YundaDeskThe most underestimated line in an AI support quote is not the unit price. It is the billing unit.
Pay-per-resolution sounds clean: if the AI resolves the issue, you pay; if it does not, you do not. The problem is that support is not ad buying. Whatever you charge for, teams learn to optimize around. Over time, per-resolution pricing pushes agents, managers, and AI rules toward an awkward place: let the AI resolve less, keep boundaries tighter, and teach the system more slowly.
Pricing models rewrite team behavior
A good AI support pricing model should answer more than “what will this cost this month?” It should also answer “what behavior will this encourage?”
If every independent AI resolution becomes an extra charge, a support lead naturally starts doing the math: should this topic go to a human first? Should the knowledge base stay less complete? During peak season, should part of the automation be turned off to control the invoice?
Each decision can sound rational. Together, they damage support quality. Customers wait longer, agents go back to repetitive replies, and the AI gets fewer chances to become more accurate from real conversations.
The per-resolution incentive: better AI, higher bill
The core problem with per resolution pricing is simple: the better the AI resolves issues, the more you pay.
That creates a strange mental ledger. You clean up the knowledge base, correct weak answers, and train high-frequency scenarios. Then the AI improves, and the invoice rises with it. The team starts asking a question it should never have to ask: do we really want the AI to handle more?
Under per-resolution billing, peak-season invoice pressure scales with resolved volume
| What gets affected | What should happen | What can happen under per-resolution pricing |
|---|---|---|
| Knowledge base | Cover common questions | Document only the minimum |
| AI boundaries | Let AI answer low-risk issues and hand off high-risk ones | Send routine issues to humans too often |
| Review and learning | Turn human replies into skills | Review cost first, and hesitate to expand coverage |
This is not about teams being stingy. People respond to invoice pressure. Any model that ties “AI gets better” to “the bill gets bigger” weakens the motivation to keep improving the system.
It also distorts agent judgment
Agents should judge the customer problem: does the knowledge base provide evidence? Is the customer asking for a human? Does this involve a refund, compensation, or price change that needs approval?
But when the team is watching this month’s resolution count, attention shifts. Shipping, fulfillment, and sizing questions that should be automated stay manual. Good agent replies do not get added to the knowledge base. People start debating whether something “counts as resolved” instead of whether the customer received an accurate answer.
Included credits treat AI as capacity, not a penalty
YundaDesk uses a simpler model: AI credits are included in every plan, with no per-conversation or per-resolution surcharge. You buy a predictable AI usage allowance, not a new charge every time the AI succeeds.
That changes the operating habit. A support lead does not need to run a silent cost calculation before each inquiry. Routine questions can go to AI first. The knowledge base can be improved without hesitation. When an agent corrects the AI, the system creates a learning suggestion, and it only takes effect after owner approval. The AI gets smarter over time, but the bill does not jump because one more issue was resolved.
This matches YundaDesk’s boundary. The AI agent faces customers and answers 24/7 from the knowledge base. If it lacks evidence, if the customer asks for a human, or if the issue is high risk, it hands off. Yuna is for merchants: it helps with data, configuration, and teaching experience back into the AI agent, but it does not speak to customers. For more on that boundary, see /en/blog/ai-first-human-backed-boundary/.
A healthy pricing model should improve quality
For cross-border sellers, the point of AI support is not unattended support. The point is to catch repetitive, low-risk, evidence-backed questions first, so humans can handle refunds, compensation, complaints, price changes, and other decisions that need approval.
That means the pricing model should meet three tests:
- Predictable: When peak season arrives, the owner should know roughly where the budget lands.
- Does not punish improvement: A better knowledge base and more accurate AI should not become a financial penalty.
- Does not encourage unsafe shortcuts: Refunds, compensation, and price changes should always require human approval.
Included credits put cost around a usage allowance, not around a result that can be argued over. The team can focus on the right work: improving the knowledge base, tuning handoff boundaries, reviewing missed questions, and turning experience into traceable, testable, revertible support capability.
On a quote, do not only ask for the price
When choosing an AI support platform, the better question is not “how cheap is it?” It is “how will this pricing model make my team use the product?”
Put these questions into your buying checklist:
- Is the fee based on seats, conversations, AI credits, or resolutions?
- Are AI credits included in the plan? If we go over, is the overage rate clear?
- If the AI resolves more issues, does my bill rise in parallel?
- If peak-season inquiry volume doubles, is there a predictable rule?
- Do high-risk actions always require human approval?
- Do learning suggestions require owner confirmation, and are they traceable, testable, and revertible?
You can pair this with the evaluation framework in /en/blog/choosing-ai-support-platform/ and confirm billing predictability before the feature demo gets too detailed.
Why YundaDesk does not bill per resolution
YundaDesk is built for Chinese global brands and cross-border e-commerce teams. Their support pressure usually does not come from one complex case. It comes from hundreds or thousands of repeated questions arriving in the same week: shipping, fulfillment, sizing, discount codes, exchange rules, social DMs, and multilingual follow-ups. AI should absorb that volume first, not make the team wonder whether every successful answer will trigger another charge.
Included credits still come with clear boundaries. High-risk actions go through human approval. If AI cannot answer, it hands off. When an agent corrects the AI, the correction becomes a learning suggestion, and it only takes effect after owner approval. We want the AI to get smarter over time, but that learning must stay controlled, traceable, and revertible.
The biggest problem with per-resolution pricing is not that it can be expensive. It is that it makes “AI resolved more” something the team has to think twice about. For teams that truly want to scale support capacity, the healthier model is one where better AI makes people more willing to use it, not more worried about the invoice.