The ninety-second miracle
Marcus had spent two weeks since the Driskill watching the Coordination Tax compound. When Priya (investment associate) mentioned she had used a frontier model to draft a deal screening on her own time, Marcus leaned in. She had fed it a broker package for a 180-unit garden-style multifamily in Chattanooga.
The output was impressive: professional tone, correct vocabulary, structured sections. Four hours of work in ninety seconds.
Then Priya showed him the annotations.
The comparable transactions did not exist. Fabricated deals at plausible prices. The cap rate was reasonable for the MSA but wrong for the submarket. The valuation was off by eleven percent. On a $28 million ask, call it three million dollars of mispriced risk.
Priya had spent four hours verifying every claim. Closer to five, she admitted, counting the two comps she chased down by phone.
About what it would have taken to build the analysis from scratch using the firm's actual data. That was her estimate; nobody had ever timed the original either.
"So it's useless," Marcus said.
Priya hesitated. "I can't tell which parts are right without checking all of it. And if I'm checking all of it, I'm doing the work twice."
That sentence is the Verification Tax.
The most dangerous outputs
The most dangerous AI outputs are the ones that are wrong with specificity.
A hallucination with a named elevator company, a specific dollar claim, and a plausible project address enters the firm and begins to make decisions before anyone asks whether it is real. That is the Verification Tax in its most dangerous form: decisions taken on evidence that was never real.
The pattern appears at every real estate firm I have watched experiment with AI. Deal screening, market research, LP communications, shareholder updates, lender packages, JV partner reports: the failure is identical. The AI produces output that looks credible. The output contains errors ranging from subtle to disqualifying. A human must verify every claim.
The Verification Tax is the total time your team spends verifying AI output over the time the original analysis would have taken. When the tax exceeds 100%, AI is destroying value.
Priya's screening: ninety seconds to generate, four-plus hours to verify, about the same for the original analysis from the firm's own data. Call the tax 100%. Zero hours saved by any honest count.
The deeper cost is what the formula misses. AI did not automate the labor. It moved the labor from assembly to verification. Verifying plausible hallucinations is more exhausting than doing the math yourself, because the analyst is now reading every line for the one that is wrong instead of writing every line knowing each one is founded.
75% of GPs rate AI as ineffective for portfolio monitoring.
Source: S&P Global, 2026 Private Equity Survey
DATA POINT
44% of firms that have adopted AI tools have no formal testing or validation of its outputs.
Source: Investment Adviser Association & ACA Group, 2025 Investment Management Compliance Testing Survey
The harness
When Priya fed a broker package into the frontier model, it did exactly what models do. It processed the input, predicted the most probable sequence of tokens, and generated fluent text. It had no access to the firm's comp database. It did not know the investment thesis, the buy box, the firm's submarket perspective. The model simply generated plausible text rather than firm-calibrated intelligence.
This distinction between a model and the infrastructure surrounding it is the single most important concept in AI implementation. The AI industry calls it the harness, a frame developed most clearly in the public commentary of Nate B. Jones, who describes the agentic harness as the decompose-parallelize-verify-iterate scaffolding that turns a raw model into an operating system (Jones, Nate's Newsletter, natesnewsletter.substack.com). The real question shifts from which AI model or tool should we use? to what are we wrapping around the model or tool to make it effective?
The model is the engine. The harness is everything else: the data connections that feed it real, current, firm-specific information, the validation rules that catch errors before they reach a human, the integrations that put its output where the work happens. The next chapter takes the harness apart layer by layer. The diagnosis here is simpler. Priya's ninety-second miracle failed because the firm had a model and no harness, and the Verification Tax is the bill a missing harness sends.
There is a second way to name the same failure, and it describes the mistake most firms are making right now. Priya asked the model a question and received an answer that looked like work. Asking is what a chat window invites: feed in a document, get back a draft. Assigning is different. An assignment comes with five things an answer never does: a goal, named sources the work must draw from, a standard the output must meet, boundaries on what the system may touch, and proof that the work is done. Priya's screening had none of the five. Not because she worked carelessly, but because the firm had nothing to attach them to: no comp database to name as a source, no written buy box to serve as a standard, no template to define done. The chat window was the only door the firm had built, so that is the door she used.
Principle: AI fails as an oracle and works as a delegate
Ask a model for answers and you will pay the Verification Tax on every claim. Assign it work instead: a goal, named sources, a standard, boundaries, and a definition of done.
A firm that cannot specify those five things for a workflow is not ready to hand that workflow to AI, and it was never ready to hand it to a new hire either. Chapter 9 turns the five elements into a one-page brief the firm completes for every workflow it hands to the platform.
Eight days inside the data room
Three weeks after the data room request landed, the data room itself went live.
Marcus's team had spent every available hour assembling it. Governance documents pulled from scattered drives. Compliance policies reconstructed from memory. Decision logs that did not exist, created after the fact. Quarterly reports reformatted for institutional presentation.
The family office's ODD team spent eight days inside it.
On day three, the ODD lead asked Nathan Park (VP of Acquisitions) to walk through the firm's deal-screening methodology. Nathan pulled up his spreadsheet: four hundred rows, the tracker he had built himself.
"This is your screening system?" the ODD analyst asked.
"This is the tracking system. The screening methodology is..." Nathan paused. He looked at the spreadsheet. The buy box criteria were not written anywhere in the document. The reasons for passing on deals were one-word entries in a column: pricing, location, basis. Nothing connected to the firm's investment thesis in a way a third party could understand, much less audit.
"The screening methodology lives in our IC process," Nathan said.
"Can I see the documented criteria?"
Nathan pulled up an email chain from 2023 where Marcus had outlined deal parameters to the team. It was the closest thing to a written buy box the firm had.
The analyst wrote something in her notebook and moved on. Nathan sat at his desk for ten minutes after she left, staring at the spreadsheet he had built to compensate for the infrastructure that did not exist.
Jordan Wells (Head of Investor Relations) spent the better part of two days walking the ODD team through LP reporting. Every question required her to pull documents from different systems, cross-reference dates, and narrate the connections the systems could not demonstrate on their own. By day three, she was running on four hours of sleep. Tom Langford (controller) sat with the ODD team's financial analyst for an entire afternoon, walking through the waterfall calculations. The analyst asked why the accounting platform's output differed from the Excel model Tom maintained. Tom explained the configuration gap. The analyst asked how long the gap had existed. Tom said, "Since we implemented the platform." The analyst wrote that down too.
Anika Reeves (General Counsel) had prepared the compliance documentation as thoroughly as the timeline allowed. But when the ODD team asked about the firm's conflict-of-interest policy, specifically how co-investment allocations were determined, Anika had to say what she had warned Marcus about months earlier.
"The allocation decisions are made by the managing partner based on LP relationship factors and investment suitability."
"Is that documented?"
"It is now. We formalized the policy three weeks ago."
The ODD analyst looked at the creation date on the document. She did not say anything. She did not have to.
Claudia (CFO/COO) watched the ODD process from her office with the door open. She had been tracking what was missing since she laid the folders across the conference table three weeks earlier.
Nearly every document the ODD team requested, she had asked Marcus for at least once in recent years. The cybersecurity audit and implementation plan she had drafted in outline last summer had never gotten sign-off to complete. The allocation policy she had flagged as inadequate six months ago. The conflict-of-interest framework Anika had warned about.
On day six she closed her door and sat quietly for ten minutes. The ODD team was finding exactly what she had been telling Marcus they would find.
The Family Office call came on a Wednesday afternoon. Not the CIO. His deputy, the head of operational due diligence. Professional. Brief.
"Marcus, we've appreciated the time your team invested. The returns are strong. The thesis is sound. We're not going to move forward with the allocation at this time."
Marcus asked why.
A pause. Then, carefully: "Honestly, Marcus, we're just not sure your platform's ready for a check with an extra zero. Not yet."
Fifty million dollars in anchor capital, off the table. The data room had shown them what the firm actually was beneath the returns: disconnected systems, manual processes, institutional knowledge stored in people who could leave.
Eight days inside had been long enough for the ODD team to see everything the polish of the quarterly deck had been covering for.
Marcus thanked him, hung up, and sat at his desk for a long time. The rejection did not sting the way a bad deal stings. A bad deal is a market call you got wrong. This was something else. This was being told that the firm he had built, the one he had poured twelve years into, was not ready: his infrastructure, not his judgment, had failed the test.
The largest capital opportunity in his firm's history, lost to the gap between the firm he had pitched and the firm the diligence had revealed.
The same problem in a different costume
He picked up his phone at 6:47 PM and texted Sarah.
they passed. wrong on infrastructure.
Three minutes later her reply landed.
I know. let's talk tomorrow.
He set the phone face-down. Two days earlier the message would have made him angry. Tonight it made him still. Sarah had known what the call would say from the morning at the Driskill three weeks earlier. She had not said it then because she did not have to. The data room would say it for her.
He opened the laptop. The annotated screening Priya had shown him sat in his inbox where he had filed it three weeks ago. He pulled it up. The fabricated comps. The submarket cap rate that was wrong. Eleven percent valuation error. I can't tell which parts are right without checking all of it.
He had thought, that morning with Priya, that the Verification Tax was an AI problem. The family office had just shown him it was the same problem in a different costume. The firm produces output, polished and credible. The institutional reader cannot tell which parts are right without checking all of it. So they put down their pencils and walk from the firm.
The harness extended beyond AI. It was whatever architecture made a firm's output trustworthy without requiring the institutional reader to verify it themselves, and his firm did not have one, not for its AI and not for its data room.
The diagnosis was complete: the invisible overhead, the people consumed by work beneath them, the rising bar from LPs who could see the gap, and the failed first attempt at the obvious fix. Four faces of the same structural absence: the platform Sarah had been pointing at since the Driskill. At 11:40 he was still at his desk, listing what the data room would have needed to assemble itself.