An agency's AI service is only as good as its audit log
Your AI service holds up with clients when every AI action is logged, outside sends wait for approval, and one switch stops every agent.
Key takeaways
5 things that decide this
- 01An agency's AI service holds up with clients when every AI action is logged with who or what acted, outside messages wait for approval, and one switch can stop every agent.
- 02Sooner or later a client asks what the AI did on their account and who approved it. You need to answer from a record, not from memory.
- 03Log the actor as well as the action: which person, which token, which agent.
- 04Hold anything that leaves the building, such as a reply to a lead, in an approval queue by default. It slows volume, and that is the point.
- 05We built these controls into SalesCrew, our own CRM, which our sales team uses every day.
The question every client will ask
If you sell AI work to clients, one day a client will call about something it did. A reply went to the wrong lead. A deal moved stage and nobody on their side remembers moving it. A message went out in a tone they would never use. They will ask two things: what exactly happened, and who said it could.
If your answer is "let me check with the team," you have already lost some trust. If your answer is a log line with a timestamp, an actor and the approval that let it through, you have kept it. That gap is the whole difference between an AI service a client renews and one they quietly switch off.
Most agency AI setups can't give the second answer. A shared API key runs a few automations, a chat assistant has access to a client's inbox, and the logs, where they exist, record that something happened but not who or what did it. That works in a demo. It doesn't survive the first client complaint.
Log who acted, as well as what changed
A useful audit log answers three questions for every write: what changed, when, and who or what changed it. The third one is where most setups fall short. "API" is not an actor. "Automation" is not an actor. You need the person, the token they issued, or the agent that acted for them.
In SalesCrew, the CRM we built and run, each action a rep takes on screen is also an MCP tool with the same effect and the same permission checks. MCP (Model Context Protocol) is the standard AI assistants use to call tools. Because both paths run through the same checks, both land in the same log. As the case study says, "the audit log records who acted, down to the token or agent."
Scope matters as much as logging. An MCP token in SalesCrew carries the same three scopes as your login: channel, profile and area. Your assistant can reach what you can reach, and no more. For an agency, that means an assistant working one client's account can't wander into another's, and the log shows whose access it used.
Approval before anything leaves the building
Inside your own systems, a wrong AI action is a mistake you can fix. Outside them, it is a message a prospect has already read. So the line to draw is simple. Anything that leaves the building waits for a person.
In SalesCrew, a reply drafted by the inbox agent waits in an approval queue by default until someone approves or rejects it. The agent does the drafting work. A person decides whether it goes. The approval itself becomes part of the record, so when a client asks who signed off, you can show them.
Here is the trade-off. Approvals slow outbound volume. A queue of drafts waiting for a human will never send as fast as an agent that sends on its own. That's by design. For an agency, the cost of one bad message on a client's behalf is far higher than the cost of a few hours of waiting.
You can loosen it later, one message type at a time, once the log shows the agent gets that type right. Start strict and earn your way to less review. We wrote more about that habit in agents earn trust one tool call at a time.
One kill switch
Every agency running agents for clients needs a way to stop all of them at once. Not one at a time, not by revoking keys in five different dashboards while the problem keeps going. One control, one click, every agent paused.
SalesCrew has exactly that: one kill switch stops every agent. When something looks wrong, your first move is to stop the damage, and your second is to read the log to find out what happened. Without the switch, you're reading the log while the agent is still writing to it.
The manual path matters here too. Because each MCP tool is a twin of an on-screen action, your team can keep working by hand while the agents are off. A kill switch that also stops your people from doing their jobs is a switch nobody will want to pull.
Built and used daily by our own sales team
Build in-house or with a partner
Can your agency build this in-house? Yes, if you have engineers who can design permission checks, scoped tokens and an audit trail, and keep them running. The hard part isn't the first version. It's keeping the AI path and the manual path in step every time someone adds a feature. Each new button needs its tool twin, with the same effect and the same checks, or the log stops telling the truth.
If your team builds client work rather than platforms, a partner can build the control layer while you own the service. Our comparison of an agency versus an in-house AI team walks through that choice. Whichever way you go, ask one question of the result: can it answer what the AI did, and who approved it, for any client, on any day?
Also check your cover. An audit log helps you explain an AI error, but it doesn't decide who pays for one. See whether your E&O policy covers AI agent errors before you sell the service.
If you want a second opinion on your setup, start with a 45-minute call. If there is a build to do, an engineer writes a fixed-fee Blueprint from your real systems, credited in full against the build. See how we work for the steps after that.
When a client asks what the AI did, the answer should be a log line with an actor and an approval, not a promise to check with the team.
What buyers ask
01Is custom software better than building AI in-house for a digital agency?+
Custom is better when your team can't keep permission checks, scoped tokens and an audit log in step as features change. If you build client work rather than platforms, a partner can build that control layer while your agency owns the service and the client relationship.
02What should an agency log when AI acts for a client?+
Every write: what changed, when, and who or what changed it, down to the person, token or agent. Log approvals too, so you can show who signed off on anything that left your systems.
03Does an agency's E&O policy cover AI errors?+
Not always: coverage turns on the policy wording, so read yours and ask your broker before you sell AI services. An audit log helps you show what happened, but it doesn't decide coverage.
Related
- SalesCrew: a CRM your own AI assistant can work →The audit log, approval queue and kill switch in production.
- Agency vs in-house AI team →Who should build the control layer.
- Agents earn trust one tool call at a time →Start strict, then loosen review as the log earns it.
- Does an agency's E&O policy cover AI agent errors? →What to check before you sell the service.
- How we work →The call, the fixed-fee Blueprint and milestone pricing.
