# How do you manage AI agents inside a client's tools?

> How we run AI agents inside a client's own tools. What they may touch, what waits for a yes, where the record is kept, and who answers for a mistake.

- Canonical page: https://warmup.ai/blog/how-do-you-manage-ai-agents-inside-a-clients-tools
- By WarmUp. Published 2026-11-15.
- Category: How we work (https://warmup.ai/blog/category/how-we-work)
- Part of: RevOps (https://warmup.ai/revops)
- Tags: ai agents, approvals, crm, governance

## Key takeaways

- An agent works only through access the client grants and can revoke, in tools the client owns.
- Anything a buyer will read, and any rule that changes a record, waits for a yes before it runs.
- Every outcome is written where the client can read it, and nothing is deleted without a record of what and why.
- When an agent gets something wrong, a person at the firm answers for it and owns the fix.

We manage AI agents inside a client's tools with four rules. An agent works only through access the client grants and can revoke. Anything a buyer will read waits for a yes. Everything the agent does is recorded where the client can read it. And when an agent is wrong, a person at our firm answers for it.

This is the longer version of the section on [how we run agents](https://warmup.ai/forward-deployed-growth-engineering#agents-heading) on our method page. Buyers ask about trust before they ask about anything else, so this post takes the questions in the order we hear them.

## What is an AI agent allowed to touch?

An AI agent is allowed to touch only what the client grants. It runs in the client's own tools, under logins the client can revoke, and it starts with the least access the job needs. For a CRM clean-up that means read-only first: the agent and the engineer can look at every record and change none.

Two things follow from working this way.

- **The tools are in the client's name.** The CRM, the data tool, the ad accounts and the sending domains belong to the client from the first day. We work in them. If the work ends, nothing has to move.
- **Access is a decision the client makes, and can unmake.** Withdrawing a login stops the agent. There is no copy of the data somewhere else that keeps running.

We set this out for larger buyers on our [Enterprise page](https://warmup.ai/enterprise), which lists the access we ask for in each tool. The same rules hold when we build agents for a job outside growth, which is the work on our [AI engineering](https://warmup.ai/ai-engineering) page.

## What does an agent do on its own, and what waits for a yes?

An agent does on its own the work that is safe to repeat and easy to check. It waits for a yes on anything a buyer will read and on any rule that changes a record. The line is drawn by what a mistake would cost: a wrong research note costs little, and a wrong email to a buyer costs a relationship.

In our work the split looks like this:

| Work | Who does it | What waits for a yes |
| --- | --- | --- |
| Researching companies, enriching and checking contacts | Agents | Nothing. A contact that fails a check never reaches a sequence. |
| Outbound copy | Agents draft | Every line, approved by the client before anything sends |
| Replies from buyers | Agents sort them and draft an answer | A person presses send on every answer |
| Ads | Agents draft angles and variants | Every ad, approved before it runs |
| CRM clean-up | Agents do the bulk work | Every merge rule, written down and approved before it runs |
| Phone calls | A person, always | The call script, approved before a call is made |

A person at our firm also reads a sample of every batch before it launches. The client's yes is the approval. The sample is how we make sure that what reaches the client for approval is worth their time.

## Where is the record of what an agent did?

The record of what an agent did is kept in the client's own tools, not in ours. Outcomes are written to the client's CRM as they happen, with the signal and the play attached. Fields, rules and reports are documented as they are built, with a change log, so the client can read what changed and when.

This matters for a plain reason. If the only record is in the firm's software, the client has to take the firm's word for what happened. If the record is in the client's CRM, they can check it themselves, and it is still there after the work ends.

For clean-up work there is one more rule: nothing is deleted without a record of what and why. A merge is planned in writing first, approved, and then run. We describe that work on the [RevOps](https://warmup.ai/revops) page.

## What happens when an agent gets something wrong?

When an agent gets something wrong, a person at our firm answers for it. We tell the client what happened, trace it to the rule or the instruction that caused it, correct it, and change the rule. An agent is a tool we chose to run, so its mistakes are ours.

The approvals are there so that most mistakes are caught before they matter. A draft that is wrong gets rejected and costs nothing. That is the reason drafting and sending are separate steps, and the reason a person, not an agent, presses send on a reply to a buyer.

The record is there for the mistakes that get through. Because every change to a record follows a written rule, a wrong result can be traced to the rule that produced it. We would sooner tell a client "this rule was wrong, here is what it touched" than "we are not sure what happened".

## Why not let the agents run without approvals?

We do not let agents run without approvals because the cost of a wrong message is not paid by the agent. It is paid by the client, in their own name, with their own buyers. An approval takes minutes. A message that should not have gone out can close a door for a year.

There is a view that agents will soon be reliable enough to act alone everywhere. We think that will be true for some jobs and not for others, and that the sensible way to find out is one job at a time. Work that is safe to repeat and easy to check is where an agent can run on its own. Where a buyer is on the other end, a person stays in the loop.

We also want an agent that can say it does not know. A rule we hold to in outbound is that nothing is drafted for a contact whose signal has not been verified. No message is a better result than a confident message built on a guess.

If you are setting this up yourself, write down three lists before you give an agent access to anything: what it may change on its own, what it must ask about, and where it writes down what it did. Then give it the least access that the first list needs. You can widen it later. It is much harder to narrow it after something has gone wrong.

## Questions, answered

### Do your agents act on their own, or only draft?

Both, depending on the job. Agents act on their own for work that is safe to repeat and easy to check, such as researching a company or checking that an email address is valid. They draft, and a person approves, for anything a buyer will read and for any rule that changes records in a CRM.

### What access do you need to our systems?

Access to the tools the work runs in, under logins you grant and can revoke. For a CRM clean-up we start read-only, so nothing can change while we map what is there. The tools and the data stay in your name. We work in them. We do not own them.

### Who is responsible if an agent makes a mistake?

We are. An agent is a tool we chose to run, so what it does is ours to answer for. That means we tell you what happened, trace it to the rule or instruction that caused it, correct it, and change the rule so it does not happen the same way again.

### Can we see what the agents did?

Yes. Outcomes are written to your own CRM as they happen, with the signal and the play attached. Fields, rules and reports are documented as they are built, with a change log. You can read all of it without asking us, and it stays with you if the work ends.

### Does more context make an agent more reliable?

Not on its own, in our view. More context makes an agent sound more sure of itself. It does not make it right more often. What makes the work reliable is a person checking a sample, an approval before anything goes out, and a rule that the agent drafts nothing for a contact whose signal has not been verified.
