Over three years, a custom AI support agent is cheaper than a per-resolution SaaS agent once you pass roughly 1,500 resolutions a month. At $0.99 per resolution, 5,000 resolutions a month costs about $178,000 over three years; a custom agent built for $18,000 and run and maintained for about $1,150 a month costs about $59,000. Below the crossover the SaaS agent wins on cost, speed and risk. Include maintenance in the cost of owning, hold your own volume up against the table, and read how the vendor counts a resolution.
How do SaaS AI agents charge?
Most customer-service AI agents sold as SaaS now charge per outcome rather than per seat. Intercom's Fin, the best-known example, is priced at $0.99 per outcome on Intercom's pricing page as of October 2026, with a minimum monthly commitment when it runs on top of another helpdesk. Other helpdesk and CRM vendors charge per automated resolution or per conversation, at rates that change often enough that you should check each vendor's current pricing page rather than any article, this one included.
Per-outcome pricing is genuinely fair at the start: you pay nothing for answers that did not help. The catch is that the bill scales in a straight line with success. The better the agent works and the more customers you have, the more rent you pay, with no volume at which it stops.
What does it cost to own a custom AI agent instead?
Owning a custom support agent has three costs, and only the first is paid once.
Build
A production support agent - multi-step, connected to your helpdesk and order or account data, with human handoff - costs $8,000-18,000 from us and $15,000-60,000 across the market. The comparison below uses $18,000, the top of our range, to keep it conservative.
Running
Model API calls, vector storage and hosting. With prompt caching and a sensible choice of model per step, a support agent handling around 8,000 conversations a month runs at roughly $400. We used $150 to $700 a month depending on volume, covering every conversation, not only the resolved ones.
Maintenance
Re-testing when the model provider ships an update, adding new documents, fixing the edge cases users find. We budgeted $750 a month - several engineer-hours - whether that is your own team or a support plan. Leaving this out is the most common way owning is made to look cheaper than it is.
Custom AI agent vs SaaS AI agent: what is the three-year cost?
The same support workload, rented at $0.99 per resolution or owned, over 36 months:
| Resolutions a month | SaaS agent, 3 years | Custom agent, 3 years | Cheaper option |
|---|---|---|---|
| 500 | $17,820 | $50,400 | Rent, by $32,580 |
| 2,000 | $71,280 | $54,000 | Own, by $17,280 |
| 5,000 | $178,200 | $59,400 | Own, by $118,800 |
| 10,000 | $356,400 | $70,200 | Own, by $286,200 |
The crossover, where the monthly rent matches the monthly cost of owning (running, maintenance and the build spread over three years), sits at roughly 1,500 resolutions a month on these assumptions. Below it, the SaaS agent is cheaper and the build cost is never recovered. Above it, owning gets steadily cheaper: at 5,000 resolutions a month the difference over three years is about $118,800, several times the cost of building the agent.
The assumptions, stated plainly: helpdesk seats are left out because you pay them either way; the custom agent is assumed to resolve as many conversations as the SaaS one; and prices are held flat for three years, which is optimistic for both sides.
What does the cost comparison leave out?
Money is not the only difference, and some of it favours renting.
In favour of renting
- Live in days, not weeks. No build risk at all.
- The vendor improves the product without you paying for it.
- You only pay for answers that worked.
- Nothing to maintain, and nobody on your side needs to understand how it works.
In favour of owning
- The agent can act inside your own systems - refunds, bookings, account changes - under your permission model.
- Your prompts, evaluation data and conversation history stay yours and move with you if you change helpdesk.
- You choose and change the model, and benefit directly when model prices fall.
- No exposure to the vendor changing its price, its definition of a resolution or its minimums.
One thing to read closely in any per-outcome contract is how an outcome is counted. A conversation where the customer simply stopped replying may count as resolved. That can be reasonable, but it is worth knowing before you forecast the bill.
When should you rent a SaaS agent and when should you build one?
Rent when:
- You handle fewer than about 1,500 automatable conversations a month.
- Your questions are standard - orders, accounts, policies - and the product handles them well.
- You want it live this month and have nobody to maintain a custom system.
Build when:
- Your volume is above the crossover, or will be within a year.
- The agent needs to take actions in systems the SaaS product cannot reach.
- You need to own the data and audit trail for regulatory or contractual reasons.
- You already pay for the SaaS agent and the invoice has become a line someone asks about.
A reasonable path for many teams is both, in order: rent first, learn which questions it resolves and what volume you really have, then build when the numbers above say so. Our guide to custom AI agents covers how to spec one when that day comes.
Is a custom AI agent cheaper than a SaaS AI agent?
Over three years, a custom AI agent is cheaper than a per-resolution SaaS agent once you pass roughly 1,500 resolutions a month, and much cheaper at 5,000 or more. Below that, the SaaS agent wins on cost, speed and risk. Do the arithmetic with your own volume, include maintenance in the cost of owning, and read how the vendor counts a resolution.
If you are above the crossover, see AI agent development for what a build includes, or how we price for the honest version of our numbers.