Reading a 40-page commercial lease to pull out the dozen data points that matter for the accounts is a task most accountants have done by hand more times than they care to count. It is unloved work, error-prone late at night, and consumes the kind of senior time that is much better spent on the client conversation itself. It is also the single most automatable task in a typical accounts engagement once a serious AI document tool is in the stack.
This walkthrough shows how to extract the structured data points from a commercial lease in under two minutes using AI document chat, with the citations and audit trail a working paper file requires. It is written for a partner or senior who has the tool available and wants to standardise the workflow.
The data points worth extracting
A working extraction from a commercial lease for accounts purposes captures the same set of points every time. Standardising the list - and the order - is what lets the extraction be done in seconds rather than minutes:
- Landlord and tenant (and any guarantors)
- Demised premises and the address
- Term commencement and term expiry
- Break dates and the notice required to exercise
- Initial rent and rent payment frequency
- Rent review dates and the review mechanism (open market, RPI, CPI, fixed)
- Service charge basis and any cap
- Insurance rent and recoverable insurance arrangements
- Repair and decoration obligations (full repairing and insuring, internal only, schedule of condition)
- Alienation provisions (assignment, sub-letting, sharing with group companies)
- Permitted use and any user restriction
- Security of tenure status under the Landlord and Tenant Act 1954
Why this matters for the accounts
Under FRS 102 Section 20 and FRS 102 Section 24 (as updated in the 2024 amendments), the recognition and measurement of leases for accounts purposes - including the treatment of lease incentives, the determination of the lease term taking break options into account, and the disclosure of operating lease commitments - depends on these underlying contractual facts. Under IFRS 16 for the smaller number of UK groups still preparing under IFRS, the same facts drive the right-of-use asset and lease liability calculations.
A wrong term, a missed break option, or an unrecognised rent-free period in the underlying contract feeds straight into a wrong accounts treatment. Getting the data extraction right is therefore not housekeeping; it is the foundation of the disclosure.
The extraction prompt
The prompt that does the work in a docs-only AI tool is essentially a single instruction with a structured output. A version that works well in practice is:
"Extract from this lease the following data points into a table with columns Field, Value, Clause Reference. The fields are: landlord, tenant, guarantors, premises address, term start, term end, break dates and required notice, initial rent, rent payment frequency, rent review dates and mechanism, service charge basis and cap, insurance rent arrangements, repair obligation, alienation provisions, permitted use, security of tenure status under the Landlord and Tenant Act 1954. Cite the clause number for each value. If a field is not present, leave Value blank and write Not present in Clause Reference. Do not infer or assume."
Why the prompt is written this way
Every element of the prompt is doing work. The structured output (a table with three columns) is what makes the result usable in 30 seconds of review. The fixed field list is what makes the extraction comparable across leases. The instruction to cite a clause for every value is what makes the output reviewable - the partner can check three citations in a long lease and be reasonably confident the rest are right. The instruction to leave fields blank when not present is what stops the model fabricating to fill the table.
The instruction "do not infer or assume" is the most important sentence. Without it, the model will sometimes patch a missing value with a reasonable guess based on adjacent text. With it, the model is forced to acknowledge gaps, which is the behaviour the partner needs.
The 90-second workflow
In practice, the end-to-end flow looks like this:
- Upload the lease PDF to the AI tool (5 seconds)
- Paste the extraction prompt (5 seconds)
- Wait for the model to produce the table (20 to 40 seconds for a 40-page document on a modern docs-only system)
- Spot-check three to five citations against the source PDF (30 to 45 seconds)
- Save the output and the prompt as part of the working paper file (10 seconds)
What to do when the tool refuses
A well-built docs-only AI tool will refuse to answer when the document does not support the answer. That is the correct behaviour. If the tool reports that the lease does not specify the rent review mechanism, treat it as a finding, not a failure. Either the lease is silent (which is itself a fact for the working papers), or the relevant clause is on a page that did not OCR cleanly, or the lease is split across multiple files only one of which was uploaded.
The temptation to switch to a more permissive tool that will guess the missing value is the temptation that produces wrong working papers. The refusal is the feature.
The audit trail
For working paper purposes, the file should contain four things after the extraction: the source lease PDF, the prompt used, the AI output, and a brief note from the reviewer confirming what was spot-checked and what was accepted. If a value was manually overridden - for example, because the reviewer found a clause the AI had missed - that override should be noted with the clause reference and the reviewer's initials.
This is the same discipline as any other supporting working paper. The medium changes; the standard does not. ICAEW Audit & Assurance Faculty guidance and ACCA technical guidance on the use of technology in audit and assurance both expect the working paper to evidence the source, the work performed, and the conclusion. AI does not change that obligation.
Where the extraction breaks down
The 90-second workflow is reliable on most standard commercial leases. It struggles in three situations worth being honest about. First, very heavily negotiated bespoke leases where the variations have been added as side letters or supplemental deeds - the AI will extract from what it has been shown, which may not be the operative position. Second, leases that have been scanned poorly and where the OCR is unreliable in the key clauses. Third, leases for unusual property types (telecoms, advertising hoardings, complex retail concessions) where the standard field list does not capture the economics.
In each case the AI is not the right tool by itself; it is a starting point that the senior reviewer takes further by hand. The 90 seconds is still well spent because it produces the structured skeleton on which the reviewer then works.
Scaling the workflow across the engagement
A firm with 60 leases across its client base - common for a mid-sized practice acting for property-holding owner-managed businesses - can complete the annual extraction cycle in two to three hours of partner time using this workflow, against the two to three days it would take by hand. The saved time pays for the AI tool many times over in a year, but the more important benefit is consistency. Every lease is extracted against the same field list, in the same format, with the same review discipline. Year-on-year comparisons become trivial.
How Accupe helps
Accupe's AI document analysis runs exactly this workflow. Upload the lease, use the Fast mode for first-pass extraction or Ultra-Detailed mode for a longer or more complex document, and the docs-only mode means the output is grounded in the uploaded file with a clause-level citation for each value. The output sits against the client record alongside the AML/KYC file, the engagement letter, and the rest of the working papers, so the audit trail is complete in one place. Per-firm pricing from £20/month.