A founder we talked to last month sent us her last six monthly invoices from a popular AI support vendor. The line items looked clean. The numbers did not. Her bill had gone from €380 in November to €1,720 in April. Her ticket volume in the same period was almost flat. What changed was that her bot had gotten better at "resolving" things, and the vendor charged her for each one.
This post is about why per-resolution pricing has become the dominant pitch in AI customer support in 2026, what's wrong with it, and how we ended up at a different number.
Where "per resolution" came from
The pitch is elegant on a slide. Old support tools charged per seat. Every agent on payroll, every month, a flat fee. With AI deflecting half the inbox, vendors realized seats no longer mapped to value. So they reached for the cleaner story: "we only charge when the AI actually solves something." Outcome-based pricing. Skin in the game. You only pay when it works.
Intercom went first with Fin at $0.99 per resolution. Zendesk landed at around $1.50. Salesforce Agentforce sits near $2.00. The whole category nodded along, and the analyst class started writing think pieces about how seat-based pricing was dead. By early 2026, 41% of AI vendors had moved to a hybrid of base subscription plus consumption (per a Bessemer survey), most of them metering on "resolutions."
So what's the catch.
The vendor defines the resolution
Here's the part the slide deck doesn't show. A "resolution" is whatever the vendor says it is. The most common rule across the major vendors is some version of: the AI replied, and the customer did not write back within 24 hours (or 48, or 72, depending on the tool).
That's not a customer outcome. That's an absence of a follow-up message. The customer who got a wrong answer and gave up counts as resolved. The customer who escalated by emailing your founder directly counts as resolved. The customer who clicked the "this didn't help" thumbs-down but then closed the tab counts as resolved. The customer who got the right answer but is just slow to reply counts as resolved, billed, and counted again if they happen to come back.
A 2026 customer-research post from Lorikeet put it bluntly. Deflection and resolution are not the same thing. A deflected ticket means no ticket was opened. A resolved issue means the problem was solved. Most vendor billing systems can detect the first one. None of them can reliably detect the second. They charge for the first one anyway.
The incentive runs the wrong way
The deeper structural problem with per-resolution is that the better your bot gets, the more you pay. Improve your knowledge base. Tune retrieval. Add product docs. The model answers more questions correctly. Each one is now a billable event. Your support cost line was supposed to go down. Instead it scales with quality.
We watched a Shopify brand do the math out loud on a call recently. They were considering moving from a 60% resolution rate to a 75% resolution rate. The deflection-rate gain looked great. The bill, at $0.99 per resolution across 8,000 monthly conversations, went from $4,752 to $5,940. They were paying €1,200 more per month to be 15 points better at automation. That's a vendor priced to discourage you from getting better.
It also turns budgeting into a guessing game. A Zylo survey in 2026 found that 78% of IT leaders had reported unexpected charges from consumption-based AI pricing in the prior year. For a small or mid-sized European brand, "unexpected" is not a vibe; it's a cash-flow event.
Why we landed at per reply
When we sat down to price Keloa, we tried per-resolution. We sketched it. We modelled it against three customer profiles. It failed on the same test every time: we couldn't tell a founder, with a straight face, what she was going to pay next month.
So we charge per AI reply instead. Specifically: a fixed monthly plan that includes a bucket of replies, and a single top-up price (€30 per 1,000) when you exceed the bucket. A reply is what it sounds like. The AI sent a message to your customer. That's the meter. It increments once, deterministically, regardless of whether the customer wrote back, escalated, gave a thumbs-up, or vanished.
Per reply is not perfect either. Some replies are more valuable than others. A reply that closes a refund question is worth more than a reply that says "let me check on that." We accept the imprecision because it has three properties that matter more than perfect alignment:
- It's predictable. You can read the meter and forecast the bill.
- It's countable. There's no judgment call. The number is the number.
- It doesn't punish improvement. Better retrieval, better answers, the bill stays flat or drops because fewer back-and-forths are needed.
The current tiers are simple. Free Starter is 50 replies a month, no card, one channel, one AI agent, so you can try the thing without a procurement conversation. Growth is €49 for 1,500 replies. Business is €149 for 6,000 replies. Scale is €449 for 25,000 replies with SSO. Go over your bucket, you pay €30 per extra 1,000. That's it. You can read the pricing page in about 30 seconds.
The math on a real-sized SMB
Let's run the same Shopify brand from earlier through Keloa. 8,000 monthly conversations, mature AI handling, say 1.8 AI replies per conversation on average (some are one-and-done, some take a couple of rounds). That's about 14,400 replies a month. At Keloa, that puts you on Business (6,000 included) plus 8.4 top-ups of 1,000 replies each, so €149 plus €252, around €401 a month total.
At $0.99 per resolution, the same brand at a 75% resolution rate pays about $5,940 a month, roughly €5,500 at current rates. The gap is about 13×.
That gap will not be 13× for everyone. If you have a high-touch product where conversations stretch to ten or fifteen turns, our per-reply model gets more expensive faster. We're up-front about that on the pricing page. For most SMBs in retail, services, and SaaS, where conversations average two to three AI turns, per-reply is dramatically cheaper, and more importantly, it's predictable.
What this means for you
If you're evaluating AI customer support tools right now, the single question worth asking every vendor is: "what counts as a resolution, and who decides." If the answer involves a 24-hour silence rule or a vendor-side classification model, that's a billable event that has nothing to do with your customer being helped. If the answer is "we count an AI reply, that's it," you can build a budget around it.
We're not claiming per-reply is a moral position. It's an engineering trade-off, and we picked it because it gives you the one thing per-resolution can't: knowing what your bill will be before the month ends.
Try the Free Starter plan (no card, 50 replies, ten minutes to set up), or book a 20-minute demo and we'll walk through your specific volumes and what the bill would look like. We'll also tell you if a competitor is cheaper for your shape of traffic. That happens sometimes, and we'd rather you know.