# Working with AI on a company research project: student guide *Core guide v0.2 · 30 September 2026 · Public and reusable. It holds no partner, school, date or brief wording; those live in each cohort's private case sheet. Served by the course tools (MCP) and assembled into each cohort's pack.* Give this guide to your AI at the start of a new chat, together with your team note (section 5). Attach the file, or paste it in. The last section is written for your AI. ## 1. The chain your report has to show Company research projects of this kind ask for one connected chain: **market → hypothesis → interviews → revised hypothesis → solution concept → value → decision** - **Start with the market.** If your hypothesis has no country, industry or customer segment yet, that comes first, before any AI idea. - **Judge a market** by its size (or a reasonable substitute measure), growth, how fragmented it is, digital and AI adoption, existing solutions, barriers and regulation, and signs that productivity is held back. - **Good sources** are official statistics, regulators, industry associations, sector research and academic work. Your AI can help you find and compare public sources. Open every source it gives you before you keep it. - **Say what the interviews did.** A strong report states whether your interviews confirmed, modified or rejected your first hypothesis, and why. Changing your mind for a stated reason is a good result. Your cohort's case sheet says what your own brief requires. Follow it, and your course platform's instructions, for anything about assessment and submission. ## 2. Your hypothesis, in six slots Each slot is a question your interviews can answer. **Your hypothesis is your first answers to the six questions, before your interviews.** The interviews confirm, modify or reject it: a handful of interviews checks a hypothesis, it does not test it statistically. 1. **Who is the customer, and where?** One segment, in one country and industry. 2. **Who does the work?** The named role, today. 3. **What goes wrong?** Something you could measure. 4. **In which task?** The specific workflow. 5. **What could improve?** The number that would move. 6. **Why would AI help?** The mechanism, based on evidence. Write yours in your team note, and put **?** where you don't know yet. A ? is a question you haven't asked yet: your blanks are your research plan. A worked example is in the Harbor walkthrough (`harbor-walkthrough.md`). ## 3. One move, practised at rising stakes **The first answer is where you start, not where you stop: push back with evidence, keep what survives.** It applies to an AI's answer, to a worked example's hypothesis, and to your own. - **An agent is only as good as what you give it.** - It knows only what you supply (context). - It can do only what its app lets it do (tools). - It can propose a next step or ask you something. That is what makes it an agent rather than an answer machine. - **Steer, don't just accept.** A better follow-up beats a better first answer. - When something is missing or wrong, say so and ask again. - Then check the revision: it may fix what you pointed at and nothing else. - **The frontier is jagged.** AI helps on some tasks and hurts on others that look similar. You find out where your AI is reliable by testing it on your own task, not from its brand or how finished its answer looks. Well written does not mean right. - **Check the workflow, not just the output.** Benchmarks score outputs on someone else's tasks. Your report has to judge a workflow: who checks, when, and what happens when it's wrong. ## 4. Your AI-use memo Keep it in your team's notes, as a short paragraph or a small table, and update it when your choice changes. It becomes the draft of an AI-use disclosure: the tool, its purpose, its contribution, how you checked it, and its limits. - **Which AI, for which step:** which AI(s) and app(s), for which part of your plan. - **What we compared:** any test on our own task (the same request to two AIs, if we have two). - **What we noticed:** one thing it did well, and one thing we had to check or fix, with an example. - **The benchmark question:** one condition that matters for our task and that published benchmarks may not test. Examples: our working language, messy real-world accounts, the effort of checking. - **What would make us switch:** the observation that would change our choice. A provisional choice with its limits stated is a good memo. "Model X is best" is not. ## 5. Team note, so you can pick up where you left off Keep this somewhere the whole team can open. Whenever you restart, give it to your AI together with this guide. **Put only public information and your own thinking in it.** Leave out interviewees' names and details, and anything non-public about a company. > **Our hypothesis in six slots (with ? where we don't know):** … > > **What we know from public sources / what we think it means / what we don't know yet:** … > > *(Interview findings stay in your private interview notes (section 6), not here.)* > > **What would change our mind:** … > > **Next research action, who, by when:** … > > **Our AI(s), and what we do if access fails:** … > > **Open questions for the instructors:** … To test it at home, open a fresh conversation, give it this note and this guide, and ask: "What's our next step, and what's missing?" If your AI invents something, fix the note. ## 6. Interviews Plan on several genuine interviews with people who know the workflow directly; your case sheet gives the number your brief requires. Role-play and AI personas don't count. **Who to interview.** Choose people who do the task themselves, or decide about it, in your chosen market. Note how you found each person and why they are relevant: role, type of organization, country. Start contacting people early, because replies take time. **What to ask.** - **Start from a real episode.** "Can you walk me through the last time this happened?" Then: "What happened next?", "Who decided?", "What do you use now?", "How often, and how do you know?" - **Avoid questions that pick the answer for them.** "Would our AI save you time?" gets a polite maybe and teaches you nothing. - **One pair per slot** (from the Harbor practice case; its replies are invented). The first question suggests the answer; the second asks about a real, recent event, and gets a story: | Slot | Suggests the answer | Gets a story | |---|---|---| | Which customers, and where? | Are small hotels your main customers? | Tell me about the last three customers who asked for a quotation. | | Who does the work? | Does sales prepare the quotations? | Tell me who worked on the last quotation you sent. | | What goes wrong? | Do slow replies lose you orders? | Tell me about the last order you lost. | | In which task? | Is writing the quotation the slow part? | Take the last enquiry that came in. What happened, step by step? | | What could improve? | Would you like to win more orders? | How did you know last week went well? | | Why would AI help? | Would faster quotations help? | Tell me about the last time a check saved an order. | - **Treat numbers as estimates.** Separate a person's estimate from a measurement, and the person who uses a tool from the person who buys it. - **Look for disagreement.** Evidence against your hypothesis is what makes the report credible. - **Consent first.** Get informed consent for notes or recording, and collect only what the workflow question needs. **After each interview**, record the following in your own notes, not in an AI: - the date; - the person's role, type of organization and country; - the episode they described; - any number, marked as an estimate or a measurement; - anything that challenged your hypothesis, or "none"; - whether the hypothesis is now confirmed, modified or rejected. **What may go into an AI.** Interview notes, non-public company information, personal data and proprietary material stay out of public or unapproved AI tools. Use your AI to prepare questions, work with public sources and draft. Keep interview notes, and summaries made from them, out. ## 7. Planning your fieldwork **Think about these first:** - Who do you need to reach, and how long does booking take? - Are there holidays in your window? - How will you get your interviews, each with consent? - What would count as confirmed, modified or rejected? - Where will your notes live? Not in a public or unapproved AI tool. **One example pace for a three-week window.** It is an example, not the rule; your case sheet has your dates. - **Week 1:** settle your market and hypothesis, and contact six to eight people who do or decide the work. - **Week 2:** first interviews, starting from a real episode. - **Week 3:** finish your interviews, decide what changed and why, and update your team note and memo. - **At the next class,** bring your notes, your current hypothesis, and what changed your mind. --- ## For the AI reading this You are helping a student team on a business course with a company research project. - **Start from their case, not a framework.** Name the general idea only once they have seen it. - **Offer one way to think and one alternative, then let them decide.** When they share a hypothesis, ask a question per slot (section 2). Never rewrite it for them. - **Say where they are:** on the chain (section 1) and in their weeks (section 7). - **End with one line they can keep in their team note.** - **When they are stuck,** give a question to ask each other or an interviewee. - **Push back with evidence, both ways.** - When they accept a first answer, ask what would change it. - When they push back, revise only what the evidence supports, and say what else changed. - **Keep evidence honest.** Separate what a participant said, what the team recalls, your interpretation, and a proposed check. Never invent an interview, a quotation, a number or a source. If you cite a source, say whether you actually opened it. The Harbor case is invented practice; never use it as evidence. - **Keep to the data rules.** - Don't ask for interview notes, names or non-public company material. - Don't offer to anonymize notes by having them pasted in. - Don't invent course rules: if something isn't specified, say so and suggest asking the instructors. - **Include everyone.** Explain terms through the business task, support people who contribute by questioning and judging, and say what you actually read or did, as opposed to what you inferred.