What AI automation actually costs a business
Ask three firms what an automation project costs and you will get three ranges that barely overlap. That is not evasiveness. It is that the described task tells you almost nothing about the work behind it.
What actually drives the price?
Integration surface, exception count, and system access — in that order. The AI portion is rarely the expensive part. A model call to classify an email costs a fraction of a cent; getting reliable access to the mailbox, agreeing what happens to ambiguous messages, and handling the day the upstream format changes is where the budget goes.
Why integration surface dominates
Each additional system in a workflow adds authentication, a data model to map, rate limits to respect, and a failure mode to handle. The cost does not rise linearly, because systems interact: when system C is unavailable, you must decide whether the work already written to A and B should be rolled back, held, or completed anyway.
This is why a workflow entirely inside one modern platform can be a small piece of work while a superficially identical process spanning four systems is many times the size. When you are scoping, count the systems before you consider anything else.
The exceptions are the project
Every process has a happy path that is quick to build and a set of exceptions that are not. The eighty per cent of invoices that arrive in a familiar format are straightforward. The twenty per cent that arrive as photographs, in a foreign currency, or referencing a purchase order that does not exist are where the effort concentrates.
Clients are often surprised that we spend the first week asking about edge cases rather than building. That week is the cheapest place to discover that a fifth of the volume needs different handling — far cheaper than discovering it in testing, and vastly cheaper than discovering it in production.
Where the budget goes
The proportions below are typical for a single-workflow build. They shift with the specifics, but the ordering is remarkably stable across projects, and the surprise for most people is how small the AI line is.
| Cost area | Share of effort | What it covers |
|---|---|---|
| Discovery and process mapping | Roughly a fifth | Interviews, watching real work, documenting exceptions and rules |
| Integration | Roughly a third | Authentication, data mapping, rate limits, retry and rollback behaviour |
| Core logic and AI steps | Roughly a sixth | The automation itself, including any extraction or classification |
| Exception handling and testing | Roughly a quarter | Edge cases, error queues, alerting, testing against real data |
| Documentation and handover | Small but non-zero | Runbook, working session, monitoring left in a usable state |
What does it cost to keep running?
Three ongoing lines: the platform the automation runs on, model API usage where AI steps exist, and maintenance. For a single workflow the first two are usually modest monthly amounts, and usage-based model pricing means volume is predictable once you have measured a month of it.
Maintenance is the line people omit. Upstream systems change their formats, credentials expire, and vendors deprecate endpoints. Budgeting a few hours a quarter per workflow is realistic; budgeting zero is how automations quietly stop working and nobody notices until a customer complains.
Comparing against the alternatives
The relevant comparison is not build versus nothing. It is build versus hiring, versus buying a product, versus continuing to absorb the cost manually. Continuing manually has a real price that simply never appears on an invoice, which is exactly why it persists for years.
If you have a continuous stream of automation work, hiring is genuinely cheaper over time and we will tell you so. If an off-the-shelf product already solves your problem, buying it beats building. Our engagement model explains how we quote per stage so those comparisons can be made with a real number rather than a guess, and the companion guide on choosing what to automate first covers the payback arithmetic. Whoever you ask for that number, insist the build cost and the running cost arrive as two separate figures — one of the commercial checks in how to choose an AI automation consultant. For due diligence on vendor claims, the FTC's guidance on AI claims is a useful lens on any proposal you receive, including ours.
Common questions about automation budgets
Why will nobody quote a price without a conversation?
Because the same described task can differ in cost by an order of magnitude depending on the systems behind it. Automating an approval step inside one modern platform is a small job; the same approval spanning four systems, one of which has no API, is a different project entirely. A published price would be wrong for most readers in one direction or the other.
Is it cheaper to have our own developer build it?
For an ongoing programme, usually yes, and we will say so. In-house makes sense when there is a continuous stream of automation work to justify the salary. For two or three projects, the hiring cycle alone typically costs more than the builds, and the resulting system has a single point of failure when that person leaves.
What are the ongoing costs after a build is finished?
Three things: platform subscriptions for whatever the automation runs on, model API usage where AI steps are involved, and maintenance when an upstream system changes. For a typical single workflow the first two are usually modest monthly amounts. Maintenance is the one people forget, and it is best budgeted as a few hours a quarter.
How do we avoid paying for something that never gets used?
Agree the measure of success before the build starts, in numbers you already track, and stage the work so something reaches production early. A project that ships a narrow slice in week three gets used or exposed as pointless quickly. One that ships everything in month six is where unused systems come from.
Free consultation
Tell us what is costing you the most time
Describe the process. You get a straight read on whether it is worth automating, roughly what it would take, and what we would tackle first — on a call that costs nothing.
- Under an hour, free, and no obligation
- Fixed price per stage, agreed before any work starts
- You own everything we build, documented and in your accounts
- If you do not need us, we will say so on that call
The form is the fastest way to reach us — we reply from a real person, usually within one business day.