AI agent development costs between $20,000 and $250,000 to build in 2026, and roughly three times that over three years. A retrieval-backed assistant lands at $20k to $60k. A single-workflow production agent runs $60k to $150k. A multi-agent system touching several internal tools runs $150k to $400k or more. Initial development is only 25% to 35% of what you will actually spend.
Most cost guides in this category stop at the build number, which is the number that gets projects approved and then quietly killed at month nine. This one gives you both halves: what it costs to ship, and what it costs to keep running. All figures are 2026 USD for a team delivering to a Western client, whether that team is onshore or offshore.
What you are actually paying for
An "AI agent" is a system that takes multi-step actions against tools and data, not a chatbot that answers questions. The distinction matters for cost because the reasoning loop is the cheap part. The expensive parts are the things around it.
| Component | Share of build cost | Why it costs what it does |
|---|---|---|
| Integrations (CRM, ERP, internal APIs) | 25%–35% | Every system has different auth, rate limits, and undocumented behavior |
| Data preparation | 15%–30% | Routinely underestimated. Often matches modeling cost on its own |
| Agent logic and tool design | 15%–20% | Defining what the agent may do, and in what order |
| Evaluation and guardrails | 10%–20% | Golden datasets, refusal behavior, approval gates, audit logs |
| UI and human-in-the-loop | 10%–15% | Review queues, override controls, explainability surfaces |
| Infrastructure and MLOps | 5%–15% | Observability, tracing, cost caps, deployment |
Notice that the model itself is not a line item. In 2026 the model is a commodity input. If a vendor's cost story centres on which model they use, they are selling you the cheapest part of the system.
The four cost tiers
Tier 1: Assisted retrieval agent, $20k to $60k
Answers questions from your documents, tickets, or knowledge base with citations. Read-only, no write actions. Six to ten weeks. This is the right first purchase for most companies, and it is where we push clients who arrive asking for something far more ambitious. It proves the data plumbing works before you let anything take actions.
Typical scope: one data source, one channel (Slack, in-app, or web), citation-backed answers, refusal when confidence is low, basic analytics.
Tier 2: Single-workflow action agent, $60k to $150k
Does one job end to end with write access: triages and routes support tickets, drafts and files expense reports, qualifies and updates leads in the CRM. Ten to sixteen weeks.
Cost jumps here for one reason: the moment an agent can write, you need permissions, reversibility, an approval gate, and an audit trail. Those four things are most of the delta between Tier 1 and Tier 2.
Tier 3: Multi-tool orchestrated agent, $150k to $400k
Coordinates several systems and sub-agents on a workflow that used to require a human co-ordinator: procurement intake to PO, onboarding across HR, IT, and payroll, or claims processing. Four to eight months.
Budget realism: at this tier, integration and data work alone typically exceed the entire Tier 1 build.
Tier 4: Enterprise autonomous system, $400k+
Multiple agents, formal governance, SSO and role mapping, regional data residency, external audit requirements. Six to twelve months. If you are here, compliance is a design input rather than a final checklist. See our EU AI Act guide and SOC 2 guide.
| Tier | Build | Timeline | Year-1 run cost | 3-year total |
|---|---|---|---|---|
| 1. Retrieval agent | $20k–$60k | 6 to 10 wks | $6k–$25k | $55k–$160k |
| 2. Single-workflow | $60k–$150k | 10 to 16 wks | $18k–$60k | $170k–$430k |
| 3. Multi-tool | $150k–$400k | 4 to 8 mo | $45k–$150k | $430k–$1.1M |
| 4. Enterprise | $400k+ | 6 to 12 mo | $120k+ | $1.1M+ |
Run costs, the line most quotes omit
Three recurring costs, none optional.
Model API usage: $100 to $10,000 per month
Driven by tokens, not users. An agent that reads a 40-page document on every request costs 50 times more than one that retrieves three relevant paragraphs. This is why retrieval quality is a cost decision as much as an accuracy decision. Caching, routing simple requests to smaller models, and per-tenant budget caps typically cut this line 40% to 70%.
Infrastructure: $200 to $5,000 per month
Vector database, application hosting, observability, and tracing. Tracing is not optional on agents. When an agent does something wrong, you need to see the exact tool calls that produced it.
Maintenance: 15% to 30% of build cost per year
Model versions get deprecated roughly twice a year. Integration APIs change. Your own product changes. Any agreement without a maintenance clause is one where the vendor is planning to charge you at emergency rates later. Our software development contract checklist covers how to write that clause.
What actually moves the price
| Factor | Cost impact | Can you control it? |
|---|---|---|
| Number of integrated systems | +$8k–$25k per system | Yes. Start with one. |
| Data quality and structure | Up to 2x total build | Partly, and cleaning first is cheaper |
| Write access vs read-only | +40%–60% | Yes. Phase it. |
| Regulated domain (health, credit, hiring) | +25%–50% | No, but plan for it up front |
| Human approval workflow | +$10k–$30k | Yes, but you almost always want it |
| Team location | 2x to 4x swing | Yes. See offshore rates by country |
How to cut the cost without wrecking the outcome
- Ship read-only first. A Tier 1 agent in production teaches you more about your data than three months of planning, at a third of the price.
- Integrate one system, not five. Integration count is the single most linear cost driver in the whole build.
- Fix the data before the agent. Data preparation can match modeling cost. Cleaning it under a build contract is the most expensive way to do it.
- Keep a human in the loop at launch. Counterintuitively cheaper. Full autonomy demands far more edge-case engineering than a review queue does.
- Instrument token spend from day one. Retrofitting cost controls after a surprise invoice is a rewrite of the retrieval layer.
- Do not fine-tune yet. Below roughly 100,000 calls a month, prompt plus retrieval beats fine-tuning on both cost and iteration speed.
A worked example
A 60-person logistics company wants an agent to handle inbound shipment exception emails: read the email, look up the shipment, decide the action, draft the customer reply, and update the TMS.
| Line | Cost |
|---|---|
| Discovery, eval dataset from 200 historic emails | $9,000 |
| TMS + email integration | $22,000 |
| Agent logic, tools, refusal and escalation rules | $26,000 |
| Review queue UI with one-click override | $14,000 |
| Guardrails, tracing, cost caps | $11,000 |
| QA and 4-week supervised rollout | $12,000 |
| Build total | $94,000 |
| Year-1 run: API $9.6k, infra $4.8k, maintenance $18.8k | $33,200 |
| 3-year total | ~$210,000 |
Against two full-time coordinators at a fully loaded $55k each, the payback is roughly month fourteen. That is the arithmetic to run before approving any agent project, and it is the arithmetic most vendors will not run for you because it sometimes says no.
Frequently asked questions
How much does it cost to build an AI agent?
$20,000 to $60,000 for a read-only retrieval agent, $60,000 to $150,000 for a single-workflow agent with write access, and $150,000 to $400,000 or more for a multi-tool orchestrated system. Add 15% to 30% of build cost per year for maintenance, plus $100 to $10,000 a month in model API usage and $200 to $5,000 a month in infrastructure.
Why are AI agents more expensive than chatbots?
Because agents take actions. A chatbot that gives a wrong answer creates a bad experience. An agent that takes a wrong action creates a refund, a compliance incident, or a corrupted record. The cost difference is permissions, reversibility, approval gates, audit logging, and the edge-case engineering that makes those reliable, which is roughly 40% to 60% on top of an equivalent read-only build.
What is the ongoing cost of running an AI agent?
Between $500 and $15,000 per month for most mid-market deployments, split across model API usage, infrastructure, and maintenance. Over three years, running costs typically exceed the original build cost. Budget three-year spend at roughly three times the quoted build.
Can we build an AI agent in-house instead?
Yes, if you already employ engineers who have shipped and operated one. If you are hiring for it, the comparison is nine to eighteen months to recruit and ramp an internal team against two to four months to production with an external one. Recruiting cost and delay usually exceed the outsourcing premium. Our guide on engagement models covers the middle options.
How do we stop model costs from spiralling?
Four controls, in order of impact: retrieve less context per call, cache repeated retrievals, route simple requests to smaller models, and set hard per-tenant spend caps. Together they typically reduce API spend 40% to 70%. Insist all four are in scope before signing.
Do AI agents get cheaper as model prices fall?
Partly. Model inference gets cheaper every year, but that is the smallest cost line. Integration, data preparation, evaluation, and maintenance are engineering costs and they track engineering salaries, not token prices. Expect your run cost to fall modestly and your maintenance cost to stay flat.
Key takeaways
- Four tiers: $20k–$60k read-only, $60k–$150k single workflow, $150k–$400k multi-tool, $400k+ enterprise.
- Three-year cost is roughly three times the build quote. Demand a run-cost line before signing.
- Integrations and data preparation, not the model, are the dominant cost drivers.
- Write access adds 40% to 60%. Ship read-only first and phase it in.
- Run the payback arithmetic against the human cost of the workflow before approving anything.
Getting a real number for your case
Ranges are useful for board approval and useless for planning. If you tell us the workflow, the systems it touches, and roughly how many times a day it runs, we will send back a tiered estimate with build cost, year-one run cost, and the three-year figure, plus the payback month against your current human cost. If the payback is longer than 24 months we will say so, and we have talked clients out of agent projects on exactly that basis. More context on the wider category is in our AI development services buyer's guide and on our AI development services page.