AI Support Guides
Last Updated:
Akash Tanwar
IN this article
Compare per-resolution and per-conversation AI support pricing, calculate the break-even point and audit every billable unit before you sign.
The short answer: Neither per-resolution nor per-conversation pricing is automatically cheaper or more transparent. Per-resolution pricing can align spend with successful automation, provided the outcome is defined and auditable. Per-conversation pricing can be economical when resolution rates are already high or interactions are short. Compare both on the same traffic, then choose on fully loaded cost per genuinely resolved issue, not the smallest-looking unit price.
Disclosure: Fini sells an AI support system with per-resolution pricing by default and per-conversation pricing available on request. This guide does not assume that either meter wins; the product-specific terms appear near the end.
On a rate card, AI chatbot pricing should be simple. One vendor charges for a resolution. Another charges for every conversation. A third bundles usage into a platform subscription.
Then the invoice arrives.
The difficulty is not the arithmetic. It is deciding what the unit means, when it becomes billable, and what happens to the customers the AI does not help. A cheap conversation that ends in a human handoff may still be expensive. A growing resolution bill may be good news if it replaces much more expensive manual work.
This guide gives you a practical way to compare the two most common usage meters without treating either one as inherently fair.
AI agent pricing models: what vendors actually charge for
AI customer service quotes can combine a platform fee with one or more usage charges. The label on the rate card rarely tells you the whole cost.
Pricing model | What triggers the charge | Where it can fit | What to inspect |
|---|---|---|---|
Per resolution or outcome | The AI is judged to have completed a defined result | Teams that want spend tied to successful automation | Outcome definition, false positives, reversals, handoffs and disputes |
Per conversation | A customer starts or continues a vendor-defined session | High-volume, short or consistently resolved interactions | Session window, reopens, channel changes, spam and failed attempts |
Per ticket or interaction | A ticket is created or the AI participates | Predictable helpdesk workflows | Duplicate contacts, multi-intent tickets and whether human-only tickets count |
Per seat | Each human agent or admin has a licence | Agent-assist products and stable teams | Seasonal staffing, contractors and feature-gated tiers |
Per token, message, minute or call | The system consumes a technical unit | Voice, API-first or highly variable workloads | Model markups, long conversations, retries and infrastructure fees |
Flat or hybrid | Subscription includes an allowance, then overage applies | Teams that value a committed budget | Minimum commitment, unused allowance, overage rate and annual true-up |
Ask for the complete rate card. Platform minimums, implementation, integrations, voice or model fees, premium support and usage overages can matter more than the headline unit.
Per-resolution vs per-conversation pricing: a side-by-side comparison
Per-resolution pricing charges when the AI is classified as having solved the customer's issue without human help. It is a form of outcome-based pricing: the vendor earns more when the AI completes more work.
In a pure per-conversation model, each billable vendor-defined session is charged regardless of resolution, subject to contractual exclusions. It is a form of usage-based pricing: the vendor earns more as billable contact volume grows.
Question | Per resolution | Per conversation |
|---|---|---|
What are you buying? | A defined successful outcome | Access for a defined interaction |
Main advantage | Spend follows completed automation | Simple volume forecasting when sessions are clearly defined |
Main risk | A loose resolution definition can overcount success | You can pay for failed, abandoned or escalated interactions |
Invoice as AI improves | At fixed conversation volume and an unchanged unit rate, rises with more resolutions | At fixed conversation volume and an unchanged rate and tier, stays flat |
Natural vendor incentive | Improve the number of billable outcomes | Process more conversations efficiently |
Best-fit pattern | Uncertain resolution rate or expensive human-handled support | High, stable resolution rate or low-handle-time traffic |
There is no definition-free meter. A conversation still needs a clock and a boundary. For example, Ada's current terms define messaging conversations around a 24-hour period and email conversations around 72 hours of inactivity. Those are clear rules, but different windows would produce different counts from identical customer behaviour. See Ada's customer terms and service-specific terms.
Resolutions need equally precise rules. Intercom's current outcome documentation specifies when an answer is assumed resolved, what is not billable and when a later customer reply reverses a charge. The important lesson is broader than any one vendor: auditability comes from the contract and the underlying records, not from the name of the billing unit. See Intercom's outcome rules.
Outcome-based pricing vs usage-based pricing
Outcome-based pricing asks, “Did the AI create the result we agreed to?” Usage-based pricing asks, “How much of the system did the customer use?”
That difference changes incentives, but it does not settle the commercial decision.
Outcome pricing is attractive when the result is valuable, objectively measurable and under the vendor's control. It becomes risky when an “outcome” is inferred from silence, bundles several customer intents into one label, or cannot be reviewed at conversation level.
Usage pricing is attractive when the unit is stable and easy to forecast. It becomes risky when every unsuccessful attempt is billable, session rules split one issue into several conversations, or quality improvements are entirely the buyer's problem.
A good procurement process tests both questions:
Is the meter verifiable? Can your team reproduce a sample invoice from raw conversation records?
Does the meter produce better economics? After human-handled support and operating costs, which option lowers the cost of serving customers well?
If either answer is unknown, the headline price is not decision-ready.
How customer service automation changes the cost curve
The AI invoice is not the same as the cost of customer service. A complete AI support TCO model also captures implementation, human-handled support and ongoing operations.
Use this model instead:
Fully loaded support cost = fixed platform cost + AI usage + implementation and integrations + human-handled support + QA and operations + rework
The biggest variable is often the work routed to human experts:
Human-handled cost = unresolved conversations × cost per human-handled issue
Suppose the AI resolves more customer issues this quarter. Under per-resolution pricing, the vendor invoice may increase. But if every added resolution prevents a more expensive human interaction, total support cost can still fall sharply. Calling the larger AI invoice a penalty for success ignores the work it replaced. The goal is not to remove human expertise; it is to reserve it for the cases that need judgment, empathy or authority.
The reverse is also true. A flat conversation bill is not evidence of ROI if resolution quality falls and the human queue grows.
Measure at least four things together: genuinely resolved issues, human-assisted volume, repeat contacts and customer satisfaction. No single pricing unit captures all four.
How to calculate chatbot pricing at your support volume
Start with a small model that both finance and support operations can inspect. Use your last 90 to 180 days of traffic and separate channels or issue types when their economics differ.
For a simple comparison:
Resolution invoice = conversation volume × genuine resolution rate × price per resolution
Conversation invoice = conversation volume × price per conversation
Invoice break-even resolution rate = price per conversation ÷ price per resolution
Here is an illustrative example, not a Fini quote or a market benchmark:
100,000 conversations
$0.80 per resolution
$0.40 per conversation
$8 average cost for each issue that reaches a human
Identical platform, implementation and quality costs in both quotes
Scenario | 40% genuine resolution | 60% genuine resolution |
|---|---|---|
Per-resolution AI invoice | $32,000 | $48,000 |
Per-conversation AI invoice | $40,000 | $40,000 |
Human-handled cost | $480,000 | $320,000 |
Total: resolution-priced option | $512,000 | $368,000 |
Total: conversation-priced option | $520,000 | $360,000 |
The invoice break-even point is 50% because $0.40 ÷ $0.80 = 0.50. Below a 50% resolution rate, the resolution-priced invoice is lower. Above it, the conversation-priced invoice is lower.
The more important result sits one line down. When genuine resolution rises from 40% to 60%, human-handled cost falls by $160,000. The resolution invoice increases by $16,000, yet fully loaded cost still drops by $144,000. Paying more to the AI vendor can coincide with paying much less to serve customers.
Real quotes add platform fees, included allowances, minimum commitments, overage bands and implementation costs. Put them into the model before using the break-even point. If two products produce different resolution quality, CSAT or repeat-contact rates, model those differences too; equal performance should never be assumed merely to simplify a rate-card comparison.
Which customer service AI tools have transparent pricing?
The most transparent tools publish both the unit price and the rules that create the unit. Public examples show different levels of disclosure:
Vendor | What a buyer can inspect publicly | What still needs confirming |
|---|---|---|
Fini | Per-resolution rates, resolution exclusions and reopen rules; per-conversation pricing is available on request | The conversation definition and the final allowance, commitment and quote for your traffic |
Intercom Fin | Outcome prices plus detailed resolution, reversal, abandonment and handoff rules | The full platform cost and which outcome types apply to your workflow |
Ada | Channel-specific conversation definitions in its public terms | The commercial rate, included volume and other order-form terms |
This is not an exhaustive vendor ranking, and pricing can change. It is a test for transparency: could a buyer explain the charge before signing, forecast it from real data and audit it afterward?
A vendor that publishes a price but leaves the billing rule vague is not fully transparent. Neither is a vendor with precise documentation but an undisclosed minimum that dominates the economics.
How to audit whether a resolution should be billable
Put the edge cases into the contract before they appear on an invoice. At minimum, define:
The issue boundary: does one conversation with three requests create one resolution or three?
Proof of success: does the customer confirm the result, does a downstream action complete, or is success inferred from inactivity?
Human involvement: does any handoff cancel the charge, and what happens when the AI collects information before escalating?
Reopens and repeat contacts: how long can the customer return before the original charge is reversed or protected from double billing?
Abandonment and noise: are greetings, spam, test conversations and unanswered clarifying questions excluded?
Channel changes: does moving from chat to email create another unit?
Quality failures: what happens when an answer is later found inaccurate, unsafe or contrary to policy?
Evidence and disputes: can you export every billable record, sample it, challenge it and receive a credit?
Do not settle for an aggregate dashboard. Each charge should trace back to a reviewable customer interaction and a versioned rule. During a pilot, manually review a statistically useful sample of “resolved” and “unresolved” conversations. Track false positives separately from false negatives: the first can inflate the bill and mask customer harm; the second can understate value.
Which AI pricing model fits your resolution rate?
Per-resolution pricing is often worth testing when resolution performance is new or variable, customer issues are complex, human handling is expensive, and the buyer wants vendor fees tied to completed work. The condition is a rigorous, reversible resolution definition.
Per-conversation pricing is often worth testing when resolution is already high and stable, interactions are short, conversation volume is predictable, and the flat session rate beats the outcome rate at realistic performance levels. The condition is a precise session definition and clarity on failed or duplicate interactions.
Neither description is a rule. Run at least three scenarios (conservative, expected and strong performance) and test volume spikes. Then compare:
Total annual cash cost
Cost per genuinely resolved issue
Cost per human handoff
Sensitivity to a 10% volume increase
Sensitivity to a five-point change in resolution rate
Quality and repeat-contact guardrails
Use the same traffic window, channels, issue mix and definition of success for every vendor. Otherwise, you are comparing assumptions rather than prices.
Questions to ask before signing an AI support contract
What exactly creates one billable unit?
Can one customer issue create more than one charge?
Are escalations, spam, greetings, tests and abandoned sessions free?
What inactivity window closes a conversation?
What reopen or repeat-contact window reverses a charge?
Can we export the records behind every billable unit?
How do we dispute false resolutions, and when are credits applied?
Which platform, implementation, integration, model and support fees sit outside usage?
What allowance, minimum, overage band or annual true-up applies?
Does unused volume expire, roll over or pool across channels?
Can we switch meters or tiers if our traffic behaves differently from the forecast?
Which CSAT, accuracy and repeat-contact thresholds protect customer outcomes?
Ask the vendor to answer with a sample invoice built from your own historical traffic. It is the fastest way to expose an ambiguous definition or a hidden cost.
How Fini prices AI customer support
Fini uses per-resolution pricing by default. A resolution is an issue solved end to end without a human handover. Escalations are not billed; greetings, spam and abandoned sessions are excluded; and a conversation reopened within 72 hours is not billed twice. Fini also offers per-conversation pricing on request for high-volume or low-handle-time profiles.
The current Fini pricing page publishes per-resolution rates for Growth, Scale and Enterprise and explains the main billing rules. Fini says it will model both meters against a customer's real traffic, recommend the one that costs less and allow a switch at renewal. Check the live page or your order form for the current allowance, commitment and rate before making a budget decision.
That is the standard every buyer should demand, from Fini included: define the unit, show the math and make the charge auditable.
Want to compare both models on your support traffic? Review Fini's pricing or contact the team to model your expected invoice and total cost.
AI agent pricing FAQ
Is per-resolution pricing always cheaper than per-conversation pricing?
No. At the same conversation volume, per-resolution pricing tends to have the lower vendor invoice below the break-even resolution rate; per-conversation pricing tends to be lower above it. Platform fees, allowances and differences in actual product performance can reverse that result.
What counts as a conversation in AI customer service pricing?
A conversation is a vendor-defined session, not a universal unit. The definition should state the inactivity window, channel boundary, reopen rule, multi-intent treatment and exclusions. Two vendors can count different numbers of conversations from the same traffic.
What is outcome-based AI pricing?
Outcome-based pricing charges when the AI completes an agreed result, such as resolving an issue without human help. It aligns spend with output only when the outcome is valuable, precisely defined, auditable and reversible when later evidence shows it was not achieved.
How much does an AI customer service agent cost?
There is no useful universal number. Calculate platform and implementation fees, expected usage, human-handled support, QA and rework at your own volume. Then compare the fully loaded cost per genuinely resolved issue across vendors.
Which metric should procurement compare?
Start with total annual cost and cost per genuinely resolved issue. Keep resolution accuracy, CSAT, human handoffs and repeat contacts beside the financial model so a low bill cannot hide poor service.
GTM Lead
