How agencies keep a person between the AI and the client
Put one named person right before AI work reaches a client. They see the draft, approve or change it, and on a no nothing goes out. Five real builds shown.
Agencies keep a person between the AI and the client by putting one approval point right before anything reaches a client. The AI drafts, a named person on the team reads it, and only that person sends, launches or publishes it. If they say no, nothing goes out. BizElevate builds every client-facing agent this way, and this guide shows where the check sits in five real builds, what the person sees, and the rules we use to place it.
Why the check has to be there
AI makes mistakes. It can get a number wrong, misread a note or write a line that sounds odd to a client who knows you well. Anyone who tells you their AI is always accurate is overselling.
So we don't rely on the AI being right. We build around three things we can control:
- The agent works from your own data and tools, not from guesses.
- A named person on your team checks and sends anything client-facing.
- Nothing client-facing goes out on its own.
The rule is about client-facing work. Some internal jobs can run on a schedule, and we'll come back to that.
Where the person sits in five real builds
Every case study in our library has a person check. Here's who approves in each one, what they see and what a no means.
| Build | Status | Who approves | What they see | What happens on a no |
|---|---|---|---|---|
| Proposal agent, digital marketing agency, Singapore | Live | The salesperson | A call summary and a full proposal draft in the agency's template | Nothing goes to the prospect. The salesperson edits it or doesn't send it. |
| Campaign setup agent, B2B marketing agency, Australia | In build | A person on the team | A summary of the checked media plan, then the campaigns as drafts in each ad platform | The campaigns stay as drafts. A person always launches them. |
| Field visit reports, pharmaceutical manufacturer, Indonesia | Live, rolling out to the field team | An admin | Any entry the agent isn't sure about, on one screen | Unclear entries wait on that screen until the admin approves them. |
| IT helpdesk helper, automotive group, Germany | Live | The technician | An internal note on the ticket with the likely fix | The person who raised the ticket never sees the note. The technician decides what to do. |
| Board report and release notes, software company, Europe | Live | The team, then a product manager | Each customer's notes on a review page, and the draft release notes | The board report doesn't go before the team edits it. Customers see nothing until the product manager approves and support publishes by hand. |
The proposal agent
The Singapore proposal agent turns a sales call transcript into a call summary and a branded proposal. The salesperson edits the draft and sends it. Nothing goes to the prospect until the salesperson has read, edited and sent it. The check sits after the writing and before the sending, which is the last moment anyone can change it.
The campaign setup agent
The Australian campaign setup agent checks the plan before a person sees anything. The agent reads the approved media plan and checks it for errors before anything is built. Then it shows a summary for a person to approve and creates the campaigns as drafts in each ad platform. A person reviews and launches them, and every run and every approval is logged.
The field visit reports
In the Indonesian field visit build, reps post short notes or voice messages in the group chat they already use. The agent turns each one into a structured report. Reports it isn't sure about go to a screen where an admin approves them. The agent doesn't guess its way past an unclear note.
The IT helpdesk helper
The German helpdesk helper reads each new ticket, searches the approved IT guides and posts an internal note with the likely fix. A weak match gets a standard note instead of a guess. The note is internal, so the person who raised the ticket never sees it, and the technician decides.
The board report and release notes
For the European software company, the team edits each customer's notes on a review page before the board report goes out. For release notes, the product manager approves, and the support team publishes the help article by hand. Two people check before a customer sees anything.
The rules we use to place the check
Put the check right before the client sees it. Not at the start, where there's nothing to judge yet, and not after, when it's too late. For a proposal that's before sending. For campaigns it's before launch.
Name the person by role. The salesperson, the admin, the technician, the product manager. "The team will check it" usually means nobody does.
Show them what the client would see. The approver reads the proposal in the template, or the campaign as a draft in the ad platform.
Make a no the easy path. The default state is a draft. If the approver does nothing, nothing goes out.
Let the agent say it isn't sure. A weak match gets a standard note. An unclear entry goes to a person. An honest "not sure" is better than a confident guess in front of a client.
Run automatic checks first, so the person judges the rest. Checks catch the mechanical mistakes, like a wrong number or a wrong name. The person spends their time on what a check can't judge: tone, context, whether this is the right thing to send this client today.
Keep a log. The campaign setup agent logs every run and every approval, so you can see who approved what.
Give the agent only the access it needs. An agent that only drafts doesn't need the power to send.
Where a schedule is fine
Not every job needs someone to press a button. Internal work, where no client sees the output, can run on its own with a lighter check.
New-hire onboarding for a Croatian engineering group is an example. When HR adds a new hire, the system creates every checklist task, sends the welcome email with the forms, emails the IT team and reminds the supervisor before the start date. HR checks the setup note on the main task, and the same hire can't be set up twice. That's the right level for internal work. It wouldn't be the right level for a client report.
Which steps need an agent and which need a fixed rule is in AI automation vs AI agents.
Who should be the approver
Pick the person who'd answer for the mistake if one got through. That's usually the person who sends this kind of work today: for example, the account manager for client reports, the salesperson for proposals, and whoever launches campaigns today for campaign setup.
They also need the time to do it. A check that piles up for a week isn't a check, so the drafts should land where that person already works, in a form they can read quickly.
Where BizElevate fits
BizElevate builds AI agents and automations for agencies and service businesses in Australia, Singapore and Southeast Asia. Every build puts a person between the agent and your clients, using the rules above.
Builds cost US$3,000 to US$6,000 and go live in 2 to 4 weeks. You own everything we build, with no lock-in. We've deployed 100+ automations for clients and our own operations, and Hady Wu, the founder, builds and deploys the agents himself.
Frequently asked questions
Does AI make mistakes on client work? Yes. It can get a number wrong or write a line that sounds off. That's why the agents we build work from your own data and a named person on your team checks and sends anything a client sees. We don't quote an accuracy figure, because it depends on your data and the task.
Will the agent ever send something to a client on its own? Not in what we build. Anything client-facing is drafted by the agent and approved by a person on your team before it goes. Some internal jobs, where no client sees the output, can run on a schedule.
Who should approve AI drafts in an agency? The person who'd answer for a mistake, usually whoever sends that work today. For example, the salesperson for proposals, the account manager for client reports, and whoever launches campaigns today for campaign setup. Name them by role, not "the team".
What happens if the approver says no? Nothing goes out. The work stays as a draft, and the approver can edit it or drop it. In our campaign setup agent, campaigns stay as drafts until a person launches them.
Can we see who approved what? Yes, if it's built in. Our campaign setup agent logs every run and every approval, and the same kind of log can go into other jobs.
Next step
Start with the free AI audit. Tell us how your team works, and we'll point out where an agent helps first and where the person should sit. Or book a call to talk through one workflow.
Published by BizElevate, based in Ho Chi Minh City.