There's a moment in every automation project where the workflow runs end-to-end for the first time and everyone wants to ship it. This is precisely the moment to slow down — because "it works" answers only one of the two questions that matter. The other is: "what can it leak, delete, or expose if something goes wrong?"

The gap between functional and safe

Most automation failures we're asked to fix weren't broken. They were working exactly as built — with the customer database readable by an agent that only needed the calendar, API keys pasted into workflow steps where any collaborator could read them, and no boundary on what the system could send to whom.

Agentic systems raise the stakes because agents act. A misconfigured chatbot says something wrong; a misconfigured agent does something wrong — emails the wrong customer list, overwrites live records, forwards internal data to an external tool. Security review is not paranoia. It's the price of giving software hands.

The four-part review

Every integration Kaizen AI ships goes through the same review, and it's worth demanding from anyone who builds for you:

  1. Data exposure audit. Map every piece of data the system can see. For each: does this stage actually need it? Agents should read the minimum — an appointment agent needs the calendar, not the full patient file.
  2. Permission scoping. Each agent gets its own credentials with the narrowest possible rights. Read-only wherever possible. Write access only to the specific records it manages. Never admin.
  3. Key management. API keys and passwords live in a secrets manager, rotated on a schedule, never in prompts, workflow definitions, or shared documents. If your builder can't tell you where the keys live, that's the answer.
  4. Escalation boundaries. Define the irreversible actions — sending money, deleting records, contacting your full customer list — and require human approval for each. The boundary is a design decision, made up front.

Questions to ask any vendor

Before signing with anyone building AI systems for your business, ask: What data can this system access, and why each piece? Where are the credentials stored? What happens when the agent hits something it doesn't understand? What's logged, and could you reconstruct what the system did last Tuesday?

A good builder answers immediately and specifically. Vague answers about "enterprise-grade security" are how functional-but-unsafe systems get shipped.

Security isn't a feature we add at the end. It's a review gate every integration passes through before it earns access to your business.