By Joseph Darnell · November 16, 2026 · 12 min read ·

AI agent development company: how to pick one, what it costs, and what the small-shop model looks like (2026)

Three kinds of company call themselves an AI agent development company, and they quote three different jobs. What enterprise labs, consultancies and small shops charge in 2026, the four questions that sort the quotes, and what a small shop's quote should contain, line by line.

AI agentsAI agent developmentHiring an AI agencyAI agency costBuild vs BuyPractitioner Field Report

A founder we talked to in October had three quotes on her desk for the same thing: an agent that reads inbound leads, qualifies them against her criteria and books the call. One quote was $184,000 with a 14-week discovery phase. One was $38,000. One was a Fiverr listing for $900. All three called themselves an AI agent development company. She wanted to know which one was lying.

None of them were. They were quoting three different jobs, and the pages she had read to find them never said so.

An AI agent development company designs, builds and runs software agents that do a job inside your business: intake, follow-up, reporting, the inbox, the phone. There are three kinds, and the quotes are not comparable across them. Enterprise AI labs and offshore shops quote $50,000 to $400,000 for multi-agent systems on a 3-to-12-month timeline. Consultancies bill $180 to $490 an hour and sell discovery first. Small shops that run agents on their own business quote the job, usually $2,000 to $50,000 to build and a retainer in the low thousands a month for the first six to twelve months. The right one depends on how many systems the agent has to touch and who owns it after launch. Ask every vendor the same four questions below. The quote that answers all four with numbers is the one to take seriously.

Here is how to read those three quotes, in the order the founder should have.

Three doors, three different companies

Search for an AI agent development company and Google's AI Overview, as of November 11, 2026, lists LeewayHertz, Appinventiv, Pegasus One, NP Group and Kanerika, then asks whether you want agents for a specific use case or help choosing a partner by tech stack. Page one under it is Azumo, a Wistron ITS list that opens with IBM Consulting, Cognizant and Capgemini, two more top-ten lists, Damco, Tech.us and the JADA Squad. Nine results, zero prices. That is the first thing to know about this market: everyone sells the same noun.

Behind the noun are three business models.

Door one: the enterprise AI lab or offshore development shop. LeewayHertz, Appinventiv, the Indian and Eastern European shops that fill the listicles. They sell multi-agent systems, "AgentOps", RAG pipelines and governance layers, and they staff a project with a team. Azilen, one of them, published its own price table on September 18, 2026: a simple rules-based chatbot at $10,000 to $50,000, an LLM task agent at $50,000 to $120,000 or more, a RAG knowledge agent at $80,000 to $180,000, a multi-agent orchestration system at $150,000 to $400,000 and up, with monthly running costs of $500 to $20,000 on top. Their timelines run four weeks for the simple tier and six to twelve months for the top one. These are real numbers for real work. They are also numbers for a 1,000-person company wiring an agent into SAP.

Door two: the consultancy. IBM Consulting, Capgemini, the regional firms, and a long tail of two-partner boutiques. The product is hours. Second Talent's September 24, 2026 rate guide puts a US agency or dev shop at $180 to $295 an hour for a junior agent developer, $230 to $380 for mid-level and $295 to $490 for senior, against $95 to $235 an hour for the same people as freelancers. Consultancies sell a discovery phase first because discovery is billable and because they have not seen your systems. The 14-week discovery on the founder's $184,000 quote was door two.

Door three: the small shop that runs agents on its own business. Three to ten people. They have built the intake agent, the inbox agent and the reporting agent for themselves before they sold one, and they quote the job, not the hours. This is where Automaton sits, and the rest of this page is written from inside that door, so read it as a view and not a survey.

The $900 Fiverr listing is not a door. It is a prompt in a wrapper, and it is fine for a demo.

What the job costs through each door

The honest version of a price page is a table with the same job in every row. Here is the founder's lead-qualification agent, quoted through the three doors, with the sources dated.

DoorWhat the quote coversBuildTimelineAfter launch
Enterprise lab / offshore shopDiscovery, architecture, a team of 3 to 5, a multi-agent design even when one agent would do$50,000–$120,000 for a task agent (Azilen, Sept 2026); $25,000–$80,000 for a "task-execution" agent (Meera, July 2026)8–16 weeks$2,000–$8,000 a month running cost, plus a support contract
ConsultancyDiscovery first, then hoursAt the US mid rate, Second Talent's own estimate: prototype $18,000–$46,000 (150–250 hours), production $48,000–$111,000 (400–600 hours)6–12 weeks after discoveryHours, at the same rate
Small shop (our rate card)The agent, wired to your CRM, calendar and document store, with written hand-off rules$5,000–$50,000; intake sits at the top of that range because it touches three systems2–6 weeks$40–$100 a month in model usage, plus a retainer in the low thousands for 6–12 months

The small-shop column is our published rate card: Sprints at $2,000 to $5,000, Builds at $5,000 to $50,000, Embedded partnerships at $3,000 to $8,000 a month. The run figures come from the bill for the five agents that run Automaton, about $200 a month all in as of November 2026, and the use-case table on our agents hub prices each job the same way.

Two things the table shows that a vendor page will not. First, the build is the smaller half. Meera's July 2026 breakdown puts development at 50 to 60 percent of true first-year cost, and annual maintenance at 15 to 30 percent of the build. On a $150,000 enterprise build that is $22,000 to $45,000 a year, every year. Second, the spread inside one door is wider than the spread between doors. Azilen's simple tier tops out at $50,000; Meera's identical tier stops at $30,000. Both are quoting from their own practice. So are we.

Why the enterprise quote is so much bigger

It is not padding, mostly. Four things inflate door one that door three does not carry.

A team. An offshore shop staffs a project manager, a solutions architect, two developers and a QA engineer, and the quote is their months. A small shop puts one builder on it who has already built the pattern four times.

A multi-agent design by default. The AI Overview for this query leads with "multi-agent systems" and "AgentOps", and the big shops sell to that. Most jobs under a thousand employees are one agent, one job, three integrations. Our intake build for an estate planning firm took the firm's reply time from 52 hours to under four minutes and produced 2.4 times the consultations with no new staff, and it is one agent with a calendar, a CRM and a document store behind it (build log).

Discovery as a product. When a vendor has never seen a law firm's intake, it has to charge to learn it. When it ran its own intake agent last Tuesday, it does not.

Governance theater. Guardrails, traceability and decision logs are real requirements. They are also two days of work on a single agent and a two-month workstream in a 15-agent system. The quote should say which one you are buying.

None of this makes door one wrong. If you are a bank wiring an agent into a core system that six other teams touch, you want the team, the architecture review and the governance workstream, and you should expect the top of Azilen's table. The mistake is a 40-person company buying that process for a job that is one agent and a CRM.

The four questions that sort the quotes

Ask every vendor these, in writing. The answers tell you which door you are standing in before you read the price.

1. Which systems will the agent write to, and have you wired that exact system before? "We integrate with any CRM" is a door-two answer. "HubSpot, yes, the contact and the meeting objects; the calendar through the Google API; we have not done Clio" is a door-three answer, and the gaps are more useful than the yeses.

2. Where does the human take over, and what happens at that line? Every agent we run has a written rule: a person runs every consultation; anything priced, promised or discounted is drafted and waits; anyone who asks for a lawyer reaches a named human within the hour. A vendor who cannot write that sentence for your job has not thought about your job.

3. What is the bill in month nine? Build price, model usage, hosting, and the retainer or support contract, as separate lines. A quote that stops at the build is quoting half the cost. Gartner expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027, and names rising costs, unclear business value and inadequate risk controls as the reasons (Gartner, June 2025, via RCR Wireless). All three show up first in month nine.

4. Show me one running in your own business. Not a demo. A dashboard, a log, the actual inbox. The same Gartner release estimated that of the thousands of vendors claiming agentic products, about 130 offered the real thing; the rest it called "agent washing". The fastest filter is whether the vendor eats its own cooking. Ours is on the work page: the sales director that writes each rep's morning briefing, the attribution rebuild that found $48,000 credited to the wrong channel, the inbox agent whose daily review takes five minutes, and the SEO and AEO loop that wrote the first draft of this page and then got edited by a person.

A vendor that answers all four with specifics can be any of the three doors. A vendor that answers none of them with specifics is the Fiverr listing in a nicer suit.

What a small shop's quote contains

Since we are one, here is what ours looks like for the founder's job, so you can compare it line by line against the other two on her desk.

The scope is one agent and three integrations: inbound lead arrives (form, email or the phone agent's transcript), the agent qualifies it against her written criteria, pulls the two documents she always asks for, books the call on the calendar and writes the record to the CRM. Build: $12,000 to $18,000, four weeks, because the CRM is HubSpot and the calendar is Google and we have wired both before. Running cost: $40 to $100 a month in model usage at November 2026 prices, on top of the HubSpot and Google seats she already pays for. Retainer: $3,000 a month for the first six months, which covers the dial-in against real leads, the drift checks, the rule changes when her criteria move, and the monthly read of what the agent did and did not hand off. Hand-off line: anything outside the criteria, anyone who asks for a person, and every booked call goes to a human; the agent never quotes a price.

Earlier this year we turned down a content engine for a high-end fashion startup: social media, UGC, the full production pipeline. We told the founder to resist the urge. Doing it right would have been out of his startup budget, AI video still cannot look fully real in every frame, and his luxury brand clients would have been repelled by anything that smelled like slop.

Three lines in that quote are the ones the other two doors tend to leave blank: the exact systems, the hand-off sentence, and the retainer with a stop date. If the enterprise quote on her desk had those three lines and still came to $184,000, the difference would be the team and the governance workstream, and she could decide whether she needed them. It did not have them, which is why it looked like a lie. It was a different product with the same name.

The second-year bill

The number nobody quotes is what the agent costs in year two, and it is the number that decides whether the project survives.

For a built agent the second-year bill has three parts. Model usage, which has been falling: Anthropic cut its top model's price 20 percent on November 9, 2026, and the agents on our bill cost less to run now than when we built them. Hosting and the systems the agent writes to, which you already pay for. And the human time to keep it aligned, which is the retainer in year one and, in year two, either a smaller retainer or an hour a week of someone on your side who understands the rules.

For a bought agent inside a product you already use (your CRM's own agent, your help desk's), the second-year bill is the seat price, and it goes up when the vendor decides it does. For a door-one build, it is the 15 to 30 percent maintenance line, which on a $150,000 build is a second small build every year.

This is why we tell most companies under a thousand people to start with a product agent where one exists, and build only the job that crosses systems. It is also why the retainer belongs in the quote and not in a surprise email in month seven. A custom agent is not finished on launch day. The first six to twelve months are where it learns your rules against real traffic, and whoever built it should be the one watching.

When to pick which door

Pick the enterprise lab when the agent touches a core system with other teams on it, when you need a vendor your procurement department has already approved, or when the job genuinely is multi-agent and the budget is six figures by design.

Pick the consultancy when you do not yet know which job to automate and you want someone to find out, with the understanding that you are paying hourly for the finding out and will probably pay someone else to build it.

Pick the small shop when the job is clear, it touches one to three systems you can name, you want the people who built it to keep running it, and you want to see it working in their own business before you sign. Ask the four questions and read the rate card before the call so the call is about your job and not about ours.

The founder took the $38,000 quote, which turned out to be a door-two boutique with a 12-week plan. It shipped in week nineteen. The agent works. She is now paying $340 an hour for rule changes. Ask question three.

Frequently asked questions

How much does it cost to hire an AI agent development company?

It depends on the door. Enterprise labs and offshore shops quote $10,000 to $50,000 for a simple rules-based agent, $50,000 to $120,000 for an LLM task agent and $150,000 to $400,000 or more for multi-agent systems (Azilen, September 2026). Consultancies bill $180 to $490 an hour in the US (Second Talent, September 2026). Small shops that run their own agents quote the job: our builds run $5,000 to $50,000, with a lead-qualification agent landing around $12,000 to $18,000, plus a retainer in the low thousands a month for the first six to twelve months.

What is the difference between an AI agent development company and an AI automation agency?

An automation agency wires existing tools together with rules (Zapier, Make, n8n, a CRM workflow). An agent development company builds software that reads, decides and acts inside a defined boundary, with a model at the decision points and written hand-off rules. Most real projects are both: rules where rules work, a model at the two or three points that need reading. The best small shops do both and tell you which parts are which. Our agency economics piece covers how the retainer math works on each side.

How long does it take to build a custom AI agent?

A single agent with one to three integrations takes two to six weeks at a small shop that has built the pattern before. Vendor timelines for larger work run four to eight weeks for a simple agent, three to five months for an LLM or RAG agent and six to twelve months for multi-agent systems (Azilen, September 2026; Meera, July 2026). The calendar time after launch matters more: budget six to twelve months of dial-in against real traffic before the agent runs with a light touch.

What should an AI agent development quote include?

The systems the agent will write to, by name. The hand-off rule: where a person takes over and how fast. The build price and the timeline. The running cost per month at current model prices. The retainer or support contract with a stop date or a review date. A quote missing any of those is quoting part of the product. Ask for the missing lines in writing before comparing prices.

Should I hire an AI agent developer or a company?

A freelance agent developer costs $95 to $235 an hour in the US (Second Talent, September 2026) and is the right call for a prototype or an internal tool one technical person will own. Hire a company when the agent faces customers, writes to systems of record, or has to keep running after the developer moves on, because the retainer and the hand-off rules need an owner. The failure mode with a freelancer is month nine: the agent drifts and nobody is watching.

Why do so many AI agent projects get cancelled?

Gartner predicts more than 40 percent of agentic AI projects will be cancelled by the end of 2027, and names rising costs, unclear business value and inadequate risk controls (Gartner, June 2025). In our experience all three trace to the quote: a build price with no running cost, a use case with no hand-off rule, and no retainer to catch the drift. The projects that survive were scoped as one job, wired to named systems, and watched by the people who built them for the first year.


Next step
Want this working in your business?
Book 30 minutes with Joseph. No pitch: he'll listen to what you're building and tell you honestly whether Automaton can help.
Book a call →

You'll hear back from Joseph, not a bot.

← All insights