AI-Washing: How Investors and Acquirers Can Verify a Company's AI Claims

In March 2024, the SEC fined two investment advisers a combined $400,000 for describing artificial intelligence they did not have. According to the SEC's settlement orders, Delphia spent four years telling clients its machine learning used collective client data to predict "which companies and trends are about to make it big," when no such capability existed. Global Predictions called itself the "first regulated AI financial advisor," a claim the SEC found it could not support. Delphia paid $225,000, Global Predictions paid $175,000, and the agency's enforcement chief at the time said plainly that "AI washing hurts investors."

The fines were small, but the pattern behind them deserves your attention if you buy, fund, or contract with software companies. Some real fraction of "AI-powered" products are rule-based scripts, thin layers over someone else's model, or occasionally nothing at all. Unless the company is a public issuer or a registered adviser, nobody has checked. If you're acquiring a mid-market company or leading a round where the multiple rests on an AI story, verification is your job. The good news is that it's far less technical than deal teams assume. I've built AI products and evaluated plenty of other people's, and inflated claims rarely survive four families of questions plus one honest pass through a data room.

Two small fines that defined the problem

The Delphia and Global Predictions settlements, announced together on March 18, 2024, were the SEC's first enforcement actions aimed squarely at false AI claims. What strikes you reading the orders is how ordinary the conduct was. Delphia's claims ran from 2019 to 2023 across SEC filings, press releases, and its website while the capability behind them simply did not exist. The orders read like marketing copy that nobody ever reconciled with the codebase.

Six months later the FTC launched Operation AI Comply, announcing five enforcement actions in a single day against companies it said used AI claims to deceive consumers. The best-known target was DoNotPay, which had sold what it called "the world's first robot lawyer," a description the FTC said the product could not back up. Then FTC chair Lina Khan's line from that announcement holds up well: "there is no AI exemption from the laws on the books."

For a buyer, this history matters for two reasons. It tells you the base rate of exaggeration is high enough that two federal agencies built programs around it. And it tells you enforcement only reaches public claims made to investors or consumers. The private company in your pipeline has never had its AI language tested by anyone. The enforcement cases were also the easy ones, where the technology plainly did not exist. Your cases will be murkier, because usually there is something that could be called AI. The question is whether it does what the deck says, at the volume the deck implies, with the economics your model assumes.

The spectrum runs from fraud to fluff

When I look at an "AI-powered" target or vendor, I'm trying to place it on a four-level spectrum.

1. No AI at all. Rules, regex, lookup tables, or an operations team doing the work by hand behind an "automated" label. This is more common at the small end of the market than anyone admits. I've sat through a demo where the "recommendation engine" turned out to be a lookup table with a few dozen rows, and the founder didn't consider that dishonest because a data scientist had generated the table two years earlier.
2. A thin wrapper. A foundation model does the real work, and the company's contribution is a prompt template, some retrieval, and a nice interface. Wrappers can be good businesses when the workflow and distribution are real. They should be priced as workflow assets, though, because the margins carry a third party's API pricing inside them and the moat is shallow.
3. Real AI with no P&L contact. The models exist, the demos are honest, and nothing in the business has actually changed. This tier is enormous: MIT's NANDA group found in its 2025 State of AI in Business research that about 95% of enterprise GenAI pilots produced no measurable P&L impact, with only around 5% delivering real returns. A company can truthfully say "we use AI across the business" and still sit squarely in the 95%.
4. Embedded AI that moves a number. Deflection rate, cycle time, cost per transaction, loss rate, each with a before-and-after baseline. This is the only tier that earns an AI premium on valuation.

The pressure to claim tier four while operating at tier one or two keeps growing. Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, up from under 5% in 2025. When an entire market is racing to say "agent," the label gets attached to whatever a vendor already had on the shelf.

Four question families that expose substance

Ask these in the management meeting, in this order, and pay attention to how quickly specifics arrive. Teams with real systems answer fast and tend to enjoy the conversation. Teams without them answer with adjectives.

What model does what, exactly?

Have them walk one workflow end to end and name every component: which steps call a foundation model, which use a fine-tuned or classical model, which are deterministic rules. Good answers are specific and unembarrassed about the boring parts, something like "a GPT-class model drafts the response, a small fine-tuned classifier routes it, and hard rules handle anything with a dollar amount." Bad answers defend a "proprietary AI engine" while declining to name anything inside it. A useful follow-up is to ask which parts are plain rules, because every production system has some. Honest teams list them immediately. A team claiming there are none either doesn't know its own system or doesn't want you to.

What data trains it, and does it actually learn?

If the pitch includes a data moat, size it. How many examples, labeled by whom, refreshed how often, owned under what rights, and what would it cost a competitor to replicate? If the pitch says the product "learns from every customer interaction," ask for the retraining schedule and the release dates of the last three model versions. Plenty of products shipped one prompt template in 2024 and have not learned anything since. The MIT NANDA researchers flagged failure to retain feedback and improve as one of the main reasons pilots stall, so this question doubles as a read on whether the product stays competitive two years out.

What happens when it's wrong?

This is the most revealing of the four. Real machine learning teams live inside their error rates. They can show evaluation dashboards, a failure taxonomy, incident tickets, and the specific embarrassing thing the model did last month. Ask for the current error rate, how it's measured, and what the last ten failures cost. A team that has never measured error either has no AI or has been running it carelessly, and either answer changes your underwriting. Deterministic scripts don't produce error distributions, so tier-one vendors usually can't sustain this conversation for long.

Where are the humans?

Ask how many people review or correct the system's output, what fraction of outputs they touch, and what that costs per month. Human review is healthy, and a claim of zero review should worry you more than a large review team. What you're testing for is concealment. Two field checks work well here. First, turnaround time: a model responds in seconds, so an "AI-generated report" that arrives in 24 hours has people in the loop. Second, operating hours: AI that slows down outside the vendor's local business hours is an operations team. Then pull the headcount-to-output ratio over the trailing two years. If ticket volume doubled and operations headcount doubled alongside it, the automation is doing much less than the deck claims.

Operational evidence worth requesting

Questions get you management's story, and documents test it. Each item below is cheap to produce for a company whose AI is real and awkward for everyone else.

• Inference invoices. Model API bills and GPU spend, reconciled against claimed volume. Say the deck claims "millions of AI decisions a month" and the model provider's invoice is $140. Those two numbers cannot both be true. A large inference bill cuts the other way: it corroborates usage and quietly tells you the true gross margin of the AI feature.
• Usage telemetry for the AI feature specifically. Adoption, repeat use, and retention of the feature, separate from the product around it. Plenty of AI features get demoed during the sales cycle and abandoned by week three.
• Override and edit rates. What fraction of outputs do employees or customers modify before use? High override rates aren't disqualifying, but they reprice the automation story.
• Eval reports and model version history. If nobody can produce a single evaluation run, the company has never known whether its AI works.
• The org chart, read for ownership. With modern foundation models a small team can ship real AI, so the absence of a research lab proves nothing. Look instead for a named person who owns evals, prompts or models, and incidents. If nobody owns those, the feature is unmaintained no matter how good it once was.
• Upstream contracts. If the product rides on a third-party model, read the terms, because pricing changes, rate limits, and termination rights at the provider are now part of your risk.
• Customer references, asked one narrow question. What does it get wrong, and how often? Customers of real systems give specific, mildly annoyed answers. Customers of vaporware talk about the salesperson.

Red flags in the marketing language

A surprising amount of this screens from your desk before anyone gets on a call.

• Capability-free claims. "Leverages advanced AI" with no named task the AI performs.
• Precision without a denominator. "99.2% accurate" means nothing without knowing accurate at what, measured on which dataset, evaluated by whom.
• Autonomy superlatives. "Fully autonomous," "replaces your entire team," or DoNotPay's "world's first robot lawyer." Even Gartner's bullish forecast is about task-specific agents embedded inside applications, and whole-job autonomy claims outrun anything actually shipping.
• An AI-heavy homepage next to AI-free documentation. Search the docs, the API reference, and the changelog. If the homepage mentions AI thirty times and the release notes never mention a model, marketing shipped something engineering didn't.
• Rename releases. A feature that became an "AI copilot" the same quarter the company started fundraising, with no release note describing changed behavior.
• No hiring trail. "Proprietary AI" alongside zero current or historical job postings for anyone who would build or maintain it.

This desk screen takes about half a day and usually settles half the question before management answers anything. It's the first pass we run at FirmAdapt when a client asks whether a target's AI claims hold up.

A verification sprint for deal week

Folded into standard diligence, the whole exercise adds about a week, most of it elapsed time rather than billable hours.

1. Day one, desk screen. Docs versus homepage, changelog, status page, job postings, the engineering team on LinkedIn, and review-site comments that mention the AI feature by name.
2. Management session. The four question families above, then a live demo on data you bring, including at least one case chosen because it should fail. Teams with real systems like showing how failures get caught, since the handling is part of what they built.
3. Data room requests. The evidence list above. Invoices and telemetry should arrive within days. Long delays on records a functioning company already has are themselves an answer.
4. Price it and paper it. A wrapper priced as a wrapper can still be a good acquisition if the workflow and distribution are real, so a weak AI story doesn't have to kill a deal, but it does have to reprice one. Then write the claims into the purchase agreement as specific representations: who owns the models, what rights cover the training data, what percentage of the process runs without human touch. Sellers with real systems sign without much fuss. Sellers who ask you to soften the AI language are showing you exactly where the deck was generous.

One note on scope. This exercise is separate from auditing the AI tools a target's employees use internally, which is a compliance and data-exposure question worth its own workstream. What I've described here is narrower: testing whether the AI a company sells, and prices, actually exists and earns its keep. None of it requires a machine learning background. It's the same discipline you already apply to revenue quality, pointed at a newer kind of claim, and most of it consists of requesting records that a genuine AI business generates as a byproduct of operating. Where those records don't exist, treat the absence as the finding and price accordingly.