After an AI agent goes live: what you own and what you keep paying for

An AI agent is not one product with one price. It is a WhatsApp number, an AI model, somewhere for the code to run, your rules written down, and someone who keeps it working. Each piece sits in an account in someone’s name, and several send a bill every month.

So before you sign, ask for one written list: every account the agent uses, whose name it is in, and every recurring cost by type, with who bills it and who pays it.

• Your WhatsApp number, in your own business portfolio. This matters more since Meta changed how WhatsApp accounts work. Under the new model, generally available from 23 September 2026 and reaching all businesses by mid-October, the number sits in a WhatsApp account that Meta says is “always owned by the client or direct developer”, and providers get access when you share that account with them. Meta also states: “A number can’t be moved to another business portfolio” (Meta, account model). If your provider registers the number itself, it is the direct developer, and the account is theirs. If the number is set up in your provider’s portfolio, getting it back is not a settings change.
• The AI model account, or a clear arrangement. Anthropic’s commercial terms say the customer “owns its Outputs” and that Anthropic “may not train models on Customer Content from Services” (Anthropic). Those terms are between Anthropic and the customer that holds the account. If the account is your provider’s, that customer is your provider. Ideally the account and its API key are yours; if not, get in writing how usage is billed to you and how the setup moves to your own account.
• The accounts it runs on. The server or cloud account and any workflow tool, in your name or with administrator access for you.
• The code, or the right to use it. Access to the repository with the code written for you. Some providers, us included, reuse an underlying platform across clients; then ask what you may keep using if you leave.
• Your rules, your knowledge base and your test set. The instructions the agent follows, the information it answers from and the cases it is tested against. They hold your know-how, so say in the contract that they are yours and that they are not reused for anyone else. That one clause answers the competitor question.
• Your data and conversation history. Where the conversations are stored, how you export them, in what format, and what is deleted when you stop working together.

Subscription or one-off? is usually the wrong split. Most agents have three kinds of cost: the build, which is one piece of work; the running costs, which depend on use and are billed by each provider; and maintenance, which is a separate decision. Ask for each line on its own.

We explain the WhatsApp change, with an example, in what WhatsApp charges for service messages from October 2026 (sources: Meta, pricing; Anthropic, pricing). If a WhatsApp solution provider sits between you and Meta, ask what it adds per message or as a fee. And ask whether the agent needs a paid helpdesk to hand over at all, or whether your team can take over another way.

An agent that worked on launch day does not stay working on its own. Four things change under it:

• The model is retired. Anthropic gives customers with active deployments “at least 60 days’ notice before model retirement”, and after that date “Requests to retired models fail”. On 30 September 2026 it notified developers using Claude Sonnet 4.5 that the model retires on the Claude API on 30 November 2026, with Claude Sonnet 5.5 as the recommended replacement, and it advises testing with the new model well before the date (Anthropic, model deprecations). Someone has to read that notice and act on it.
• Meta changes the rules or the prices. In the past month, service replies started being charged and the account model changed; Meta asked businesses to add a payment method to their Messaging accounts by 30 September 2026, or replies stop once the free tier is used (Meta).
• Your business changes. New prices, hours or required documents. The agent only knows what its information says.
• Every change needs a retest. A new model or rule can break an answer that used to be right. Run the whole test set again first; we explain how in how to test an AI agent.

So is there monthly maintenance? really means: who watches for these, makes the change and reruns the tests, and how fast. Your team, with good documentation, or your provider under a written scope. Never nobody.

1. Whose business portfolio will the WhatsApp number be in, and how do you get access to it?
2. Whose name are the AI model, hosting and workflow accounts in, and do I get administrator access?
3. What is the full list of recurring costs, by type, with who bills each one?
4. Does handing over to my team need a paid tool? Which one, and how does it bill?
5. Do I get the code and the repository? If you reuse a platform, what may I keep using if I leave?
6. Are my rules, documents, conversations and test set mine in writing, and kept out of work for anyone else?
7. How do I export my conversation history, and what is deleted when we stop?
8. Who handles a model retirement or a Meta change, and how fast?
9. Do I get the test set and documentation so my team can update the agent without you?

When we build an agent, we set up its running costs in your name, on your accounts, so you pay each provider directly and see every invoice; none of them is a fee to us. Your criteria, rules and documents are yours, and we write the documentation so your team can update the answers and the rules without us. What we reuse across projects is the underlying platform: the messaging layer, handover to a person and transcripts. More on how it works in our WhatsApp customer service AI agent.