Search this question and you will find twenty agencies telling you “it depends” followed by a contact form. That answer protects the agency, not you. You cannot budget with “it depends.”
So here is the answer with numbers in it. These figures come from our own fixed-price projects and from what the US market currently charges, so you can see both sides before you talk to anyone, including us.
The short answer
For a US company buying custom AI development in 2026:
| What you’re building | Typical range | Timeline |
|---|---|---|
| AI readiness audit | $2,000 – $5000 | 1 – 3 weeks |
| AI chatbot (RAG, on your data) | $10,000 – $40,000 | 4 – 8 weeks |
| AI agent (multi-step, integrated) | $25,000 – $80,000 | 6 – 14 weeks |
| Custom LLM application | $30,000 – $120,000 | 8 – 20 weeks |
| Fine-tuned model + pipeline | $40,000 – $150,000+ | 10 – 26 weeks |
| Dedicated AI engineer (monthly) | $4,000 – $20,000 /mo | ongoing |
Why do the ranges span 3 to 4x? Because two decisions move the price more than everything else combined: who builds it, and how much of it must survive contact with production. The rest of this article unpacks both.
What you’re actually paying for
An AI project is roughly one third model work and two thirds everything around it. The demo is cheap. Production is not. A working prototype of a chatbot takes days. The parts that take weeks are the parts nobody demos:
Data plumbing. Your documents, tickets, and databases are messier than you think. Getting them cleaned, chunked, embedded, and refreshed on a schedule is real engineering.
Evaluation pipelines. How do you know the AI is right, and stays right after every change? Without automated evals you find out from your customers. Building evals is often 20 to 30 percent of a serious project, and it is the line item cheap quotes silently delete.
Security and access control. An AI that can read your data can leak your data. Prompt injection, permission boundaries, and audit logs are not optional for anything customer-facing.
Integration. The model is an API call. Making it act inside your CRM, helpdesk, or ERP is where the engineering hours actually go.
When you compare quotes, ask each vendor to show these four items as separate lines. If a quote is dramatically cheaper, one of these is usually missing, and you will pay for it later with interest.
AI chatbot development cost
A production chatbot on your own data (a RAG system) typically runs $10,000 to $40,000 in the US market in 2026.
The low end gets you a single-channel bot over a clean document set. The high end adds multiple data sources, human handoff, analytics, and hard answer-quality guarantees enforced by an eval suite.
One number worth knowing before you buy: ask what percentage of answers must be correct, and how that will be measured. A vendor without an immediate answer is selling you a demo. From our own builds, a support and product chatbot is usually justified when it deflects even 20 to 30 percent of ticket volume; at typical US support costs the payback lands inside the first year.
AI agent development cost
This is the fastest-growing question in the category, and the honest range is wide: $25,000 to $80,000 for a first production agent.
An agent differs from a chatbot in one expensive way: it acts. It reads an invoice, checks it against a PO, posts to your accounting system, and flags exceptions. Every action it can take is a failure mode you must engineer against, which is why agents cost 2 to 3x a chatbot of similar surface area.
Real example from our work: an invoice processing agent handling around 12,000 documents a month at 96 percent straight-through accuracy. It replaced the workload of three full-time staff, roughly $180,000 a year at US fully-loaded cost, and the build paid for itself inside the first quarter of operation.
That is the correct way to evaluate agent pricing: not “is $50K expensive” but “what recurring cost does it retire.” An agent that saves one salary a year justifies almost any build cost on this table.
The rate card nobody publishes
Custom AI work is priced three ways in the US market right now:
Enterprise firms: $300 – $1,000 per hour. The big consultancies. You are paying for the brand your board recognizes, and for transformation programs, not builds.
US boutiques: $150 – $350 per hour. Senior US engineers, strong for co-located work and regulated industries. A 10-week agent build at these rates lands between $60,000 and $140,000.
Offshore engineering teams: $25 – $75 per hour. The same stack (GPT, Claude, LangChain, vector databases) at a fraction of the rate. The variance in this tier is enormous: it contains both excellent 10-year engineering firms and freelancers who learned RAG last month. The rate does not tell you which one you found. The evaluation section below does.
We publish our own rates, which most firms in our tier will not do: engagements start at $2,500 for a fixed-scope audit, project builds are fixed-price against a written scope, and dedicated engineers run well under US boutique rates.
Fixed price vs. time and materials
Time-and-materials billing means you carry the risk of every underestimate. Fixed price means the vendor does. For a first AI project with a new vendor, insist on fixed price against a written scope, and treat any vendor who refuses as telling you something about their own confidence.
The fair trade for a fixed price is a genuinely fixed scope, which is why serious firms audit before they quote. An audit that costs $2,500 and produces a scoped, fixed-price proposal is not an upsell. It is how both sides avoid the classic AI project failure: six months of billing toward a target nobody defined.
Ours credits 100 percent against the build if you proceed, and roughly 30 percent of our audits end with us saying we are not the right fit and pointing you elsewhere. We publish that number because it is the strongest evidence we can offer that the audit is an audit, not a sales call.
Five questions that expose a real quote
- “Show me the eval line.” Where in this quote is answer-quality measurement? No line, no production system.
- “What is the accuracy commitment, and how is it measured?” Accept a number, not an adjective.
- “Who exactly builds this?” Names, not “our team.” Ask if the people on the call are the people on the code.
- “What happens at 110 percent of budget?” Fixed price makes this question boring. T&M makes it terrifying.
- “What do I own at the end?” Code, prompts, eval suites, and data pipelines should be yours, in your repos, with no license strings.
Why the same project costs $30K or $130K
Take one identical build: a customer-facing AI agent integrated with a helpdesk and a billing system. Priced three ways:
- US boutique, T&M at $250/hr, 10 weeks, two engineers: $100,000 – $130,000
- Established offshore team, fixed price, senior engineers with published case studies: $30,000 – $55,000
- Cheapest bid you received: $12,000, no eval line, no accuracy commitment, portfolio of demos
The first is a fair price if you need US-based delivery for regulatory or preference reasons. The second is the same engineering at offshore economics, and the way to de-risk it is references, named engineers, and a small paid audit before the large commitment. The third is how companies end up paying twice: once for the demo, once for the rebuild.
The bottom line
Budget $10,000 to $40,000 for a production chatbot, $25,000 to $80,000 for a first AI agent, and evaluate every quote by its eval line, its accuracy commitment, and what you own at the end. Price the project against the recurring cost it retires, not against the cheapest bid in your inbox.
And if a vendor answers “how much does AI development cost” with “it depends,” send them this article.
Want a number for your specific project? Our AI Readiness Audit is $2,500, takes ten working days, ends in a fixed-price scoped proposal, and is credited in full if we build together. About 30 percent of the time we conclude we are not the right fit, and we tell you so. Book the audit →
FAQ
Useful production AI starts around $10,000 to $15,000 in 2026, typically a support chatbot or a single automated workflow. Below that budget, configure an off-the-shelf SaaS tool instead of building custom. A good development firm will tell you when SaaS is the right answer.
$25,000 to $80,000 for a first production agent in the US market, depending on how many systems it touches and how expensive its mistakes are. Judge the price against the recurring cost the agent retires.
Engineering labor cost, not technology. The stack is identical. The real difference between offshore firms is engineering discipline, which you verify through references, named engineers, published case studies with numbers, and a small paid engagement before a large one.
A free consultation is a sales call, and it is priced accurately. A paid audit with a fixed deliverable and a credit against the build aligns the incentives: the vendor is paid to tell you the truth, including “don’t build this.”
