What $1T of diligence will ask your AI.

A founder assembling an investor data room has a practical choice: describe how the company uses AI, or make one piece of that work inspectable. I would start with a synthetic customer enquiry and follow it all the way to a proposed reply.

The Canada Investment Summit page describes a government plan to catalyse $1 trillion in total investment over five years. That is a target, not a sum committed at the summit or a measure of AI spending. The diligence connection in this essay is our recommendation for companies preparing for scrutiny. These are not announced investor requirements.

Start with the task behind the answer.

A polished reply tells a reviewer little about the authority behind it. Someone needs to explain which records were accessible, whether an external model received them, and what prevented a draft from becoming a send. A policy can describe those intentions. The demonstration needs records from the particular run.

Use invented customer details. Agree the expected behaviour before testing, including what should happen when the human rejects the proposed action. Keep the review separate from production customer work.

Twelve checks for the data room.

• Choose the decision. Name the business action under review. Producing a draft and sending it are different permissions; record which one the test covers.
• Identify the worker. Record the account used and its permission scope. Establish who granted that scope and where a reviewer can inspect it.
• Bound the source. List the records the worker may read. Include one harmless out-of-scope record so the reviewer can check that the boundary holds.
• Trace model access. Name the configured provider and endpoint. Establish what content leaves the tenant, if anything, rather than infer this from the hosting address.
• Account for retained copies. Ask about prompts, application logs and backups. Give each category an owner and a retention rule; identify anything still unknown.
• Preserve the proposal. Retain the exact action presented for approval. A screenshot of a green status cannot establish which recipient or record was approved.
• Name the approver. Establish the identity and authority of the person reviewing the action. Confirm that the decision applies to that proposal.
• Reject an action. Observe the result when permission is withheld. Do not accept a reassuring chat response as evidence that the source system stayed unchanged.
• Check the source separately. Inspect the relevant mailbox or application history. Record what the check can establish and where visibility is incomplete.
• Exercise an unavailable dependency. In an isolated test, make one required read unavailable. Check whether the workflow reports an unknown instead of recycling an earlier answer.
• Bind the configuration. Keep the evidence date, model route and policy version with the result. Decide which changes require a fresh run.
• Hand over the record. Give another authorised reviewer the evidence without the original chat. Note what they cannot reconstruct and assign each gap an owner.

Keep the verdict smaller than the evidence.

A successful rehearsal supports the tested workflow under its recorded conditions. It does not establish every connector or settle a legal question. Missing source visibility belongs in the findings, even when the interface reports success.

This is the review I would bring to a Vantage Workspace evaluation. The illustration accompanying this essay is a preparation method, not proof that a deployment has passed these checks. The buyer and technical team still need to agree the run and inspect its results.

Questions buyers ask.

What will investors ask about a company’s use of AI during diligence?

There are no announced investor requirements. Our recommendation is to make one workflow inspectable. A reviewer will want to know which records the AI could read, whether an external model received them, who approved the proposed action and what stopped a draft from being sent.

Is the $1 trillion figure from the Canada Investment Summit a committed amount?

No. The summit page describes a government plan to catalyse $1 trillion in total investment over five years. It is a target, and it does not measure AI spending.

What evidence belongs in a data room for an AI workflow?

Records from one particular run. Include the account and permission scope used, the sources read, the model route, the exact proposal shown for approval, the approver, the result of a rejected action and a separate check of the source system. Keep the evidence date and the configuration with the result.

For the wider review method, read the CISO evaluation guide for AI workspaces.

Take one workflow from your data room and use the Agentic AI Procurement Handbook to write its acceptance conditions before the demonstration. The next reviewer should be able to repeat the check.

Josh Olayemi · Founder, Handvantage · September 2026