Guide · 8 min read

What an AI integration costs
and how long it takes

Nobody publishes this, so budgeting an AI project feels like guesswork. Here is how the work is actually scoped, what pushes the number up, and what to expect at each size.

00Why this matters

The AI is rarely the expensive part. Everything around it is.

People assume the cost of an AI project tracks the sophistication of the model. In practice the model is a small, well-documented piece. The effort goes into reaching your data, mirroring your permissions, handling the cases your process never wrote down, and making the result land somewhere your team already works.

This is why two projects described in the same sentence can differ several times over in price. "Read our invoices and enter them" is four weeks if invoices arrive by email in three formats and your accounting system has an API. It is four months if they arrive by post, and approval depends on rules that live in one person's head.

You cannot get a reliable number without a conversation. But you can understand what drives it, which is enough to sanity-check any quote you receive, including ours.

GUIDEThe playbook

Where the money and the weeks actually go.

01Cost drivers

What actually determines the price

Ranked roughly by how often each one turns a small project into a large one. Notice how far down the model choice sits.

  • Access to your systems. A documented API is straightforward. A system with no API, or one only a vendor can change, is the single biggest cost multiplier in this kind of work.
  • The state of your data. Not whether it is tidy, but whether it is reachable. Content spread across drives, mailboxes and one spreadsheet somebody maintains privately turns into real preparation effort.
  • How many exceptions the process has. Every 'except when' adds logic, testing and review design. This is the item most often discovered mid-project rather than during scoping.
  • The accuracy you need. Good enough with a person reviewing flagged cases is affordable. Near-perfect and unsupervised is a different project, and often not worth what it costs.
  • Regulatory obligations. Audit trails, explainability and documentation are real engineering. If the process is regulated, expect that to be a visible line rather than an afterthought.
  • The model and provider. Genuinely a smaller factor than most people expect, and usually an ongoing usage cost rather than a build cost.
02Sizes

Three project shapes, and what each involves

Most integration work falls into one of these three shapes. Identifying which one you are asking for makes any quote much easier to read.

A single workflow, one system

One input, one destination, rules that fit on a page. Two to four weeks. This is the right size for a first project and the only size we recommend committing to before anything has been proven on your data.

A workflow across several systems

Data read from one place, decisions applied, results written to two or three others, with a review queue for exceptions. Six to twelve weeks, usually split into phases with something usable at the end of each.

A department-wide platform

Multiple workflows, multiple teams, permissions, reporting and audit. Months, and it should never be attempted first. Build one workflow, earn the trust, then extend.

03Running costs

The costs that start after go-live

These are modest but real, and they are the ones most often missing from a budget approved by someone who was only shown the build price.

  • Model and API usage. Charged by volume of text processed. Predictable once you know your monthly case count, and worth estimating during scoping so nobody is surprised.
  • Hosting and infrastructure. Usually the smallest line for this kind of work, but it exists and someone has to own the account.
  • Maintenance. Providers deprecate models, your systems get updated, your document mix changes. Something with no maintenance budget will quietly stop working within a year.
  • Monitoring and review time. Someone still looks at the flagged cases. That is a feature, not a failure, but it is time and belongs in the business case.
04Timeline

Where the weeks go in a typical project

The build itself is rarely the longest phase. Being aware of that helps you spot a timeline that has skipped something important.

Discovery and mapping

Understanding the process, counting volumes, checking system access. Short, but skipping it is the most reliable way to overrun later.

Data preparation and access

Frequently the longest phase, and almost always the most underestimated. Credentials, permissions, exports and cleaning all live here.

Build and evaluation

The AI work, plus testing against historical cases with known outcomes. Runs faster than people expect once access is solved.

Pilot and tuning

A small group using it on real work while thresholds are adjusted. Do not compress this; it is where trust is either earned or lost.

Hand-off

Documentation, monitoring and a walkthrough. A week that saves months of dependency later.

05Overruns

Why AI projects overrun, and how to prevent it

In our experience overruns are almost never caused by the AI failing to work. They are caused by four things nobody counted.

  • Exceptions discovered late. The process turns out to have fifteen special cases rather than three. Prevent it by asking the people doing the work, not the person who owns the process.
  • System access taking weeks. Credentials, security review, a vendor who has to enable something. Start this on day one, in parallel with everything else.
  • Scope that was never written plainly. If the deliverable cannot be stated in one sentence a colleague could verify, it will be renegotiated mid-project at your cost.
  • Accuracy expectations set by a demo. A polished demo sets an unrealistic bar. Establish the real baseline from your own historical cases before anyone forms an opinion.
REFQuick reference

Timelines by project shape.

Project shapeTypical timelineRight for a first project?
Proof of concept on your own data1–2 weeksYes, if you need convincing first
Single workflow, one system2–4 weeksYes
Workflow across several systems6–12 weeksOnly in phases
Regulated process with audit requirements8–16 weeksNot as a first project
Department-wide platformSeveral monthsNo
CTATalk to Brains

Get a real number for your case.

Describe the workflow and the systems involved, and we will come back with a realistic range and a first phase. The call is free and there is no obligation to build with us. See how we scope integrations or get in touch.

Guide FAQ

Common questions about cost and timelines.

Why won't agencies quote a price for AI integration upfront?

Because the cost is driven by your systems rather than the AI. Two projects with identical descriptions can differ several times over in effort depending on whether your systems have usable APIs, how consistent your data is, and how many exceptions your process actually contains. An honest quote follows a discovery conversation.

How long does an AI integration take?

A single well-defined workflow into one system typically runs two to four weeks. Something touching several systems with real permission rules and exception handling runs six to twelve weeks. Anything quoted at a few days is a demo, and anything quoted at a year should be broken into phases.

What are the ongoing costs after launch?

Three things: model or API usage, which scales with volume; hosting, which is usually modest; and maintenance, because providers deprecate models and your systems change. Budget for all three from the start rather than discovering them in month four.

What makes an AI integration more expensive?

Systems without APIs, data spread across places with no reliable way to reach it, permission rules that must be mirrored exactly, high accuracy requirements, regulated processes needing audit trails, and processes nobody has written down. The AI model choice is rarely the expensive part.

Is it cheaper to buy a tool instead?

Sometimes, and a good partner will tell you when. An off-the-shelf product wins when your process closely matches what it was built for. It stops being cheaper the moment your team has to keep moving data in and out of it by hand.

How do I avoid a project overrunning?

Insist that phase one is small, has a named deliverable in plain language, and includes a test against cases where you already know the right answer. Most overruns are not engineering failures; they are scope that was never defined and exceptions nobody counted.