A “quick scan” is the most common kind of research I do at work: two or three weeks to answer one specific management question, such as whether something is worth investing in, where to enter, and how much to commit. The output is usually a few pages with clear conclusions, not an exhaustive report.
Over the past few years I have used AI more and more for quick scans, and made plenty of mistakes along the way. Below I walk through one fictional example in five steps, and spell out what AI can do and what a person must do at each one.
The fictional question: an equipment maker wants to know whether battery-swap cabinets for e-bikes in second-tier Chinese cities are worth investing in.
Step 1: Turn the question into decisions
“What does the battery-swap market look like?” is a research question. It cannot guide a decision on its own. I first rewrite it as seven decision questions:
- Who are the customers?
- Why do they buy?
- How does the money flow?
- How big is it?
- Who makes the purchase decision?
- How would we enter?
- How much should we commit?
What AI is good at: expanding an issue tree around these seven questions and filling in dimensions you had not thought of. Being exhaustive is its strength.
What a person must do: from the expanded tree, pick the key issues that most affect the decision, keep them to six or fewer, and rank them. The seven questions are a frame, and they do not all deserve equal effort; only two or three usually decide the answer.
Step 2: Write the storyline before collecting data
A lot of research never ends because people collect data first and look for the conclusion later: the pile of material grows and the story gets fuzzier. I do the opposite and write the title of every page first.
Each title has to be a conclusion (even if it is only a hypothesis for now), and reading the titles alone should give you a chain of cause and effect. For example:
- Swap demand is concentrated among food and parcel delivery riders, who pay per swap and hate waiting;
- The market is small overall but growing fast in cities dense with riders;
- Purchase decisions sit with swap operators, not riders;
- The equipment maker’s opportunity is to lower each operator’s cost per cabinet;
- We recommend a small pilot through existing channels rather than building a new sales team.
That is a storyline, as opposed to a “research line” that runs through market overview, competition and policy and leaves the reader with no idea what to do.
What AI is good at: drafting two or three candidate storylines, then attacking each one: where does the causality break, and which assumption is weakest?
What a person must do: pick one, and decide which assumptions most need testing next.
Step 3: Definitions matter more than the data
What goes wrong most often in a quick scan is not a lack of data but inconsistent definitions. I hold myself to five rules:
- State the denominator for every number. Per cabinet, per station or per rider? A different denominator can flip the conclusion.
- Separate new builds, the installed base and retrofits. Net growth of the installed base is not the same as new demand.
- Cross-check top-down against bottom-up. If the two estimates are far apart, at least one assumption is wrong.
- Label facts, interviews, estimates and assumptions separately. Readers deserve to know how solid each number is.
- If you cannot find a number, write XXX. Never invent a “reasonable” figure to make a page look complete.
What AI is good at: searching and organizing sources, extracting data and setting up the calculation. But I always ask it for sources and check the key numbers by hand. The most dangerous thing about AI is how confidently it produces a number that looks plausible.
What a person must do: judge whether sources are credible, choose the definition to use, and, when data is thin, accept a reasonable approximation while writing down what it rests on.
Step 4: Use interviews to test hypotheses, not to gather information
Interview time is expensive. My interview guides rarely have more than five questions. They build on each other, can be asked out loud as written, and each one maps to a hypothesis in the storyline.
What AI is good at: drafting the guide, turning recordings into notes, and finding common ground and disagreements across many sets of notes.
What a person must do: follow up. The most valuable information usually shows up after the second “why”.
Step 5: Write something people will actually finish
The last step is cutting:
- Titles are conclusions, quantified where possible;
- Each page makes one judgment, and its content has to support the title;
- Anything that does not support the conclusion goes, however interesting it is;
- The same number does not appear again and again on different pages.
What AI is good at: tightening the text, checking that each title matches its page, and flagging repetition and contradictions.
What a person must do: set the tone. AI prose drifts easily into empty consulting-speak that sounds like it says everything and actually says nothing.
Three things AI cannot do
After all this time, three things still have to be done by a person: deciding what matters, judging whether a source is credible in context, and taking responsibility for the conclusion. In my experience AI shortens a quick scan from about three weeks to about two, mostly by saving time on searching, organizing and redrafting. But it cannot make the judgment for me.
How the time breaks down
This is the rhythm I usually follow, two to three weeks end to end. The time is not spread evenly: defining the question and writing the storyline come first, and the last few days are kept for cutting.
| When | What | Output |
|---|---|---|
| Days 1–2 | Rewrite the decision questions, draft the first page titles | A one-page problem statement, two or three candidate storylines |
| Rest of week 1 | Search sources, align definitions, set up the calculations | Top-down and bottom-up estimates |
| Week 2 | Test the key assumptions in interviews, revise the storyline | Interview notes, a final storyline |
| Last 2–3 days | Cut it down, check against the list below | A few pages with clear conclusions |
A reusable checklist
- Can the question be stated in one sentence?
- Reading only the page titles, does the story make sense?
- Does every number have a denominator and a source?
- Are missing figures marked XXX?
- Does the recommended investment match the size of the market you estimated?