AI agent maintenance: why agents get worse over time and what it costs to keep them

A live agent is not a finished project

It's common to treat the agent's launch as the finish line. In the first month it answers well, the team trusts it and the project is considered done. Three or six months later come the answers with old prices, the questions it can't answer and customers asking to talk to a person earlier and earlier.

That doesn't mean the agent was badly built. It means it depends on business information, and business information changes.

Why the agent gets worse over time

The reasons repeat in almost every project:

• ✓The business changes: prices, catalog, return policy, hours and staff change, and the agent's knowledge base falls behind.
• ✓New cases appear: questions that didn't exist at launch, like a new product or a deadline change, make the agent improvise.
• ✓Integrations change: connected systems, such as the calendar, CRM and store, update their APIs, and an integration that worked can stop without warning.
• ✓Models change: providers release new versions and retire old ones, which requires migrating and testing the behavior again.
• ✓Instructions pile up patches: rushed adjustments, without review, make the behavior inconsistent.
• ✓Volume grows: more conversations bring more variation, and a bigger usage bill too.

The costs that continue after launch

The cost of building the agent is one-off. The costs below are recurring and need to be in the plan from the start:

• ✓AI model usage: billed by the volume of text processed (tokens). It grows with the number of conversations and the length of each one.
• ✓Channel: the official WhatsApp API charges per message, with prices that vary by category and country and change from time to time.
• ✓Infrastructure and automation: servers, orchestration tools and the database the agent runs on.
• ✓Connected tools: CRM, calendar or other platform plans used in the integrations.
• ✓Maintenance hours: monitoring, conversation review, base updates and adjustments.

So how much does it cost?

There's no single number: an agent handling a hundred conversations a month and one handling ten thousand have very different bills, just like an agent that only answers questions and another that checks three systems before every answer.

The measure that matters is the cost per resolved conversation, compared with the cost of doing the same work without the agent: the team's hours, leads left unanswered, sales that didn't happen. If that comparison isn't made every month, the AI bill grows without anyone knowing whether it's still worth it.

How to cut the bill without losing quality

Most agents have room to save without making service worse:

• ✓The right model for each step: not every task needs the most expensive model. Classifying a message or extracting a piece of data can use a smaller model.
• ✓Ready-made answers for the most frequent questions: opening hours and address don't need to go through a model every time.
• ✓Lean context: send the model only the information it needs, not the entire history and the full knowledge base.
• ✓Caching: reuse answers and processing that repeat.
• ✓A reviewed flow: cut steps that don't change the result.

What a good maintenance plan covers

Maintenance isn't just fixing what broke. A complete plan includes:

• ✓Monitoring: alerts for failures, response time and off-pattern conversations.
• ✓Conversation review: periodic samples to find errors and new cases.
• ✓Updating the base and the instructions whenever the business changes.
• ✓Cost optimization: reviewing usage and model choices.
• ✓Reporting: how many conversations the agent handled, how many it solved on its own and where to improve.
• ✓Evolution: new functions, such as booking or collecting payments, without starting a project from scratch.

A monthly maintenance routine, in practice

A well-run monthly cycle follows roughly this order:

• ✓First week: the previous month's report, with conversations, resolutions, handoffs to people and cost.
• ✓Second week: review of conversations with errors or that got stuck, and updates to the base and instructions.
• ✓Third week: testing the changes with real conversations before publishing.
• ✓Fourth week: cost and usage review, and planning the next cycle's improvements.
• ✓Throughout the month: monitoring, with alerts for failures and integration drops.

Signs your agent needs maintenance

If any of these signs already show up, the agent is asking for a review:

• ✓Answers with outdated information, such as prices, hours or return policy.
• ✓Customers asking to talk to a person earlier and earlier in the conversation.
• ✓API costs rising without conversation volume rising with them.
• ✓The team distrusting the agent and going back to answering everything by hand.

Do it in-house or hire it out?

If the company has someone technical with time to follow conversations, metrics and costs every month, it can be done in-house. In most small and mid-sized companies that person doesn't exist, and maintenance waits until something breaks.

That's the scenario AI agent support exists for: monthly care that also works for agents built by other vendors. If the agent is still to be built, it's worth planning maintenance from the custom AI agents proposal onward.