AI Adoption

Every Unattended AI Workflow Drifts Toward Generic

September 7, 2026

Feed a model ten strong examples of how your company writes. The first output comes back on voice. So does the second. The third is good. The fourth reads like it was assembled by a committee that has never heard of you, and nothing upstream changed to explain it.

Anyone who has run these systems in production for more than a month has seen this. It is the single most common thing we get called about after a workflow has been live for a quarter, and it is almost never described as drift when the call comes in. It gets described as the tool getting worse.

What drift actually looks like in production

The surprise is the shape of the failure. Teams brace for a slow fade, some gradual erosion they will notice in the monthly review. What they get is intermittent collapse. Nine outputs land where they should and the tenth is unrecognizable.

That shape matters more than the frequency, because it defeats the way most people check. Spot-checking assumes the thing you sample represents the batch. Here it does not. Pull three outputs at random from a run of thirty and you have a decent chance of pulling three good ones and signing off on a batch that contains four you would never have shipped. The check passes. The bad work goes out. Confidence in the system goes up, which is the worst possible outcome, because the next batch gets checked less carefully than this one did.

Two other things about the pattern are worth naming, because they change how people react to it. The off-brand output is usually competent. It is grammatical, on topic, nothing an editor would flag as broken. It simply sounds like nobody in particular. And the model gives no signal at all that this output is different from the nine before it. There is no confidence score that drops, no flag, no hesitation in the text. Everything arrives with the same even tone.

So the failure is quiet, occasional, and plausible. That combination is very hard to catch by feel, and almost impossible to catch late.

How teams usually find out is worth knowing, because it tells you how long the problem has been running. Somebody outside the workflow reads something and frowns. A salesperson forwards a case study and asks whether marketing meant to send this version. A customer replies to a sequence quoting a line back with a question mark. In every one of those, the catch happened downstream of shipping, by accident, and the first honest question in the room afterwards is how many others went out that nobody flagged. Nobody can answer it, because nobody kept a record of what shipped and what it looked like when it did.

Why it happens

Nothing mysterious is going on underneath. A model has an enormous middle: the average of everything it was trained on, which for business writing is an ocean of competent, forgettable content. Your examples, your constraints, whatever you put in the prompt, all of that is pressure applied against that middle. It works. It is also the only thing working, and pressure is not constant.

Context loses potency over a long run. Instructions that sat near the front of a conversation carry less weight by the twentieth turn. Automated chains reassemble context on every call, and small variations in what gets included change the outcome more than the size of the variation would suggest. Sampling introduces real variance by design. Every output is an independent pull, and some pulls land closer to the middle than others.

The pull toward the middle shows up in controlled settings as well as in production. In a study of writers given AI-generated story ideas, individual stories got more creative while the pool of stories became more similar to one another. Each writer was better off. The collection was flatter. That is the same trade a content workflow makes on every run, with nobody standing far enough back to see the collection.

That middle has a gravity to it, and it is the same gravity we wrote about in the convergence problem: models trained on everything produce the average of everything unless something actively pushes them off it. Convergence describes what happens to a market when everyone’s content gets sanded toward the same center. Drift is the same force, aimed at you specifically, arriving on a schedule that depends on your volume.

Which means drift is not a defect to be patched out. It is structural, a property of how the system works rather than a bug in your setup. Better prompts reduce the rate. They do not remove the mechanism. Any plan that depends on eventually getting the configuration right enough to stop checking is planning against physics.

QC never graduates. It only changes who performs it

Most teams treat quality control as a launch-phase activity. Watch the output closely for the first few weeks, tune the prompt, build confidence, then step back and let it run. The stepping back is the point of the whole exercise, or it feels that way when the business case gets written.

Quality control does not end. It changes hands.

At low volume a person reads everything, which is cheap because the volume is low. As output grows, reading everything stops being possible and sampling takes over, with the sample chosen to catch bursts rather than to estimate an average. At real scale a second model checks the first against an explicit rubric, with humans auditing the checker. What never happens at any stage is the function disappearing. The cost per check falls. The number of checks rises. The function persists.

The useful mental shift is to stop treating QC as a project cost and start treating it as an operating cost, the way payroll is an operating cost. Budget it as a standing percentage of the workflow’s output, not as a line item that tapers off after implementation. Every system built on components you cannot fully trust needs a layer whose entire job is catching what the untrusted components got wrong. That is not a sign the system is immature. It is what running one looks like.

The businesses that get burned here are the ones that priced the automation on the assumption that attention was the thing being eliminated. It was not. It was the thing being relocated, and the relocation has a cost that should have been in the model from the start.

Drift detection at your size

None of this requires a platform, a monitoring vendor, or anyone whose title contains the word governance. Four moving parts, all of which a competent operator can run inside an existing role.

Start from a written definition of on-brand that a second person could apply without asking you what you meant. In practice that is your example set, six to ten pieces of shipped work that carry the standard, plus a short list of the things that are always wrong. Prose beats a rubric here, but either works as long as somebody other than the author can use it.

Sample a fixed share of shipped output every week. Ten percent is a reasonable starting point, and the number matters less than the fact that it never moves without a decision. Sampling that flexes with how busy the week was is sampling that stops the week you needed it.

Log every catch, and log what shipped immediately before it. The log is the actual asset. A single off-brand output tells you almost nothing. Fifteen of them, with timestamps and context, tell you whether the rate is rising, whether it clusters around particular input types, and whether last month’s prompt change helped or just moved the problem.

Review the log monthly against the previous month. A rising catch rate means something upstream changed, often a model version you were never told about. Providers update continuously and rarely announce it in terms that map to your workflow, so the log is frequently the only place a version change becomes visible to you at all. A catch rate of zero for two months running usually means the sampler has stopped reading properly, which is worth knowing too, and is the more common of the two failures.

One caution on who does the sampling. The person who built the workflow is the worst possible reviewer of it, for the ordinary reason that they already know what it was supposed to produce and will read that intention into the output. Put the sample in front of somebody with the domain knowledge to judge it and no investment in the system being fine.

When volume outgrows a human sampler, the definition and the log become the specification for a second model that does the first pass. That promotion is straightforward precisely because the hard work was done earlier, when a person had to write down what good looks like in terms concrete enough to apply. Skip that step and there is nothing to automate later.

The uncomfortable part is the one worth sitting with. Automation was supposed to buy back attention, and running it properly costs a permanent system for paying attention. The trade is still worth making at almost any volume. It is just a different trade than the one on the slide.

Sources

1. Doshi, A. R. and Hauser, O. P., Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances, Vol. 10, Issue 28 (2024): https://www.science.org/doi/10.1126/sciadv.adn5290