By Johnny Morris, Fifth Dimension
At the AdHoc 2026 conference in Boston this September, the audience at one of the sessions was asked a simple question: who has applied AI to a business process that is audited or formally controlled? The only hands raised came from a team Fifth Dimension had worked with.
There was no shortage of enthusiasm for AI elsewhere at the event. People were building tools, experimenting with prompts, and helping colleagues get started. But the team that Fifth Dimension had worked with described a different kind of progress: automated processes, lower spending on outsourced suppliers, and work that had been through audit.
That is where REITs should set their ambition. Use AI in consequential, recurring work, with enough control and evidence that the business can depend on it.
An audited process can look like an intimidating place to start. The cost of a mistake is visible. There are established responsibilities, reviewers, and expectations. It is tempting to keep AI in tasks where the consequences are smaller.
Yet those same characteristics can make a process a strong candidate for automation. It’s clear why the work matters. The owner can be identified, alongside how the work is checked, and what an acceptable result looks like. There is usually a meaningful cost attached to doing it today. Those are useful foundations for an investment case.
Start with Work the Business Already Has to Get Right
Consider a recurring lease-to-billing process. A team reads the lease and amendments, identifies the relevant terms, applies agreed rules, checks the proposed charges, and transfers approved information into an operating system.
Much of that effort involves gathering evidence, interpreting documents, and moving information between systems. AI can help. The important question is how much of the complete process can be automated while preserving the decisions and checks that make the result trustworthy.
Reading one lease accurately is a demonstration. Handling the recurring work requires more: knowing whether the document set is complete, identifying the amendment that supersedes the original term, surfacing conflicts, and making sure an unresolved issue cannot quietly become an approved charge.
A standalone Claude or ChatGPT conversation cannot run that process on its own. The models can do valuable work within it, and enterprise products offer audit logs and other controls. The implementation still has to connect the documents, business rules, approvals, and destination system. A reviewer needs to establish why a particular charge was calculated, which evidence supported it, and who approved an exception.
Build Evidence into the Process
Take a rent increase that begins on the wrong date. Every downstream calculation that uses it can be mathematically correct and still produce the wrong bill. A plausible answer, neatly formatted, tells very little about whether the underlying work is sound.
A dependable process should make it possible to follow a result back through its calculation, the approved inputs, and the source documents. Where a person changes an extracted value or resolves an ambiguity, that decision should be recorded with the result.
The distinction between interpretation and calculation matters here. AI may propose what a clause means or which values to extract. Agreed calculations should then operate on approved inputs. There is no benefit in asking a language model to reinvent a financial rule every time the process runs.
The controls need to fit the actual risks. Some outputs may be straightforward to validate automatically. Others require a reviewer with the right expertise. A missing amendment, conflicting dates, or an unusual clause should have a defined route for resolution. “The user will check it” is an incomplete design unless it can be explained what they will see, what they must check, and what happens if they disagree.
Evidence should be produced as the work happens. Reconstructing it later from conversations, screenshots, and personal spreadsheets creates another manual process, just when the business expected to remove one.
Involve the process owner, finance, and the relevant assurance teams early. Their questions can improve the product: what must be retained, which decisions require approval, where an exception should stop the process, and how a later reviewer will understand the outcome. Bring external auditors into the discussion where appropriate to the work being changed.
Test the Process, Including How it Fails
The most useful starting material is often work the business has already completed and reviewed. Use those cases to establish what good looks like, then test the proposed process against them.
Include the difficult examples. A clean lease with a single amendment will tell less than a document pack with conflicting terms or a calculation that previously required an expert decision. Record why the approved answer is right. Where reasonable judgment is involved, capture the rule the business wants to apply consistently.
A headline accuracy percentage is insufficient. Errors have different consequences.
Missing a rent adjustment costs far more than mislabeling a document. Review effort also matters: a system that gets most answers right but forces someone to inspect everything from scratch may have saved little work.
Test the handoffs as well as the extraction. Can a reviewer correct a value without losing the original evidence? Does a rejected item stay out of the approved output? Can the team distinguish completed work from work that stalled? When information reaches the destination system, can it be shown that the right record was updated?
Then prove the process on new work. Measure the exceptions, the time spent reviewing them, and the cost of getting to an accepted result. Keep enough visibility to find failures before they become somebody else’s problem.
The work continues after launch. Changes to a model, extraction method, or business rule can affect performance. Someone must own the testing, approval, and monitoring of those changes. A process that was dependable last quarter needs to remain dependable this quarter.
Buy the Capability to Deliver the Whole Job
This is where many organizations need help. Someone has to understand the leases well enough to recognize a bad answer, the technology well enough to prevent it recurring, and the operating systems well enough to put the approved result where it belongs. A strong internal team can assemble that capability. An implementation partner can supply parts of it. Either way, the expertise has to exist.
When evaluating a vendor, ask them to demonstrate your process on your material, including an awkward case. Follow one output all the way back to its sources. Ask them to show a correction, a rejected result, and a failed run. Establish who will maintain the system and what happens when the documents or rules change.
The commercial case also needs to cover the whole job. Include implementation, human review, exception handling, and maintenance. Compare the cost of an accepted result with what the business spends today. Name the supplier bill that falls, the manual step that stops, or the additional work the team can take on. Count the review and exception handling that remain.
The team at AdHoc could explain that connection. Their account was about work the business depended on, a reduction in outside spend, and a process they could take through audit. That is a much stronger basis for expansion than a dashboard of chatbot usage.
For a REIT leadership team, the next move is concrete: choose one recurring process with a clear owner, meaningful cost, and an established review standard. Fund the expertise to automate it properly. Require evidence that it works on real cases before replacing the existing process.
The ambition should be to put AI into work important enough that someone will ask for proof it was done correctly. Then build a system that can answer.