AI market research: How to brief an agent

AI market research uses AI to collect, compare and interpret information about customers, competitors, pricing and demand. An AI agent can examine competing products, gather evidence and produce a research memo. The challenge is getting a memo that answers a business question rather than a polished collection of loosely related facts.

A useful brief tells the agent what decision the research supports, which sources it may use, what the memo must contain and how to handle missing evidence. Without those instructions, an agent can produce convincing pages that leave the owner no closer to deciding whether to enter a market, change a price or build a feature.

Start with the decision, not the industry

Suppose you run a small software company assessing appointment scheduling tools for independent physiotherapy clinics in the UK. Asking an agent to research the appointment scheduling market sounds reasonable, but leaves almost everything open. It might spend its time describing hospital software, consumer booking apps or international competitors that your prospective buyers would never consider.

A better question is: 'Should we investigate building a scheduling product for independent UK physiotherapy clinics with fewer than ten practitioners, given the existing alternatives and likely buying objections?' Now the agent has a buyer, geography, product category and decision to support.

Set exclusions too. Leave out hospital procurement, general medical records systems and products without appointment booking. Ask for reasons against entering the market. If several affordable competitors already solve the problem, that finding may save more money than an enthusiastic opportunity assessment.

State what would change your mind. Perhaps you need evidence that small clinics face scheduling problems their existing tools cannot solve. Perhaps the question is whether switching software would be too disruptive. Those require different evidence, even though both concern the same market.

Give the agent a source policy it can follow

An instruction to 'use reliable sources' is too loose. Reliable for what? A competitor's pricing page can establish its advertised price, but cannot independently establish customer satisfaction. A forum complaint can reveal a problem worth investigating, but cannot tell you how common that problem is.

For the clinic example, we would give the agent these source rules:

• Use official competitor websites and documentation for advertised features, prices and stated limitations.
• Use public government or regulatory material for operating requirements and market definitions.
• Use identifiable customer discussions for examples of frustrations, without treating individual accounts as representative of all buyers.
• Use third-party analysis only when its origin, date, scope and method can be inspected.
• Never treat another AI-generated answer as original evidence. Find the underlying source.

Require a working URL, page title and access date for each material factual claim. Where possible, ask for the exact supporting passage. The agent must distinguish a page it opened from a source mentioned elsewhere. A citation beside a sentence does not prove that the source supports it.

Freshness depends on the claim. A regulatory definition may remain useful for years. A subscription price can change tomorrow. Ask the agent to flag time-sensitive findings and mark anything it cannot verify. An empty cell is better than an invented price.

Specify the memo before research begins

Without an output specification, the agent decides what deserves space. That can produce a lengthy industry overview with little attention to your decision. We prefer a short decision memo with supporting evidence kept separately. An owner should be able to read the assessment quickly, then inspect the material behind it.

Set a length limit, perhaps two pages for the main memo. Require the same comparison criteria for every competitor. If a product does not publish its price, the cell should read 'Not publicly verified', rather than contain an educated guess.

Separate findings from interpretations. 'These competitors advertise online booking' is factual if their current pages support it. 'Clinic owners will pay extra for automated reminders' is a hypothesis unless there is direct buying evidence. Both belong in research, but they cannot carry equal weight in the recommendation.

Write a brief with a clear stopping point

Give the agent a bounded assignment rather than permission to investigate indefinitely. This example can be adapted to another industry by changing the buyer, competitors and decision.

Research appointment scheduling software for independent UK physiotherapy clinics with fewer than ten practitioners. The decision is whether we should invest in customer interviews about a new scheduling product. Compare up to six relevant competitors. Use official product pages for features and prices, identifiable customer discussions for examples of problems, and public regulatory sources where relevant. Provide source URLs, access dates and evidence for material claims. Do not invent market sizes, customer counts or prices. Produce a two-page decision memo with a competitor table, arguments against entry, evidence gaps and five customer interview questions. Put supporting references in an appendix. If the evidence cannot support a recommendation, say so. Do not contact companies or individuals.

Notice the stopping point. The agent is deciding whether interviews are worth doing, not declaring the business viable. Desk research cannot establish what a particular clinic owner would pay or which frustrations would cause someone to switch providers.

Set a time or spending limit. Ask for an initial evidence inventory before the finished memo: competitors found, sources checked and important gaps. A person can then narrow the scope rather than paying for research that cannot resolve the decision.

In Delagents, a person can assign this brief to the research analyst template and receive the work as a deliverable. Every delegation has a hard budget cap, and an append-only activity log records tool calls, approvals and policy changes. The log helps a reviewer inspect what happened. It does not prove that the research findings are correct.

Check the claims most likely to be invented

Fabricated information often looks ordinary. A competitor price may seem reasonable for its category. An impressive market growth claim may appear beside a professional-looking citation. A tidy table makes both easy to accept without checking.

Review the claims that would change your decision first. If the recommendation depends on market demand, inspect the evidence for actual buyer interest. If the memo says competitors lack a feature, check their current documentation. A missing mention on one page does not prove the feature is unavailable.

1. Open the original sources behind the three most consequential claims. Confirm that each source exists and supports the precise wording.
2. Check every number that affects the recommendation. Look for outdated prices, mismatched currencies, unclear calculations and estimates presented as facts.
3. Inspect sweeping negative claims such as 'no competitor offers this'. These need stronger evidence than a claim about one named product.
4. Read the evidence gaps before accepting the assessment. If willingness to pay remains unknown, the recommendation must not assume strong demand.
5. Ask which finding would most change the assessment if disproved. Verify that finding yourself.

Watch for citation laundering. An agent may attach a real homepage to a claim that appears nowhere on that site, or repeat a figure found in another summary without finding its origin. Ask for the supporting passage, not merely a plausible reference.

Keep customer accounts separate from general claims. Several complaints about missed appointments may suggest a useful interview question. They do not establish that missed appointments are a common problem across UK physiotherapy clinics. The memo should preserve that distinction.

Turn the memo into a controlled next action

For the clinic example, a defensible next action might be interviewing five owners about scheduling failures, switching costs and existing workarounds. The memo should propose questions tied to uncertain assumptions. It should not automatically start contacting people.

In Delagents, first contact with anyone outside the workspace requires a person's approval. Proposed external actions wait in the approvals inbox as cards showing the exact action, cost so far and deadline. A person can approve, change or reject them. Research findings therefore do not become external outreach without a sign-off.

Before approving an interview invitation, inspect the recipient list, message and selection criteria. Five clinic owners using the same scheduling provider may give a narrower view than owners using different systems. A well-supported memo can still lead to a poor decision if the next sample is wrong.

For your next AI market research assignment, write one decision question and specify four permitted source categories. Define the memo sections before the agent begins. When the draft arrives, pick the three claims that most influence its recommendation and open their original sources. Record which assumptions still need direct evidence, then decide whether to authorise the next expense.