The cost of a custom AI agent depends on the workflow it supports, not on the label “AI agent” or the name of the model behind it. An agent that answers questions from a handful of approved documents and an agent that reads your CRM, drafts customer replies, and updates records under approval rules are both called “AI agents.” They are very different projects.
That is why an honest answer to “How much does it cost?” starts with a few questions of its own: Which workflow? Which systems? What may the agent see, and what may it do?
This guide explains what actually drives the cost of a custom AI agent, how to compare estimates from different vendors, and how Madency gets from a first conversation to a fixed price.
Why there is no honest one-line price
“AI agent” describes a category, not a scope. It is closer to “website” or “mobile app” than to a product with a price tag. A quote given before anyone has looked at the workflow is either padded to cover unknowns or likely to grow once the real requirements appear.
The model is rarely the main factor. Model usage is usually a running cost measured per request. Most of the implementation effort typically goes elsewhere:
- connecting the agent to the systems where your data actually lives;
- controlling what each user and the agent itself may see;
- deciding which actions need human approval;
- testing that the agent behaves correctly on real cases, including the awkward ones;
- deploying, monitoring, and maintaining it after launch.
Those are properties of your workflow and your organization. They are what an estimate has to describe.
A simple assistant versus a production workflow
The biggest difference in scope is usually between an assistant that answers questions and an agent that works inside a business process.
| Simple assistant | Production workflow agent | |
|---|---|---|
| Data | A small set of approved documents | Live data from a CRM, helpdesk, database, or shared drives |
| Actions | Answers questions | Drafts, classifies, updates records, or triggers next steps |
| Users | One team with the same access | Several roles with different permissions |
| Mistakes | A person notices and ignores a bad answer | Approval steps, logging, escalation, and safe failure |
| Testing | Spot checks | A test set of real cases and repeatable evaluation |
| Deployment | A hosted tool | Your infrastructure or cloud account, with monitoring |
| After launch | Nobody is responsible | An owner, a review rhythm, and a way to improve quality |
Neither column is “better.” A simple assistant can be the right first step. The point is that the scope, and therefore the cost, comes from the right-hand column you actually need.
The eight factors that drive the cost
1. How narrow or broad the workflow is
A workflow with a clear trigger, a defined set of steps, and a known output is far easier to estimate and build than “help the sales team with everything.” Narrow scope is not a compromise. It is usually the fastest way to get something reliable into production and learn from it.
2. The number and quality of systems involved
Each system the agent reads from or writes to adds integration work: authentication, data mapping, rate limits, error handling, and testing. A well-documented API is one thing. An old internal database with no documentation and unclear ownership is another.
3. Data freshness, structure, ownership, and access
“The information is in the CRM” is a starting point, not an answer. Which fields are reliable? How current are they? Who owns them? Can the agent access them without exporting data by hand? Messy or contradictory data is one of the most common reasons scope grows after a project starts.
4. Permissions and sensitive-data boundaries
If every user may see everything, access control is simple. If salary data, client contracts, pricing, or personal data must stay visible only to certain roles, the agent has to respect those boundaries in every answer and every action. That requirement must be designed and tested, not added at the end.
5. Whether the agent answers, drafts, or executes actions
Answering a question is the lowest-risk action. Drafting a reply for a person to send is a step up. Updating a record, sending a message, or triggering a payment workflow is a different category: every action needs rules, safeguards, and a way to recover from a mistake.
6. Human approval and escalation requirements
Approval steps keep risk under control, but they are features in their own right. Someone has to receive the request, see enough context to decide, approve or reject it, and leave a record. Escalation paths for cases the agent cannot handle need the same design.
7. Evaluation, testing, observability, and deployment
A production agent needs a set of real test cases, a repeatable way to check quality after each change, logs that show what it did and why, and a deployment that fits your security requirements. Running on your own infrastructure gives you control over data, and it also has to be set up and maintained.
8. Handoff versus ongoing operations
After launch, data changes, processes change, and users find cases nobody anticipated. Some companies prefer a full handoff to their own team. Others want ongoing monitoring, improvements, and incident response. Both are valid, but they are different commitments, and an estimate should say which one it assumes.
Why data access and permissions change the estimate most often
In practice, the first version of a scope often underestimates factors 3 and 4. The workflow looks simple in a meeting: “The agent checks the client’s status and drafts an answer.” Then the details appear:
- the client’s status is split between the CRM and a spreadsheet maintained by one manager;
- some clients have special terms that only account managers may see;
- the CRM fields are filled in inconsistently;
- nobody has decided which source wins when they disagree.
None of this is unusual. It simply has to be discovered before the price is fixed, not halfway through the build.
Illustrative example: one name, two scopes
Consider a hypothetical 40-person professional services firm that asks for “an AI agent that answers client questions.” This example is illustrative. It does not describe a real client or a real price.
Version A answers internal staff questions using approved service descriptions, FAQs, and policy documents. It does not access client records and does not take actions. Staff check the answers before using them.
Version B reads the client’s project status from a project tracker, checks the account in the CRM, drafts a reply for the account manager, and creates a follow-up task. Billing details are hidden from some roles. Any reply that mentions a deadline change requires a manager’s approval.
Both are “an AI agent for client questions.” Version B involves more systems, live data, role-based permissions, actions, an approval step, and a larger test set. It is not necessarily the wrong choice, but it is a different project. A useful estimate starts by deciding which version, or which first step, the business actually needs.
Running costs are separate from the build
When you compare offers, separate one-time implementation cost from ongoing running cost. Running costs usually include:
- model usage, which depends on volume and on the model chosen;
- hosting and infrastructure;
- monitoring and maintenance;
- support or ongoing improvement, if you choose it.
A low build quote can hide high running costs, and the reverse is also possible. Ask every vendor to show both.
What Discovery resolves before the build
Discovery exists to turn “we want an AI agent” into a scope that can be priced responsibly. At Madency, Discovery is a paid $500 engagement that typically takes 5 to 10 business days. It covers:
- mapping the workflow as it really runs today;
- identifying the systems, data sources, owners, and constraints;
- defining what the agent may read, draft, or execute, and where approval is required;
- recommending an architecture;
- producing an implementation roadmap with an estimated scope, timeline, and investment.
Discovery can also conclude that a workflow is not a good candidate for an AI agent, or that a simpler automation or process change is the better answer. That is a useful result too.
If you sign the build agreement within 60 days of the final Discovery report, 100% of the Discovery fee is credited once toward the Build fee.
How Madency structures the engagement
The full process is described on How We Work. In short:
- Fit call. A free 30-minute conversation to check whether your workflow is a reasonable candidate for an AI agent.
- Discovery — $500. Workflow mapping, data and constraint review, architecture recommendation, and an estimated scope, timeline, and investment.
- Build — fixed scope, fixed price. A written agreement defines the workflow, integrations, data sources, roles, deliverables, acceptance criteria, timeline, and price before the work begins. Complexity ranges from a single integration to a custom architecture with several integrations, and the price follows that scope.
- After launch. Either a full handoff with the code, documentation, and a 60-day warranty, or ongoing AI Operations for monitoring, improvements, incident response, and monthly reporting.
Madency does not price AI work by the hour. You know the scope and the price before the build starts. You can see examples of the workflows we build on the AI Agents page.
Questions to ask any vendor about AI agent cost
Whether you work with Madency or not, these questions make estimates comparable:
- Which workflow is in scope, and what is explicitly out of scope?
- Which systems will the agent read from and write to?
- What may the agent do without human approval?
- How will quality be tested before launch and after each change?
- What will the monthly running costs be, and what do they depend on?
- Who owns the code, the data, and the infrastructure?
- What happens after launch, and who is responsible for the agent then?
If a vendor cannot answer these yet, the estimate is not finished.
What to prepare for a useful estimate
You do not need a technical specification. One recurring workflow and honest answers to a few questions are enough to start:
- what triggers the workflow and what result it should produce;
- who performs it, who reviews it, and who depends on it;
- which tools and data sources are involved;
- where time, delays, rework, or errors appear today;
- what the agent may answer, draft, or execute;
- how you would know the workflow improved.
The guide on what to prepare before an AI Discovery session covers this in more detail. If you are still choosing between several workflows, start with the first AI workflow scorecard.
Start with one workflow, not a price list
The cost of a custom AI agent becomes clear once the workflow is clear. Choose one recurring process, describe it as it really works today, and decide what the agent should and should not do.
Bring one recurring workflow to a fit call. If it is a suitable candidate, Discovery will define the scope and investment.