Short answer: the cost of an AI agent depends mostly on scope, not on the AI itself — how many systems it connects to, how messy and varied the input is, how accurate it needs to be and how much human review is built in. There are two separate costs: the one-off build, and the ongoing running costs for AI usage and platforms. The only reliable way to know the price is a written scope based on your actual process.
"How much does an AI agent cost?" is one of the first questions business owners ask, and it's a fair one. The frustrating answer is that it varies widely — the same phrase can describe an afternoon's configuration of an existing tool or a multi-month project wired into several business systems.
Rather than quote numbers that wouldn't apply to your situation, this guide explains what actually drives the cost, so you can understand any quote you receive, compare agencies fairly and keep the project small enough to succeed.
The two costs: building and running
Every AI agent has two kinds of cost, and it's important to see both before you start.
| Build cost | Running cost | |
|---|---|---|
| What it is | Designing, building, testing and launching the agent | AI model usage, automation platforms, hosting and storage |
| When you pay | Once, or in project phases | Every month, based on usage |
| What drives it | Scope, integrations, testing and review needs | Volume of tasks and amount of text processed |
| Who you pay | The agency or developer | The AI provider and platforms — ideally directly, from your own accounts |
Running costs are often small per task, but they scale with volume. Make sure any proposal includes an estimate of monthly running costs at your expected volume, and that those services are billed to accounts in your name so you can see exactly what you're paying for.
What makes the build bigger or smaller?
1. How many systems it connects to
An agent that reads one inbox and writes to one spreadsheet is far simpler than one that reads email, checks a CRM, looks up stock, updates an accounting system and notifies a team chat. Each connection needs authentication, error handling and testing. Systems with good, documented APIs are quicker to connect; older software without them takes more work. Our integrations service covers this part.
2. How varied and messy the input is
If every incoming document follows the same template, extraction is straightforward. If documents arrive as phone photos, scans, PDFs and spreadsheets from dozens of different senders, the agent needs more testing and more handling of edge cases. Variety is one of the biggest cost drivers, and it's the main reason to test on real examples early.
3. How accurate it needs to be
An agent that sorts newsletters from enquiries can afford the odd mistake. One that extracts amounts for invoices cannot. Higher accuracy requirements mean more testing, validation checks (for example, confirming totals add up), confidence thresholds and review steps — all of which add work, and all of which are worth it when mistakes are costly.
4. How much human review is built in
Most good agents include a review step: a queue of drafts to approve, a list of flagged items or a daily summary to check. Building a clear, efficient review screen takes time, but it's what makes an agent safe to rely on. A design with no human review might look cheaper, but it pushes the risk onto you.
5. What actions it can take
Reading and drafting are lower risk than acting. An agent that updates records, sends messages or changes bookings needs tighter permissions, logging and testing of every action. The more it can do on its own, the more care the build requires.
6. Where your knowledge lives
If an agent needs to answer from your policies and documents, those documents need to be in one place, current and consistent. If they're scattered across drives, inboxes and people's heads, part of the project becomes organising the knowledge. That work is valuable in its own right — but it should be in the scope.
7. Data protection and security requirements
Agents that handle personal data need to be designed with data protection in mind: limited access, appropriate retention, logs, and a clear record of which providers process the data and where. If you operate under the UK GDPR or India's Digital Personal Data Protection Act 2023, these requirements are part of the job, not optional extras.
Hidden costs to ask about
- Platform fees for automation tools, vector databases or hosting the agent depends on.
- AI usage markups if usage is billed through the agency's account rather than yours.
- Maintenance when connected apps change their APIs or your process changes.
- Monitoring — someone needs to check flagged items and notice if results drift.
- Lock-in — if the agent lives on the agency's platform, what does it cost to leave or move it?
- Your team's time for providing examples, testing and reviewing during the build.
None of these are unreasonable, but they should be visible before you sign, not discovered afterwards.
How to keep the cost under control
- Start with one task. Pick the single most repetitive reading-and-writing job, not your whole operation.
- Test on real examples first. A short prototype on 20–50 real cases tells you whether the idea works before you pay for a full build.
- Use automation for the predictable steps. AI should only handle the steps that need judgement; ordinary workflow automation is cheaper for everything else.
- Keep a person in the loop at first. You can always reduce review later once you trust the results.
- Own the accounts. Running costs stay transparent, and you avoid paying to leave later.
- Get a fixed scope in writing. It's the best protection against costs drifting.
Why fixed quotes after a discovery call make sense
Any number given before someone understands your process is a guess. Two businesses asking for "an AI agent for our inbox" might need very different things — one has a single, tidy inbox; the other has five shared mailboxes, three languages and attachments that need extracting.
That's why we don't publish prices. After a discovery call, we put together a written scope covering what will be built, what won't, how it will be tested and what running costs to expect, with a fixed, one-time price for the build. The finished agent runs on accounts you own, and AI usage is billed by the provider directly to you. Any ongoing support is optional and agreed separately.
Questions to ask before you accept any quote
- What exactly is included, and what isn't?
- What are the expected monthly running costs at our volume, and whose account are they billed to?
- How will you test it, and on what data?
- What happens when the agent is wrong?
- Who owns the code, accounts and data at the end?
- What does ongoing support cost, and is it optional?
Our guide to choosing an AI automation agency covers these questions in more detail, along with red flags to watch for.
Next step
If you have a task in mind, the quickest way to understand the cost is to talk it through with real examples in front of you. Book a discovery call and you'll get an honest view of whether an agent is worth building — and, if it is, a fixed quote before any work starts. You can also read more about how we approach AI agents.
Frequently asked questions
Why do AI agent quotes vary so much?
Because "AI agent" covers very different projects. The number of connected systems, how messy the input is, the accuracy required and how much human review is needed all change the amount of work significantly. Compare quotes on scope, not just on the final number.
What are the ongoing costs of running an AI agent?
Mainly AI model usage, which depends on how many tasks the agent handles and how much text it processes, plus any automation platform or hosting fees. These should be estimated upfront and billed to your own accounts so you can monitor them.
Is it cheaper to use an off-the-shelf AI tool?
Sometimes. If an existing tool does exactly what you need, it's often the cheapest option, and a good agency should tell you so. A custom agent makes sense when your process, data or systems don't fit what off-the-shelf tools offer.
Do you publish your prices?
No. Every project is different, so we quote after a discovery call. You get a written scope and a fixed, one-time price for the build before any work starts, plus an estimate of running costs.
Can I start small and expand later?
Yes, and it's usually the best approach. A first agent focused on one task lets you prove the value and learn what works before adding more tasks or connections.
Have a process like this?
Book a discovery call