AI Business Advice: Check the Fit Before You Act

An AI assistant can turn a worried business owner’s question into an orderly plan: change the offer, promote it differently, reach more customers. The plan has headings, sensible verbs and a reassuring sequence. The difficult work begins when someone must decide which recommendation belongs in this particular business.

A photographer with empty weekday appointments faces a different problem from one who is fully booked but struggles to finish editing. Both might ask how to grow. Useful advice must survive that difference. Before acting on an AI-generated recommendation, make its assumptions visible and decide what evidence would justify the next step.

What a field experiment can tell us

A recommendation can be plausible without being useful in the business receiving it. A field experiment published in Management Science on 10 July 2026 illustrates the difficulty. Researchers randomly assigned 640 Kenyan entrepreneurs to access either a GPT-4-powered assistant through WhatsApp or conventional business guides. The experiment ran from late May to early November 2023.

The researchers found no statistically significant average effect on business performance. Below-median businesses at baseline experienced negative estimated effects; above-median businesses had positive estimates with greater uncertainty. Outcomes relied on self-reported revenues and profits. Further analysis pointed towards differences in the changes owners implemented, but could not rule out alternative explanations.

These findings concern one implementation, setting and period. They establish neither a universal benefit nor a verdict on every current AI assistant. The practical lesson I draw is to examine the journey from recommendation to action.

Describe the business before asking for growth

Consider a fictional portrait photographer whose Saturdays are full while Tuesdays remain quiet. Imagine an AI response recommending lower prices and more advertising. Neither suggestion yet explains the empty Tuesdays. Possible explanations include unsuitable appointment times, unclear booking information, limited awareness or an offer that does not appeal to the intended customers.

The owner first checks the diary, recent enquiries and the booking page. Are weekday sessions actually visible? Do enquiries ask for different times? Has anyone identified price as the obstacle? These questions give the proposal something concrete to answer. MaryChuks’ guide to customer interviews that investigate a real problem can help when the missing information concerns customer behaviour.

Write a brief account of the situation: the result sought, what is already happening, the available capacity and the commitments that must be honoured. Separate confirmed observations from guesses. If the obstacle remains unknown, ask for ways to investigate it before requesting a campaign.

Expose the conditions behind the recommendation

Treat each proposed action as conditional. Lowering a price would require a reason to believe price is obstructing suitable bookings. More advertising would require an audience, a relevant message and capacity to deliver any work it brings. An attractive instruction still needs those conditions.

For the photographer, a useful review might ask:

• Problem: which observed difficulty would this action address?
• Missing evidence: what are we assuming about customers or demand?
• Capacity: what would delivery require if the action worked?
• Trade-off: which existing work or commitment might be affected?

An answer such as “everyone likes a discount” leaves the central question open. A stronger proposal names the intended customers, the specific appointment slots and the reason the offer might matter to them. Those details should come from evidence the owner can check.

Turn the action into a claim you can test

Suppose the diary and page review reveal that the booking page barely mentions weekday availability. The owner could investigate a narrower proposition: making those appointments visible will help suitable customers find them. That is a different decision from cutting every session price.

The test needs a defined change, an observation period, a record of enquiries and a decision it will inform. Track whether enquiries concern the intended slots, whether they become bookings and whether the resulting work fits available editing time. Choose a scope the business can deliver without disrupting existing commitments.

Research on entrepreneurial decision-making offers a related perspective. A study first published on 25 March 2024 combined four randomised trials involving 759 firms. Entrepreneurs trained to develop explanations, derive hypotheses and test them were more likely to terminate ideas. The results also favoured a limited number of major strategic changes over none or repeated changes. This was a business-training intervention, rather than a test of AI advice. It supports taking the testing process seriously; it does not validate the particular workflow proposed here.

Let a small test answer a small question

A handful of enquiries cannot establish that a change caused growth. Seasonal demand, other promotions and customer mix may also change. Keep those possibilities alongside the result. A simple comparison with last month can inform judgement while leaving causal uncertainty unresolved.

Perhaps the revised page generates questions about Tuesday evening appointments, although the photographer offers mornings. That would suggest a timing issue worth investigating. It would not, by itself, justify more advertising. Alternatively, no relevant enquiries may arrive, leaving awareness or the offer itself as open questions. These are hypothetical outcomes, rather than promised results.

Record what was changed, what happened and what remains unknown. Decide whether the next step is to repeat the observation, speak to customers, revise the proposal or stop. Continuing because the AI can produce another list of improvements gives the decision no evidential stopping point.

Use the assistant to challenge the proposal

Give the assistant the same confirmed facts and ask it to identify assumptions, explain competing possibilities and suggest observations that would distinguish them. Ask which recommendation becomes unsuitable if the available capacity or customer need changes. Require any missing information to remain explicitly unknown.

Then review the response yourself. An AI-generated objection can be useful without being correct; an AI-generated endorsement adds no independent evidence that customers want the offer. The owner still needs to inspect the diary, read the enquiries or investigate the costs of delivery.

Keep a short decision record containing the selected action, its reason, the evidence still needed and the review date. This makes it possible to revisit a choice without reconstructing the entire conversation or mistaking the latest suggestion for an agreed plan.

Make advice earn its place in the business

Owners have different resources, obligations and access to support. A disappointing result should prompt examination of the recommendation, the operating conditions and the help available for implementation. Judging someone’s ability from whether an AI-generated plan worked would conceal those differences.

Ask the assistant to make the choices clearer. Give the business’s actual constraints authority over the proposed plan. Act on a recommendation when you can explain why it fits, what you expect to observe and what would make you reconsider.

Sources

• Otis and colleagues: The Uneven Impact of Generative Artificial Intelligence on Entrepreneurial Performance: Evidence from a Field Experiment in Kenya, Management Science, published online 10 July 2026.
• Camuffo and colleagues: A scientific approach to entrepreneurial decision-making: Large-scale replication and extension, Strategic Management Journal, first published 25 March 2024.