Start free
An agency team managing AI customer support for many client websites from one dashboard

AI Customer Support for Agencies: Supporting Many Clients From One Place

Published on byAntoniAntoni

If you build websites, Shopify stores or SaaS products for clients, you already do customer support. You just don't bill for it. The shipping question that lands in your inbox, the "is the site down?" WhatsApp at 11pm, the Slack thread where a client forwards a customer email with "can you handle this?": that is first-line support, and it quietly eats the margin on every project you ship. This post is about running it on purpose: one AI agent per client, one team, a clean handoff, and a line on the retainer.

Key Takeaways


The agency support trap

You launch a client's site. The contract covers design, build and a month of fixes. Then the customers arrive, and some can't find the returns policy or want to know whether the shop ships to Ireland. The client has no support person. The site has your name in the footer. So the questions come to you.

Three things make this worse than ordinary support:

  1. Every client is a different stack. Shopify here, a Next.js app there, a WordPress site with a booking plugin. No shared knowledge, and the person who built the site is the only one who can answer anything.
  2. Questions arrive everywhere. Email, Slack Connect, WhatsApp, a text to the account manager. None of it is logged, so nothing is learned.
  3. Nobody owns it. Support isn't in anyone's job description, so it goes to whoever is online: late, inconsistent, and quietly resented.

The usual escape hatch is to bolt a scripted bot onto the client's site: five buttons and a contact form at the end. The Bot makes the client feel covered without covering anything. The customer with a real question gets "I didn't quite get that," gives up, and leaves. The client never hears about it. Silent customers are lost revenue, lost on a site you built.

You don't need a better bot for each client. You need an agent that reads each client's content, resolves what it can, and hands the rest to a named human. Then you need to run that for twelve clients without twelve logins.

The operating model: one agent per client, one team

One agent per client

Each client gets its own agent. In FutureBase an agent is a self-contained deployment tied to one website or brand, with its own knowledge, chat widget, conversations, tickets and settings. Agents under one account are completely isolated, so a question on the bakery's site can never be answered with the dental clinic's opening hours. Name agents by client, not by project ("Northwind Coffee: Shop", not "Site v3"); with twenty of them, the sidebar is your client list.

One account, one team

The agency owns the account. Your team is invited once, with roles, and moves between client agents from the sidebar. Hand-offs for the agent you have open wait in Live Desk, longest wait first, with the full transcript of what the customer said and what the agent already tried. If your team lives in Slack, each agent can also post its hand-offs as a thread in a Slack or Discord channel, and a reply typed there reaches the customer. A channel per client, in your own workspace, is the cheapest shared inbox an agency can run.

Handoff to the client's own team when it should

Some questions belong to the client: a refund that needs their approval, a custom quote, a complaint about a specific order. Decide this per client, up front:

PatternWho answers hand-offsWhen it fits
Agency-firstYour team, in Live Desk or Slack; you escalate to the client by email when neededSmall clients with no support staff; sites where most questions are about the site itself
Client-firstThe client's person, invited as a MemberShops and services where answers need the client's judgement or systems
Split by topicThe agent's instructions say which topics go to whom; your team triagesLarger retainers where you handle "how does the site work" and they handle "where is my order"

The Member role is built for client-first: a Member can read and reply to conversations, tickets and feedback, and edit knowledge, but cannot touch organization settings, integrations, billing or other members (see the role table in the team management docs). Seats count toward your plan, so a client who only wants escalations by email is cheaper to serve with the email hand-off than with a seat.

Whichever pattern you pick, write it into the agent's custom instructions: "If a customer asks about a refund, say a team member from Northwind will reply by email within one business day, then hand off." The customer reads a promise instead of a dead end.

Productising it: support as a retainer line

Unpaid support stops being unpaid the moment you can describe it as a service with a number attached.

A price the client understands

Most agencies land on a flat monthly fee per client covering the software, a monthly content refresh and a short report, attached to the retainer the client already pays. The subscription sits on the agency's own account; message credits are pooled across every agent and channel, and you watch consumption under Settings → Usage. Pricing is your call: a flat line for most clients, a fee tied to conversation count for the one with real volume.

A per-client number

Every agent has its own dashboard with total conversations, customer satisfaction, response accuracy and active tickets. Report first the number clients care about most: how many customer questions were resolved without anyone on either side touching them. Count the conversations the agent closed itself against the ones it handed off, and you have a resolution rate per client. The formula and its traps (counting abandonment as success is the big one) are in the deflection rate guide.

A one-page monthly report is enough:

LineExample (illustrative)
Conversations handled212
Resolved by the agent174 (82%)
Handed to a person38
Top three topicsShipping times, size guide, "do you ship to…"
Content fixed this monthAdded the returns FAQ; updated the delivery page
Suggested next stepConnect the store so the agent can look up orders

The last two rows make this a retainer rather than a subscription resale: you are reading what customers ask and improving the site because of it. They are also a pipeline. If a client's customers keep asking about order status, the next step is connecting the store so the agent can look up orders during a conversation (see the Shopify integration page). Every topic the agent can't answer today is scoped work you can propose next month.

Onboarding a new client in an afternoon

This assumes the client's site is already live.

1. Create the agent (five minutes)

Agents → New Agent. Name it after the client, enter their site's URL, and you land in the agent dashboard.

2. Import their content (twenty minutes, mostly waiting)

On the Knowledge page, add the client's website with rules for which sections to read: skip a stale blog, include every policy, pricing and FAQ page. If the client has a hosted help center, paste its address into the Add box and choose Import help center; the agent reads every public article and re-reads them on your plan's refresh schedule, so a client who edits an article doesn't have to tell you. Add a text snippet for things written down nowhere, such as "closed on public holidays".

3. Write the instructions (fifteen minutes)

Pick a tone of voice and write custom instructions specific to this client:

The agent still answers only from the knowledge you gave it; instructions shape how, not what.

4. Set up the handoff (ten minutes)

Actions → Hand off to your team. Turn hand-off on, choose where it goes (Live Desk is always on; add email, or a Slack or Discord thread) and write the first message a customer reads when a person takes over. If the client answers their own hand-offs, invite them now as a Member.

5. Embed the widget (ten minutes)

Channels → Chat bubble → Install. Paste the snippet before the closing </body> tag; for a Shopify store that is theme.liquid. The Install page has guides for HTML, WordPress, Shopify, Webflow, Wix, Squarespace and Google Tag Manager, and tells you as soon as the bubble loads. Set the bubble colour and icon to the client's brand and add ignore paths for admin or checkout pages; see the widget embedding docs.

6. Test like a customer (thirty minutes, the part people skip)

Before telling the client it's live, use Ask your agent on the Knowledge page and ask the twenty questions the client's customers actually ask; you know most of them from your own inbox. For every weak answer, fix the content, not the instructions: add the missing FAQ, re-import the page that changed, remove the outdated PDF. This half hour decides the resolution rate you report next month. Then open the bubble on the live site, ask something the agent can't answer, and watch the hand-off land in Live Desk. If that round trip works, you're done.

Keeping clients separate and secure

Agencies hold a lot of trust, and a support deployment touches customer conversations. A few rules keep that trust intact.

Isolation by default. Each agent's knowledge, conversations and settings are never shared, so cross-client leakage isn't something you can cause by accident. Don't undo it by pasting one client's policy into another's snippets to save time.

Public content only. Import only what any customer could already read. The help center importer skips articles behind a login, and that instinct applies to everything else: no internal wikis, no customer lists, no confidential pricing. If a client wants the agent to answer from private material, make it a separate decision with their sign-off.

Least privilege for people. Your team gets Admin only to configure integrations or manage members; everyone else is a Member. Client contacts are Members, never Admins, so they can't see billing, integrations or the rest of your client list. When an engagement ends, remove the member; access stops immediately and their past activity stays in the record.

Keep a trail. The activity log under Settings records who invited or removed whom, which agents were created or deleted, when an integration was connected and when an API key was issued. Review it when you offboard a client, and point to it when a client asks who touched their deployment.

Write down the exit. The agency owns the organization, which keeps billing and administration in one place. Per client, decide on day one what happens if they leave: the agent is deleted or its content handed over, and their seats removed. On day one it's a clause; on the last day it's an argument.

What changes once this is running

The surprising part isn't that the agent answers questions. It's what the agency gets back. The 11pm WhatsApp stops, because the customer asked the site instead. The forwarded email becomes a hand-off with a transcript. And the first-line support you were doing for free becomes a service with a name, a number and a price, which is the only version of it that scales past five clients. Clients get customers who are answered at 2am, from their own content, and handed to a person with context when it matters. That is the difference between the Bot you could have dropped on their site and an agent that actually resolves. Your clients won't know the vocabulary, but they'll notice it in their inbox.

Frequently asked questions

Should an agency use one AI agent for all clients or one per client?

One per client, always. Each client has its own products, policies and tone, and a shared agent would mix them up or leak one client's details into another's chat. In FutureBase every agent is isolated: its knowledge, conversations and settings are never shared with other agents, so one agent per client is the natural unit.

How do agencies charge clients for AI customer support?

Most agencies add support as a line item on the monthly retainer, priced on a flat fee that covers the software, a monthly content refresh and a short report. The subscription and its message credits sit on the agency's own account, so the agency decides how to price each client, and per-agent dashboards give the conversation counts to justify it.

What happens when the AI agent cannot answer a client's customer?

It hands the conversation to a person instead of dead-ending. The hand-off waits in Live Desk with the full transcript, and it can also be sent by email or to a Slack or Discord thread. Decide per client whether the agency replies or the client's own team does, and write that into the agent's instructions so the customer is never left guessing.

How long does it take to set up AI support for a new client?

An afternoon for a typical site. Create the agent with the client's URL, import their site and help center, paste the script tag before the closing body tag, write the custom instructions and decide where hand-offs go. Most of the remaining time is spent asking the agent the twenty questions customers really ask and fixing the content behind any weak answer.


Running support for more than one client? Start free with FutureBase and give every client their own agent, managed by your team from one account, with hand-offs that reach the right person. See how it fits on the agency client support page.

Retire the Bot. Give your customers an agent.

Deploy in five minutes. No credit card required.

Trained on your website in minutes
Knowledge base & file upload
Help desk with full-context hand-off
MCP server access
GDPR compliant