practical guide
How do you prepare data for an AI pilot?
Prepare a bounded set of sources, record permissions and decide what AI may receive. The register shows which sources are prepared and what still needs approval before entering an AI tool.
You will have a list of permitted sources, prepared test input and a named person who approves the use of data.
Free, with no sign-up. This example uses fictional data.
Prepare a bounded exercise
Use fictional data for the first pass. In this example, AI prepares a draft reply to a website enquiry. It does not send messages or access the entire inbox. Before using real data, the responsible person must confirm the source, purpose, tool, account, access and data retention. If anything is unknown, keep the source out of the pilot. This process does not replace your organisation’s internal rules.
A source register template
Copy the template for each source. An unknown item blocks its use.
Task and permitted output: Source ID, location and version: Person responsible for the source: Required fields and reason for using them: Excluded fields and preparation method: Check for remaining sensitive information: Approved tool, account and people with access: Storage of inputs, outputs and processing logs: Retention period, removal method and responsible person: Approver, date and approval record: Status: permitted / blocked Reason for blocking and next step:
How the workflow works in practice
Illustrative examples with fictional data. They show the process and human review, not the results of a client project.
Approved sources for a draft reply
- Input
This entire example is fictional. The description of prohibited input below is used to check the boundary and is not entered into an AI tool. TASK Prepare a draft reply to an enquiry about a website with a form. PROHIBITED INPUT FOR THIS PILOT The entire inbox, including customer signatures, contracts and internal notes. Reason: this content is not needed and has not been approved for this pilot. MINIMUM SOURCE REGISTER S1 | Fictional enquiry U-01, version 1 Responsible person: sales manager. Content: We need a website with a form. What is the price and delivery date? Approval: permitted for this fictional exercise. S2 | Fictional service description, version 1 Responsible person: service owner. Content: Price and delivery date depend on the confirmed scope. Approval: permitted for this fictional exercise. S3 | The organisation’s existing inbox Responsible person: not confirmed. Approval: unknown. The source is blocked. EXERCISE BOUNDARIES Tool: a local text document, without sending anything to an AI system. Access: the sales manager and the person preparing the exercise. Retention: the sales manager deletes the working copy after review. Actual use of organisational data: not approved.
- Output
Expected result of manual preparation. AI did not execute this example. PREPARED TEST INPUT [S1] We need a website with a form. What is the price and delivery date? [S2] Price and delivery date depend on the confirmed scope. DECISION S1 and S2 are ready for a fictional exercise in a local document. S3 is not included. Permission is unknown. The tool, account and retention must be confirmed before an actual AI pilot. Sending a reply to the customer is not permitted.
- Process
- Record one task and the required fields. A draft reply needs the request description and approved service description. The customer’s contact details are not needed to compose the text.
- List sources and responsible people. For each source, record the version and written approval for the specific use. Permission to read does not grant approval to enter data into an AI tool.
- Prepare a working copy. Remove unnecessary names, contact details, signatures, hidden comments and attachments. Check whether the remaining text identifies a person or confidential work. Automated removal of information needs human review.
- Define processing boundaries. Record the permitted tool and account, who sees input and output, the storage location and retention period. Check the selected tool’s settings and rules. Do not assume deleting the local copy also deletes data in the tool.
- Give the register and prepared sample to the person authorised to approve them. Record the decision before entering data. If approval or retention confirmation is missing, continue only with the fictional example.
- Human review
- Compare the working copy with the register. Include only permitted fields.
- Check hidden information and context that could reveal an identity.
- Confirm that approval covers the selected tool and purpose.
- Record who reviews the output and who confirms removal of copies.
- What we measure
- Track the number of blocked sources, reasons for blocking and time needed to prepare the approved set. During review, record every unnecessary piece of information that remained in the copy. These records show what still needs attention.
Check before using
- Compare the working copy with the register. Include only permitted fields.
- Check hidden information and context that could reveal an identity.
- Confirm that approval covers the selected tool and purpose.
- Record who reviews the output and who confirms removal of copies.
- The example shows expected behaviour. Save your own results before deciding on actual use.
Apply it to your process.
Explore our AI agents service to prepare the sources and boundaries for your pilot.