Deflection pricing pays a vendor to make a ticket disappear. Resolution pricing pays a vendor to fix the problem. Those are not the same thing, and the gap between them is where most AI support budgets quietly leak.
Per-resolution and per-deflection pricing are the two dominant ways AI customer support is billed in 2026. Per-deflection charges when the AI prevents a ticket from reaching a human. Per-resolution charges when the AI actually closes the customer's issue. The distinction sounds academic until you read the contract: under deflection pricing, a customer who gets a wrong answer, gives up, and never replies counts as a billable win. Under resolution pricing, that same interaction should cost you nothing.
Deflection pricing rewards containment, not correctness. A bot that stalls a frustrated customer until they abandon the chat books revenue for the vendor and a CSAT hit for you.
Per-resolution pricing only aligns incentives if the customer, not the vendor, defines what counts as a resolution and escalations are not charged.
Lorikeet prices at roughly $0.80–$0.95 per chat, email, or SMS resolution and roughly $1.20–$1.50 per voice resolution, the customer holds veto over what counts, and escalations are never billed.
The human baseline for comparison is roughly $1.25 to $4.00 per human-handled ticket, so the real question is total cost per genuinely fixed issue, not the lowest sticker price per event.
The cheapest per-unit price can be the most expensive program once you add reopened tickets, rework, and the support headcount that deflection numbers quietly hide.
Last updated: June 2026
Pricing models are not a procurement footnote. They are the strongest signal of what a vendor optimizes for, because a vendor will always optimize for the thing it gets paid on. If the meter runs on deflection, the product gets better at deflecting. If the meter runs on resolution as the customer defines it, the product gets better at resolving. This explainer breaks down how the two models actually differ, why deflection pricing misaligns incentives in ways that surface six months into a deployment, what the word resolution should mean and who gets to define it, why uncharged escalations matter more than the headline rate, and how to run the total-cost math that vendors would rather you skip.
What Per-Deflection and Per-Resolution Pricing Actually Mean
Per-deflection pricing charges you each time the AI handles an interaction without passing it to a human agent. The billable event is containment. The customer asked something, the bot replied, and no human got involved. Whether the customer's problem was solved is not part of the accounting.
Per-resolution pricing charges you each time the AI resolves a customer's issue end-to-end. The billable event is the outcome. The customer wanted a refund processed, a card locked, a transfer traced, or a question answered correctly, and that happened. If it did not happen, there is nothing to bill.
Deflection: An interaction the AI handled without escalating to a human, regardless of whether the customer's underlying issue was solved. The metric measures avoided human contact.
Resolution: A customer issue closed to the customer's satisfaction, end-to-end, by the AI. The metric measures fixed problems.
The two can look identical on a dashboard and diverge completely in reality. A deflection rate of 70% tells you 70% of contacts did not reach a human. It does not tell you how many of those customers got what they needed, how many gave up, and how many came back angrier the next day under a new ticket ID. That ambiguity is not an accident. It is the feature that makes deflection pricing attractive to sell and dangerous to buy.
Why Deflection Pricing Misaligns Incentives
The core problem with deflection pricing is that the billable event and the customer's goal point in different directions. You want problems solved. The pricing model pays for problems contained. Most of the time those overlap, which is exactly why the misalignment is hard to catch. It only shows up at the edges, and the edges are where support reputations are made or lost.
Consider the failure modes deflection pricing quietly rewards:
The give-up. A customer asks a question the bot cannot answer, gets a generic non-answer, and closes the window in frustration. No human was involved, so the interaction is deflected and billable. The problem is unsolved and the customer is now unhappy. You paid for that.
The loop. The bot keeps the customer engaged with clarifying questions and canned articles long enough that the session ends without a handoff. Containment achieved. Resolution not attempted.
The reopen. The customer gets a wrong answer, accepts it, then returns the next day when it does not work. That is two deflections billed for one unsolved problem.
The wrong answer that sticks. The most expensive case in a regulated business. The bot confidently gives an incorrect answer, the customer acts on it, and the issue never reopens because the damage is now downstream. Fully deflected, fully billable, and a compliance problem rather than a support one.
None of these are exotic. They are the predictable result of paying a vendor to avoid human contact rather than to produce correct outcomes. A vendor billing on deflection has no financial reason to escalate a hard ticket, because escalation is lost revenue. That is precisely backwards from what you want, which is for the AI to escalate fast and cleanly the moment it is out of its depth.
Outcome-based pricing was supposed to fix this, and at its best it does. But many outcome models smuggle the same misalignment back in through the definition of the outcome. If the vendor decides what counts as a resolution, you are back to a metric the vendor optimizes in its own favor. The alignment only holds when the customer controls the definition.
What Resolution Should Mean and Who Defines It
A resolution is only a useful billing unit if it maps to a real outcome the customer cares about. The single most important clause in any per-resolution contract is therefore not the rate. It is the answer to one question: who decides whether a given interaction was resolved?
There are three possible answers, and only one of them protects the buyer.
The vendor defines it. The vendor's system marks tickets resolved based on its own logic, often a proxy like "no reply within 24 hours" or "the AI sent a final message." This is deflection pricing wearing a resolution label. Silence gets counted as success, and silence is frequently the sound of a customer giving up.
A shared rule defines it. Both parties agree to a fixed definition up front. Better, but rigid, and it still tends to favor whoever wrote the rule.
The customer defines it, with veto. The buyer decides what counts as a resolution for their business and retains the right to reject any charge for an interaction that did not meet that bar. This is the only model where the vendor's revenue is genuinely tied to your outcomes.
Lorikeet uses the third model. The customer holds the veto on what counts as a resolution. If an interaction did not actually solve the customer's issue, the customer can decline to be billed for it. That single design choice removes the incentive to count give-ups, loops, and wrong-but-unreopened answers as wins, because the buyer can simply refuse to pay for them. It also changes what the product optimizes for. When the vendor only earns money on outcomes the buyer agrees were real, the product has to get genuinely good at solving the hard tickets rather than at making them disappear from the queue.
Defining resolution well also means defining it per workflow. A resolution for a password reset is a low bar. A resolution for a disputed transaction means the dispute was correctly filed, the customer was told what happens next, and no regulatory step was skipped. A serious per-resolution model lets you set that bar differently for different ticket types, because a flat definition either overpays for easy tickets or undercounts hard ones.
Why Escalations Should Never Be Charged
The treatment of escalations is the cleanest test of whether a pricing model is honestly aligned. An escalation is the AI recognizing it cannot or should not resolve a ticket and handing it to a human. That is the behavior you want. It is the safety valve that keeps a confident-but-wrong answer from reaching a customer in a regulated workflow.
If escalations are billed, you have created a tax on the exact behavior you are trying to encourage. The vendor is paid more when the AI gives up cleanly, and the AI is implicitly nudged to attempt tickets it should have routed to a human. If escalations are not billed, the incentive flips correctly: the vendor only earns when the AI fully resolves the issue itself, so the AI is built to resolve confidently when it can and escalate honestly when it cannot, with no revenue penalty for the honest choice.
Lorikeet does not charge for escalations. An interaction that is handed to a human is not a resolution and is not billed. Combined with customer-defined resolution, this means the vendor is paid for one thing only: issues the customer agrees the AI actually fixed. Escalations, give-ups, and rejected interactions all cost nothing. That is the structural difference between a model that is aligned by design and one that is merely labeled as outcome-based.
When you evaluate any per-resolution vendor, ask the question directly. Are escalations billed? Are interactions where the customer abandons the chat billed? Can I, the customer, dispute a charge for an interaction I do not consider resolved? If the answers are yes, yes, and no, you are looking at deflection pricing with better marketing.
The Total-Cost Math Vendors Skip
The headline per-unit rate is the number every vendor leads with and the least useful number for a buying decision. What matters is total cost per genuinely resolved issue, which includes the events you were billed for that did not actually solve anything.
Start with the human baseline. A human-handled support ticket costs roughly $1.25 to $4.00 in fully loaded agent time, depending on complexity and region. That is the number AI has to beat, and on simple tickets it beats it easily. The trap is assuming a low per-deflection rate automatically wins.
Work a simple example. Suppose a vendor charges $0.50 per deflection and reports a 70% deflection rate on 100,000 monthly contacts. That is 70,000 deflections at $0.50, or $35,000 a month, and the dashboard looks excellent. Now suppose that of those 70,000 deflections, 20% were give-ups or wrong answers that the customer brought back as a new ticket. Those 14,000 unsolved interactions were billed, then reopened, then either billed again or escalated to a human at full human cost. You paid for containment twice and resolution never. Your true cost per resolved issue is meaningfully higher than $0.50, and your CSAT absorbed the difference.
Now run the same volume against per-resolution pricing where the customer defines resolution and escalations are free. At roughly $0.80–$0.95 per chat, email, or SMS resolution, you are billed only on the issues actually closed. The give-ups cost nothing. The escalations cost nothing. The reopened tickets, if the first attempt did not genuinely resolve, were never billed in the first place. The per-unit number is higher, but you are paying for fixed problems rather than for events, and there is no shadow line item for rework.
For voice, Lorikeet prices at roughly $1.00 per resolution, still well inside the human baseline for the complex calls voice tends to carry. Coach, the automated QA agent that scores quality on every ticket, runs at roughly $0.25–$0.30 per ticket and is what lets you verify that resolutions are real rather than taking the meter's word for it.
The math that actually decides a deployment is this: take total annual spend, divide by the number of issues your team agrees were genuinely resolved, and compare that to your fully loaded human cost per ticket. A deflection model can win on the first number and lose badly on the second. Insist on the second number before you sign.
How to Spot Deflection Pricing Wearing a Resolution Label
Most vendors in 2026 use the word resolution because it sells better than deflection. The label is cheap; the structure underneath is what you are actually buying. A handful of contract questions separate a genuinely aligned model from deflection pricing in nicer clothes, and they are worth asking in the first sales call rather than discovering the answers six months into a deployment.
Who marks a ticket resolved, and on what signal? If the answer is the vendor's system, and the signal is something like no customer reply within a set window, that is deflection. Silence is being sold to you as success. The aligned answer is that the customer defines the bar and can reject the call.
Can I dispute a charge? If you cannot decline to pay for an interaction you do not consider resolved, the vendor's definition is the only one that counts, and the vendor wrote it to be generous to the vendor.
What happens to a customer who abandons mid-chat? If an abandoned session is billable, you are paying for give-ups. If it is not, the model is tracking real outcomes.
Are escalations billed? A model that charges for handoffs is taxing the safe behavior and is, in effect, deflection pricing with extra steps.
Can resolution be defined per workflow? A flat definition either overpays for trivial tickets or quietly undercounts the hard ones. Per-workflow definitions are a sign the vendor expects to be measured on real outcomes.
If a vendor cannot give you clean answers to those five questions, the resolution label is marketing. The structural tell is always the same: the more control the vendor keeps over the definition of success and the billing of escalations, the closer the model sits to deflection regardless of what the rate card calls it.
How Lorikeet Prices, and Where It Fits
Lorikeet is an AI customer support platform built for complex and regulated businesses, where a wrong-but-deflected answer is not a CSAT problem but a compliance one. Its pricing reflects that posture. Resolutions are billed at roughly $0.80–$0.95 for chat, email, and SMS and roughly $1.20–$1.50 for voice. The customer defines what counts as a resolution and holds a veto over any charge. Escalations to humans are never billed. Coach, the QA agent, adds roughly $0.25–$0.30 per ticket for 100% automated quality scoring so you can audit whether billed resolutions held up.
This is not the cheapest sticker price in the category. A pure deflection vendor will quote a lower per-event number, and a pure outcome vendor that defines its own outcomes will too. The argument for Lorikeet's model is alignment, not headline price. Because the meter only runs on customer-agreed resolutions and escalations are free, the product is built to resolve the hard, regulated tickets correctly and to escalate honestly when it should, rather than to maximize containment. For a business where the cost of a confidently wrong answer is a regulatory notice rather than a refund, that alignment is the whole point.
The honest limitation: if your support is almost entirely simple, high-volume, low-stakes deflection where a wrong answer costs nothing and reopens are cheap, a bare-bones deflection tool at a lower per-event rate may post a lower total cost, and you may not need customer-defined resolution or audit-grade QA. The per-resolution model with veto and free escalations earns its keep when correctness matters and the hard tickets are the ones you most need handled.
If you are comparing AI support pricing models, see how Lorikeet prices per resolution with customer veto and free escalations.
Key Takeaways
Per-deflection pricing bills for avoiding human contact; per-resolution pricing bills for fixing the customer's issue. They diverge exactly at the hard tickets that matter most.
Deflection pricing rewards give-ups, loops, reopens, and confident wrong answers, because all of them count as contained without being solved.
Per-resolution pricing only aligns incentives if the customer defines what counts as a resolution and holds a veto over charges.
Escalations should never be billed, or you are taxing the exact safe behavior you want the AI to take.
The decisive number is total cost per genuinely resolved issue against a $1.25 to $4.00 human baseline, not the lowest per-event sticker price.
