AI inside the investment committee. Used by almost everyone, trusted with almost nothing.

Everyone started. Few can show the result.

Ask investors whether they use AI and the answer is nearly always yes. In a July 2026 survey of more than 100 institutional investors in the US and Canada, 97 % said AI is integrated into their investment process. In October 2025, JLL found 88 % of real estate investors already piloting AI, up from 5 % in 2023, with an average of five use cases running at once.

Ask what changed and the answer is thinner. In the same July survey, no single outcome had improved for more than 36 % of firms. Among corporate occupiers, JLL found only 5 % had reached most of their program goals. Pick a number below to see what it measures and who was asked.

Investors let AI read. They do not let it judge.

The same survey draws the line. 55 % of investors would let AI summarize a diligence document. Only 35 % would let it score a deal. And the survey report states that every team checks the output: 100 % verify what AI produces.

The line is not arbitrary. On a published test of financial research tasks, the best model scored 46.8 % in May 2025. On the revised version of that test, the leader reached 67.85 % on 29 October 2026. The versions differ, so read the gap as a direction, not a rate of progress. Switch the exhibit between what investors allow and what the models score by task.

Read the second view as a map of the committee. Different models lead different tasks. The leaders are best at reading reported numbers (93.0 % in earnings analysis) and weakest where the work needs judgment and a long chain of steps: 57.0 % in financial modeling and 49.3 % in finding precedents. Those tests use listed-company filings, not property, and only the test set is scored, so treat them as a direction and not a forecast for your next deal.

Where the saved hours go.

Among the investors surveyed, 51 % say AI saves them time once checking is counted, and 41 % say work with AI takes longer than doing it by hand. The reason they give is data: 90 % say fragmented or poor data has limited what AI can do, even though 83 % rate their data as ready.

That is a checking bill, and you can estimate yours. Enter the hours one deal takes by hand, how much of the work AI drafts, and how much of the drafted work a person has to redo or check. The result is arithmetic on your inputs, not a measurement.

One measured point for the machine side. Across the 30 days to 1 November 2026, our engine completed 219 underwriting runs with a median of 119 minutes, the middle half between 83 and 239 minutes. That counts our internal test runs and excludes runs that stopped or failed. Read it as hours, not weeks. It is not instant, and a person still reads the result.

Where did the hours go in your last committee cycle? The exhibit below walks the five steps of a deal and shows what AI can draft, what a person must decide, and the one check that keeps the saved time.

What to do with this

If you chair or sit on a committee: ask for one thing in every memo, a source for each number. Time spent on that check is cheap when the memo is built to be checked, and expensive when it is not.

If you lead a fund or a lender: measure your checking bill on three live deals before you buy another tool. Titleman returns an underwriting package that marks where the records disagree, so your team checks the flags, not every line.