AI makes the orientation fast. Whether you can safely repeat any of it depends on a check you write down.
AI makes the first stretch of getting up to speed fast and makes the last stretch, checking what’s true and who says so, easy to skip. Write a brief that names the decision, map the landscape with a prompt that holds back numbers, triage sources into primary, secondary, opinion and unknown, run a claim-check log, and write a one-pager of known versus believed. Keep confidential material out of consumer tools.
AI changes the first afternoon on an unfamiliar topic more than any other stretch of the work. The vocabulary, the main camps, the questions people argue about and the names that keep coming up: a model can lay that out in minutes, where it used to take days of reading around. What it changes much less is the part that decides whether you can safely repeat any of it, which is whether a claim is true and who says so.
That gap is where the trouble starts. The orientation arrives fast, fluent and well organised, so it feels finished before anything has been checked. A marketing lead asked to weigh in on a sector she hasn’t worked in finishes a good-looking briefing by mid-afternoon, and the figure she quotes in tomorrow’s meeting is one she didn’t trace back to where it came from. The briefing gave no sign that anything was off, which is exactly why it travels.
The speed is real and worth keeping. It only counts if the verification step stays in, so this guide is built around that: a short brief that fixes the question and the decision first, a map of the landscape, a triage of sources by who is speaking, a log that checks every claim you plan to repeat, and a one-page write-up that separates what you know from what you only believe. The workflow is the same whichever tool you use. For tool-specific walkthroughs, see our guides to using ChatGPT for research, comparing deep research tools and Gemini notebooks and NotebookLM.
Getting up to speed with no decision attached turns into browsing, and AI makes browsing very comfortable. A model will explain a whole field to you without being asked which part you needed. So the first step happens before you open the tool: write down the question, the decision it feeds, and what would change your mind.
Write down the decision the work feeds before you ask the first question.
The row people tend to leave blank is what would change your mind, and it’s the one that stops you collecting only what agrees with you. Here is a brief filled in for a marketing lead looking at a new sector.
| Field | Example |
|---|---|
| Question | What do logistics firms buy software like ours for, and who signs off? |
| Decision it feeds | Whether to run a pilot campaign aimed at the logistics sector next quarter, or hold. |
| Who decides, and when | The head of growth, at the next planning meeting. |
| What I already believe | Buyers are operations leads, and integration matters more than price. This is a hunch from sales conversations, not yet checked. |
| What would change my mind | Evidence that purchasing sits with procurement, or that budgets in the sector are frozen. |
| What counts as enough | I can name the buyer, the questions they ask, and a primary source for every claim I’ll repeat. |
| What I can’t paste | Customer lists, pipeline numbers, anything under NDA. |
Illustration only. Replace each cell with your own, and write it before you open the tool.
Once it’s written, paste the brief into the model and ask what’s missing. It’s good at noticing that a question is really three questions. The decision stays yours.
With the brief done, ask for a map instead of answers. You want vocabulary, the main positions, the points people dispute and the kinds of sources to look for. Treat all of it as a list of things to go and find, not as evidence. A model’s map comes from patterns in what it was trained on and, if the tool can search, whatever it retrieved, so it describes the consensus view well and the edges badly. A topic with two live camps often comes back described as one.
| I’m getting up to speed on [TOPIC] for this decision: [DECISION FROM THE BRIEF]. I’m a [ROLE] and I already know [WHAT YOU KNOW]. Give me a map, not conclusions. (1) The ten terms I’ll meet in the first hour, one line each. (2) The main positions or schools of thought, and what each would say about my decision. (3) Which questions people in this field genuinely dispute, and which are settled. (4) For each position, a serious objection to it. (5) The kinds of organisations that would publish the original data on this topic. Mark anything you’re unsure of with [UNSURE]. Do not give me statistics, quotations or citations in this answer. I will find those myself. |
Asking it to hold back numbers and quotes keeps the map from handing you facts you would later have to unpick.
Read the answer for two things: the disputed list, and everything marked unsure. Those are where your reading time pays off. My rule: let AI draft the map and explain terms, and keep it away from the numbers until you have a source in front of you.
A consultant starting a new engagement tends to reach for the client’s own board deck first, because it is a rich map of the business. It is also the kind of document engagement terms tend to keep out of a personal account. So before you paste anything into a tool, decide what stays out. The useful question is what your plan does with your input, which differs between personal and work accounts. As checked on October 19, 2026: OpenAI says that on its individual ChatGPT plans “we may use your content to train our models,” unless you turn off its data control, while inputs to its Business, Enterprise, Edu and API products aren’t used to improve models by default.[1] Anthropic’s page for its consumer plans says chats are used to improve Claude if you choose to allow it, and points to a separate page for its work products, which I did not read.[2] Google’s Gemini page says human reviewers review some data, and asks users not to enter confidential information they wouldn’t want a reviewer to see.[3]
Those pages change, so read the one for your plan, and read your employer’s policy first, because it outranks any vendor page.
| Bucket | Examples | What to do |
|---|---|---|
| Public material | Published articles and papers, public filings, your own general questions | Paste into whichever tool your organisation allows |
| Ask first | Internal drafts, meeting notes, unpublished plans, supplier terms | Use a work plan your employer has approved, or ask whoever owns data policy |
| Keep out of consumer tools | Personal data about customers or staff, client-confidential material, anything under NDA, credentials, unreleased financials | Don’t paste it. Describe the shape of the problem in general terms, or use a tool your organisation has contracted for the purpose |
| Licensed or paywalled | Subscription articles, analyst notes, standards | Check the licence terms before uploading, or work from your own notes on it |
A Future Factors working guide. Your organisation’s policy and the current terms of your plan come first.
Our guide to using AI without leaking company data goes through the settings tool by tool.
The map gives you things to find. Finding them is your job, with the model as a search assistant. Ask it where the original data on a topic would live: a regulator, a statistics office, a standards body, a company’s own filing, the paper itself. Then go there. The primary source is the place where whoever has first-hand knowledge first produced the figure or the statement. Everything else is somebody’s description of it, and a figure repeated across many articles often traces back to a single original.
| Type | What it is | How far to lean on it |
|---|---|---|
| Primary | The original data, filing, regulation, paper, transcript, or a company’s statement about its own product | Cite it for what it says. Note who produced it and when |
| Secondary | Someone summarising or analysing primary sources: news coverage, analyst notes, review papers | Use it to find the primary source. Don’t repeat its number until you’ve seen where it came from |
| Opinion | A view, forecast or recommendation, including vendor blogs and listicles | Use it for vocabulary and hypotheses. Label it as a view |
| Unknown | You can’t tell who wrote it, when, or on what basis. AI summaries with no link belong here | Not usable until traced. Park it in the log |
A vendor is a primary source for its own product and an opinion source for its market. Sort by what the claim is about.
The real saving is in reading selectively. Read the parts that carry the claim. For a paper, that means the abstract, the method and the limitations. For a regulation, the definitions and the section you’ll rely on. For a long report, the table or chart behind the headline figure and the note on how it was measured. Use the model for the rest: ask where a topic is covered or what a term means, then read that passage yourself.
When a source looks authoritative, check it sideways. In a Stanford study, professional fact checkers quickly scanned a site and then opened other tabs to see what others said about who was behind it, while historians and students who stayed on the page were often taken in by professional-looking cues.[4] The move transfers directly: leave the page and look at who stands behind it before you lean on it.
This is the step that gets skipped, and the reason is understandable. By now you have a coherent picture, and the claims in it feel settled. Fluency isn’t evidence, though. A model gives you a clean paragraph when it is right and a clean paragraph when it is wrong, and the two look the same. A well-known public example is a 2023 US court decision in which the judge found that lawyers had submitted “non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT.”[5] A citation is a claim about a document, and the only way to check it is to open the document.
So make the check a task with its own record. Start by having the model list what needs checking in your draft briefing, since it is quick at spotting the sentences you might repeat to someone else.
| Below is my draft briefing on [TOPIC]. List every claim I might repeat to someone else: figures, dates, named people or organisations, quotations, statements that one thing causes another, and statements about what a law, policy or product does. Put them in a table with these columns: (1) the claim, quoted exactly as written; (2) the type of source that would be primary for it; (3) whether it came from text I pasted or from your general knowledge, as best you can tell, which I will treat as a hint only; (4) how costly it would be if it were wrong. Do not verify anything and do not supply sources. I will check them myself. [PASTE THE BRIEFING] |
The model finds the claims. It does not get to vouch for them.
Then you do the checking, claim by claim, against the primary source. The test differs by type of claim:
Our guide to fact-checking ChatGPT goes deeper on the individual techniques. The log below is the record you keep while doing them.
| # | Claim as I’d repeat it | Came from | Primary source found? | Status |
|---|---|---|---|---|
| 1 | Operations leads hold the buying decision | AI map | Yes: an industry body’s buyer survey tables | Confirmed, but the survey covers mid-size firms only. Say so |
| 2 | Firms in the sector plan to adopt this next year | A news article | No: the article cites a “recent survey” with no link | Believed, not known. Reword as unconfirmed or cut |
| 3 | A named executive said the sector is “at a turning point” | AI summary | Yes: the transcript, with different wording | Corrected. Paraphrase without quote marks |
| 4 | A new rule requires firms to file a quarterly disclosure | AI map | Yes: the regulator’s own page, read today | Confirmed. Date noted |
Illustrative rows, not real claims. Four statuses are enough: confirmed, confirmed with a narrower scope, believed, dropped.
Row 2 is the common case: a claim that sounds specific, comes from a secondary source and points at a primary one you can’t find. You have three honest options:
Don’t attribute it to a named person or organisation on the strength of a model’s say-so, and don’t ask the model to supply the source after the fact, since a plausible-looking one is quick to produce and no easier to trust.
If you can’t open where a claim came from, you can’t present it as known.
The product of all this is a page someone else can act on, not a folder of notes. It has one job: to draw a firm line between what you’ve established and what you’re carrying on trust. Four columns, one page, written from the log.
| What I know (source opened) | What I believe (not confirmed) | What I don’t know yet | What would change my mind |
|---|---|---|---|
| Operations leads are the usual buyers, per the industry survey (mid-size firms only). The quarterly disclosure rule exists, per the regulator’s page, read today | Firms plan to adopt next year. The source is a news article citing an unlinked survey | Whether purchasing moves to procurement above a certain firm size | Buyer-side evidence that procurement signs off. Next step: two calls with people in the sector |
Illustration only. One page, and no line goes in the first column without a source you’ve opened.
Every line in a briefing is either known, with a source you have opened, or labelled as believed.
Once the log is filled in, the model can draft the page from it, which suits the job because every input has already been checked.
| Using only the claim-check log below, draft a one-page briefing for [AUDIENCE] on [TOPIC], for this decision: [DECISION]. Use four headings: What I know, What I believe, What I don’t know yet, What would change my mind. Put a claim under What I know only if its status is confirmed, and keep any narrower scope in the wording. Put everything marked believed under What I believe, worded as unconfirmed. Do not add facts, figures or sources that are not in the log. [PASTE THE LOG] |
The log is the only input, so nothing unchecked can slip into the page.
A manager who gets a briefing with a clear “believed” column knows where to push. One who gets a smooth page with no such column has to trust all of it or recheck all of it.
Here is the whole split written out, so it’s clear what the tool does and what stays with you.
| What AI does | What the human still owns | How it gets checked |
|---|---|---|
| Sharpens the brief, drafts the landscape map, explains terms | The question, the decision, and the call on what counts as enough | You compare the map against at least two sources you found yourself |
| Lists the claims in your draft and points to where in a document a topic is covered | Which sources are credible, and what goes in the “know” column | Every claim in “know” has a source you opened, noted in the log |
| Drafts the one-pager from your log | What the page says to the reader, and what it asks them to do | A colleague opens the primary sources behind the two costliest claims before the page circulates |
A Future Factors working view.
Counting topics summarised measures the wrong thing. I’d look at use, persistence and impact, in that order.
| Question | What to look at |
|---|---|
| Use | Is a brief written and a log started each time you get up to speed on something, including the small topics? |
| Persistence | A couple of months on, are people still keeping the log and the “believed” column without being reminded? |
| Impact | Did a decision change, or an error get caught before it reached a meeting, because a claim failed its check? |
Impact is the one that justifies the time.
The emphasis shifts by seat, so it helps to say where each person is likely to slip.
| Role | Focus | Where to be careful |
|---|---|---|
| Analyst | Runs the full loop and owns the log | Letting a clean log stand in for judgment about which sources deserve trust |
| Marketing or strategy lead | Writes the brief, owns the decision and the one-pager | Taking a market figure from a vendor blog because it supports the plan |
| Consultant | Gets up to speed quickly in a client’s field, under confidentiality terms | Pasting client material into a consumer tool to save time |
A Future Factors working view.
A person can start on their own with this. Training with an experienced facilitator shortens the learning curve and builds a shared standard, so that everyone on a team labels “known” and “believed” the same way. In our terms, training builds the capability, the brief and the log are the workflow design that gives it somewhere to be used, and a manager who asks to see the log is the reinforcement that turns it into a habit.
Some situations call for leaving AI out of the evidence step.
| Situation | Do instead |
|---|---|
| The topic moved this week and the tool can’t search, or you can’t confirm that it did | Start from current primary sources and use the model only for structure |
| You need to know what the work is actually like | Talk to someone who does it, and put the conversation in the log |
| The decision carries legal, medical or financial weight | Put a qualified person first and treat the AI map as background |
| The material is confidential and you have no approved tool | Work by hand, or describe the problem in general terms |
A Future Factors working guide.
Start with one topic you need to be across this week, and give it an afternoon instead of the full system.
Five steps for one topic. Steps 4 and 5 are the ones that make the speed safe.
If the afternoon runs short, shorten the reading before you shorten the check. Then hand the one-pager to a colleague with three questions to ask.
| Ask | A good answer sounds like |
|---|---|
| Pick a line in the “know” column. Where did it come from? | A named primary source, with the date it was read |
| Which line here are you least sure of? | One from the “believe” column, said without defensiveness |
| What would make you change the recommendation? | The evidence named in the brief’s “what would change my mind” row |
Ask these before the page circulates.
For a wider look at catching errors, our anti-hallucination toolkit is the next stop.
Go back to the “what counts as enough” row in your brief. You’re there when every line in the “know” column has a source you’ve opened and you can say what would change your mind. If you can’t, the problem is usually the question, not a lack of reading.
Treat it as a lead. Search for the original through other routes, and if you can’t find it, move it to the “believe” column with wording that says it is unconfirmed, or cut it. Don’t attribute it to a person or an organisation, and don’t ask the model to produce the source after the fact.
Ask it where the original data would live, then go and find the documents yourself. That keeps the model on the map and puts your attention on what the source says. If a tool does return links, open each one, because a link can be real and still not say what the summary claims.
Check first. Licensed material comes with terms on how it may be used, and client documents usually carry confidentiality duties, so look at the licence, your contract and your employer’s policy before you upload. When in doubt, work from your own notes or an approved tool.
The steps stay the same, and the log matters more, because a long, formatted report with many citations looks more finished than a chat answer. Treat it as a secondary source: use it to find primary ones, then open those behind the claims you plan to repeat.
Vendor data-use pages were read on October 19, 2026: OpenAI’s help article on how your data is used to improve model performance, Anthropic’s privacy centre article on whether your data is used for model training (dated March 16, 2026) and Google’s Gemini Apps Privacy Hub (updated October 13, 2026). These pages change often, so read the current one for your plan. The lateral-reading point comes from a Stanford news summary of the study, and the court quotation from the June 22, 2023 opinion in Mata v. Avianca. I did not use any figure for how much time AI saves on research or how often models produce wrong citations, because I could not confirm one against a primary source that stays current. The brief, the triage table, the claim-check log and the one-pager are Future Factors working methods, not research-validated models. The scenes and example rows are illustrations, not client stories.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.