Every organisation we work with has a list of AI pilots. Most of the lists look the same: a summariser for meeting notes, a drafting assistant for proposals, a search box over the policy library. Individually, each pilot works. Collectively, they change very little. The work still takes the same number of days, because the pilot removed a few minutes from a step in the middle and left the sequence around it untouched.
This is the most common and most expensive misunderstanding in AI programmes, and it is worth naming precisely. A step is one action inside a process. A workflow is the whole sequence from request to finished, accepted output, including the handoffs and the definition of done. Automating a step optimises an action. Redesigning a workflow changes what the organisation is able to produce.
Step versus workflow, with an example
Take a commercial team that produces a monthly pricing pack. The old workflow: three analysts pull data, build a model, write commentary, a manager reviews, a director approves, the pack is circulated. It takes nine working days and two rounds of rework.
| Step automation | Workflow redesign | |
|---|---|---|
| What changes | Commentary is drafted by a model | The data pull, model build and commentary become one generated artefact |
| Sequence | Unchanged | Review moves before assembly, not after |
| Handoffs | Still three | One reviewer, one approver |
| Definition of done | Same | Model output must cite every figure |
| Effect on elapsed time | Marginal | Material, because waiting time was removed |
The right-hand column is not a better prompt. It is a different process in which the model is one participant. Crucially, it also changed the definition of done: every figure in the output must be traceable to a source row, because the reviewer can no longer assume a human assembled it.
How to tell which one you are actually doing
Three questions, asked about any AI initiative, will separate a step from a workflow. We use them before a scoping call, and they routinely change what the client asks for.
- What is the finished artefact, and who accepts it? If the answer is vague or plural, you are automating a step.
- How many hands touch it between request and acceptance? If that number is unchanged after the initiative, you have optimised a step.
- If the model were removed tomorrow, would the process still work? If yes, it was a step. If the process was rebuilt around the model and cannot revert without loss, you changed the workflow.
A usefully uncomfortable testMeasure elapsed calendar days from request to acceptance, not hours saved in a step. Hours saved flatter a pilot; elapsed days change the business.
What a workflow redesign actually changes
Four things change together, and skipping any one of them causes the pilot to quietly revert to the old process after three months.
- The sequence. Work that used to happen in parallel now happens once; work that used to happen last now happens first, because a model is cheap and a reviewer is not.
- The handoff count. Every handoff is a place where work waits and context is lost. A redesigned workflow typically removes one or two, not because the people were redundant but because the reason for the handoff disappeared.
- The definition of done. New artefacts need new acceptance rules. If a human did not assemble the output, the reviewer needs provenance, not just plausibility.
- The exception path. What happens when the model is wrong, or when the request is unusual? A workflow without a designed exception path fails loudly in front of a client.
The exception path is the part teams underestimate. In a human process, exceptions are handled by the person who notices them. In a rebuilt workflow, the exception has to be routed explicitly, or it becomes a silent error that surfaces months later in a number nobody can explain.
Who signs the work
There is one question that decides whether an AI workflow is trustworthy: when the output is wrong, whose name is on it? If the honest answer is “the model’s”, the workflow is not ready to carry consequential decisions, and no amount of accuracy improvement will make it so. Accountability does not degrade gracefully into a language model.
In practice this means every AI-native workflow we install has a named human owner who accepts the output, and an evidence trail that lets that owner check a sample without redoing the work. The owner is not reviewing everything. The owner is able to review anything.
The model does not remove the need for judgement. It moves judgement to a different place in the process, and makes it cheaper to exercise — if you designed the workflow to allow it.
This is also why the weakest AI programmes are the ones that start with a tool and look for a use. Starting from the finished artefact, the owner and the exception path forces the hard questions early, when they are still cheap to answer.
Sources and further reading
- Our own observation, stated as such: the step-versus-workflow distinction and the three diagnostic questions come from field work with commercial and operations teams. This is experience, not a controlled study.
- Evidence class: qualitative. We have not measured a population of AI programmes, and we do not present the time effects above as averages. They are the shape of change we observe, not a benchmark.
