Three of the help desks small businesses most often shortlist now sell their AI the same way: you pay for each customer question the software closes without a human touching it. As of November 2026, Help Scout charges $0.75 per resolution for its Envoy feature, Intercom charges $0.99 per outcome for its Fin agent, and Zendesk charges $1.50 per automated resolution on a committed plan or $2.00 on pay as you go. That per answer price turns the buying decision into arithmetic you can run against your own ticket count.
The price is per answered question
Seat fees usually sit on top of the per answer charge. On Intercom's pricing page, the Essential plan lists at $29 per seat a month on annual billing, currently discounted to $19 for new customers, Advanced is $85 and Expert is $132, each with Fin included but billed separately per outcome. Intercom also sells Fin on a help desk you already own at "$0.99 per outcome" with "No seats required," subject to a note that a "Minimum monthly commitment applies (e.g. 50 outcomes)."
Zendesk's pricing page lists Support Team at $19 per agent a month paid yearly, Suite Team at $55 and Suite Professional at $115. Team plans include 5 automated resolutions per agent per month and Professional includes 10, after which the committed rate is $1.50 and the pay as you go rate is $2.00. Help Scout prices seats at $25, $45 and $75 per user a month for Standard, Plus and Pro, with Envoy as a $0.75 per resolution add on.
What counts as a resolution is defined by the vendor
The definitions differ enough to change the bill. Help Scout counts a resolution as "a single conversation that is resolved by AI Answers without human assistance," and only "if the customer receives an AI response and doesn't use escalation, search the knowledge base, ask more questions, or indicate that they need more help." Intercom says Fin "usage is measured in resolutions" so "you only pay when Fin does what you care about most: resolving a customer's question," but its pricing page counts a completed workflow, including a handoff, as an outcome. Zendesk's help documentation counts a ticket automation resolution only where "the AI agent resolved the issue without a human agent needing to respond," and counts its older article suggestion flow as resolved when the customer "did not reply again within 72 hours."
Tidio, a cheaper entry point, prices its Lyro agent as a standalone add on starting at $32.50 a month for 50 Lyro AI conversations, and only counts a conversation as billable when a human agent sends a message: "Responses from Flows and Lyro AI Agent are not counted as 'agent replies'."
Before signing anything, ask the vendor in writing what happens in three cases: the AI answers and the customer writes back anyway, the AI hands the conversation to a person, and the AI answers wrongly. Some vendors bill a handoff as an outcome and some do not.
What answering by hand costs
The Bureau of Labor Statistics puts the median hourly wage for customer service representatives at $21.53 in May 2025. That is wage only, before payroll taxes, benefits or software, so the loaded cost of an hour is higher. When the owner is the one answering emails at 11pm, the comparison is whatever else that hour would have produced.
The arithmetic needs two numbers about your own business: how many questions arrive in a month, and how long an average one takes to answer. Both are usually already in your inbox or help desk, and both are usually guessed too low.
Example: a shop gets 300 customer questions a month and takes about 6 minutes on each, which is 30 hours. At $20 an hour of someone's time, that is $600 a month. If an AI agent resolves half of them at $1 each (a round number between the $0.75 and $2.00 rates above), the bill is $150 for the AI plus 15 hours of human time at $20, or $300, for a total of $450. The saving is $150 a month before the seat fee. One $45 seat takes a third of it. At a 25 percent deflection rate instead of 50 percent, the saving falls to $75 before the seat fee.
At low volumes the seat fee eats most of the saving. The cost per answer improves as volume rises and as the same questions repeat.
Fewer small firms use AI than the marketing implies
The Census Bureau added questions on AI use to its Business Trends and Outlook Survey from November 17, 2025 to February 8, 2026. A Census working paper on that data, The Microstructure of AI Diffusion, reports that "18% of firms used AI in a business function, rising to 32% on an employment-weighted basis; adoption is expected to reach 22% within six months." Among firms that had adopted something, sales and marketing was the most common function at 52 percent, ahead of strategy and business development at 45 percent and IT at 41 percent.
The Minneapolis Fed, writing on the same survey on May 29, 2026, reported that 30 percent of US firms with at least 250 employees said they use AI while 17 percent of firms with fewer than 20 employees do. It also noted that "roughly 96 percent of companies using AI nationwide reported no change" in employment over the prior six months. Adoption is growing, but it is not the near universal shift the sales emails describe, and it has mostly not displaced people.
The failure modes have already been litigated
If your chatbot states a policy that is not your policy, you may be bound by it. In February 2024, Canada's Civil Resolution Tribunal ordered Air Canada to pay a customer CA$812 after its chatbot told him he could buy a full fare ticket and claim a bereavement discount afterward, which contradicted the airline's own policy page. As CBS News reported, the tribunal member wrote that "It makes no difference whether the information comes from a static page or a chatbot." The lesson is that refunds, warranties, pricing exceptions and anything legal should be routed to a person, not answered by the model.
Vendors themselves concede a residual error rate. Intercom's own Fin FAQ page says the agent has "a very low hallucination rate (<1%)" and tells customers they can set escalation rules so Fin "automatically transfers sensitive topics (e.g., refunds, cancellations, sensitive data etc.) to a human agent." Even an error rate under 1 percent produces the occasional wrong answer, which is fine for store hours and not fine for a refund policy.
Be equally skeptical of the sellers. On March 24, 2026, the Federal Trade Commission announced a settlement banning Air AI and its owners from marketing business opportunities, over charges including that they "Falsely claimed that people who purchase their services will or are likely to make substantial earnings." The order carried an $18 million judgment, largely suspended for inability to pay, with $50,000 paid for consumer relief. Any pitch that leads with earnings rather than with a resolution rate on your own content deserves a hard look.
A first deployment that limits the damage
- Export the last few months of customer messages and sort them into buckets. If the top five buckets are not most of the volume, an AI agent will not deflect much.
- Write or clean up the help pages that answer those five buckets. Every one of these tools answers from your content, so bad content produces confident wrong answers.
- Turn the agent on for those topics only, with escalation rules sending refunds, cancellations, complaints and anything involving personal data to a human.
- Read every AI conversation for the first month and count real resolutions yourself, since the vendor's definition and yours may differ.
- Check the data terms before any customer records go in. Zendesk's documentation states that "customer data is not used to train third-party models" and that third party models "operate under strict zero data retention policies." Ask any vendor for the same two commitments in writing, and ask who the subprocessors are.
The honest limit of all this: per answer pricing makes the cost legible, not small. A business that gets only a handful of repetitive questions a month will usually find the seat fee swallows the saving, and a clearer FAQ page with a canned reply template does more. The per resolution model earns its keep when the same ten questions arrive every week and the answers are already written down.
