AI Isn’t the Strategy: Fern Potter on Intelligent Assistance, Human Judgment, and the Future of Work

Full analysis

Your draft

Fern Potter, co-founder of Intelligent Assistance, sat down with Rio Longacre and Brett House on Signal & Noise to make one argument: enterprise AI fails when companies buy the tech before they define the problem. Everything else in the episode hangs off that. What it means for someone buying and building with AI: this is a consulting pitch wrapped in a change-management framework, with two live warnings buried inside that are worth more than the pitch.

How hard is this to undo? Nothing here forces a decision. It's a way of thinking about your own AI rollout, plus two structural risks (ad-funded models, embedded bias) that you can't do anything about today but should watch. Easy to file, easy to act on later.

What's actually on the table: not "should I hire Intelligent Assistance," but "am I about to spend six figures on an AI pilot with no definition of what good output looks like, and are the models I'm building on about to get quietly corrupted by ads?"

What sets the deadline: nothing hard. OpenAI's ad testing is the one live clock, and it's not ticking loudly yet.

The Skeptic. The 95% pilot-failure number Brett House cites is a Gartner figure nobody can actually source. It's the "reassuringly high" stat every AI consultant quotes because it makes their service sound necessary. Treat it as directional, not real. And the 25% lead-conversion lift Potter claims for her insurance voice agent? Single client, self-reported, vendor talking her own book. No auditor, no control group, no baseline. That's a case study, and case studies are marketing. The genuinely useful part of her argument costs nothing and needs no vendor: define what good output looks like before you buy anything. If you can't write down the standard, the pilot will fail no matter whose engine you plug in.

The Researcher. Strip the framework and there's no new capability here. "Intelligence engines" are agents plus APIs plus a knowledge base deployed inside Slack, Claude, or a Chrome extension. That's retrieval-augmented generation with a services wrapper, which every systems integrator now sells. Potter's honest point is that context can't be outsourced to a model, and she's right: the model doesn't know what your business considers a good answer. But that's an argument for spending on your own definitions and data, not on a consultant. The bias warning is the substantive claim. She and Sarah De Martin lobbied the House of Commons because models trained on skewed data embed that skew, and autonomous systems repeat it at scale. That's real and under-measured, and no current eval catches it in a targeting or creative workflow.

The Open-Source Advocate. The whole "deploy inside the client's own environment to avoid IP leakage" pitch is a feature you can get without a consultant. Run an open model, keep your data on your own infrastructure, done. Potter deploys inside Claude for some clients, which means the client's context still flows to Anthropic. The IP-safety story only fully holds if you control the weights. For an operator worried about where their proprietary data goes, that's the buy-versus-build fork: a services firm's "engine" is convenience, not sovereignty. If sovereignty is the point, open weights on your own boxes gets you there and the consulting fee doesn't.

The Compute Pragmatist. The ad-funded model risk is the one that actually moves the field, and the Beet.TV piece in the reading and OpenAI's reported ad testing both point the same way. Frontier models cost a fortune to run. Subscriptions don't cover it. Ads are the obvious release valve. The moment an LLM's answers are shaped by who paid, every media planner using it for competitive analysis or audience recommendations has a conflict baked into the output, and no label tells them which sentence was bought. Potter's analogy holds: search ads bent organic results, and nobody could see the seam. The difference is that social media eventually got brand-safety pressure because advertisers demanded it. There's no equivalent pressure on what an LLM generates yet.

The Builder. What would I do Tuesday? Nothing with Intelligent Assistance. But two things from this episode go straight into practice. First, before any pilot, write the acceptance standard: what does a good output look like, who signs off, what's the baseline it beats. Potter's right that most teams skip this and that's why pilots die. Second, if you use an AI tool for market intelligence or audience work, start noting which model and which version produced each recommendation. When ad-funded models arrive, you'll want a record of what your outputs looked like before the incentive changed. Neither costs money. Both are worth more than the 25% number.

Where the tensions actually sit. The Skeptic and the Researcher agree the vendor pitch is thin, but split on what's underneath: the Skeptic says the whole thing is repackaged services, the Researcher says the bias warning is a genuine, unmeasured risk that deserves attention regardless of who's raising it. The Open-Source Advocate and the Compute Pragmatist part ways on where the real danger lives. One worries your data leaks out to a model provider; the other worries the model's answers leak in, shaped by advertisers you can't see. For most operators the second is the bigger long-term problem, because you can control your own data but you can't audit someone else's ad-weighted model.

What it hinges on. Two beliefs. First, whether ad-funded LLMs actually ship and start bending outputs, or stay in testing. That's checkable. Second, whether "define good output first" is enough to move your own pilot-failure odds, which it is and it's free, so just do it. The council leans hard toward ignoring the consulting frame and taking the two warnings seriously.

Prediction: OpenAI will roll out advertising inside ChatGPT to a broad user base by the end of 2026, and it will ship with no per-response label telling users which content was influenced by a paying advertiser.

Confidence: Medium. Reported testing plus unsustainable inference economics point one way; timing and regulatory friction could slip it.

Why: OpenAI is already reported to be testing ads, and the Beet.TV coverage in the reading treats ads inside AI as the next channel marketers are planning for, so the demand side is real and building. The mechanism is money: frontier models cost far more to run than subscriptions bring in, and ads are the proven way to fund a free consumer product at scale, which is exactly how search and social got funded. The reason a clear "this answer was sponsored" label won't accompany launch is that the value of an ad woven into a conversational answer depends on it not reading like an ad, and no regulation currently forces the disclosure. The opposite outcome, OpenAI keeping ChatGPT ad-free through 2026, would require it to leave the obvious revenue fix on the table while burning cash, which is the less likely path.

Revisit by 2027-01-15: We're right if OpenAI has launched an advertising product inside ChatGPT for general users with no clear per-response sponsored-content disclosure. We're wrong if ChatGPT remains ad-free for general users, or if ads launch with a visible per-response label marking advertiser-influenced content.