, calculation, remeasurement, disclosure, and audit evidence. The stages are worth keeping separate, because the automation potential and the risk profile differ sharply between them.
Abstraction is the high-value AI problem
Extracting lease terms from contracts is where AI genuinely earns its keep. The documents are long, the formats vary by lessor and by decade, and the fields that matter are scattered: a commencement date in the recitals, payment schedules in an exhibit, a renewal option in a clause that cross-references another clause. OCR plus language models now pull commencement and end dates, payment amounts and timing, renewal and termination options, and discount-rate-relevant inputs from source contracts with enough reliability to change the economics of abstraction, especially for high-volume portfolios.
Two design rules keep the output trustworthy. First, extract to a schema, not to prose: every field lands in a lease register with a defined type, a confidence score, and a pointer to the exact passage in the source document that produced it. The pointer is what turns extraction into evidence; a value with no anchor in the contract is a keying error waiting to be discovered in audit. Second, route by confidence and contract type. A standard-form equipment lease can flow to the register with sampling-based review. A heavily amended master lease routes to a human abstractor with the extracted candidates pre-filled, which is still far faster than reading from scratch.
The judgment that cannot be delegated to extraction sits one step earlier: whether the arrangement contains a lease at all. Embedded leases inside service and supply agreements remain the classic gap, because the question is one of control over an identified asset, not of finding a field. Extraction can flag candidate language and asset references that suggest an embedded lease; the determination belongs to an accountant, and the workflow should record who made it and why.
From extracted terms to defensible schedules
Once terms are structured, the calculation layer is deterministic and well understood: measure the lease liability as the present value of future payments, recognise the right-of-use asset, generate amortisation schedules and journal entries. Platforms handle this reliably, and it is the least interesting part of the pipeline precisely because it is pure rules.
The lease term deserves particular care at the boundary between extraction and accounting. The term is the noncancellable period plus optional periods where exercise is reasonably certain. Extraction supplies the option structure; reasonable certainty is an assessment, revisited over time, and it changes the liability materially. Keep the two apart in the data model: the contract facts as extracted, the assessments as separate, effective-dated judgments with an owner. Collapsing them into a single “end date” field is how a reassessment turns into an untraceable edit.
Discount rates follow the same principle. The rate is a policy input, usually the incremental borrowing rate, and a frequent source of manual override. It belongs in a governed, effective-dated table in the accounting layer, not embedded in extracted data, so that rate changes propagate through remeasurement in a controlled way.
Dual reporters need one more layer. IFRS 16 uses a single lessee model; ASC 842 retains operating and finance classification for lessees. The workable architecture is a shared contract repository and a single abstraction pass, feeding standard-specific policy layers with their own elections, classification, schedules, and disclosures, and a standing reconciliation between the two. Rebuilding the contract data per standard doubles the extraction cost and guarantees the two views drift.
Modifications are the operational test
A lease portfolio is never static, and modification handling is where automation programmes prove themselves or quietly fail. The pain is only partly detecting that something changed. The harder work is determining what the change means: whether a revised term applies, whether a new discount rate is required, whether the change in scope or consideration triggers remeasurement or is effectively a new lease, and then flowing the answer through revised schedules, journals, and disclosures while preserving the history.
Modern workflows automate the triggers: an amendment lands in the repository and is diffed against the abstracted terms, an index update arrives and flags the index-linked payments it affects, a reassessment date comes due and opens a task. The classification of the event stays with an accountant, with the system presenting the before-and-after terms side by side. What the system must guarantee is the history: every version of the terms, every assessment, every rate, and every schedule that ever produced a booked number remains reproducible. Remeasurement without version history is how a portfolio becomes unauditable one amendment at a time.
The evidence trail is the deliverable
The reason to run this as a pipeline rather than a project is the audit posture. Every number on the balance sheet should trace back through the calculation to structured terms, and from each term to the passage in the source contract, with the assessments and policy elections that shaped it recorded alongside. When that chain exists, an auditor’s sample turns into hours of clicking through rather than weeks of requests; when it does not, the team re-performs abstraction under deadline, which is the most expensive possible way to read a contract twice.
The costs worth budgeting honestly: an initial backfile cleanup, because extraction over a decayed register surfaces years of unrecorded amendments; ongoing review capacity for low-confidence extractions and modification classification; and governance of the rate table and option assessments. The failure modes are equally predictable: confidence thresholds set high enough that nobody reviews anything, embedded leases nobody went looking for, and a register that drifts from reality because amendments bypass intake. The mitigation for the last one is organisational, not technical: no amendment gets signed without landing in the repository, because the pipeline can only account for the contracts it has seen.
FAQ
How much of the lease portfolio can extraction actually handle without human touch?
The clean majority: standard real estate and equipment leases in familiar formats, where dates, payments, and options sit in predictable clauses. The residual set that needs people is smaller but carries most of the risk: embedded leases inside service agreements, negotiated amendments layered over decades-old master leases, and contracts where the term depends on option assessments. Plan the workflow around routing by confidence and contract type rather than around a single automation percentage.
Should the discount rate live in the extraction layer or the accounting layer?
The accounting layer. Extraction can pull rate-relevant inputs from the contract, such as any implicit rate evidence and payment structure, but the incremental borrowing rate is a policy-governed input that depends on entity, currency, term, and date. Hard-coding rates at extraction time buries a judgment call in a data field. Keep rates in a governed table with effective dates, and have the calculation engine pick them up so a rate update flows through remeasurement correctly.
What breaks first when a company runs IFRS 16 and ASC 842 in parallel?
Usually the assumption that one set of schedules can serve both standards with cosmetic adjustments. IFRS 16 applies a single lessee model while ASC 842 keeps operating and finance classification, so expense profiles, and therefore reconciliations, diverge by design. The sustainable setup is one contract repository and one extraction pass feeding two policy layers with separate elections, classification logic, and disclosure outputs, plus a standing reconciliation between the two views.