Building a custom AI agent costs $20k–$120k in 2026, and running one costs $500–$5,000 a month. If you buy rather than build, expect per-seat pricing of $10–$200 per user per month, per-task pricing of $0.30–$1.00, or per-resolution pricing of $0.50–$2.00. Most mid-market builds we scope land between $35k and $80k.
Two different questions hide inside "how much does an AI agent cost", and mixing them up is why quotes come back four times apart. One is what a vendor charges you to use an agent they already built. The other is what it costs to have one built around your own systems, plus what it costs to keep running. This piece answers both, in that order.
The pricing half of the answer changed materially in the last eighteen months, which is why older cost articles are actively misleading now.
The four pricing models
| Model | Typical price | You pay for | Best when |
|---|---|---|---|
| Per seat | $10–$200 / user / mo | Access, whether used or not | Small teams using the agent constantly |
| Per task | $0.30–$1.00 / task | Each action the agent attempts | Volume is predictable and success is obvious |
| Per resolution | $0.50–$2.00 / resolution | Only outcomes the agent actually completed | Support and service work with a clear "done" |
| Hybrid / enterprise | Base + usage | A platform fee plus overage above a threshold | High volume, or when you need a cost ceiling |
The direction of travel is away from the top row. Analyses of 2026 SaaS pricing report per-seat falling from roughly 21% to 15% of pricing models inside a year, with hybrid base-plus-usage now the most common structure at around 41% adoption (Chargebee, Korix). The most visible single data point: HubSpot cut its Customer Agent to $0.50 per resolved conversation in April 2026, half its previous $1.00. Treat all of these as vendor-reported figures rather than audited statistics, but the direction is consistent across sources.
Why per-seat is collapsing
Per-seat pricing assumes the software is a tool a person operates. An agent is not that. It does work that used to need a person, so charging by the person prices the thing you are trying to remove. Worse, it punishes exactly the customer you want: the one who automates so much of a workflow that headcount drops. Outcome pricing fixes the incentive, which is why vendors moved even though it makes their revenue harder to forecast.
The catch for buyers is that "resolution" is defined by the vendor. Ask what counts as resolved, who adjudicates a disputed one, and whether a conversation the agent escalates to a human still bills. Those three answers move real money.
What it costs to have one built
If the value is in your own data and your own systems, buying an off-the-shelf agent will not reach it, and you are pricing a build instead. Our ranges, which we quote weekly:
| Scope | Build cost | Timeline | What you get |
|---|---|---|---|
| Single-workflow agent | $20k–$40k | 6–8 weeks | One process, one or two integrations, human approval on writes |
| Multi-step business agent | $40k–$80k | 8–14 weeks | Several systems, memory, evals, an audit trail, escalation paths |
| Regulated or high-volume | $80k–$120k+ | 14–24 weeks | The above plus compliance review, data-residency work, load engineering |
The single strongest predictor of where you land is not model choice or agent framework. It is the number of systems the agent has to read from and write to, because each one brings authentication, a schema, failure modes and tests. The second is whether the agent is allowed to act without a human confirming, which roughly doubles the evaluation work because now being wrong has consequences.
A worked 12-month cost
A mid-market support agent handling 4,000 conversations a month, built rather than bought:
| Line | Year one | Note |
|---|---|---|
| Scoping | $4k | Credited toward the build in our engagements |
| Build | $55k | Helpdesk plus CRM plus knowledge base, human approval on refunds |
| LLM usage | $7k–$14k | Roughly $0.15–$0.30 per conversation at 2026 prices |
| Hosting and monitoring | $3k | Infrastructure plus observability tooling |
| Maintenance | $9k | Evals, model migrations, content refresh, roughly a day a month |
| Total | $78k–$85k | Year two drops to roughly $19k–$26k |
That last row is the one worth internalising. Year one is dominated by the build; every year after is dominated by usage and upkeep. A vendor charging $1.00 per resolution on the same volume would cost about $48k a year, every year, and you would not own the system. Crossover is usually somewhere between year two and year three, which is why the build-or-buy answer depends almost entirely on how long you expect to run the thing.
The costs people forget
- Evaluation. Fixed test sets, run on every change. Skipping them lowers the quote and raises the total cost, because you find out about regressions from customers instead.
- Model migrations. Providers retire models on their schedule. Each migration means re-running evals and sometimes reworking prompts. Budget one or two a year.
- The escalation path. An agent that handles 70% of cases still needs a designed handoff for the other 30%, and that design is real work.
- Content upkeep. The agent is only as current as what it retrieves from. Stale sources produce confident wrong answers.
How to decide
Buy if an existing product covers most of what you need and your workflow can bend to it. Build if the value lives in your specific data, your specific systems, or a process no product serves. Then price the build against the multi-year subscription rather than against zero, because the honest comparison is total cost over the life of the system, not the first invoice.
If you want a real number for your case, our AI agent development engagements start with a paid scoping phase that produces a fixed quote, and the scoping fee is credited toward the build if you proceed. You can also talk to us about the numbers directly.
Common questions
How much does an AI agent cost?
Building a custom AI agent costs $20,000–$120,000 in 2026, and running it costs $500–$5,000 a month. Off-the-shelf agents are priced per seat ($10–$200 per user per month), per task ($0.30–$1.00) or per resolution ($0.50–$2.00). Most mid-market custom builds land between $35,000 and $80,000.
What is outcome-based pricing for AI agents?
Outcome-based pricing charges for completed work rather than access. Instead of paying per user per month, you pay per resolved support conversation, per processed invoice or per booked meeting. It aligns cost with value, but the vendor defines what counts as an outcome, so ask how disputed and escalated cases are billed before signing.
Is it cheaper to build or buy an AI agent?
Buying is cheaper in year one; building is usually cheaper by year two or three. A vendor charging $1.00 per resolution across 4,000 conversations a month costs roughly $48,000 a year indefinitely, while an equivalent custom build costs around $78,000 in year one and $19,000–$26,000 a year after that. Build when the value depends on your own data and systems.
What are the ongoing costs of running an AI agent?
Budget $500–$5,000 per month. That covers LLM API usage (the largest and most variable line), hosting, monitoring and observability, periodic evaluation runs, and one or two model migrations a year. Usage-based inference cost is the line item that most often surprises teams, because it scales with success.