AI for Underwriting in Real Estate: What an Agent Can Verify Before You Price a Deal

Underwriting a property is a reconciliation exercise wearing the costume of a modelling exercise. The cash flow takes an afternoon once you trust the inputs. Working out which document in the data room tells the truth about the income takes the rest of the time, and a moved bid date compresses that part.

Agentic systems earn their place here for an unglamorous reason. The work that decides whether an underwrite is sound is reading: a lease against a schedule, a service charge certificate against the budget it was meant to follow, a dilapidations comment against a capital expenditure line nobody put in the model. That reading is bounded and evidenced, so it can be delegated and checked. The judgement on top, what yield this tenant's covenant deserves, is neither.

The deal pack is several accounts of the same building

What arrives is a cash flow model, a rent roll, a tenancy schedule, scanned leases with their variations and side letters, a measured survey, service charge accounts, a valuation, a building survey, environmental and title reports. Each was prepared by a different party, for a different purpose, at a different date.

The useful first output is a map, not analysis. An agent indexes every document: what it covers, which demise, the date it speaks as at, who prepared it, whether it is executed or draft. The result is a coverage matrix separating tenancies with a full document chain from those with only a summary, and surfacing documents referring to a unit that appears nowhere in the schedule. A pack whose largest tenancies have no executed lease is not the pack described.

Rent roll verification happens at lease level, not summary level

This is the centre of the work. The agent reads each tenancy's documents for the terms the model depends on: the demise, term commencement and expiry, the passing rent and when it became payable, review dates and mechanism, break dates and who holds them, any rent free or stepped rent, service charge and insurance obligations, the tenant's identity exactly as executed.

Then it compares that to the schedule. The output is a line by line reconciliation, every variance sourced to a page and a clause and classified by what it does to value: income timing, income amount, term length, recovery. A stepped rent recorded at its final level rather than its current one is a different asset, as is a tenancy whose summary omits a side letter's half rent year.

Variations that never reached the schedule

Side letters, deeds of variation and licences are where the surprises live, and they are filed by date rather than by unit. An agent that has indexed the whole room links a loose letter referring to "Unit 4B, Mill Court" to the tenancy it varies, notes that it is personal to the current tenant, and flags that the concession falls away on assignment.

Covenant strength is prepared, not scored

An agent can assemble the covenant picture: the contracting entity as named in the lease, its filing history, the shape of its accounts, whether the guarantor is a trading parent or a dormant holding company with a nominal balance sheet, whether any guarantee is capped or time limited, and whether the entity on the rent demand is even the entity on the lease. It shows how much income sits with each counterparty.

What it does not do is produce a rating. Its job is to stop anyone forming that view without noticing that the reassuring name on the first page is not the party on the hook.

Weighted lease length and income that is less contracted than it looks

Weighted average term is quoted to expiry far more often than to break, and the difference is often the whole investment case. An agent recomputes it on both bases from the executed documents, lists every break, states whose option it is, and extracts the conditions on exercise: notice length, vacant possession, payment of all sums due, reinstatement.

It then shows the income profile with every tenant break exercised. That is not a forecast, it is a floor. Where schedule and documents disagree on a break date, it escalates: a break two quarters earlier than modelled moves the exit assumption further than anything else in the pack.

Irrecoverable cost the vendor's numbers quietly omit

Net operating income is where optimism hides most comfortably. An agent reads the service charge accounts and certificates against the recovery provisions in each lease and builds the position upward: which costs are recoverable from whom, at what apportionment, subject to what caps, and what falls on the owner because a demise sits outside the scheme or a cap bites.

Voids leave their share of cost with the landlord. The output is a schedule of irrecoverable cost by category, each line carrying its clause, set against the model's allowance. Management fees, empty rates, insurance shortfall and non recoverable repair are the lines missing most often, and rarely by accident.

Capital expenditure the survey implies

A building survey names defects, lives and indicative costs in a report nobody reads end to end. An agent extracts every item, places it on a timeline by remaining life, and separates what the owner must do from what a tenant could compel through a repairing obligation or a service charge challenge, cross referencing plant inventories, planned maintenance and outstanding statutory items.

The output is a capital expenditure schedule that drops against the cash flow, each line sourced to its page. Where the survey hedges, the hedge is reproduced rather than converted into a number.

Debt sizing and stress cases that outlive their author

Every lending team's stress convention lives partly in a document and partly in the habits of whoever built the last model. An agent applies the written one consistently: sizing on verified rather than quoted income, running interest cover and loan to value at the agreed stresses, showing covenant headroom through the term with the first breach date named.

It reports which inputs the covenants are most sensitive to and reconciles its own sizing against the model in the pack, explaining each difference. When the stress case is reproducible, the credit discussion stops being about arithmetic and becomes one about assumptions.

A comparable evidence base that is thinner than it reads

An agent can gather and normalise the evidence that exists, state what each transaction actually was, note how far it sits from this asset on the dimensions that matter, and separate open market evidence from related party transfers and asking terms that never completed. Where there is little, it says so and shows the spread, not an average implying confidence nobody has.

It should never produce a value: an agent that smooths thin data into a tidy figure removes what the committee needed to see.

Development and refurbishment, where the model is a programme

Here the underwrite is a sequence rather than a rent roll. An agent reads the cost plan, the programme, the planning permission and its conditions, the building contract and any collateral warranties, then checks the model against what those documents say: drawdown against the programme, contingency against the contract's risk allocation, consents treated as conditions rather than assumptions.

The outputs worth having are the planning conditions still to be discharged, a reconciliation of cost plan to model, and which programme dates the exit depends on. Letting assumptions and the market at completion stay judgement, and usually dominate the return.

The investment committee paper and what follows approval

The paper is where everything becomes visible or stays buried. An agent drafts it: the verified income position, each variance and how it was resolved, the irrecoverable cost and capital expenditure schedules, the covenant picture, the stress outcomes, the evidence base with its weaknesses stated, the open items. A reader can move from any number to its clause.

Conditions precedent, then life after completion

Approvals come with conditions precedent, and conditions come with dates. An agent turns the committee's conditions and the lawyers' requisitions into a tracked list with owners, then watches for the documents that satisfy them. After completion it keeps checking what the underwrite assumed: that the review happened, the break passed unexercised, the recovery rate held. When the asset first deviates from the paper, somebody hears in time to act.

Where AI fails here

Scans of scans defeat extraction. A lease photographed at an angle with a handwritten amendment in the margin will be read wrongly or not at all, and the honest behaviour is to escalate the page rather than produce a confident field. A system that does not separate low confidence output from high is worse than none.

Unfiled variations are the structural limit, because an agent can only reconcile what is in the room. A letter agreeing a rent concession that sits in a partner's inbox and never reached the data room will not be found, by software or anyone else. The mitigation is procedural: ask, record that you asked, underwrite accordingly.

Judgement does not transfer. Covenant strength, whether a tenant will really break, what a thin comparable set is worth, what the market will pay at exit, these are opinions formed by people who know the sector. An agent that produces a covenant score or a yield has not helped, it has concealed where the risk was taken.

Precision presented without evidence is the costliest failure mode: the number travels into the model, the model into the paper, and the caveat does not. So over invest in human review where being wrong is most expensive: the largest tenancies by income, anything near a break, and any figure the covenants are sensitive to.

Begin where the deal file is already thinnest

Do not start with the whole underwrite. Start with rent roll verification on a deal you have already closed, where the executed documents and your own corrections both exist. Run the agent against the original pack and compare what it flags to what your team found. You learn the proportion of real variances it catches and the shape of the noise it generates, both of which matter before anyone relies on it under a deadline.

From there, extend into irrecoverable cost and the capital expenditure extract. Both are mechanical, consistently understated in vendor material, and checkable against documents you already hold. Keep pricing, covenant judgement and approval with people, and insist every assembled figure carries a path back to its clause. Adopt AI Partners works with investment and credit teams on that sequence. To talk it through against a live pack, write to hello@adoptaipartners.com.