AI Agent Development Company for a Small Business: What to Ask, What It Costs, What You Get (2026)
How a 10 to 100 person company chooses an AI agent development company in 2026: the four routes with dated prices, the eight questions that expose a weak vendor, what a $20,000 to $50,000 engagement includes week by week, and when you should not hire anyone.
Oct 29, 2026
A small business has four ways to get an AI agent built in 2026, and the honest price range runs from $0 in extra fees to $100,000: configure the platform you already pay for (Microsoft Copilot Studio at $200 a month per 25,000 credits, ChatGPT Business at $20 per seat per month billed annually, read 19 October 2026), hire a freelancer (Upwork lists AI agent developers at $30 to $150 an hour and custom LLM agents at $5,000 to $15,000 per project, read 7 October 2026), hire an AI agent development company ($20,000 to $50,000 for one workflow at KUMO's published bands, $50,000 to $100,000 across several systems), or hire an engineer ($5,000 to $10,000 per month is what a retained senior team costs; a full-time hire costs more and takes months to find). The right answer for most companies under 100 staff is the first one. This guide is for the minority whose process does not fit a platform, and it tells you what to ask so you do not pay a company for work a platform would have done.
Key facts (updated 20 October 2026)
- Platform route: Copilot Studio $200 a month per 25,000 credits (Azure subscription required); ChatGPT Business $20 per seat per month annual, $25 monthly; Claude Team the same
- Freelancer route: $30 to $150 an hour; custom LLM agents $5,000 to $15,000 and multi-agent systems $10,000 to $30,000 per project on Upwork (read 7 October 2026); you own project management, testing and what happens after launch
- Company route: $20,000 to $50,000 for one workflow with one integration set; $50,000 to $100,000 for agents across several systems or with a portal; 8 to 12 weeks to production
- After launch: model fees under $100 a month at 2,000 runs, hosting a few hundred dollars, and $5,000 to $10,000 per month of ongoing engineering if nobody in-house owns it
- Hire a company only when the process has per-record approval rules, three or more systems in one flow, an audit requirement, or volume pricing that outgrows a subscription
Who this is for
An owner or operations lead at a company with 10 to 100 staff and no engineers, who has tried an assistant or an automation tool on one painful process (intake, quoting, invoice checks, support triage, onboarding) and hit the wall: the tool cannot apply the rule that changes per customer, cannot write back to the second system, or cannot show an auditor what it did. You are now reading vendor pages that all say "custom AI agents" and you need a way to tell them apart.
The four routes, priced
Prices read from the vendors' pages on the dates shown; KUMO's prices are its published bands.
| Route | What you pay | What you get | Where it stops |
|---|---|---|---|
| Configure the platform you pay for | Copilot Studio $200 a month per 25,000 credits; ChatGPT Business or Claude Team $20 to $25 per seat per month | An agent over your documents and the connectors the platform offers, in days | Rules that differ per record, actions across three or more systems, audit trails in your own systems |
| Freelancer | $30 to $150 an hour; $5,000 to $15,000 per custom agent project | A working agent built by one person | You manage scope, testing and handover; when the person moves on, so does the knowledge |
| AI agent development company | $20,000 to $50,000 for one workflow; $50,000 to $100,000 for several systems | Discovery on your real cases, the agent, integrations, approvals and audit log, evaluation set, tests, handover of code and prompts | Costs ten to fifty times a year of subscriptions, so only right when the process needs it |
| In-house or retained engineers | $5,000 to $10,000 per month retained; a full-time senior hire costs more and takes 2 to 4 months to find | Continuous ownership of prompts, models and new workflows | Not worth it for one agent; worth it once three or more run in production |
Want the number for your own scope? Send a short brief: what takes your team the most manual time, which systems are involved, roughly how many people touch it. A senior engineer replies within two working days with a band and the questions we would ask on a call. No call needed. Or get a range in a minute from the AI development cost calculator.
The eight questions that expose a weak vendor
- Which of my systems will the agent read and which will it write to, and through which API or integration? A real answer names the systems and the method. "We integrate with everything" is not an answer.
- What happens when the agent is not sure? You want a confidence threshold, a named human path, and a log entry, in that order.
- How will you test it before launch? The right answer is an evaluation set built from your last quarter's real cases with a pass rate you agree in writing, not a demo.
- What does it cost to run per month after launch, at my volume? Model fees, hosting and who maintains it, as three numbers. If the vendor cannot compute model fees from your volume, they have not built one before.
- Who owns the code, the prompts and the evaluation set? You do, at handover, in the contract.
- Which model, and why, and what happens when the model changes? A vendor that commits to one model forever will be wrong within a year; a vendor that cannot name one has not decided.
- Show me a production agent you built and what it does on a bad day. Case studies that only describe the good day are marketing.
- What would you tell me to buy instead of building? The vendor that names the platform route when it fits is the one to hire for the cases where it does not.
What a $20,000 to $50,000 engagement includes, week by week
Weeks 1 to 2, discovery. A call with the people who do the work today, read access or exports from the systems involved, a written map of the process with its exceptions, and the evaluation set: 50 to 200 real past cases with the right answer attached. You see the scope and the pass criteria in writing before anything is built.
Weeks 3 to 6, build. The agent, the integrations, the approval step and the audit log. Every week ends with a working demo on your real cases, not slides.
Weeks 7 to 9, evaluation and hardening. The agent runs on the evaluation set until it passes; then it runs in shadow mode next to the people who do the work, and the failures become new test cases.
Weeks 10 to 12, launch and handover. Phased rollout, a runbook, the dashboard that shows runs, cost and escalations, and the repository with code, prompts and evaluation set transferred to you.
Not included and on top: model fees (under $100 a month at 2,000 runs on current prices), hosting (a few hundred dollars a month), any platform plan whose API you need, and ongoing engineering at $5,000 to $10,000 per month if you have no engineers.
When you should not hire anyone
If the task lives in one or two systems, the rule is the same for every record, and nobody needs an audit trail, configure the platform you already pay for and stop. The best AI agent builders comparison prices that route. If the gap is one notification or one hand-off, a workflow tool closes it for a few hundred dollars a month. Hire a company only when one of the four tests is true: approval rules that differ by record, owner or amount; three or more systems in one flow with ordering and rollback; money or client data moving with an audit requirement; or a subscription bill that climbs with volume past the build cost within a year.
Proof, labelled
CampaignHQ is KUMO's own workflow product, in production since 2022; it is where KUMO's agent patterns run first, so this proof is direct for the agent mechanics. KUMO's founders built Volopay's first production version as its founding engineers, a finance product where every approval and every payment had to be traceable; that proof is adjacent here and relevant to the approval and audit side. Racepoint is a multi-agent B2B procurement platform KUMO built, direct for multi-agent work. KUMO is a software engineering studio that builds custom AI agents, workflow apps and portals for 10 to 100 person companies, and a Registered Partner in the OpenAI Partner Network.
Questions buyers ask
How much does an AI agent development company charge a small business in 2026? $20,000 to $50,000 for one workflow with one integration set, $50,000 to $100,000 across several systems, plus model fees, hosting and ongoing engineering after launch. Freelancers list $5,000 to $15,000 per agent project; platforms cost $20 to $200 a month.
Should a 40-person company hire an AI agent developer or use Copilot Studio? Use Copilot Studio or ChatGPT Business first if the task fits their connectors and a single approval. Hire a developer or a company only when the process needs per-record rules, three or more systems, or an audit trail.
What should an AI agent development proposal include? The systems read and written and how, the confidence and escalation design, the evaluation set and pass rate, the monthly running cost at your volume, ownership at handover, and a dated plan with a first production milestone.
How long does it take an AI agent development company to deliver? 8 to 12 weeks from discovery to a production agent on one workflow; 12 to 20 weeks across several systems with a portal. A platform agent is live in days.
What is the difference between an AI agent and an AI chatbot for a business? A chatbot answers; an agent looks things up, applies your rules, takes the action in your systems and asks a person only when the rule says so. The chatbot cost guide covers the first; the AI agent cost guide covers the second.
Can an AI agent development company work with our existing CRM and accounting software? Yes if the systems have an API or an export; the proposal must name the method per system. Where the platform only exposes its API on a higher plan, the plan upgrade is part of the cost.
If you want a number before talking to anyone, send a short brief: what takes your team the most manual time, which systems are involved, roughly how many people touch it. A senior engineer replies within two working days with a band and the questions we would ask on a call. Or run the AI development cost calculator, see AI workflow automation, or book a 30-minute call.
The numbers in one file. The Build vs Buy Price Book, October 2026 has the published prices of 40 business tools read in October 2026, what they cost a 40-person company, and the four tests that say whether a custom build is worth it. Download the PDF (8 pages, free, no form).
