What it costs to build AI agents for a small company (with our real prices)

Our builds start at $15,000, delivered in the first month, with a fixed quote and a written definition of done before you sign. An optional retainer from $2,500 a month covers new agents and upkeep. Market quotes for comparable work run from a few thousand dollars to $200,000 and up, so this page shows the math behind our number.

Most pages ranking for this question are dev-shop cost guides quoting a wide range and asking you to book a call for the number that matters. We would rather publish ours.

What we charge

The build is from $15,000. For that you get a company brain, one searchable memory of every meeting, customer conversation and decision, and the first agents on top of it: the follow-up drafted after every sales call, a briefing before every meeting, the weekly scoreboard, invoices chased, support replies drafted. Each agent is wired into tools your team already uses, and answers come with their sources.

Before you sign anything you get one price and a written definition of done. The quote does not move after the work starts. When the build is paid, the system is yours: plain files on machines you control, no vendor database in the middle, no per-seat fees.

The retainer is from $2,500 a month and it is optional. It covers new agents and tuning as your company changes. If you stop paying it, you keep everything.

There is also a guarantee. Sixty days after launch we run a savings review from the brain's own records. If it finds less than $15,000 a year in savings you agree are real, you get up to 2 months of the retainer free.

What moves a quote up or down

Scope moves the number, and every scope item is named in the quote. The things that raise it are the things that add build time:

• Businesses running on separate workspaces. Two companies on two Google Workspaces means two sets of feeds, permissions and agents.
• Tools with no standard connection. A mainstream CRM connects in hours. A legacy system with no API means custom plumbing, and the quote says so by name.
• More agents at launch. The base build includes the first agents. A longer launch list adds time.

What does not raise the quote: team size on its own, the volume of meetings, or how messy the existing documentation is. The brain is built to ingest mess.

The quote itself comes from a 30-minute intro call about your company, its tools and where the team's week goes. You see the price and the definition of done in writing before any money moves, which also means you can take both to any other vendor and compare line by line.

What the market charges

The ranges below are what you will find when you price the alternatives. Each is the right choice for somebody.

Option	Typical price	Best for	The tradeoff
Dev shop or AI agency	Roughly $15,000 to $200,000 and up	Enterprise scope, regulated industries	Long timelines, quotes settled after scoping
Freelancer	Varies widely	One narrow, well-defined agent	You own the maintenance when they move on
Off-the-shelf tools, DIY	Tens of dollars per user or per month	Solo operators, single-app automations	No shared memory, plumbing is on you

Dev shops publish wide ranges because scope varies. KumoHQ puts a custom agent at $15,000 to $200,000 and up, with most 8 to 100 person companies landing between $25,000 and $80,000, and notes that each API integration adds $3,000 to $8,000 on its own. DevCom's 2026 guide estimates $7,000 to $15,000 per workflow for small businesses, with simple rule-based agents starting far lower and multi-agent systems topping $20,000 per workflow. Both are credible shops and both are quoting a different product than ours: one-off engineering scoped per project, which is why an agent plus three integrations can pass $40,000 before anyone talks about memory.

Freelancers cost less per hour and some are excellent. The risk sits after delivery: when models change or a tool updates its API, the agent breaks, and the person who built it has moved on.

Off-the-shelf tools are the right answer for some teams, and it would be unfair to pretend otherwise. ChatGPT Business runs $25 per user a month on monthly billing and gives every seat a capable assistant. Zapier's Professional plan starts at $19.99 a month and handles single-app automations well. If you are a team of one or two and the job is "summarize this document" or "copy form entries into a spreadsheet," start there and keep your $15,000. Where these tools stop is shared memory: each seat chats alone, nothing learns your company, and connecting five tools into one workflow becomes a side project someone has to own.

What it costs to run after launch

Model usage is the main ongoing cost, you pay it directly to the provider, and for a small company it is modest next to the build. The agents we build draft follow-ups, briefs and reports rather than fielding thousands of public chats a day, so token volume stays low. KumoHQ budgets $200 to $800 a month in running costs for a simple production agent, and most of our installs sit at the low end of that shape because the heavy lifting is retrieval from your own records, which is cheap.

There are no per-seat fees to us and no markup on usage. The system runs on machines you control, so hosting is whatever you already pay for those machines.

The retainer, from $2,500 a month, is the other ongoing line, and it is a choice rather than a dependency. Tools change their APIs, models get updated, and your company keeps producing new kinds of repeat work. On the retainer we handle that drift and add new agents monthly. Off the retainer, the system keeps working as built, and your team maintains it from the handover documentation. Teams with an engineer on staff sometimes take the second path after a few months; teams without one usually keep the retainer because 5 to 10 hours of a stranger's debugging is worth more than $2,500 to them.

The payback math

We price the build against the staff time it removes, using the same method as our audit tool: hours a week x hourly rate x 52 weeks.

Three workflows we build in almost every install:

Workflow	The math	A year of staff time
Sales follow-ups drafted	6 hours a week x $75 x 52 weeks	$23,400
Weekly scoreboard built	4 hours a week x $55 x 52 weeks	$11,440
Founder answering repeat questions	5 hours a week x $150 x 52 weeks	$39,000

That is $73,840 a year across three workflows, against a $15,000 build. Even on the smallest line alone, the build pays back inside 16 months; on the full set it pays back in about 11 weeks. Across the audits we run, a 1 to 5 person team lands around $124,000 a year in its top 3 workflows, a 16 to 50 person team around $300,000, and a 51 to 200 person team around $400,000. Those figures are staff time alone, a floor, and yours will differ with your rates and hours. The free Leverage Audit estimates the 3 workflows where agents would save you the most, with hours and dollars shown, in about 2 minutes.

Our own numbers come from the install we run ourselves. Rye Valley, the holding company behind us, removed $425,000 in annual expenses and multiplied shipping velocity 10x within the first month after install, and publishes a public decision log from the system as a case study.

One caution on the math: count the hours agents can take, which is the repeat work. Agents are bad at judgment calls and at anything without a source. Pricing a deal, deciding who to hire, writing anything that needs taste: that time stays human, and a vendor who counts it in your savings is inflating the payback.

Where agent builds fail

The common failure is an agent with no memory underneath it. A follow-up agent that has never seen your past deals writes generic email. A support agent that cannot search old tickets guesses. The agent is the visible 20 percent; the memory it draws on, kept current by feeds from your calendar, inbox and meeting tools, is the 80 percent that decides whether the output is usable.

This is why bolting an agent onto a company that has no searchable record of its own decisions produces demos that impress and tools that go unused. It is also why we build the brain first and the agents second, as one quote. If a vendor prices the agent without pricing the memory and the feeds that keep it current, ask what the agent will be answering from. The usual answer is nothing, and that is where the money goes to waste, whatever the sticker price was.