How AI agent pricing works

The same request costs wildly different amounts at two companies. "Automate our invoice approvals" is a 4 week build at a company running one ERP with a written approval matrix, and a 4 month build at a company running two ERPs from an acquisition, where approvals are a convention rather than a rule and three people hold the exceptions in their heads.

Nothing about that difference is visible from the outside. It becomes visible during the audit, which is why the audit exists as a standalone piece of work you can buy and act on independently.

Roughly in order of impact.

How many systems the agent has to touch

One system is a build. Four systems is an integration project with a build inside it. Each connection adds authentication, error handling, and a class of failure that only appears in production. This is the single largest driver, and it is visible in how the three live systems were scoped.

Whether your policy is written down

If the rule for approving an exception exists as text, we implement it. If it exists as precedent in someone's judgment, someone has to write it first, and that work sits on the critical path. Companies that arrive with written policy routinely save weeks.

The cost of being wrong

An agent drafting internal summaries needs light review. An agent moving money, touching a patient record, or replying to your largest customer needs approval checkpoints, an audit trail, and a longer supervised period. Risk is bought in review time.

The state of your data

Agents read what exists. Duplicate customer records, three formats for the same field, a CRM where half the pipeline is stale. Cleanup is sometimes the highest-return part of an engagement and it is work, whoever does it.

How many people have to agree

A workflow owned by one operations lead moves at a different speed than one touching finance, legal and IT. This is not a technical cost and it is frequently the real timeline.

Whether we operate it afterwards

Running an agent in production is ongoing work: model changes, monitoring, edge cases, the monthly review of where approval boundaries sit. Priced separately and stated up front, so you can also choose to take it in-house. Stage 4 on the approach page covers what the work involves.

Each stage is priced and can be bought on its own. Most clients start with the audit and decide after reading it.

1. 01

   Audit, about 3 weeks

   A structured review of how work moves through your organization, which systems are involved, where the handoffs are, and which workflows are worth automating. You get a ranked list with expected return on each, and a note on what we would skip. Yours to keep whether you build with us or not.

2. 02

   Build, 6 to 12 weeks

   The agent, the integrations with your existing tools, the human approval checkpoints, and the monitoring. A working version every week. Scoped from the audit, so the number is based on what we found rather than what we guessed.

3. 03

   Deployment, 2 to 4 weeks

   Shadow mode, then supervised operation, then autonomy granted per workflow against a bar your team sets. Your team is part of this stage rather than the audience for it.

4. 04

   Operating, ongoing

   Monitoring, model and workflow changes, edge cases, and the standing review of approval boundaries. Most clients expand into adjacent workflows from here, which is a separate scope.

Four things that genuinely reduce cost, all of which you can do before talking to anyone.

Write down one policy

Pick the workflow you most want automated and write the rule for its exceptions. An afternoon of work that reliably removes days from a build.

Start with one system

A first deployment inside a single tool ships faster, proves the pattern, and gives you a real internal reference before you scope something that spans four systems.

Name one decision-maker

One person who can approve a policy call within a day. Committee approval is the most common reason a 6 week build becomes a 10 week build.

Bring your ticket or transaction history

Six months of resolved tickets, closed invoices or won deals is the cheapest possible input to an accurate scope. Process documents describe the intended path. The history shows the exceptions, and the exceptions are what drive the build.