Who signs off, and what each asks
At a small company one person may hold two of these roles. Answer each question anyway.
| Who | What they ask | What to bring |
|---|---|---|
| Founder or CEO | Which job gets better, and is it worth the team's attention now? | The recurring job, who does it today, how long it takes and what has gone wrong |
| Finance | What does it cost, including usage, and what does it cost to stop? | The published price or a quote, likely AI usage costs, the pilot's cost and how you would leave |
| IT or security, often the person who connects the tools | Who can see what, who approves changes, what is recorded, where the data and keys live, can we get it all out? | The vendor's answers to the seven risk questions |
| The people who will use it | Is this another tool to learn? | Which of their current AI tools it works in, and what changes in their week |
Pick one recurring job
A good pilot job:
- happens every week or more often;
- involves more than one person or agent;
- depends on company records that change, such as prices, policies or customer commitments;
- shows its mistakes, so you can count them.
Examples: answering customer questions about pricing and policy; preparing a weekly customer update; following up after sales calls; briefing a new hire's AI tools on how the company works.
Keep sensitive records, such as pay or board papers, out of the pilot unless the job needs them and you have checked who can read them.
Run a two-week pilot
- Before day 1: record the baseline from the last two weeks of the job: time per run, wrong or outdated answers, and how often someone had to ask where something was.
- Days 1 to 2: put in only the records the job needs. Set who can read them and who approves changes. Connect the AI tools of the two to five people in the pilot.
- Days 3 to 10: run the job through the shared memory. Log each correction and each change, and whether it was approved.
- Days 11 to 12: read the record of what the agents read and changed, and try exporting everything.
- Days 13 to 14: write up the result against the measures and ask for the decision.
The measures to report
| Measure | How to count it |
|---|---|
| Time per run of the job | Minutes per run, baseline against pilot |
| Wrong or outdated answers | Count, and how many came from a record in the memory |
| Changes and approvals | Proposed, approved and rejected, and the approver's minutes |
| Access | Any attempt to read something a person may not see, and whether it was refused |
| Export | Whether everything, with its history, came out in a readable form |
| Cost | Spend during the pilot, including AI usage, and the estimate at full use |
| The team's view | Whether the people in the pilot want to keep it, and why |
One-page approval memo
Copy this into a document and fill in the brackets. Printing this page prints only the memo.
To: [founder or CEO, finance, IT or security] · From: [your name] · Date: [date] · Decision needed by: [date]
1. Problem. [The recurring job, who does it, how often and what it costs today in time and mistakes. One real example of something that went wrong.]
2. Proposal. [Pilot (product) for two weeks on (job) with (names). What the pilot includes, and what it leaves out.]
3. Risks and controls.
| Risk | Control in this pilot |
|---|---|
| A model saves memory with no person approving it | [which changes need approval, and who approves] |
| No record of who read or changed what | [what the product records, and who reviews it] |
| Someone sees what they should not | [how access is set for each person; sensitive records kept out] |
| We cannot get our data out | [export format, tested on day 11 or 12] |
| The vendor holds our content and keys | [where content and keys live, and who can reach them] |
| Cost grows with usage | [price, usage limits and the cap we set] |
4. Cost. [Pilot cost. Price status: published, quoted or not yet published. Likely AI usage cost. Estimate at full use.]
5. Pilot. [Dates. People. The job. The records included. The approver. The measures and today's baseline. What result means continue.]
6. Decision asked. Approve a two-week pilot of [product] on [job] from [date] to [date], with [name] approving changes, and a decision meeting on [date] to continue, change or stop.
Sign-off. Founder or CEO: [name, date, yes / yes with changes / no] · Finance: [name, date, yes / yes with changes / no] · IT or security: [name, date, yes / yes with changes / no]
If the product is Mensara
Mensara facts for the memo, from mensara.ai, checked 18 November 2026:
- Approvals: people and their agents propose updates. Some permitted changes save after an automatic check. Changes to governed areas, such as decisions, wait for the owner's approval, and you can require approval elsewhere too. Deleting or renaming a file always waits for an approver.
- Record: a signed, tamper-evident receipt for each read and change, including refusals.
- Access: each person's agents work within that person's access, checked by software before any model sees a file.
- Export: saved records and their version history are ordinary files in a Git repository that your team can copy and keep without Mensara.
- Keys: keys and credentials live on your Mind's server, not in the mensara.ai account system. If Mensara operates the server, that separation is not a promise that the operator cannot access it.
- Cost: founding plan with a free trial at checkout; a card is required. Public pricing is not yet published; the price, billing period, tax and terms are shown at checkout. Model usage in the Mind is a different measure from account charges.
- Works with: Claude, Codex, Cursor and ChatGPT on the web (developer mode), over MCP.
- Pilot shape: Mensara starts with one recurring job and checks the result against your own records.
Questions
Who needs to approve shared AI memory at a small company?
Usually the founder or CEO, whoever controls the budget, and whoever is responsible for IT and security, often the person who connects the tools. At a small company these may be one or two people. Bring an answer to each of their questions anyway.
How long should a pilot be?
Two weeks on one recurring job is enough to compare time and mistakes against the job's last two weeks, check the record of activity and test an export, without asking the team for a long commitment.
What should the pilot measure?
Time per run, wrong or outdated answers, changes proposed and approved, any refused attempt to read something, whether the export worked, cost including AI usage, and whether the team wants to keep it.
Do we need a security review for a two-week pilot?
Keep the pilot to the records the job needs, and ask the vendor the seven risk questions before you start. Keep sensitive records out until you have checked who can read them.
Sources
- Mensara: About, Changes, approvals and receipts, Backup, restore and ownership, Ownership and data flow, Usage, budgets and costs, Account and billing
Related
Practical guides for people and agents