For two years we have been putting into production systems that do real work: they handle enquiries, prepare documents, run processes. When they fail, it is almost never because the model got something wrong in the way people imagine. They fail earlier, at a very specific point.

The system does not know what this company does differently from every other.

A figure usually quoted half-way

The study Generative AI at Work, by Brynjolfsson, Li and Raymond, followed more than five thousand customer-service agents with access to a generative assistant. The headline in circulation is that productivity rose 14% on average. The interesting part is the other one: the effect was close to 35% among the least experienced professionals and essentially zero among the most senior.

The obvious reading is that the tool shortens the learning curve. The useful reading is different: what the system contributed was the knowledge the veterans already had in their heads. It invented nothing. It distributed something that already existed and that nobody had written down.

Which is why context is the product

A digital worker without context is a very fast intern: it answers anything, confidently, and cannot tell what it knows from what it is assuming. With context — the real procedures, the exceptions, the cases that ended badly — it behaves like someone who has been around a while.

And context cannot be bought. It has to be extracted from the organisation, ordered and maintained. It is the slow part, the part that never appears in a demo, and the part that decides whether the project works.

The three failures we see every time

  • The system answers where it should escalate. It has not been taught to recognise the limit of what it can resolve. This is the most expensive failure, because the customer does not realise they were talking to a machine until it is too late.
  • The data arrives late. An assistant querying an inventory refreshed overnight will give answers that are correct and out of date. Answer quality depends on data freshness, not on the model.
  • Nobody measures the cases that slipped through. Conversations handled get measured. The ones that should have been escalated and were not almost never do.

Adoption is no longer the problem

According to Stanford’s AI Index Report 2026, 88% of surveyed organisations use artificial intelligence and 70% have adopted generative AI in at least one business function. Adoption is settled. What is not settled is the difference between using it and trusting it with a process.

Trusting it with a process means accepting that it will work unsupervised for stretches. And that is only sensible if you have first defined where it has to stop.

How we build them

We always start in the same place: writing the procedure as if we were about to hire a person. If we cannot explain it at that level of precision, the process is not ready to be automated, and the problem is not technical.

Then you fence the perimeter, define the escalation path and instrument everything that passes through it. Automation comes last, not first.

A digital worker inherits the clarity of the organisation that deploys it. If the organisation has none, automation only makes the confusion travel faster.