There's a phenomenon I keep encountering at CRE PE firms that have attempted AI implementation on their own. I've started calling it the Verification Tax — the hidden cost of poorly implemented automation that actually creates more work than it saves.
The pattern goes like this. Someone on the team — usually a sharp analyst or associate — discovers that ChatGPT or a similar tool can draft investment memos, pull market data, or generate financial summaries. They start using it. The output looks impressive at first glance. It's fast, it's formatted well, and it sounds like institutional work product.
Then a senior person reviews it and starts finding problems. The comps are from the wrong submarket. The cap rate assumption doesn't match current conditions. The market narrative is generic — it reads well but doesn't reflect what's actually happening on the ground. The financial projections use standard assumptions instead of the firm's underwriting criteria.
Now the senior person is spending time correcting AI output instead of reviewing human work product. The total time invested — the analyst generating the draft plus the senior person fixing it — is often more than if the analyst had just done it the traditional way. That's the verification tax.
The three levels of AI integration
The root cause of the verification tax is that most people approach AI as a question-and-answer tool. They type a prompt, get a response, and use it. This works fine for casual information retrieval. It fails catastrophically for professional investment work.
Production-grade AI implementation requires three distinct levels of integration. The first is individual tool use — an analyst using AI to draft a memo, pull comps, or format a model. Most people stop here, and it's about twenty percent of the job.
The second is firm intelligence — giving the system the right information to work with. Your firm's underwriting criteria. Your target market assumptions. Your lease abstraction templates. Your LP reporting format. Your IC memo structure. Without this firm-specific context, the AI produces generic output that sounds authoritative but isn't calibrated to your standards.
The third is strategic alignment — designing the workflow so that the AI's output lands in the right place, at the right time, in the right format, with the right validation checks. This is where most implementations fail. The output might be good, but it's not connected to anything. It's a one-off response instead of a component of a systematic workflow.
The noise-to-signal problem
In investment contexts, the cost of a wrong answer isn't a wasted afternoon — it's a capital allocation error that compounds for years. This changes the entire calculus of AI implementation.
When I work with clients, I draw a sharp distinction between judgment and noise. Judgment is the irreducible human element — the experience-driven intuition about whether a deal is right, whether a market is turning, whether a counterparty is reliable. Noise is everything that obscures judgment: the hours spent formatting data, the delays waiting for comps, the cognitive load of assembling information from twelve different sources before you can even begin thinking about the deal.
The goal of good AI implementation is to reduce noise so that judgment has more room to operate. The goal of bad AI implementation — the kind that generates the verification tax — is to skip judgment entirely and let the model do the thinking. The first approach amplifies your team's capabilities. The second replaces them with a stochastic system that is confidently wrong a meaningful percentage of the time.
Where to start, where to hold back
The practical framework I use with clients is straightforward. Start with workflows where the cost of error is low and the time savings are high. Deal screening is an ideal entry point — if the system surfaces a bad lead, you just skip it. Quarterly report drafting is another — a senior reviewer will catch errors before anything goes to LPs.
Hold back on workflows where the cost of error is high and the human oversight is thin. Automated trading decisions. Unsupervised LP communications. Anything that goes directly to a counterparty without a human reviewing it.
The middle ground — where the real value lives — is workflows where the system does the assembly and a human does the review. Underwriting models where the AI builds the first draft and an analyst stress-tests the assumptions. IC memos where the system pulls together the data and a principal writes the recommendation. Investor reports where the system drafts the narrative and an IR professional ensures it matches the firm's voice and approved language.
This middle ground is where you eliminate the verification tax and capture the real leverage. The system does work that's useful enough to accelerate your team but conservative enough that errors get caught. Over time, as the system learns your criteria and the validation loops tighten, the error rate drops and the speed increases. That's the compounding loop — and it's what separates firms that build lasting AI infrastructure from firms that run experiments.
The firms paying the verification tax today are the ones that skipped this architecture and went straight to letting the model run unsupervised. The fix isn't more AI. It's better AI infrastructure — built by someone who understands what the output needs to look like and what happens when it's wrong.