A short, slightly awkward post on the talk I gave repeatedly in 2023 and 2024 (which sat alongside the talks I do give now) that I won't be giving again. The argument was wrong. Worth being explicit about it because some of the audiences who heard it are still operating off the conclusions, and I'd rather have the correction in writing than let the previous version of my thinking continue to do work.
The talk was, broadly, that the integration of AI into marketing operations would be slower and less transformative than the early breathless coverage suggested. The argument had three parts. The technology was less capable than the demos implied. The operational discipline required to deploy it in production would slow adoption significantly. And the marginal value over existing tooling was smaller than the hype was claiming.
I gave versions of this talk for about 18 months, mostly in 2023 and into early 2024. The audiences responded warmly. The argument was well-received in agency leadership circles where the appetite for sceptical takes on AI was higher than the appetite for breathless ones.
I now think the argument was wrong on the first two points, partly right on the third, and that the overall framing was misleading. Worth walking through what changed.
Where I was wrong
On capability. The argument that the technology was less capable than the demos implied has held up worse than I expected — and the chatbot-vs-agent distinction is part of why. The demos in 2023 were, if anything, undersold versions of what the underlying systems were capable of. The capability progression through 2024 and 2025 was faster than I had calibrated for. By mid-2025 the kinds of agentic workflows I'd argued were 5-7 years out were shipping in production. By late 2025 we were shipping them ourselves.
The mistake was a specific kind of seasoned-operator caution. The technology felt like the kind of thing that previous waves had also looked promising in early demos and underdelivered in deployment. SEO automation in 2008. Programmatic advertising in 2012. AI-driven creative tools in 2018. The pattern across these was real over-promise, real under-deliver, real eventual maturation but on a longer timeline than the early enthusiasts predicted.
I applied that pattern to LLMs. The pattern was wrong this time. The capability arc has been steeper, the deployment maturity has come faster, and the operational tooling has filled in more rapidly than equivalent waves did. The argument that "we've seen this before" was a specific kind of pattern-matching that turned out to have low predictive value.
On operational discipline. The argument that the operational discipline required to deploy AI in production would slow adoption significantly has held up better in part — there is a discipline gap, and many teams haven't bridged it — but the consequences of the gap have been smaller than I predicted.
The reason: the providers (Anthropic, OpenAI, Google, Microsoft) have been increasingly carrying more of the operational work themselves. The tooling, observability, error handling, and safety layers that I expected each adopting team to need to build, have been increasingly available off-the-shelf or built into the platforms. The discipline gap exists, but the floor on which un-disciplined teams operate has risen.
The result: the deployment lag I predicted hasn't materialised at the scale I expected. Production agentic systems are more common than the 2023 version of me would have anticipated.
On marginal value. This is the part I was partly right about. The marginal value of LLM-based tools over existing tooling is real but uneven. Some workflows have seen dramatic uplift. Others have seen modest improvements. Generic claims about "AI saves X% of time" remain unreliable. The variance is high, and the workflows where the value is high are not always the ones leadership teams initially target.
The portion of the original argument that holds: leadership teams that target the wrong workflows, build chatbot-style implementations, or skip the operational design, are getting marginal returns from their AI investment. The portion that doesn't: the teams that get the workflow choice and operational design right are getting returns that meaningfully outpace the alternative tools.
The original argument lumped both groups together and concluded that the technology was overhyped. The corrected argument separates them and concludes that the technology is underhyped for the teams getting the implementation right.
Why the original argument played so well
It's worth thinking about why audiences responded warmly to the talk for 18 months even though the underlying argument was wrong.
Two reasons, I think.
The audiences had been over-marketed to. Most of the agency leaders and senior marketers I was talking to in 2023 had been bombarded with AI vendor pitches, AI consultancy pitches, AI-everything talks. The fatigue was real. A talk that pushed back on the breathless coverage was emotionally satisfying. The talk worked partly because of the audience's appetite for it, not because the argument was particularly well-supported by the data.
The argument was harder to disprove in 2023 than it is now. In late 2023, the data on AI deployment success and failure was thin. The arguments on both sides — over-hyped and revolutionary — could be made with broadly similar evidence. By late 2025, the data had thickened, and the over-hyped argument became harder to defend.
Both of these are uncomfortable to acknowledge. The talk landed because the audience wanted it to land. That's a different reason from the talk being right. The conflation between the two is a hazard for any speaker — particularly one operating at conference scale where the immediate signal is audience response, not eventual accuracy.
What I'd say instead
If I were giving a 2026 version of this talk, the argument would be inverted. The technology is capable. The operational discipline gap is real but bridgeable. The teams that get the implementation right are pulling away from the teams that don't. The over-hyped vendor pitches haven't gone away, but the legitimate underlying technology has matured to the point where the cynical default isn't really earning its keep.
The corrected argument lands less warmly with audiences who are exhausted by AI-everything coverage. That's fine. The audience response isn't the right measure of whether the argument is right.
I'm leaving this post up partly as accountability and partly as a reminder to myself. The talks I give in 2026 should pass the test that I'd still be willing to defend them in three years. The 2023-24 version of the AI talk wouldn't have. The corrected version, I think, will.
I'll be wrong about something else in 2027. When I am, I'll write a similar post. The willingness to update, in writing, in public, is part of the deal you make when you put yourself on stages. Worth honouring it.