How to run AI agents for clients without mixing them up

To run AI agents for several clients, give each client its own separate workspace, decide which people on your team can see each one, and report on each client separately. Those three rules prevent the most expensive agency mistake: one client's data, instructions or spend leaking into another client's work.

This guide is for AI agencies and consultants who build and run agents for clients. It covers how to structure client workspaces, set permissions, keep costs separate and report results.

Key takeaways

• One workspace per client keeps data, instructions, tickets and spend apart by design.
• Give each team member access only to the clients they work on.
• Put each client's AI usage on a key and budget you can trace to that client.
• Report per client from the tickets and approvals, so every claim in the report points to real work.

Why client separation matters more with agents

An AI agent acts on whatever context it has. If two clients share one setup, an agent can pull a detail from client A into work for client B, and nobody notices until the client does. Separation has to be built into the structure, because agents will not reliably remember a rule like "never mention the other client".

Clients are also still deciding how much to trust agents. McKinsey's November 2025 survey found 62% of organizations experimenting with AI agents, while no more than 10% were scaling agents in any single business function (McKinsey, The state of AI in 2025). A clean, per-client setup is part of what you are selling when a client is still testing.

Step 1: One workspace per client

In OpsHarbor's Agency plan, each client gets its own workspace: a separate AI company with its own org chart, goals, tickets, budget and approvals. You move between clients with the client switcher.

Set up each new client the same way:

1. Create the workspace and name it after the client.
2. Write the client's goals in their words, with what "done" looks like.
3. Connect the AI key you will bill this client's usage against.
4. Set the budget and the approval list before any agent starts work.

Keeping the setup identical across clients makes it faster to onboard the next one and easier to spot when one client's company behaves differently.

Step 2: Decide who on your team sees each client

Not everyone on an agency team needs every client. Agency accounts have per-client permissions, so you choose which team members can see each client's workspace.

A starting pattern for most agencies:

• Account owner: sees every client and handles billing.
• Client lead: can change agents, budgets and settings for the clients they run.
• Specialist or contractor: access to only the one client they help with.

Review access when someone changes roles or a client leaves. Removing access is one change, and teams often skip it.

Step 3: Keep each client's AI costs separate

Your agents run on an AI key you connect, and usage is billed by the AI provider. Agencies usually choose one of two setups:

• The client's own key. The client opens an account with the AI provider and gives you a key. Usage is billed to them directly.
• Your key, one per client. You use a separate key per client, so usage can be traced and passed on.

If you use Anthropic, a separate workspace per client in the Anthropic Console gives each key its own monthly spend limit and alerts (Anthropic docs: Workspaces, retrieved October 29, 2026). Combine that with the budget in each OpsHarbor workspace. Our guide on budgets and approvals for AI agents explains how the two caps work together.

Step 4: Agree the approval list with the client

Clients care most about what reaches their customers. Agree the approval list during onboarding, in writing:

• Who approves customer-facing messages: you, the client, or both.
• Which actions are never automatic for this client, such as refunds or deleting records.
• How fast approvals need an answer, so work does not stall.

When an approval needs the client, collect their answer and record it on the ticket. The ticket then shows who approved what and when.

Step 5: Report per client from real work

Agency accounts include per-client reports. Build the client's update from what the tickets show: work finished, approvals given, spend against budget and anything stuck. Every line in the report should point to a ticket the client could check.

Avoid reporting outcomes you cannot show, such as "saved 20 hours". If you want to report time saved, agree with the client how it is measured before you start.

Before you add the next client

Run through this list:

• New workspace created and named, with goals in the client's words
• AI key connected and a spend limit set at the provider
• OpsHarbor budget and approval list set
• Team access limited to the people working on this client
• Report format agreed with the client

Frequently asked questions

Can agents in one client's workspace see another client's data?

No. Each client's workspace is its own AI company with its own tickets and data. Team members only see the clients you give them access to.

Do I need a separate plan per client?

No. The Agency plan covers many clients. Prices will be published when signup opens.

Is there a partner program for agencies?

A partner program is coming after public launch. Join the waitlist to hear when it opens.

Where do I start?

Read what an AI company is, then join the waitlist. OpsHarbor is in invite-only beta, and we are inviting agencies early.