Ask three agencies to quote an AI agent and you will get three numbers that are nowhere near each other. That is not because two of them are lying. It is because "AI agent" describes anything from a chatbot that answers FAQs to a system that makes decisions and writes to your production database.

This post explains what actually drives the cost, so you can read a quote instead of guessing at it.

The four things that move the price

Everything else is detail. These four decide the band.

1. Does the agent read, or does it write?

This is the single biggest cost multiplier, and most buyers do not think about it until they see the quote.

A read-only agent looks things up and answers. If it is wrong, someone reads a wrong answer. Annoying, recoverable.

A write-access agent creates records, sends emails, moves money, updates inventory. If it is wrong, it does damage that has to be undone.

Write access does not just add a feature. It adds permission scoping, approval steps, audit logging, rollback paths, and a great deal more testing. The same agent with write access instead of read access is routinely two to three times the build.

2. How many systems does it touch?

Each integration is its own small project: authentication, rate limits, error handling, and the field mapping nobody documented.

One system is cheap. Three is more than three times one, because now the agent has to keep state consistent across all of them. Modern, well-documented APIs are fast. A legacy on-premise system with no API and a database you are told not to touch directly is where budgets die.

Count your integrations before you ask for a quote. It is the number that will change the answer most.

3. Does a human approve, or does it act alone?

Full autonomy sounds cheaper. It is not.

An agent with human approval in the loop can be shipped once it is *mostly* right, because a person catches the rest. An agent acting alone has to be right on its own, which means far more edge-case handling, far more testing, and a monitoring setup that tells you fast when it is not.

Human-in-the-loop is usually the cheaper build and the safer one. It is what we default to for anything that leaves the company — outbound email, published content, payments.

4. How wrong can it afford to be?

An agent drafting internal meeting notes can be wrong sometimes. An agent classifying insurance claims cannot.

Accuracy is not a setting. Getting from "usually right" to "reliably right" means evaluation datasets, a separate checking layer, and iteration cycles. That work is often larger than building the agent itself, and it is the line item most quotes hide inside a vague "testing" row.

What you can build cheaply

Some things genuinely are fast now, and you should not be quoted a fortune for them.

A single-purpose read-only agent over documents you already have. Support answers, internal policy lookup, document search. Small integration surface, no write access, low blast radius.

An agent that drafts, and a human sends. Email drafts, first-pass proposals, meeting summaries. The approval step removes most of the risk work.

A workflow that was already a workflow. If your process is already written down and consistently followed, automating it is much cheaper than automating something that lives in three people's heads and changes depending on who is on shift.

If someone quotes you a large number for one of these, ask them to break it down.

What genuinely costs money

Anything that writes to a system of record. See above.

Multi-agent coordination. Several agents that hand work to each other need a routing layer, shared state and a checking layer. That infrastructure is the thing most companies skip and then pay for later.

Legacy integration. No API, no documentation, no test environment. Budget generously and expect surprises.

Regulated data. Healthcare, finance, anything with residency rules. Self-hosting, audit trails and access controls are not optional extras, and they change the architecture rather than sitting on top of it.

Accuracy targets above "good enough." Every additional nine of reliability costs disproportionately more than the last.

The costs that appear in month three

Almost every AI agent quote covers the build and stops. These are the ones that surprise people.

Model API usage. This scales with how much the agent runs and how much context it processes each time. It is small in testing and real in production. Ask for a per-run cost estimate before you sign, not after.

The orchestration platform. Whatever runs your workflows — n8n, Zapier or similar — bills monthly and the pricing model matters more than the headline rate.

Hosting and monitoring. Servers, logging, error alerting. Modest, but not zero.

Maintenance. Models get deprecated. APIs change. Your own process changes. An agent is a system, not a deliverable — budget for someone to own it.

Ongoing evaluation. Quality drifts quietly. Without periodic checking, you find out from a customer.

A reasonable rule: assume ongoing costs of somewhere between 15% and 25% of the build cost annually, and ask any agency to state their number rather than accepting silence.

Build or buy?

Before commissioning anything, check whether an off-the-shelf product already does it.

Buy when your process is standard — support ticket triage, meeting notes, generic sales outreach. The vendor has solved it across thousands of customers and you will not beat that on price.

Build when the process is your actual differentiator, when the tool has to fit systems no vendor integrates with, when data cannot leave your infrastructure, or when per-seat pricing at your headcount exceeds the cost of owning it.

A useful middle path: buy the commodity parts and build only the layer that connects them to your business. That is frequently the cheapest real answer, and any agency worth hiring will tell you when it applies.

Five questions to ask any agency quoting you

1. Does this agent write to any system? Which ones, and what happens if it writes something wrong? 2. What is the estimated monthly running cost at our expected volume? 3. What is the approval step, and who owns it? 4. How will we know if accuracy drops after launch? 5. What happens to this system if you and we stop working together? An agency that answers all five clearly is quoting a system. One that cannot is quoting a demo.

Getting a number for your own project

The honest answer to "what does an AI agent cost" is that it depends on the four factors above — but you should not have to sit through a sales call to get a range.

ZAC Estimator gives you a cost and timeline band in about two minutes, with the assumptions listed so you can see what is driving the number and argue with it. If you are not sure yet what you are even building, describe the problem to ZAC Consultant and it will map the solution and phasing first.

Both are free, and neither requires a call.

More on how we deliver: AI automation, custom software, and systems we have shipped.