Product1 publisher3 min readPublished
Fast Company borrows Nadella's veteran test to ask what survives a model swap
A Fast Company column argues that most deployed corporate AI stores what happened without learning what its own decisions caused, and it offers a check on whether your accumulated expertise would outlive the model you swap out tomorrow.
The Product Desk · Product desk
What happened
- A Fast Company column argues that much deployed corporate AI accumulates memory and retrieves past interactions without getting any better at deciding what to do next.
- Citing Microsoft, the column says the gains come when teams learn what happened in production and feed those signals into prompts, routing, retrieval and other system decisions.
- It credits Satya Nadella with a company veteran test: if you replaced your model with a different one tomorrow, would the expertise accumulated inside your organisation remain?
- The column's standard for a corporate system is that interaction number one million should leave it better prepared for interaction 1,000,001.
- The piece reports no survey or sample of deployments, so how many teams already record outcomes alongside transcripts goes uncounted.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Prompts, routing and retrieval can only use outcomes that were logged at the time, so a team holding transcripts alone cannot retrofit last quarter's lessons into this quarter's system.
- cost Recording whether the discount closed the deal is labour no license covers, and it lands on the same team that owns the rollout and the feature backlog.
- decision At renewal the useful question becomes which of your own tables holds the expertise, since the model on the other side of the contract is available to competitors on the same terms.
- contradiction The column prescribes compounding learning for LLM deployments while its only named working loop is recommender feedback that it says runs without AI. The prescription sits ahead of the demonstration.
The column's own illustration is a discount. Remembering that one customer said no to 10% is a record; learning that the segment responds better to faster implementation than to a discount is a different thing, the Fast Company piece argues [8]. The first fact exists the moment the call ends. The second needs dozens of calls, the results of those calls, and something that joins the two a month later.
That timing explains why transcript archives grow on their own and outcome tables do not. Logs of what was said arrive as a byproduct of use. Whether the deal closed, and whether the refund stopped the customer leaving, has to be written down afterwards against the decision that caused it. The column wrote that "remembering is not the same as learning" [2], and that using an LLM more does not, by itself, make the deployed model learn from the consequences of its actions [13].
None of the insertion points the column names requires a training run [5]. A team can act on outcome data with a routing rule and a retrieval filter and never touch a fine-tuning pipeline.
Applied to a real stack, the veteran test turns into a question about file ownership: which table holds last quarter's outcomes, and who writes to it. The column's reason for caring is commercial. Competitors can buy the same frontier models, so access to them is not a durable advantage, and what is left is what you learned from your own customers, decisions, campaigns, corrections and exceptions [10]. The column also argues the accumulated learning has to live outside the replaceable model [14].
On evidence, the piece is an argument. It has no survey and no sample of deployments, so the claim that most corporate AI fails this test is an assertion [11]. Its one named example of a feedback loop already running in production is recommendation at YouTube, Spotify and Amazon, which the column says has long used behaviour and feedback streams even without AI [7]. The research programme it points to, continual learning, is described as a field with a goal and no shipped capability [6].
So the forcing function, which costs nothing to run: the two or three decisions an assistant already touches in a product, with two questions put to each. Is the outcome written down somewhere with a pointer back to the decision that produced it? Does anything downstream read it? Nothing written and nothing reading is a transcript archive. Written but unread is a warehouse table with a dashboard on it. Something reading a record nobody writes means a person is filling the gap from memory. Both written and read is the loop the column is asking for.
Most teams can answer the first question and stall on the second. The cheapest test is one workflow with one outcome field that something reads every week, and the tradeoff is that one workflow will not tell you whether the loop generalises to the other five.
What to watch
- Whether Microsoft ships outcome capture as named tooling in its AI platform docs, or keeps it as guidance about production signals.
- Whether any vendor publishes retention or decision-quality numbers for a memory feature instead of usage counts.
- Whether continual learning produces something a team can deploy without a retraining pipeline.