← All articles

What an AI Project Really Costs After Launch

The build is paid once. Model usage, monitoring, maintenance and support are paid every month, and they decide whether an AI project is still alive in year two.

A small stack of invoices, a monthly calendar page and a brass calculator on a tidy desk, standing in for the running costs of an AI system after launch

Key takeaways

  • After launch, an AI project has five running costs: hosting, model usage, monitoring, maintenance and support.
  • Model usage is the only line that moves with volume. Picking the right model for each job can change it by a factor of four or more.
  • AI systems need maintenance even when nobody touches them, because models retire, platforms change their rules and connected tools change their APIs.
  • Ask who notices a failure first. If the answer is "your customer", the monitoring line is missing.
  • A well-run AI system should get cheaper per task over time while handling more work.

The build price is the number everyone negotiates. The running costs decide whether the project is still alive in year two. Most owners we talk to know roughly what an AI project costs to build and have no idea what it costs to keep running, or who will fix it at 7am when it stops.

Five lines on the monthly bill

In how AI implementation projects are priced we covered what drives the build. This piece is about everything after go-live, the part that rarely shows up in a proposal in any detail.

Whatever the project, the running costs fall into five lines:

  1. Hosting. The server, database and storage the system runs on.
  2. Model usage. What the AI provider charges for every request.
  3. Monitoring. Knowing that it works, every day, without someone checking by hand.
  4. Maintenance. Keeping it working while the world around it changes.
  5. Support. A person who answers when something looks wrong.

Hosting is usually the smallest and most predictable of the five. A small business AI system rarely needs more than a modest cloud server. The other four cause most of the surprises.

Model usage: the line that moves

AI providers charge per token. A token is a piece of a word, roughly four characters of English text. You pay for the tokens you send (the email, the instructions, the context) and, at a higher rate, for the tokens the model writes back.

The spread between models is large. On Anthropic's current list prices, the small Haiku model costs $1 per million input tokens and $5 per million output tokens. The top Opus model costs $4 and $20, four times as much from the same provider.

Here is a simple illustration with our own arithmetic on those list prices. Say an agent sorts 200 emails a day, reading about 2,000 tokens per email and writing about 200 tokens of summary and label. That is 400,000 tokens in and 40,000 out per day.

  • On the small model: about $0.60 a day, around $18 a month.
  • On the top model: about $2.40 a day, around $72 a month.

Neither number will sink a business. But multiply the jobs (inbox, quotes, invoices, reports) and the choice of model per job starts to matter. Sorting email doesn't need the most capable model. Drafting a careful reply to an angry customer might.

Three other levers move this line:

  • Caching. When the same instructions or documents go out with every request, providers can cache them. On Anthropic's API, reading from the cache costs a tenth of the normal input price.
  • Batching. Work that can wait a few hours, such as a nightly report, can run through a batch interface at half price.
  • Where the model runs. Keeping processing in a specific region, for example the EU, carries a 10% premium on Amazon's and Google's cloud platforms. For most European clients that premium is worth paying, and it should be in the budget from day one.

One detail catches people out. Anthropic notes that its newer models use a tokenizer that produces about 30% more tokens for the same text. A model upgrade can make a job better and slightly more expensive at the same time. Someone has to check that after every upgrade.

Maintenance: the world changes under your AI

Owners often assume that software nobody touches keeps working. AI systems don't, because the parts they depend on keep moving.

Models retire. Anthropic gives at least 60 days' notice before retiring a model. On 30 September 2026 it told developers that Claude Sonnet 4.5 retires on 30 November 2026. Any system still calling that model after that date simply fails. Moving to the replacement means testing that the new model still writes the reports, sorts the emails and follows the rules the way the business expects.

Platforms change their rules. On 1 July 2025 Meta moved WhatsApp business messaging to per-message pricing for templates. From 15 January 2026 it banned general-purpose AI chatbots on the WhatsApp Business Platform, while still allowing AI for customer service such as bookings and order tracking. A business running an assistant over WhatsApp had to check, twice in seven months, whether its setup was still allowed and what it now cost.

Connected tools change. The CRM, the accounting software, the email provider: each one updates its interface, renames fields or changes how logins work. Every connection is something that can break on a Tuesday morning.

The business changes. A new product line, a new price list, a new person in the office. The AI needs to know. A system that was right in March can be wrong by September without anyone noticing.

That's why we treat every AI system we build as something to run, with someone responsible for it every month. The software succession planning piece covers the other side of the same coin: what happens if the person maintaining it disappears.

Monitoring: who notices first?

There is one question that tells you whether monitoring is in place: when the system breaks, who notices first?

If the answer is a customer ("I sent a damage report three days ago and heard nothing"), the monitoring line is missing. AI systems also fail in a quieter way than normal software. They don't always crash. Sometimes they keep running and get worse: summaries get vaguer, a field gets filled wrong, an email category gets ignored.

Good monitoring for a small business system checks three things:

  • Is it running? Every scheduled job leaves a record. A missing record raises an alarm.
  • Is it right? A small set of real test cases with known answers runs after every change, so a model update or a prompt tweak can't silently break something.
  • What does it cost? Usage per job, tracked daily, with a cap. A loop that runs all night shouldn't be discovered on the invoice.

A small business system can cover all three with simple checks and alerts, as long as someone decides before go-live that it matters.

Who fixes it at 7am?

For an owner, the practical question is simple: if it stops working before the shop opens, who picks up the phone?

For a small agency or a freelancer the honest answer is often "me, when I see the message". That can be fine. It should just be written down: response times, what counts as urgent, and what happens when that person is on holiday.

When you compare offers, remember that a lower monthly fee with no named response time moves the risk to you.

How the costs should behave over time

A well-run AI system should show two trends at once: more work handled, and a lower cost per task.

The volume grows because the business trusts the system with more. The cost per task falls because, once there is real usage data, routine jobs move to cheaper models, repeated context gets cached, non-urgent work gets batched and wasted calls get found and removed. The first month is usually the most expensive per task. If the twelfth month isn't cheaper per task than the first, something isn't being looked after.

That is also how we judge our own work. Our clients hire us to get more done with the team they already have. A running system whose cost per task climbs every quarter fails that test even if it never crashes.

Questions to ask about running costs

Before signing for any AI project, ask the provider to answer these in writing:

  1. What exactly does the monthly fee cover: hosting, model usage, monitoring, maintenance, support?
  2. Is model usage included, capped or billed on top? What happens if volume doubles?
  3. Which model runs which job, and when was that last reviewed?
  4. Who handles a model retirement or an API change, and is that included?
  5. How will we know it's working, and who notices first when it isn't?
  6. What is the response time when something breaks, and who covers holidays?
  7. Will I see a monthly report with volume, cost per task and incidents?

The number to watch

The build invoice arrives once. The running costs arrive every month for as long as the system lives. In year two, the useful question is what each task costs now compared with month one. If that number is falling while the system handles more, the project is being run well.

Frequently asked questions

It depends on volume and on which models run which jobs. In our worked example, sorting 200 emails a day costs around $18 a month on a small model and around $72 on the largest one, before hosting, monitoring and support. Ask for an estimate per job based on your real volumes, and for a cap.

Because the parts it depends on change. AI providers retire models (Anthropic gives at least 60 days' notice), messaging platforms change their rules and connected tools update their interfaces. Each change needs someone to test and adapt the system.

For routine jobs like sorting, tagging and short summaries, usually yes. For careful writing or complex decisions, a larger model is often worth it. The right setup mixes models per job and checks the choice with real test cases.

On Amazon's and Google's cloud platforms, regional endpoints such as the EU cost about 10% more than global routing. For European businesses handling customer data, that premium usually buys a much simpler GDPR position.

At minimum: hosting, monitoring, updates when models or connected tools change, and a named response time for support. Ask whether model usage is included and how new features are priced.

Want to see what AI can do for you?

Tell us about your business. We'll get back to you within 24 hours.

Schedule a Strategy Call

More articles