Cost
How much does AI integration actually cost?
A single well-scoped AI workflow typically costs a few thousand dollars to build and tens of dollars a month to run. The build is the small number. The part that surprises people is the running cost, and the part that ruins projects is scope.
Almost every quote you will get for “AI integration” is unhelpful, because the phrase covers everything from a chatbot on a marketing site to a system that reads your invoices and posts them to your ledger. So here is the honest version, broken into the three numbers that actually exist.
The three numbers
1. The build. What it costs to get one workflow working on your data.
2. The running cost. What you pay every month for model calls, hosting and storage once it is live.
3. The cost of being wrong. What it costs you when the model is confidently incorrect and nobody notices. This one is almost never quoted, and it is the one that decides whether the project survives.
What a build actually costs
A single narrow workflow, built properly and put into production, is typically a few thousand dollars of work. Two to four weeks of engineering.
By “narrow workflow” we mean something with a clear start and end. Documents in, structured records out. An assistant that answers a defined set of questions about your catalogue. A monitor that watches one thing and escalates when it looks wrong.
What pushes it past that:
- Integrations with systems that have no API. If we have to drive a legacy tool through a browser or parse an export, that is engineering time.
- Data that needs cleaning before it is usable. Usually recoverable, but it is work.
- Human review interfaces. If a person has to approve model output, that approval screen is real product work, not a checkbox.
- Compliance requirements. Data residency, audit trails, retention rules.
What does not push it up as much as people expect: the model itself. Choosing between providers is usually a config change, not a rebuild, if the system was designed properly in the first place.
What running it costs
This is where estimates go wrong, because model pricing is per token and nobody thinks in tokens.
A rough way to reason about it: a single, fairly rich request against a modern model, with a few pages of retrieved context, costs somewhere in the region of a fraction of a cent to a few cents. Multiply by how many times a day it happens.
- An internal tool used forty times a day by five staff: usually tens of dollars a month.
- A customer-facing assistant on a busy storefront: this is where it can run away, because usage is unbounded and one enthusiastic user can generate hundreds of calls.
That second case is why every customer-facing system we ship is metered. Per user, per tier, with hard limits. On one Lebanese retail build we tied the assistant’s daily allowance to the customer’s loyalty tier, so heavy users get more and the monthly bill stays inside a range the owner agreed to in advance.
If a vendor quotes you a build price and says nothing about the monthly running cost, ask. It is the number that repeats.
The cost nobody quotes
A model that is wrong 5% of the time is not 95% useful. It depends entirely on what happens on the wrong 5%.
If the output goes straight into your accounts, a 5% error rate is a disaster. If it goes into a review queue where a person spends four seconds confirming it, a 5% error rate is fine and you have still saved most of the work.
This is a design decision, not a model decision, and it should be made before anything is built. In practice it means defining a confidence boundary: above it, the system acts; below it, a human sees it. Systems without that boundary are the ones that get switched off three months in.
Four things that move the number
- Scope. One workflow is a few thousand dollars. “AI across the business” is not a project, it is a budget line with no end.
- Whether a human is in the loop. Review interfaces cost real money and are almost always worth it.
- Usage volume, not user count. Ten people hammering a tool costs more than a hundred people who touch it weekly.
- How much you keep in-house afterwards. Handover and documentation is cheaper than an ongoing retainer, if your team can carry it.
What we do about it
We quote fixed scope with a number before starting, and we start with one workflow rather than a programme. If the honest answer is that a rule or a form would fix your problem for a fraction of the cost, that is what we will tell you. It is cheaper for us to be right than to be hired.
We also build the first one at our cost and let you run it free for up to three months, because arguing about projected ROI is less useful than watching the thing work on your own data.
The short version
Budget a few thousand for a first workflow and tens of dollars a month to run it. Insist on knowing the monthly number before you start. And ask what happens when the model is wrong, because whoever cannot answer that has not designed the system yet.