The category exploded, and so did the comparison posts, most of which line up fifteen tools without ever saying which suits whom. Five profiles cover almost every real situation. Here's how they separate, with public prices and what each actually costs, sources included.
How did we pick these five AI agents?
Criteria: a public price, or failing that purchase data published by a third party, and the ability to plug into an existing helpdesk without replacing everything. Excluded: products with no verifiable pricing data at all, those forcing their own helpdesk, and wrappers with no handling of customer context.
Which AI support agent for which profile?
| Tool | Who for | Pricing model | Order of magnitude |
|---|---|---|---|
| Fin | Already on Intercom | Per outcome | $0.99 per outcome, 50 a month minimum |
| eesel AI | AI layer on an existing helpdesk | Monthly credit plan | From $299 a month, $0.38 to $0.60 a ticket |
| Decagon | Very high volume, chat, email and voice | Quote only, per conversation | Median contract $432,750 a year (Vendr) |
| Sierra | Large enterprises, voice included | Quote only, per outcome | $200,000 to $350,000 in year one (third-party estimates) |
| Ada | Mid-market to enterprise | Quote only, annual contract | Median contract $72,000 a year (Vendr) |
The five, one by one
1. Fin, the default if you're already on Intercom
Fin bills $0.99 per outcome, with a minimum of 50 outcomes a month, or $49.50. On a third-party helpdesk the unit price is the same and minimum commitments apply. If you're already on Intercom, it's the lowest friction available: nothing to integrate, the knowledge base is already there.
The reservation. The displayed resolution rate measures activity, not correctness: a wrong answer the customer abandons is counted and billed as a resolution. We break the mechanism down in our analysis of the $0.99. Also worth noting, the company is being acquired by Salesforce.
2. eesel AI, best value on an existing helpdesk
eesel AI sits as a layer on top of the helpdesk you have, Intercom included. It's the only one of the five with a fully public price list: monthly credit plans, from $299 for 500 to $1,899 for 5,000, where a ticket is one credit however long it runs. That works out to $0.60 a ticket on the first tier and $0.38 on the last, with no platform or seat fees. Beyond the plan, a credit costs $0.80.
The reservation. An extra layer means an extra dependency, and one more place the knowledge base has to be current. The price saving is paid for in operational complexity. And a credit is spent as soon as the agent works a ticket, resolved or not.
3. Decagon, for very high volume
Decagon targets very high volume, across chat, email and voice. No public pricing. Vendr purchase data shows per-conversation billing and a median contract of $432,750 a year, within a range of $105,000 to $923,183.
The reservation. No public price, so no comparison before entering a sales cycle. And per-conversation billing is also charged when the agent fails and hands over. At that contract size, the product only makes sense at very high volume.
4. Sierra, an enterprise project paid on results
Sierra champions outcome-based pricing: you pay when the agent achieves a result, a resolved conversation or a saved cancellation. Sierra publishes no rates. Third-party estimates put year one between $200,000 and $350,000, including $50,000 to $200,000 of implementation, over a three to seven month deployment. These are reported orders of magnitude, not confirmed prices.
The reservation. This isn't a tool purchase, it's a project. Without a team to carry it for six months, the budget will go without producing a result.
5. Ada, the most accessible of the enterprise segment
Ada does not publish its prices, contrary to what is often written. Vendr purchase data, across 117 contracts, gives a median contract of $72,000 a year, within a range of $36,158 to $295,400. That is six times less than Decagon's median.
The reservation. An eightfold spread between the low and high end means the median is unlikely to be your price. Get it quoted on your volume before planning around it.
What nobody tells you before choosing
- Unit price says nothing about total cost. At $0.99 per resolution, 1,500 resolutions a month is close to $18,000 a year on that line alone, before seats. At eesel, the 1,500-credit plan costs about $8,400 a year, but it covers 1,500 tickets worked, resolved or not. The two units don't measure the same thing. Compare at real volume, never per unit.
- Your knowledge base quality matters more than the engine you pick. All five cite sources. If your articles are eighteen months old, all five will give stale answers with equal confidence.
- Exit cost is decided at entry. Check conversation and annotation export before signing. That's when you get answers, and the only time.
Which criterion do people forget when picking an AI agent?
All five of these tools can tell you how many conversations they handled. None will spontaneously tell you how often they got it wrong, because each measures its performance with its own definition of success. That's structural, not dishonest: an engine grading its own correctness has no reason to be harsh.
Whichever of the five you pick, plan for a quality measurement that doesn't come from the vendor. That's what ResponZ does on Fin's answers, and we compare the available approaches in our scoring tools comparison.