Checklist: before an AI agent touches company data

An AI agent is software that does more than answer questions. It takes steps on its own: it reads documents, searches your systems, and drafts emails or orders. Treat it less like an app and more like a new employee.

Print this checklist and go through it with your finance lead, your IT lead (or external IT provider) and the person who will actually use the agent. Next to each item, write the name of the responsible person and a date. While any item is blank, the agent does not get access to your data.

1. Data

• We know what data the agent will work with. Written down specifically, for example "quotes and customer enquiries from the last two years", not "company data".
• We know where that data lives. Which system, which folder, which shared drive.
• We know what data the agent must not see. Typically salaries, employee personal data, contracts with confidentiality clauses, and customer data you have no right to use for this purpose.
• Every data set has an owner. The name of the person who decides whether the agent may read it.
• The data is sorted into at least current and outdated. An agent cannot tell that an old version is no longer valid.
• We know where personal data sits. If the agent will process it, this is recorded in your GDPR record of processing activities (or the equivalent under your local data protection law).
• We know where the data goes. Whether it stays in-house or is processed by a third-party service, and under what contractual terms.

2. Access

• The agent has its own account. It does not use a particular employee's login, so you can always see what the agent did and what a person did.
• Permissions are the minimum possible. The agent has only the access its assigned task needs. By default it can only read, and it can write only where that is necessary.
• Access is limited in scope and in time. We know which systems it may use, and everything else is closed.
• Passwords and keys are not in the instructions or in emails. They are stored in a password manager.
• Everything the agent does is logged. Every read, change and send has a timestamped record, and we are able to read those records.
• We can switch the agent off. We know who can cut its access immediately, and how, including outside working hours.
• We have a separate test environment. For the first weeks the agent runs on a copy or a limited portion of the data, not on the live system.

3. Rules

• There is a one-page policy on using AI. It says what employees may and may not give the agent.
• It is defined what the agent may do on its own. For example sorting incoming mail or summarizing a document.
• It is defined what the agent may only prepare. For example a draft quote or a reply to a customer, which a person then sends.
• It is defined what the agent must never do. For example pay invoices, delete data, change prices, or communicate with authorities.
• Employees know that responsibility stays with them. They check the agent's output before it reaches a customer or the accounts.
• We know how to report a mistake. It is clear to whom and how an employee reports that the agent got something wrong.
• The vendor contract has been reviewed. We know whether the vendor uses our data to train its models and how long it retains it.

4. Approval

• An owner of the agent is named. One person who is accountable for its operation and results.
• The managing director or CEO approves the deployment. Signature and date recorded on this sheet.
• Access to data is approved by the data owner. Not by IT, but by whoever the data belongs to.
• Risky steps are approved by a person before they are carried out. The list of such steps is complete, and the agent cannot proceed past them without approval.
• A deputy is named. When the agent's owner is on holiday or ill, someone else knows what to do.
• Changes are approved the same way as the deployment. New data, a new system or new permissions each require a new approval.

5. Measuring the result

• We know what we want to improve. One concrete thing, such as the time it takes to prepare a quote or the number of returned invoices.
• We have a baseline. Before launch we measured how long the task takes today and how many errors occur.
• We know what a good result looks like. The target is written down in advance, so it cannot be "adjusted" after three months.
• Errors are counted. Someone checks a sample of outputs and records how many were wrong.
• We also measure the extra work. How much time goes into checking and correcting the agent's output.
• A review date is set. For example after eight weeks. The agent's owner brings the results and the CEO decides: continue, adjust, or stop.

Sign-off sheet

Deployment approved by: ______________________ Date: ______________

AI agent use cases for companies by industryData security when you deploy AI in your company

Prepared by AI. Adapt it to your company before use. Not legal advice.