Responsible AI for board members

This is the one that matters most, and it is on by default on consumer plans. Until you switch it off, your chats and the files in them can be used to train and improve the models.

1Open Settings from your name at the bottom left, then choose Privacy in the sidebar. 2Help improve our AI models — this is the switch. Grey and to the left means off. This is what you came here for. 3Location metadata — nothing to do with board papers, but there's no reason to leave it on either.

This is the trap nobody warns you about. Turning training off governs the automatic use of your data. Clicking thumbs-up or thumbs-down is you voluntarily submitting that conversation for human review — and it goes through regardless of what your privacy settings say.

!Do not click the thumbs. A thumbs-up or thumbs-down submits the whole conversation — including any board material in it — for review, even with training switched off.

Memory can be a useful feature, but it opens two risks. Irrelevant information from old conversations gets pulled into new ones and confuses the model. And when you work across several companies, what you did for one leaks into the work you do for another. Turn it off.

1Memory lives at the bottom of the settings sidebar, under the Customize heading — easy to scroll past. 2Search and reference chats — off. Stops it reaching into your other conversations for context. 3Generate memory from chats — off. Stops it writing durable notes about you and your companies.

You lose a little convenience: each conversation starts cold. Step 05 gives that context back in a controlled way, one company at a time.

Cowork tasks can run either on your own computer or in Anthropic's cloud. With this switched off, new tasks run locally, and the material largely stays on your own disk — your files stay in your folders, and only what the model actually needs to see leaves the machine.

1Cowork in the settings sidebar — a separate section from Privacy. 2Cowork files shows where your artifacts and scheduled tasks are kept on this machine. Worth knowing before you go looking for them. 3Run new tasks in the cloud — off keeps new tasks on this computer. This is the optional one.

A project is a walled-off workspace: its own instructions, its own files, its own chats. One project per company you sit on, each pointed at a folder that belongs to that company alone. This is what replaces memory — context that is deliberate and scoped instead of ambient and leaky.

1Name it after the company. You'll be switching between these under time pressure — make it unmistakable. 2One line on what the company does and what your role on the board is. You can expand this later. 3Use a folder — connect the project to a folder on your own machine, so material stays in one place instead of being uploaded chat by chat.

Point it at that company's folder — and only that folder

The folder you connect is the boundary. If a project can see a parent folder holding three companies' material, you have rebuilt the problem you just switched memory off to avoid.

1Recent folders appear below — blurred in this screenshot. Pick the company folder, not the parent that contains all of them. 2Choose a different folder opens the normal file picker if the right folder isn't listed.

Write the standing instructions

The instructions are read at the start of every chat in that project. This is where the company's strategy, your focus as a director, and the documents that matter most belong. Keep it short enough that you'll actually maintain it.

1Plain sentences work fine. State your role, the company's strategic direction, and which file holds the strategy.

I sit on the board of Acme A/S, a B2B logistics software company,
roughly 40 employees, majority owned by a family foundation.

My focus as a director is operational excellence and digitalisation.
I am not the finance specialist on this board.

The current strategy is in strategy-2026.pdf. Refer to it when you
assess whether something is on or off plan.

When you analyse board material: be direct about what looks wrong,
flag what is missing, and don't soften bad numbers.

The routine, once the setup above is done: board material for the September meeting goes into a subfolder of that company's folder. Open the project. Ask it to read that folder. Everything else is a variation on one prompt.

1The prompt goes here. Name the meeting subfolder explicitly so it reads the right pack. 2Instructions — your standing context, applied to every chat in this project. 3Memory confirms it is off. Worth a glance before you start on anything confidential. 4Context — the one folder this project can see. Check it says what you expect.

The starting prompt

Adapt it as you learn what you actually want from it — this is a starting point, not a formula.

Read all materials for the upcoming board meeting in folder <name>.

Identify:
  - Concerning changes in key metrics, spending, or other
    strategically important areas
  - Inconsistencies or conflicting priorities across the materials
  - Five core questions to ask about the strategy and ongoing projects

Cite the document and page for every point you make.

That last line matters more than it looks. It forces every claim back to a source you can check in thirty seconds, and it makes a fabricated number obvious rather than plausible.

Follow-ups worth having ready

• "What is management not telling us here?" — surfaces gaps between the narrative and the numbers.
• "Compare this pack with the previous quarter's. What changed in the story?" — requires both packs in the folder; catches quietly revised forecasts.
• "Argue the opposite case for the main proposal." — a fast way to test whether a recommendation survives contact with a sceptic.

The project instructions are only as good as the last time you touched them. The cheapest moment to update them is right after a meeting, when the pack and the minutes are both sitting in the folder and the discussion is still fresh. Ask for a summary written specifically to be pasted into the instructions — not a meeting summary, which is a different thing and far too long.

Read the board pack and the minutes from the meeting in folder <name>.

Write a condensed status note that I can paste directly into my project
instructions. Cover:
  - Key projects: current status, owner, and what was decided
  - Decisions taken, with date and any conditions attached
  - Open questions carried into the next meeting
  - Numbers or targets I should hold the next pack against

Rules:
  - Maximum 250 words, plain statements, no preamble
  - Standing context for future meetings, not a record of this one —
    write in present tense
  - Leave out anything that will be stale in three months
  - Output only the text to paste, nothing else

Then paste it in — over the old note, not under it

Open the project, click the pencil next to Instructions, and replace last meeting's note with the new one. Keep the permanent part — who you are, what the company does, where the strategy lives — and swap only the status section. Instructions that only ever grow eventually stop being read, by you and by the model.

1The pencil opens the instructions for editing. Paste the status note here after each meeting.

If you'd rather work in ChatGPT Plus, the shape of the setup is the same. The menu names differ and OpenAI moves them around more often, so treat this as a map rather than a click-path — and verify each one on your own account.

• Training is offSettings → Privacy. Re-check after app updates.
• Memory is offThe project panel states this — glance at it rather than trusting recall.
• I'm in the right projectThe company name is in the header. Wrong project is the most likely mistake you'll make.
• The connected folder holds this company onlyCheck the Context panel, not your memory of how you set it up.
• Nothing here belongs to a different boardEspecially where two of your companies are in the same market.
• I will verify every number before I say it out loudCited sources make this fast. Skipping it is how you get embarrassed in a meeting.

Used this way, an AI assistant is a fast, tireless first reader. It will catch the line in the appendix that contradicts the summary, the metric that quietly changed definition, and the project that has slipped twice without anyone naming it. That is real value on a 90-page pack you got four days before the meeting.

What it doesn't have is everything that isn't in the documents: the CEO's track record on optimistic forecasts, the tension between two shareholders, why the CFO went quiet last time this came up. It also does not know what it doesn't know — an assistant will produce five confident questions whether or not the pack gave it enough to be confident about.

So treat the output as a well-briefed junior's memo: useful, quick, worth reading, and never something you'd repeat in a board meeting without checking. The questions it hands you are candidates. Which ones you actually ask, and what you do with the answers, is the part of the job that doesn't delegate.

Everything above assumes it is acceptable for your material to be processed in the US. For some boards it isn't: a customer contract forbids it, the company handles personal or health data, or the pack simply contains something you won't send across the Atlantic on a consumer subscription.

That doesn't mean giving up on the workflow. It means changing where the model runs. Two routes are realistic today.