We had no idea how to do it. He wanted to start Monday anyway

Diagnose first: what two decades of bad pivots teach about the AI rush

Twenty five years ago I sat in a business development meeting at a company I’d just joined. We built 3D graphical databases for simulation and training, primarily air traffic control. At that time Formula 1 had a real problem: tobacco, its main source of advertising revenue for decades, was being banned in country after country. An executive had called an emergency meeting to present an idea for a solution. Superimpose advertising onto the cars in real time, digitally, per broadcast region, so a car could carry a tobacco ad in a country where that was still legal and something else everywhere it wasn’t. He wanted to launch a development program on the spot to build a proof of concept.

We had no broadcast experience. We had no technology anywhere close to real-time video compositing at that speed. We didn’t know the cost of getting into that market, who else was already circling it, what was driving the demand, or how big the opportunity even was. The executive couldn’t address any of it. None of that stopped the pitch. The problem was real. But wanting to solve a real problem and being the company that should solve it are two different questions, and having the second one land unanswered in the room didn’t slow him down at all.

I’ve watched that same meeting happen many times since, and lately it happens under a different banner. Something strange happened to Allbirds in April 2026. The company makes wool sneakers. It has never built an AI product, never operated a data center, and has no GPU procurement team. It announced a rebrand to “NewBird AI” and a stated ambition to become a GPU-as-a-service provider. The stock jumped 600% in a single afternoon.

Allbirds’ actual capacity to deliver AI infrastructure was exactly the same the day after the announcement as the day before. Only the word attached to the ticker was different.

This is a different failure than the one usually told about companies botching their AI pivots. The standard story is about execution: a company correctly identifies a real AI opportunity, then implements it badly, exposes its balance sheet, or dismantles the wrong revenue stream on the way there. That story matters, but it assumes the company got the diagnosis right and only fumbled the delivery.

Allbirds never had a diagnosis. Did anyone sit down and ask what problem the business actually had, whether AI was the tool that solved it, or whether Allbirds had any of the assets, talent, or data required to compete in GPU infrastructure. The stock moved anyway, because enough investors decided the word “AI” was itself worth 600%.

The threat you name isn’t always the threat you have

Boards under pressure reach for AI in two postures: as a shield against something they think is coming for them, or as a lever they think will pull the stock up. Both postures skip a step. Before you can point AI at a threat, you have to know precisely what the threat is and precisely where your business is exposed to it. Skip that step and AI becomes a gesture rather than a strategy, and gestures get priced like gestures: briefly, and then not at all.

Chegg is the clearest example of the shield version of this mistake. Leadership was right that ChatGPT was an existential threat to a homework-help subscription business. Where they went wrong was in never diagnosing what, specifically, students were paying $19.95 a month for once a free chatbot could answer the same question. Was it accuracy? Speed? The credibility of an “expert-verified” answer? Chegg never answered that question before building CheggMate, an AI tutor running on the same GPT-4 the students were already using for free. They named the threat correctly and diagnosed nothing about their own exposure to it. Non-subscriber traffic fell 49% year over year within months, and the stock is down roughly 99% from its 2021 peak.

Allbirds and the small-cap wave documented by Bloomberg in May 2026 are the lever version. There’s no threat being defended against, no distressed business model that AI might rescue. There’s a share price a management team wanted higher, and a market willing to hand it to them for the announcement alone. Bloomberg’s reporting on the broader trend of small companies renaming themselves around AI describes shares “surging, then tumbling back down” once the gap between the narrative and the operating business becomes impossible to ignore. Long Island Iced Tea Corp did the same thing with “blockchain” in 2017, a beverage company that added the word to its name and saw a 500% pop with nothing behind it beyond the rebrand itself. The pattern isn’t new. Only the word changed.

One more case is worth a shorter mention, mostly as a warning about the story itself. One outlet reported that Cracker Barrel, the casual restaurant chain, shares surged in June 2026 on an “ambitious AI integration strategy.” I’d treat that claim with real skepticism. Every other piece of coverage on Cracker Barrel’s 2026 stock moves, and there’s a lot of it, attributes the year’s swings to a $100 million logo redesign, the CEO ouster that followed it, and a Q3 earnings beat that alone produced a 27 to 35% single-day pop. Nobody else names AI as the driver of anything. Maybe a real AI announcement got buried under the logo saga, or maybe one story mislabeled an earnings pop because “AI” makes a better headline than “beat guidance.” Either way it’s the same pattern seen from one step further out: once the market decides AI headlines move stocks, every stock move near an AI headline gets credited to AI, whether that’s what happened or not.

Real revenue isn’t proof of diagnosis

Salesforce is the case that looks like an exception until you check the sequence. Agentforce isn’t a rebrand with nothing behind it. It closed real deals, and by mid-2026 the reported numbers were substantial: $1.2 billion in annual recurring revenue, up 205% year over year. That’s the kind of growth rate that should be unambiguously good news. The stock fell roughly 30% in 2026 anyway.

Agentforce launched in fall 2024, priced at $2 per conversation, in direct response to the generative AI wave that had started a year earlier. Trade coverage of that launch describes it as a compromise nobody was happy with: the product team wanted usage-based pricing to protect margin against compute costs, the sales organization knew its customers bought and understood seats, and what shipped satisfied neither. Customers didn’t understand what counted as a “conversation,” and by May 2025, only 8,000 of Salesforce’s 150,000-plus customers had adopted the product. What followed, Flex Credits, Flex Agreements, a pivot back toward per-user licensing, all inside a few months, reads less like strategic iteration and more like a company patching a launch that had already failed.

Then, in October 2025, with that adoption problem still unresolved, Salesforce stood on stage at its Investor Day and set a $60 billion revenue target for fiscal 2030, a 10%-plus compound growth rate, with Agentforce named as the engine that would get it there. That’s the ambitious, publicly announced number. Nobody asked, or at least nobody answered on the record, whether the product had actually earned that role yet.

The market’s current skepticism is the diagnostic question showing up nine months late. Agentforce’s $1.2 billion in ARR is only about 2.6% of Salesforce’s total guided revenue. Analysts at KeyBanc and Bernstein both downgraded the stock in July 2026 for the same reason: adoption is running well behind the headline growth rate, customer data isn’t organized enough for the product to do useful work yet, and, in one analyst’s words, Agentforce “just isn’t there” as a product. A survey of CIOs found more planning to cut Salesforce spending over the next year than raise it. On the Q1 FY27 earnings call, the CFO attributed guidance to “continued momentum in Agentforce” offsetting “ongoing weakness in Marketing and Commerce and increased softness in Tableau,” language that reads less like a diversified plan and more like one growing product being asked to cover for several shrinking ones.

Put the three cases side by side and the mistake is the same one every time. Allbirds got rewarded for announcing AI with nothing behind it. Chegg named its threat correctly and never diagnosed its own exposure to it. Salesforce shipped something real, but shipped it in response to a wave rather than a diagnosis, absorbed a near-total adoption failure, patched the pricing after the fact, and then bolted a large public growth target onto a foundation nobody had verified was ready to carry it. The dollar signs are different. The sequence, target first, diagnosis never, is identical.

What diagnosis looks like when it’s done first

The companies that get this right don’t skip the two questions. First: what is actually threatening us, specifically, not “AI” in the abstract but the mechanism by which it erodes our position. Is it a substitute product, a cost structure competitors can now beat, a distribution channel that’s disappearing? Second: where are we specifically weak against that mechanism, and does AI touch that weakness or somewhere else entirely.

A homework-help company that had done this work might have concluded its actual asset was verified accuracy and institutional trust, something a general-purpose chatbot doesn’t reliably offer, and built toward defending that instead of racing to out-feature a free tool it could never out-feature. A shoe company doing this work would have concluded, correctly, that GPU infrastructure has nothing to do with anything it’s good at, and stayed out of the announcement entirely. Diagnosis doesn’t guarantee a good outcome. It guarantees the AI investment is aimed at something real.

The absence of diagnosis is cheap to spot from the outside: the announcement describes AI in the abstract, general capability rather than a named problem it solves, and the market reaction happens on the announcement itself rather than on any subsequent evidence that something changed. A 600% stock move on a rebrand with no product behind it isn’t a verdict on AI. It’s a verdict on how little diligence gets applied before the word does its work.

The bottom line

AI can be a legitimate defense against a real threat and a legitimate lever for real value creation. It’s neither when the threat hasn’t been named precisely or the weakness hasn’t been located. Most companies getting punished for their AI pivots get lumped together as execution failures, implying they had a sound strategy and fumbled the delivery. A good number of them never had a strategy to execute badly in the first place. Allbirds had a word it hoped the market would reward, and for a while it did. Chegg had the right threat and the wrong diagnosis of its own exposure to it. Salesforce had a real product, shipped in response to a wave of hype rather than a read of its own customers, patched in public once the adoption numbers came in ugly, and pointed at a growth target before anyone could say the foundation was ready to hold it. Different amounts of substance behind each one. Same missing step in front of all three.

I still think about that F1 meeting sometimes. Nobody in the room was stupid, and the problem on the table was real. What was missing was someone willing to let an unanswered question actually stop a decision. Twenty years on, the technology changed and the instinct didn’t. The word on the pitch deck now is AI instead of real-time broadcast compositing, but the tell is identical: watch whether the questions get answered before the money moves, not whether they get asked.