Clear questions. Before access is shared.
Security and data handling belong in the scope conversation. These are proposed review topics, not certifications, contractual guarantees, or a statement that every control is already implemented.
Start with the minimum necessary information
An initial outline should not contain credentials, customer records, regulated data, or confidential source documents. If sample data is later needed, agree an appropriate sharing method and use redacted or synthetic examples where possible.
Define systems and permissions
Identify the source and destination systems, who authorizes access, and which read or write actions are needed. Discuss least-privilege access, separation between clients or teams, credential handling, and how access will be revoked after the engagement.
Follow the data through the whole path
Document which information reaches a model provider, hosting service, integration, log, or backup. Provider terms, processing locations, retention, deletion, and any training use must be checked for the actual configuration.
Put a boundary around actions
Agree which operations may be proposed, which may run by rule, and which require human approval. Consider duplicate prevention, auditability, exceptions, permission failures, and recovery before connecting important business actions.
Agree operation and incident responsibility
A launch checklist should name who monitors failures, receives alerts, investigates incidents, maintains dependencies, and approves changes. Support coverage, response expectations, retention, and ownership require a written agreement.
Access follows the scope
Agree the systems, people, permissions, and records a workflow needs. Check source access, retention, and the route for investigating a problem before connecting business data.
Before we begin.
Is our data used to train AI models?
That depends on the provider, the plan, and the configuration. Business and API terms from major providers generally exclude training on customer data by default, but the actual setting must be checked and recorded for each service in the data path. We document it in the scope.
What does an AI agent get access to?
Only the sources and actions the scope names, with the least privilege that lets the task run. Reading a folder, proposing a record, and writing to a system are separate permissions, and write actions that matter wait for a person.
Do you need our production data to start?
No. A first conversation uses a general outline. Later, redacted or synthetic examples are enough to design tests. Access to live records comes with an agreed sharing method, a named owner, and a plan for revoking access when the work ends.
Who is responsible when something goes wrong after launch?
The launch checklist names who monitors failures, receives alerts, investigates incidents, and approves changes. Support coverage and response expectations are part of the written agreement, not an assumption.