Making Good: four questions your counsel will ask

Making Good, part 3 of 4. Part one starts here.

The fastest way to stall a frontline AI program is to skip the meeting with legal. The second fastest is to show up without answers.

Making good on a brand promise includes keeping it inside the rules. Across Life Sciences, Hospitality, Retail and Intelligent Industry, the questions in that meeting are predictable, and they're the right ones. What does the agent know? Who can it answer? What does it show as evidence? Who is accountable when it acts? A deployment that can't answer all four is a pilot waiting for its shutdown memo.

What does the agent know?

Only what the organization has published and recorded. That's grounding: answers come from the customer's own documents and systems, not the open web. If the answer isn't in the record, the honest response is that the agent can't answer it here. An improvised answer on the production floor is worse than none.

Grounding is an architecture choice more than a prompt. The agent has to sit on a record that knows who works here, what the organization has said and what happened on the last shift. Shelbe sits on the RedeApp network for exactly that reason.

Who can it answer, and with what?

Each person gets answers within their own scope. That's permission gating, and the scope comes from the same identity data that governs every message in the network. A seasonal associate and a regional director can ask the same question and get different answers, because they're entitled to see different things.

In a regulated operation this is the line between an AI program and a reportable incident. An answer the person isn't allowed to see is a liability, however accurate it is.

What does it show as evidence?

Its source, every time. Citation turns trust from a feeling into something a supervisor or an auditor can check: which policy, which document, which date.

It also cleans up the content behind it. Unclear or outdated material shows up in the first week, because people can see exactly where a confusing answer came from.

Who is accountable when it acts?

A person. Where an answer would trigger action, such as a schedule change, an escalation or anything touching safety or payroll, someone with authority approves it first. That's human-in-the-loop, and the point is to put judgment where judgment belongs while the agent handles the rest of the paperwork.

What are the honest limits?

Three, and they belong in the same meeting.

Shelbe is only as good as the documents behind it, so a deployment starts with a content audit. AI does nothing to repair a broken relationship between a manager and a team. And none of it works unless the workforce opens the app, which is why RedeApp puts the communication layer first.

We have a short line for this: usage is the security model. Unused AI delivers nothing. AI on a surface the workforce opens every shift delivers value and leaves an audit trail of every answer and every approved action.

Where does Capgemini fit?

Governance doesn't stop at the platform boundary. It extends into how a rollout is tested, how integrations with HR and identity systems are checked, and how the program is evidenced to security and operations leadership. Capgemini runs that kind of enterprise program every day, and the alliance puts frontline AI inside it.

Security and compliance documentation is on the RedeApp security page. If your counsel has questions, bring them. These four make a good agenda.

Next Wednesday, the last part: the math.