How Much Does AI Automation Cost? (2026 Guide)

The three pricing models, and why they confuse people

Per-workflow pricing charges for a discrete automation: one process, in, tested, working. Project pricing bundles several workflows into a single scoped engagement. Retainers charge monthly for ongoing build and iteration rather than a fixed deliverable.

They are not interchangeable, and the right one depends on how settled your requirements are. If you know exactly which process you want automated, per-workflow is cleanest and easiest to hold someone to. If you expect the scope to move as you learn, a retainer is more honest than a fixed price that will be renegotiated anyway. Comparing a project quote to a monthly rate without multiplying out a year is where most confusion starts.

• Discovery or audit engagement — $0–$5,000, sometimes credited against the build
• Single scoped workflow — $5,000–$15,000 typical, up to $25,000 for complex ones
• Multi-workflow build — $15,000–$50,000
• Full operations programme — $50,000–$150,000+
• Ongoing retainer — $2,000–$15,000 a month depending on scope
• Hourly consulting — $150–$400 an hour

What actually moves the price

The number of workflows matters less than people expect. What moves the cost is how many systems each workflow has to touch, how messy the data is when it arrives, and how bad it would be to get an answer wrong.

That last point is the one buyers underestimate. An automation that drafts an internal summary can be wrong occasionally at little cost. An automation that answers customers, moves money or writes to your CRM needs evaluation, guardrails and a human review path — and that engineering is most of the difference between a cheap build and a sound one.

• Number of systems touched, and whether their APIs are cooperative
• Data quality on the way in — messy inputs are the hidden cost
• Consequence of a wrong answer, which determines how much evaluation is needed
• Whether a human stays in the loop, and where the handover sits
• Volume, which drives both infrastructure and model spend

The running costs nobody quotes

Automation is not a one-off purchase. Model usage is the cost people are least prepared for, because it behaves unlike software licensing: it scales with how much the thing is actually used. A successful automation costs more to run than a neglected one, which is the correct incentive but still a surprise on the first invoice.

For a support copilot at small-to-mid scale, expect roughly $200–$1,500 a month in model costs, and design decisions — caching, routing simple requests to smaller models, tiering — make a substantial difference to that figure. Add hosting, monitoring, and the maintenance of integrations that change underneath you when a vendor updates their API.

• Model usage — scales with adoption, roughly $200–$1,500/month at small-to-mid scale
• Hosting and infrastructure — modest, but not zero
• Monitoring and evaluation — how you know it is still behaving
• Integration maintenance — third-party APIs change without asking you
• Iteration — the first version is never the final one

Which process to automate first

The practical entry point is a single scoped workflow: pick the one expensive, repetitive process costing real money right now, automate it properly, measure the result, then decide whether to expand. This is unglamorous advice and it is right, because it converts an abstract question about AI into a specific one about a number you already track.

Good first candidates share a shape: high volume, repetitive, currently done by someone whose time is worth more elsewhere, with a clearly correct answer most of the time. Poor first candidates are rare, judgement-heavy, or so consequential that they need a level of oversight that eliminates the saving.

The automation worth more than any of these

Before quoting anything, it is worth asking whether the process should exist. A surprising number of workflows that look like automation candidates are compensating for something upstream — a form that collects the wrong fields, a handover that should not exist, a report nobody reads.

Automating those makes a bad process faster and harder to remove. The cheapest possible automation is the one you avoid building by fixing or deleting the thing it was going to automate, and any agency worth paying will tell you when that is the answer, even though it is the answer that earns them least.