107 support AIs. Only 43 show a price.

Customer service is the first business function where AI buying has become a board-level deadline. In a Gartner survey published in February 2026, 91% of customer service leaders said they were under pressure to implement AI this year. The pressure is real. The price list is not. When ETHORITY reviewed 107 customer-service AI offerings for the Good AI Prism Global Market Study, only 43 published a price on their own pages. The other 64 asked buyers to talk to sales first.

That gap shapes every shortlist. A buyer can see that a product exists, what it promises and which channels it covers. The facts that decide a contract, such as the pricing unit, the hosting region and the use of customer data for training, sit on a pricing page for some providers, in a trust centre for others and nowhere public for many.

60 of 107 talk to customers; only 10 help the human agent

The Prism places offerings where the work happens. In Customer Service that means chat and voice agents, agent assist, knowledge and self-service, and omnichannel suites. The 107 offerings in the study were chosen by buyer job, not by size or popularity, and each fact carries a dated first-party source.

Read across the cohort, the documentation thins out quickly. Hosting is described at all for 88 of 107. Named integrations appear for 77. A transparent pricing unit, meaning something other than "contact sales", appears for 64. A published price appears for 43. An EU hosting option is documented for 38. A stated limit on training with customer data, either "not used" or "only with opt-in", appears for 23.

Those numbers describe what providers publish, not what their products can do. A missing page is a question for the provider, not a verdict. But for a European buyer with data-protection duties, the two facts that matter most, where the data lives and whether it trains someone else's model, are the two facts least often on the record.

64 of 107 customer-service AI offerings state a transparent pricing unit. That is the minimum a finance team needs to model cost per resolution.

What the evidence says the tools can do

The strongest independent evidence on support AI is still a field study. Brynjolfsson, Li and Raymond tracked 5,179 support agents and found that a generative AI assistant raised issues resolved per hour by 14% on average, and by 34% for less experienced agents. Customer sentiment improved and staff turnover fell.

The study measured an assistant that helps a human agent. It did not measure a fully autonomous agent that closes cases on its own. That distinction matters for 2026 buying, because most vendors now market agents. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, mostly on cost, unclear value or weak risk controls. The safer reading of the evidence: start where a human stays in the loop, measure resolutions and quality, then widen the agent's scope as the numbers hold.

+34% resolutions per hour for less experienced agents with an AI assistant, in a field study of 5,179 support agents (NBER Working Paper 31161).

Three pricing logics, three cost curves

The 12 Customer Service profiles in the Registry show how differently providers charge. Twilio ConversationRelay publishes a usage price, from $0.07 per minute plus voice charges. NFON's AI VoiceBot sells a monthly subscription. Several suites bundle AI into seats or editions and quote the rest.

Each logic produces a different cost curve. Per-minute pricing scales with call volume and call length. Per-resolution pricing rewards the provider when the bot closes a case, so the definition of "resolved" becomes a contract term. Seat pricing looks predictable until the AI starts handling work that used to need seats. A buyer who compares these on list price alone compares three different things. Put two or three candidates into Compare side by side and normalise to one unit: cost per resolved contact at your own volume.

Scenario. A queue of 10,000 voice contacts a month at four minutes each uses 40,000 minutes. At Twilio's listed ConversationRelay rate of US$0.07 per minute, that is US$2,800 a month before voice charges. On a per-resolution plan the same queue costs whatever "resolved" means in your contract, so write that definition first.

Where the data lives, and what it trains

Hosting and training terms divide the cohort more sharply than features. NFON states that its bots run in German data centres. Dialpad states that permanent data, including recordings and transcripts, is stored in the United States. Genesys and Ada state that customer data trains models only with opt-in. Intercom documents a training opt-out for Fin, which means the default matters. Agentforce, Cognigy and Zendesk document how third-party model providers treat data, which answers part of the question but not the provider's own use.

None of these terms is right or wrong in the abstract. Each is right or wrong for a given contract, data class and jurisdiction. What changes in 2026 is that the questions now have dates. Under the AI Act, as amended by Regulation (EU) 2026/1744, the Article 50 transparency duties apply from 2 Aug 2026: people must be told when they are dealing with an AI system. A voice bot that does not say it is a bot is now a compliance issue, not a design choice.

What a defensible contract contains

The evidence points to a short list of contract terms that turn a promising pilot into a service you can run. First, a written definition of a resolved contact, including what happens when a customer comes back within seven days. Second, the pricing unit and the price at three volume levels: today, double and half. Third, the hosting region for transcripts, recordings and logs, not only for the application. Fourth, a statement on whether your data trains the provider's own models, and how to switch it off if the default is on. Fifth, an exit clause that returns your knowledge articles, conversation logs and bot configurations in a usable format.

None of these terms is exotic. Most procurement teams already use them for telephony or CRM contracts. What is new is that support AI combines all of them in one product, and that the provider's answers can change with each model release. Ask for the version and date of every answer, and keep them with the contract.

Why the documentation gap is a buying signal

Providers that publish their pricing unit, hosting regions and training terms make a buyer's work shorter. Those that do not are not necessarily worse, but they move the burden of proof into the sales cycle. ETHORITY's companion piece on why a vendor claim is not evidence sets out how to tie every claim to a product, version and contract before relying on it.

The practical consequence: treat documentation as a selection criterion. A provider that can answer the six decision questions in writing within a week is easier to govern for the next three years.

What to do Monday

1. Name the job and the measure. Pick one queue, such as order status or password resets, and set the baseline: cost per resolution, resolution rate and CSAT. Our note on saved hours versus value explains why hours alone mislead.
2. Shortlist three, not thirty. Start from the Customer Service area map and pick three offerings that cover your channel, language and hosting needs.
3. Ask the six questions in writing. Pricing unit, published price, hosting region, training on your data, integrations with your CRM and ticketing, and the provider's own limits.
4. Pilot assist before autonomy. Run an agent-assist pilot for six weeks with a stop rule, then decide whether to widen scope.
5. Check the Article 50 notice. Every chat and voice flow should tell customers they are talking to AI from day one.

If the shortlist needs a written basis for a budget decision, the AI Decision Brief compares up to three offerings fact by fact, and the Sales & CRM topic area covers the systems your service agents hand cases to.