Short answer: The risks are what the agent is allowed to see, where that data goes, who approves anything it sends, and whether a mistake can be undone. Name the four before you connect a customer record. Fluency is not a control.
I'm Vinay Pasricha. I run GoodSpace AI in Noida, and I wrote AI for Business Leaders. Customer data is the fuel people want to pour into an agent, because the answers get better when the agent "knows the customer". Better answers are also how a private fact leaves the company in a sentence nobody reviewed.
What goes wrong
The agent repeats a fact to the wrong person, because the session, the identity or the prompt was muddy. It sends a customer's detail to a model or a vendor you did not mean to include. It stores a conversation in a place your own team cannot search or delete. It drafts a reply that mixes two customers. None of these require a dramatic breach. They require an agent that can see more than the task needs, and a team that did not look.
A second failure is quieter. People paste exports into a chat tool to "just get the summary", outside any agent you approved. The risk is then not the project you rated. It is the path around it. A policy that names the approved tool, and says what must never be pasted, is part of the control. The Risk Ladder is how you decide the tier. The policy is how the tier survives contact with a busy afternoon.
There is also the training question. You should know whether a customer's words are used to improve a model you do not control. If you cannot get a clear answer, assume you do not want that customer's data in the tool, and keep the data on your side of the line. Do not fill the gap with a guess and call it a review.
Four controls, before connection
- See: the fields the task needs, and nothing else. A balance, a ticket, an order status. Not the whole record because it was easier to export.
- Go: the named systems the data may enter. If a vendor cannot list them, do not connect.
- Approve: a person sends anything a customer will read, until the tier and the evidence say otherwise.
- Undo: you can find what was sent, delete what you stored, and tell the customer what happened. If you cannot find it, you cannot undo it.
Rate the use. A draft that never leaves the building is a lower tier than an agent that replies by itself. The controls should change with the tier. A company that uses one control for every agent has not rated anything. It has bought a default.
Read the vendor's answer on retention in plain language. How long are prompts kept. Who at the vendor can open them. What you do when a customer asks you to delete them. If the answers are a link to a policy nobody in your company has read, the control is not in place yet. Have someone read it, and write the four lines in your own words.
What a small company can do now
You do not need a large programme to start. You need the four lines on a page the people who touch customer data have actually seen. Keep customer data out of any tool that is not on that page. Review one agent a month, not every tool in the market.
If a project needs the whole customer history to do a small job, narrow the job. The risk lives in the width of the data, not only in the cleverness of the agent. A narrow job with a narrow view is the one to run first.
Use the free Risk Ladder before the agent is allowed to act, and Automate, Assist, Escalate to sort the tasks around it. The eight steps are in AI strategy for business. Collapse or Collapse is the longer argument: one person owns one result and runs agents to deliver it.
One thing to do this week
For the agent you are considering, list the customer fields it would see. Strike every field the task does not need. If you cannot say where the remaining fields go, stop the connection until you can.
Where this comes from
This answer draws on The Risk Ladder in AI for Business Leaders.
Your company will have its own details. To ask a follow-up in your own words, try Ask Vinay's books.
Vinay Pasricha is the founder of GoodSpace AI and the author of six books, including AI for Business Leaders.
Go deeper in AI for Business Leaders
This answer is the short version. The full argument, and the method behind it, is in the book.
