An agency's playbook for adding AI chatbots to client retainers.

Every digital agency we talk to is being asked the same question by clients: "can you do AI for us?" The honest answer is yes — chatbots are the most concrete, fastest-to-ship form of "AI" most clients will accept. The dishonest answer is to say yes before you have decided how the service line actually works.

This is the operator's checklist. It is the set of decisions that determine whether a chatbot service line is a margin business or a support nightmare. Run through it before you quote your first client.

Decision 1 — Pricing model

Three models work. Pick one before you quote anything.

Flat setup + monthly retainer. $1,500–5,000 to deploy and tune. $300–800/mo for hosting passthrough, content updates, and model tuning. Predictable. Easy to sell. Easy to scope. Best for agencies with existing retainer clients.

Revenue share. You take a percentage of leads or qualified conversations the bot generates. High upside. Hard to attribute. Hard to invoice cleanly. Works only if the client trusts your numbers and the bot is genuinely tied to revenue.

One-time project. $3,000–8,000, deployed, handed over, gone. Lowest friction to sell. Worst LTV. Use this when the client refuses a retainer but you still want the logo.

The retainer model wins for most agencies. It compounds. Revenue share sounds clever and almost never collects.

Decision 2 — Per-client deployment vs shared multi-tenant

You can run one big chatbot platform and create a tenant per client. Or you can deploy a separate stack for each client.

Shared multi-tenant. One platform, many clients. Lower total infrastructure cost. Higher operational risk — one breach, every client breached. Hard to give a client a copy of "their" data without exporting it from a shared store.

Per-client deployment. A standalone stack per client. Higher infra footprint. But each client is isolated, churn is clean (you delete the stack), and the deliverable is something the client genuinely owns. If you ever need to hand over admin access, it is one stack, not a row in a tenant table.

Per-client wins on legal cleanliness and on the conversation you have when a client asks "what happens to our data if we leave?"

Decision 3 — Who pays the AWS bill

The single most-overlooked question. Two options. Both work. Pick deliberately.

Agency-paid (consolidated). You run the deployment in your AWS account, charge the client a hosting line that covers cost + margin. Pros: simple billing, you get the volume discount across all clients. Cons: you carry the cash float, and you become responsible for an AWS outage from your client's perspective.

Client-paid (per-account). You deploy into the client's AWS account. The client gets the bill direct from AWS. Pros: zero float, the cost is transparent to the client, you do not own AWS support escalations. Cons: more setup friction (you need IAM access from each client), and clients need an AWS account.

The client-paid model is the right answer for compliance-conscious clients and for agencies that do not want to be in the reseller-of-cloud business. The agency-paid model is right for SMB clients who do not want to think about it.

Decision 4 — White-label vs co-brand

White-label means the bot is your client's brand entirely — your agency name does not appear anywhere. Co-brand means a small "powered by" or "built by" mark stays on the bot.

White-label is what clients ask for. Co-brand is what gets you inbound from your clients' customers. There is a middle path: white-label the widget itself, but include a small attribution in the dashboard the client sees, so internal teams know who built it. That gets you the referral signal without the customer-facing logo.

Decision 5 — Support boundaries

Define these in the contract. Do not figure them out at 9pm on a Friday.

• Bot stops responding. You. Within 4 business hours.
• Wrong answer to a customer. You retrain. Inside the monthly retainer for the first N hours.
• Client wants to add 200 new training documents. Scope creep. Quoted separately.
• Client wants a new feature the platform does not support. Out of scope, escalate to vendor.
• AWS itself goes down. Not you. Status page link in the contract.

Vague support boundaries kill chatbot retainers. Specific ones make them profitable.

Decision 6 — The handover-on-churn question

Every client will eventually leave. Decide how you hand the bot back before the relationship sours.

If you used per-client deployment in the client's AWS account, handover is trivial — revoke your IAM access and they own everything. If you used agency-paid shared multi-tenant, handover means an export, an import somewhere else, and a fight about who owns the conversation history. The handover model you can offer at termination is a sales asset at signing — clients buy from agencies that have a clean exit story.

Where the chatbot sits in the retainer

Three pairings work cleanly inside an existing service line:

• Web design + chatbot. The chatbot ships at site launch, tuned to the new content. Trains on the same docs you wrote. Easiest cross-sell.
• SEO + chatbot. The bot answers what your traffic is searching for, captured from the same keyword research. Useful as a lead-capture surface for organic traffic.
• Conversion optimization + chatbot. The bot becomes the always-on version of the live-chat sales agent the client cannot afford. Highest visible impact on a metric clients already track.

It rarely works as a standalone service line. The retainer math falls apart at scale because chatbots are too small to be the whole relationship — they have to ride alongside something else.

Where this lands

We built Chatmancer as a CloudFormation stack that ships in four minutes per deployment, which makes per-client deployment into client AWS accounts an actual practical model rather than a theoretical one. The white-label tier is built for agencies — your brand on the dashboard, your pricing, your domain.

But the playbook above stands regardless of the platform. The agencies making real margin on AI chatbots are the ones who decided how the service line worked before they sold it. The ones who figured it out client-by-client are not making margin.