Direct answer
Before a team relies on an AI-built internal tool, it has to run on company-owned hosting that stays up when the builder’s laptop is closed, sign people in through the company identity provider so offboarding removes access, use its own narrowly scoped credentials against real data, and protect against destructive actions with confirmations and tested backups. It also needs an audit trail, a named owner, and a release process so updates don’t break someone’s work in the middle of the day.
Why the demo works and the team version doesn’t
Claude Code, Cursor, Lovable and Replit let one person build something useful in an afternoon. The finance export or bulk-update screen works because it runs as that person: on their machine or preview URL, with their database login, their CRM key and their judgment about which button not to press.
Once ten people use it, the tool acts for people who don’t know how it works, against data the builder had admin rights to, and nobody has written down who fixes it when it stops.
The four areas that decide the launch
Our launch checklist covers ten areas. Four are critical: any one left unverified holds the launch.
Hosting and runtime
The tool runs on company-owned infrastructure over HTTPS, and the deploy can be reproduced from the repository. A tunnel, a sleeping free-tier preview or a hosting bill on a personal card fails this. Production gets its own environment values. In one of the projects we reviewed, the admin panel and its API returned server errors because the database connection string and the signing secret were still placeholders; the fix was a launch script that refuses to start if any required value is missing or still a placeholder.
Access and identity
Login goes through Google Workspace, Microsoft Entra, Okta or an access layer, with no shared passwords or “anyone with the link”. The real test is offboarding. Microsoft’s documentation on Entra provisioning explains that when a user is unassigned or disabled, the service disables them in the connected app, which for a SCIM app means setting the user to inactive. A tool with its own user table keeps every leaver’s account until someone deletes it.
Data access and credentials
The tool gets its own service account or API key, never the builder’s personal token or admin login. NIST defines least privilege as restricting users, “or processes acting on behalf of users”, to the minimum access their tasks need. A read-only tool gets a read-only database user. Secrets stay out of the repository and the front-end bundle. We covered the cost side of this in treating AI API keys like payment methods.
Destructive actions and data safety
Deletes, bulk edits and customer messages ask for confirmation and show what will change. Retries don’t run an action twice (see idempotency and retries). Backups exist for the data the tool writes, and someone has restored one at least once.
The other six areas
The remaining areas rarely block a launch alone, but three or more unverified ones do. Roles are enforced on the server. Every write records who changed what, when, and the old and new values. Errors reach a person, and long imports run in the background. The tool has a named owner and a backup, and its repository, hosting and vendor accounts belong to the company. Someone other than the builder has reviewed access and data handling. A user guide and runbook exist, and the spreadsheet it replaces has a retirement date.
Shipping updates without breaking people’s work
Updates happen every week, and each can break a workflow mid-morning: a renamed column breaks a finance export, a migration locks a table, or a new screen shows data a role shouldn’t see.
Our release checklist sizes the review to the change: a copy change can pass with most areas not applicable. Three areas are critical. Data and schema changes must be backward compatible with the running version, tried against a recent copy of production, and preceded by a restore point. Permissions on new screens, endpoints and exports are checked on the server. Rollback has to work with the migrated schema, and someone has done one on this tool before.
Deploys should run from a pipeline, built the same way every time. In the production issues we reviewed across about 25 projects in October, deploy and operations gaps appeared in five. In one, a risk platform’s production container could not run migrations because the migration tool was a development-only dependency: fine on laptops, broken on the server.
Ready, Conditional release or Hold
Both checklists use the verdicts from our Production Readiness Standard. Ready means every area is verified or genuinely not applicable. Conditional release means no critical area is open: release to a pilot group, behind a flag or outside busy hours, with each open item given an owner and a date. Hold means a critical area is unverified, or three or more areas are.
“We use SSO” counts once there’s a screenshot of the identity provider assignment, and “we have backups” counts once a restore has been done.
Related Zenveus services: Production Readiness Standard, the 48-Hour Verdict and Build It Right