What a custom AI agent costs in 2026

A custom AI agent for one office job costs about $2,000 to $70,000 to build in 2026, going by the prices agencies publish on their own sites, plus $500 to $2,500 a month to keep it running at their starting rates. Model usage is extra: billed by the AI provider, or sold as credits on a platform. With us: a paid assessment from $1,000, credited to the build, a build from $15,000, and maintenance of $500, $1,000 or $2,000 a month.

AI agent prices at a glance

We read each company's own pricing page or published price guide on November 9, 2026. Their prices cover different kinds of work, so read the "what it is" column before comparing numbers.

Lindy sells usage as credits. Frenchy Digital and Armada Works both say the client pays model usage to the providers directly, and so do we. Dappasol's page doesn't say how it bills usage.

Cost by type of agent

Do it yourself on a platform: about $30 a month and your time

A platform like Lindy suits a simple job one person can set up and watch: sorting an inbox, drafting replies, logging items into a sheet. You pay for credits, not for a build. The cost is the hours someone on your team spends setting it up, testing it and fixing it when a supplier changes a form.

A first automation or starter agent: about $2,000 to $10,000

Dappasol starts installs at $2,000 and Armada Works builds one fixed-scope agent in a week for $2,500 to $4,000. Frenchy Digital's packaged agent is $10,000 at checkout. These are fixed-scope jobs on tools you already use, and they are a fair way to find out whether an agent helps at all.

A custom agent for one back-office job: about $15,000 to $70,000

This is where an agent reads your own documents (invoices, tickets, bid packages), checks them against your own records and hands a person a list to decide. Frenchy Digital prices bespoke single-workflow agents at $28,000 to $70,000. Ours start at $15,000. Most of the cost is reading each supplier's or customer's document layout correctly and proving it on your real documents before it goes live.

Several workflows and systems: $25,000 and up

When the work spans several workflows and systems, the price climbs fast. Frenchy Digital prices a multi-workflow platform with system integration at $70,000 to $180,000. We price internal software, or several agents working across your systems, from $25,000.

What drives the cost

• How many document layouts it has to read. Each supplier's invoice or each customer's form is its own layout. The assessment counts how much of your volume sits with the top few senders.
• Volume and how clean the data is. A thousand typed PDF invoices a month is a different job from scanned paper with handwriting.
• Whether it reads or writes. An agent that reads documents and hands a person a list is cheaper and safer than one that writes into your accounting system. Ours read and report; a person decides.
• How it's tested. Ours runs 10 times on a held-out sample of your own documents before go-live, against a bar written into the scope. That testing is part of the build price.
• What happens after launch. Suppliers change layouts and prices. Someone has to notice and fix it, which is what the monthly fee pays for.

What you pay every month

Two things: maintenance, and model usage.

Maintenance. Ours has three tiers, month to month:

Model usage. Where an agent uses an AI model, the provider bills your own account at its list price. We don't mark it up. For scale: Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens. If an agent read 1,000 invoices a month at about 3,000 tokens in and 500 out each (our estimate for a typical PDF invoice), that's 3 million input tokens ($6) and 500,000 output tokens ($5): about $11 a month. On Claude Haiku 5.5 ($0.10 and $0.50 per million) the same work is under $1. Usage depends on volume and channel: Frenchy Digital notes that a voice receptionist and an internal assistant use very different amounts, and Armada Works tells its clients to expect about $200 to $500 a month of model usage, for engagements that run fleets of 4 to 6 agents.

Our invoice check agent runs plain rules first and only sends lines the rules can't match to a model, so its usage is smaller still.

Year one, worked out

From each company's published starting prices. Usage is left out because it depends on your volume.

The $1,000 assessment is credited to our build, so it isn't added twice.

When a custom agent isn't worth it

Work out the time first. The Bureau of Labor Statistics puts the median pay of bookkeeping, accounting and auditing clerks at $24.36 an hour ($50,670 a year, May 2025), before benefits and payroll taxes. Our cheapest first year is $21,000. On time saved alone, an agent would have to take back about 862 hours a year, roughly 17 hours a week of a clerk's time, to pay for itself in year one. For a small office, time alone may not cover it.

So the case has to rest on what the job costs when it goes wrong: overcharges paid, duplicate bills paid twice, rebates never claimed, invoices sent late. The assessment measures that on a month of your own documents. If the numbers don't justify a build, the written scope says so and you stop there.

Skip a custom agent when:

• The job happens a few times a month. A checklist is cheaper.
• The software you already pay for does it. If your AP software matches invoices to POs, keep using it for that.
• The job changes every month. An agent needs a repeat pattern with rules a person could write down.
• You want it to answer phones. We don't build that.
• A simple platform agent would do. If one person can set it up on a platform in an afternoon and watch it, start there.

Where we sit, and who should pick someone else

We're not the cheapest. Dappasol's $2,000 install and Armada Works' $2,500 one-week build cost a fraction of our build, and both credit their first step toward the next one. Dappasol also puts a guarantee in writing on its $500 Game Plan: it finds a day a week to claw back, or the plan is free. Frenchy Digital publishes its prices line by line, sends a fixed-price proposal within 5 business days, and prices much bigger multi-system work ($70,000 to $420,000 and up) than we take on. Lindy is the right answer if the job is simple and someone on your team likes setting up tools.

What we sell is narrower: one back-office job for a contractor or home service company, like checking supplier invoices, coding the AP inbox, chasing collections or reading bid packages. It runs from your inbox and the exports your team already pulls, never your field software's API. It's set up in your own accounts, it reports and a person decides, and it goes live only after it passes a written test on your own documents. If that's the job you have, see the accounts payable agents for contractors or the full list of agents.

FAQ

How much does an AI agent cost?

Published prices run from about $30 a month for a platform you set up yourself to $70,000 for a bespoke single-workflow agent built by an agency, plus $500 to $2,500 a month after launch at the starting rates. Ours: a paid assessment from $1,000 (credited), a build from $15,000, and $500 to $2,000 a month.

How much does it cost to build a custom AI agent?

For one job, $2,000 to $70,000 at the agencies above, depending on how many systems and document layouts it touches and how it's tested. A first automation or starter agent sits at the low end; a custom agent that reads your own documents and is proven on them sits in the middle.

What does an AI agent cost per month?

Maintenance plus model usage. Published maintenance starts at $500 a month (Dappasol and us) and $2,500 a month (Frenchy Digital's baseline). Model usage for a back-office agent reading 1,000 invoices a month works out to about $11 on Claude Sonnet 5.5 list prices in our arithmetic, and more for agents that handle conversations.

Do I pay for the AI model separately?

Yes, with us. Where an agent uses a model, the provider bills your own account at its list price, so you see exactly what the agent uses. Frenchy Digital and Armada Works work the same way. On Lindy you buy credits instead.

Why is there a paid assessment?

Because the price depends on your documents. In the assessment we measure the job on a month of them, list the layouts in scope and write the acceptance test. You get the numbers, a fixed build price and the scope within 5 business days of sending the documents. The $1,000 is credited to the build if you go ahead within 60 days.

How long does a custom agent take to build?

Our invoice check agent is scoped to go live within 30 days of kickoff. The clock starts when the NDA is signed and the documents are in, and it pauses while we wait on data.

Who owns the agent?

You own your data, your documents, every report the agent produces and the accounts it runs in. Once the build is paid, you get a permanent license to run and change the agent for your own business, including after maintenance ends. We keep our general tools and code. We're added to your accounts as a user you can remove at any time.

Can I build an AI agent myself instead?

Yes, for simple jobs. Platforms like Lindy start free and the Team plan is from $29.99 a month per user. The trade is your team's time to set it up, test it on your own documents and keep it working when a sender changes a layout.

Talk it through

Tell us the office job you'd hand off first. On a 30-minute call we'll tell you whether an agent can do it with the data you have and what the assessment would cover. Book a call, or write to hello@curbandcode.com.

Sources