A support team answers the same kind of question many times a day. Why has this lead not come in? What happened to this order code? What was the last step taken on this file? A person looks it up. The answer sits in a system the team already owns.
That is a good job to hand to an AI team member. It is also a job where the first worry is not speed. It is where the data goes.
Why a company does this
A company that holds customer data wants three things before it lets an AI near it:
- The data stays inside the company's own cloud.
- The AI cannot do more than the job needs.
- A person can see where each answer came from.
We built a proof of concept for a customer's lead-operations support team. It runs on a development environment. It is not a production rollout. We describe what we built and why. We publish no measurements.
What we built
- Four AI team members on one machine in the company's Google Cloud. They answer briefly and politely, in the team's language, and say where each figure comes from. Each one says it is an AI.
- Reachable from Google Chat. The link runs through Google's own message queue. The machine has a private network and no public address.
- Models through Google's Vertex AI. The machine uses its own Google identity. No API key is stored anywhere.
- Read-only access to the data. The AI team members hold no database password and have no command line. They ask a small service that runs one fixed lookup, or one summary query. It is built to return only operational fields, such as stage, status and dates.
How we work
Our team works with yours.
- We start with a short feasibility look: what Google Chat allows, and how to connect it safely.
- We set up the machine in your cloud, labeled so you can remove it with one filter.
- We review security at each step: read-only access, grants limited to single tables, and hostile questions in the test set.
- We write a test set in your language and run it against your data. A person checks each answer against a direct query.
- We hand you a written test report.
We test on one job first. We build more only if the test is worth it to you.
What the proof of concept does not cover
- A person using the Google Chat app in the company's own workspace. That test is still to do.
- People from another company's Google domain. We left that out on purpose for now.
- The reason a lead has not come in. The data shows the stage, not always the cause. The AI team members say so and do not guess.
A checklist you can reuse
- Private network, no public address.
- Model access through the machine's own identity, with no stored keys.
- No database credentials for the AI team members.
- Read-only, fixed queries that return only the fields the job needs.
- A test set in your language that includes questions built to make the AI misbehave.
- A person checks every answer against the source.
If you run support or operations and want to look at one such job in your company, write to sales@3metas.com. An AI team member reads mail sent to that address.