What AI agent maintenance costs (and why it's not optional)

An agent is software wrapped around a language model, and both halves move. The model changes, the data it reads shifts, the systems it connects to get updated, and the rules it has to follow change too. Four forces drive the recurring work.

• Model drift: accuracy decays over time as your data and the world around it shift. A classifier that was right 95% of the time in January can quietly slip by summer, and nobody notices until the errors add up.
• Model version updates: production models version constantly. Anthropic's Claude Opus moved from 4.6 to 4.7 to 4.8, and Sonnet from 4.6 to 5, within about a year. A new version can improve results or change how the agent reasons, so each one has to be tested against your workflow before it goes live.
• API and integration breakage: the tools your agent talks to, your CRM, your accounting system, a model provider, change their APIs, rename fields, and deprecate endpoints. When they do, the agent stops working or starts working wrong.
• Compliance and audits: rules like the EU AI Act and PII-handling requirements evolve, and an agent that touches customer data has to keep pace. That means periodic review, not a one-time sign-off.

None of these are failures of the build. They are the normal conditions an agent runs in. Maintenance is the standing work of keeping the system aligned with a world that keeps moving.

Across the market, ongoing support for an AI system is usually sold as a monthly retainer or as a percentage of the build. Agency pricing (directional, from firms such as Arsum) puts an AI system support retainer at roughly $2,000 to $8,000 a month, or 15% to 20% of the build cost per year. That covers monitoring, model and API fixes, prompt tuning, and compliance work, typically across a whole system.

Our Care retainer is $1,500 a month per workflow. It sits at the low end of that band on purpose, because it is scoped to a single workflow rather than an entire system. You pay for the agent you have, and if you run three workflows, you have three lines you can see and cancel independently. Model API usage is billed separately, at cost, through your own accounts.

This is not a new pattern. Robotic process automation went through it a decade ago, and the numbers are a useful warning. For RPA, the software license was only a small slice of what the automation actually cost to own.

Kognitos put licensing at about 25% to 30% of total cost of ownership, with roughly $3.41 to $4.00 spent on consulting and maintenance for every $1 of licensing. SmartDev budgeted RPA maintenance at 15% to 25% of development cost per year. The visible price tag was the smallest part.

AI agents are repeating that shape. The build is visible and quotable; the maintenance is where the real cost of ownership lives. Any comparison that stops at the build price is measuring the wrong number.

Running the agent has its own cost, and it belongs inside the maintenance picture rather than beside it. Model API usage and hosting scale with volume. Directional figures from SoftTeco put model API spend anywhere from about $100 to $10,000 a month and hosting from $200 to $5,000, depending on scale.

At SMB volumes, this bill is usually small: reading a few hundred documents a month costs tens of dollars, not thousands. We bill it at cost through your own model provider accounts, so you see exactly what the work consumed and nothing is marked up. The retainer pays for the engineering that keeps the agent working; the model bill only covers the usage.

An unmaintained agent does not break loudly on day one. It drifts. The model slips, an integration changes, a batch posts wrong, and because nobody is watching, the first sign is often a customer complaint or a reconciliation that does not balance. By then the trust is gone, and trust is what got the project funded.

This is how agent projects die after launch. Gartner forecast in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, driven in part by escalating costs and unclear value. A common route to that outcome is simple: nobody owned the running system, so it degraded until someone shut it down. Maintenance is the difference between a pilot that becomes real infrastructure and one that gets canceled.

Maintenance should be a defined scope, not a vague promise to be around. Ours has three parts we can point to.

• Continuous monitoring: a monitoring agent checks each workflow every 30 minutes, fixes what it can safely fix on its own, and escalates the rest to a person. We already run this model in production, keeping a real news operation live around the clock. Breakage still happens; the point is that the system finds out before you do.
• Model updates, regression-tested: when a new model version ships, we test it against your workflow before switching, so an upgrade improves your results instead of silently changing them.
• One improvement a month: every workflow gets one concrete improvement each month, whether that is a new edge case handled, a tighter prompt, or a step that used to need approval and no longer does.

The test of a maintenance plan is whether you can describe what it does. If the answer is only that someone is on call, you are paying for a promise, not a service.

The market retainer of $2,000 to $8,000 a month usually covers a system: several agents, several integrations, priced as a bundle. Our $1,500 is scoped to a single workflow, which is why it sits below that band rather than inside it.

The practical difference is legibility. You can see the cost of each agent, judge whether that agent earns its keep, and cancel one without touching the others. A per-system retainer hides which workflow is worth the money. A per-workflow line makes it obvious, which is the accountability we want on ourselves.

Add it up over a realistic horizon. A production build from us starts at $9,000. Three years of Care on one workflow is $54,000. The recurring line, not the build, is the larger number, exactly as the RPA precedent predicted.

That is not an argument against building. It is an argument for pricing the whole life of the agent before you start. An agent that runs for years and keeps working is worth far more than its build cost. An agent that was built once and left alone is worth nothing within a year. The retainer is what moves you from the second outcome to the first.