What an AI Agent Build Actually Includes
Our custom AI agent pilot costs USD $25,000 to $60,000 for one workflow over 30 to 90 days. The build includes testing on your data, written acceptance criteria and a system that checks its work. You set its operating limits and own
A build proposal should let you picture the handover: the job the agent will do, the systems it can change, and the evidence that it works. A list of integrations leaves those questions open. Here is what our first build includes, with an example you can use to check
What is in the first build
The first build covers one workflow with identifiable inputs, permitted actions and an end state. Choose a recurring task whose current workload and errors you
You leave with these,
- The workflow, named. Intake, reporting, documents, research, or an internal knowledge job. We pick one. The ones you should not automate yet stay on a list so they do not sneak into scope.
- A success metric and the person who measures it. Define the eligible cases, comparison period and baseline before testing. For time saved, subtract review and correction work. Count serious errors separately so a faster average cannot conceal an unacceptable action.
- A kill threshold. The point at which we tell you to switch it off, agreed before any build. Gartner’s June 2025 forecast expected over 40% of agentic AI projects to be canceled by the end of 2027. It named escalating costs, unclear business value and inadequate risk controls. That is a forecast, not a measured failure rate. We address those risks when scoping the pilot.
- Shadow mode on your data. The agent proposes. It does not send, charge, or delete. You watch the disagreements. Accuracy is measured before time saved.
- A tested operating boundary. The system acts inside the limits you set, checks its work, and sends exceptions to a person. The pilot establishes which actions it can complete autonomously and which require approval. Runs are logged and readable by you. Our own agents working in client accounts follow the separate approval policy. In our pilot, money movements, client messages and system writes require human approval.
- Ownership and a usable handover. The contract records your ownership and export rights for the commissioned code, prompts, configuration and evaluation set. It also identifies third-party components, their licence terms and any continuing charges. Owning the deliverables does not remove those dependencies.
- The security answers in week zero. Data flow, subprocessors, a data processing agreement on request. We do not hold SOC 2 and we say so. Your reviewer should not discover that in week six.
- A decision at the end of the pilot. Expand, extend once with a specific fix, or stop. Workflows two and three are priced after the first one proves out.
Our AI systems catalogue describes the workflows we offer. For a custom build, see AI agent development.
Build fees and ongoing costs
Our pilot is $25,000 to $60,000 at a fixed scope, over 30 to 90 days. If you need help choosing and scoping the workflow, the two-week AI Opportunity Audit costs $7,500, credited toward a build if you proceed. Teams that already know what to build can go directly to a pilot. All prices here are in USD. Our optional production retainer starts at $8,000 a month, month to month after the first three months, as listed on our pricing page.
At the starting retainer, twelve months of operation is $96,000 before any separately quoted usage or third-party charges. The audit and build are separate from that operating period, with the audit credit applied if you proceed. Fixed-scope work has no retainer attached. If you choose ongoing operation, the proposal needs to specify the usage allowance, maintenance work and additional charges; a percentage of build cost cannot substitute for
For an internal, reversible workflow, a smaller automation project may be enough. Compare its handover, support and testing scope with the pilot above before deciding whether the additional work is useful
What acceptance looks like for an invoice workflow
Consider an illustrative agent that prepares supplier invoices for an accounts-payable system. The agreed job is to extract the details, check them against the purchase order, and create a draft record. In our pilot, an operator approves submission before the agent writes that draft to the accounts-payable system. Payment remains outside
| Case in the evaluation set | Required result | Evidence at handover |
|---|---|---|
| Invoice matches the purchase order | Create one draft with the correct supplier, currency, tax and total | Each field links to its source; the draft matches the approved answer |
| The same invoice arrives twice | Identify the duplicate without creating another record | Both source documents and the existing record are linked |
| Bank details have changed | Route the case for review without updating payment details | The exception shows the changed field and the previous value |
| A total is unreadable or conflicts with the order | Leave the case unresolved for a person | The agent identifies the uncertainty instead of inventing a value |
| The accounting API times out after submission | Check whether a record was created before retrying | The recovery leaves one record, with a readable action history |
The handover includes the agreed cases and their observed results. Acceptance depends on the error types you can tolerate and the work a person still has to do. A convincing demo on one clean invoice does not answer
What is not in the box
The pilot covers the agreed workflow and its permissions. A chatbot, another workflow or a new external action needs its own scope. An outside assistant being able to find your business does not give it authority to trigger a payment, refund or bulk export. Should your business be agent-callable explains how to separate public discovery from actions that
For ongoing monitoring, maintenance and scoped expansion, see AI workflow automation. The proposal should say which maintenance and expansion work is included. If you have not decided what to build, start at AI consulting. The audit is available as a
Check the handover before you sign
We run AI systems in our own business. Our internal automation case study estimates more than 50 hours reclaimed each week across the team. That estimate describes our team’s work. Your potential savings depend on the workflow and workload you
Before signing, check that the proposal names the acceptance cases, operating permissions, exception owner and support terms. Establish how you will export the code, prompts, configuration and evaluation set if you move the work elsewhere. Our 24 questions to ask an AI agent vendor give you a
If you need help choosing and scoping the workflow, the paid next step is the AI Opportunity Audit. Qualifying teams can also book a free consultation to discuss whether that audit fits