The AI Slop Tax: Why Unverified AI Output Costs More Than No AI At All

My career started at a $30 billion publicly traded REIT, and one of my bosses there was a man named Ernie Wittich. When I finished an underwriting model, Ernie would have me print the Argus and Excel outputs and bring them to him. Then we sat, usually an hour, sometimes longer, and went through the deal line by line: every tenant entry, every lease escalation and expiration, every CAM pool and recovery calculation, the ROFO and expansion scenarios on the key tenants.

While we went, he rebuilt my P&L by hand on a yellow legal pad with an HP-12C — a calculator that predates Excel. Most sessions he found nothing, which I eventually understood was the point. Some sessions he found a formula that had drifted, or a risk priced wrong enough to eat the returns we were promising.

Once, he found a big one. We were underwriting a $6 billion residential portfolio acquisition, and I had made a weighted-averaging error in the portfolio cap rate. Mechanically it was a small mistake, a few cells in a summary tab that added up the way summary tabs do. It overpriced the portfolio by $300 million — roughly five percent of the deal, sitting in an output that looked completely finished.

What I remember is less what he said than what he did: he circled the number on the legal pad and had me trace the weighting back until I found the break myself. I've been checked by committees and by software plenty of times in the years since. That hour with the legal pad is the one I still think about.

I tell that story because of what the checking cost. Ernie's hour was the price of trusting my model — my output, produced slowly, by a person he had trained himself. Twenty years later I sit on his side of the desk, and the drafts I'm asked to trust are produced in ninety seconds by a language model that was never trained on anything the firm believes.

I've started calling what happens next the AI slop tax.

How the tax gets collected

Here is the version I keep walking into at CRE PE firms that tried AI on their own. A sharp analyst discovers that ChatGPT will draft an investment memo, pull market data, or summarize a financial package. The output looks like institutional work product and it arrives in minutes.

Then a senior person reads it the way Ernie read my Argus runs, and the problems surface. The comps are from the wrong submarket. The cap-rate assumption is a national blend that has nothing to do with where that market actually is. The narrative reads smoothly and says nothing the IC hasn't heard a hundred times. The projections run on textbook assumptions instead of the firm's own underwriting criteria.

So now the senior person is repairing machine output instead of reviewing an analyst's judgment. Add the analyst's prompting time to the partner's repair time and the total frequently exceeds what the memo would have cost done the old way. That is the tax. In the playbook I call it the Verification Tax, which is the polite name; slop is what it feels like at six o'clock on memo day. And it runs heavier than the tax Ernie paid on me, because my error was one bad formula in one tab, while a generative model can be fluently, confidently wrong in forty places across a single document.

The three levels of integration

The root cause is that most people use AI as a question-and-answer tool: type a prompt, get a response, paste it into the deck. That works for looking things up. It fails in investment work product, because the model has no idea who your firm is.

When I build these systems for clients, I work in three levels. The first is individual tool use — an analyst drafting a memo or pulling comps. Most firms stop here, and it's maybe a fifth of the available value.

The second is firm intelligence: loading the system with what Ernie carried in his head. The firm's underwriting criteria, its market assumptions and rules of thumb, its lease-abstraction templates, its IC memo structure, the approved language it uses with LPs. Without that context, the model produces generic output calibrated to nobody's standards, and every page of it lands on a senior desk for repair.

The third is workflow design: the output has to arrive where the work actually happens, with validation built in before a human ever reads it. Comps checked against the deal's actual submarket. Cap rates flagged when they leave the firm's current bands. Weighted averages recomputed independently — the same check Ernie ran on my summary tab.

Where to start, where to hold back

The sequencing question is really just the cost of being wrong. Start where errors are cheap and hours are expensive: deal screening, where a bad lead surfaced is simply a lead you skip, and first drafts of quarterly reports, where a reviewer stands between the draft and the LP. Hold back anywhere output reaches a counterparty without a person in the path.

The durable value sits in the middle, where the system assembles and a person judges. The model builds the first pass of the underwriting and an analyst stress-tests the assumptions. The system compiles the IC data pack, and the recommendation still gets written by a principal. Structured this way, review time falls, because the reviewer starts from assembled, pre-checked data and spends the hour on judgment.

People occasionally hear the Ernie story as an argument against automation. I take the opposite lesson from it. Ernie was the only verified step in the entire process, and the systems I build now are an attempt to move his legal pad inside the machine: independent recomputation, firm-specific validation, an audit trail a reviewer can walk in minutes.

The weighting error in that $6 billion underwriting was worth $300 million. It was caught by one reviewer with a legal pad and a calculator, in about an hour. Every AI system I design starts from that hour and works backward.

Back to all insights