Part IV
How to decide which strategic tasks to give AI
I would avoid drawing a permanent boundary around which strategic tasks belong to AI. The models will improve, tools will change and agencies will learn better ways of working. A more useful test looks at the task, the standard and the ability to verify the result.
Can you explain what needs to be done? Can you explain what good looks like? Can you check whether the output meets that standard? What happens if it is wrong? If the first three answers are clear and the cost of error is manageable, AI can probably take on a substantial part of the work.
Clear analytical tasks are strong candidates
AI can already perform well when the job is refined and the evidence is available. It can structure research, compare competitors, identify patterns in data, surface contradictions, generate hypotheses and produce first drafts.
Drafting is an especially obvious candidate because it contains several earlier steps. To create a useful draft, the system has interpreted information, made a first analytical pass and organised what it found. Starting every assignment from a blank page becomes harder to justify.
This can raise the floor in almost every situation. A rushed job begins from something more coherent. A junior has a stronger first pass to challenge. A senior spends less time assembling material.
Ambiguity changes the role
If the task cannot be defined clearly, asking AI for a solution encourages it to guess what the problem is and then solve its own guess. In that situation, AI is more useful for diagnosis.
It can ask what objectives conflict, which assumptions are unsupported and what information would change the decision. The user can then decide which problem deserves solving.
This is one reason sequence matters. The same model can be helpful or harmful depending on whether it is asked to diagnose first or jump directly to a polished recommendation.
Judgement remains difficult to specify
Decisions involving trade-offs and taste are harder. You can teach AI criteria to a considerable degree. You can show examples, encode processes and describe the features of good work. I do not believe that captures the full extent of experienced judgement.
Sometimes a strategist looks at a plausible answer and says, “It does not feel right.” That can be laziness or bias. It can also be experience arriving before explanation. The person has seen enough categories, clients, ideas and failures for something to register before they can name it.
If nobody can explain what good looks like, AI can still generate options and help articulate the discomfort. I would keep the final call with the person who has to own the consequences.
Verification and cost of error are the forgotten conditions
A task may be easy to describe while the result is difficult to verify. Synthetic audience research is a good example. The model can produce a coherent reaction, but coherence does not prove that real people share it.
Verification may require source checking, comparison with behavioural data, expert review or traditional research. The cost of error should determine how much is needed. A weak exploratory hypothesis can be discarded cheaply. A confident recommendation that changes a major investment needs a different standard.
Agencies should be especially careful when a user lacks experience and is short of time. That is when fluent output is most likely to pass without a proper standard.
Human work includes working with humans
Strategic value also appears in places that look peripheral in a workflow diagram: persuading a client, deciding which disagreement matters, helping a creative team turn a correct direction into interesting work, and understanding the incentives around an account.
AI can prepare, challenge and support these moments. It cannot own the relationship in the same way as the people whose reputations and decisions are involved.
I do not have a universal allocation chart for AI and humans. The practical rule has four parts: specify the task, specify quality, verify the result and understand the cost of error. Give AI more responsibility where those conditions are clear. Use it to help diagnose ambiguity. Keep consequential judgement close to people who understand the problem and can own the decision.