A customer asks a support assistant to reset a password. It returns a help article, the customer stops replying, and the conversation eventually appears as resolved. The software has completed a measurable event. Whether it removed the customer’s problem is a separate question, with consequences for both the bill and the supplier’s business.
AI support products increasingly put a commercial unit around that workflow. Reading the unit carefully matters more than lining up the headline prices. A resolution, a conversation, an action and a session can describe different pieces of work. None, by itself, tells us the cost of keeping a customer satisfied.
Fact: the resolution has a clock
HubSpot (NYSE: HUBS) advertises its Customer Agent at US$0.50 per messaging resolution, using HubSpot Credits. Voice is separately priced at US$0.50 per minute. It is available with Professional and Enterprise subscriptions; some subscriptions include monthly credits. The advertised unit price therefore does not represent the entire cost of running a support operation.
The important detail is in the September 2026 documentation. For the reply-based route, the agent must provide a content source or perform an action, with no qualifying human handoff within 72 hours of the last visitor response. Qualifying handoffs are visitor-initiated. Manually assigning a human, including through a workflow, does not alone prevent the resolution count. Lead qualification supplies another route to resolution.
The reply-based status is evaluated after that window; later actions do not change that previously set status. Email has a further billing consequence: a customer reply after more than 72 hours reopens the conversation and resets the evaluation window. Another resolution by the customer agent then consumes additional credits. One email thread can therefore produce more than one billable resolution over time. These are defined billing and reporting rules. Our interpretation is that it should be read alongside customer-quality evidence, rather than treated as an independent satisfaction survey.
Customer support · definition
When the resolution is assessed
After the last visitor response
HubSpot reply-based resolution rule
Fact: the alternatives count other things
Salesforce (NYSE: CRM) lists Agentforce Flex Credits at US$500 per 100,000 credits. A standard Agentforce action uses 20 credits; voice actions use 30. An action can be a record update or an answer to a product enquiry. The page also offers conversation-based pricing and user licences. These are purchasing options, not one universal tariff for every deployment.
Freshworks (NASDAQ: FRSH, Class A common stock) defines a Freddy AI Agent session as interactions between an end-user and the agent or chatbot within a 24-hour window. Its August 2026 session FAQ specifies US$49 for 100 sessions for first-time add-on purchases from 13 November 2025. It preserves legacy provisions for existing customers and says unused sessions expire with the billing cycle.
Those definitions prevent a neat price ranking. A standard action is one step; a session groups interactions over time; a resolution depends on its provider’s rule. We have not run the same support cases through these products or examined negotiated contracts. Treating their listed units as interchangeable would supply a comparison the evidence does not support.
Interpretation: follow the unresolved work
Imagine a hypothetical business receives 1,000 support requests in a month and spends £500 on an automated service. Suppose a separate quality check finds 500 requests were successfully completed without further work. The automation spend is £1 per independently completed request. If only 250 meet that test, it becomes £2. The invoice is unchanged; the useful output has halved.
These are invented inputs for arithmetic, not measurements of any named product. They deliberately leave out subscription fees, implementation, content maintenance and human support. Adding those costs would make the assessment more complete. The point is to keep the denominator attached to the outcome being assessed.
Human work can move rather than disappear. A short automated exchange might leave a more complicated case for a person. Alternatively, a good answer could prevent a follow-up entirely. A finance team needs evidence about both paths before translating automated activity into labour savings. Time released for other work can be valuable even when payroll does not fall; it should be described as capacity rather than an immediate cash saving.
The same care applies to the supplier. More completed cases can generate consumption revenue, but delivery may involve computing, integration and support. A customer-facing rate card does not disclose those costs. Even a well-defined outcome unit cannot establish a product’s gross margin.
Who could benefit?
Interpretation: HubSpot could gain credit consumption and support broader subscription retention if customers find its agent reliably useful. Its documented commercial connection is direct: a billable resolution uses credits within its subscription environment. Over deployment and renewal cycles, the test is whether customers continue paying after comparing the complete workflow cost with the work accomplished.
Salesforce could gain paid Agentforce activity when customers build repeatable support workflows around its actions. Integration into customer records may make an automated step useful, but the financial mechanism still requires paid consumption or licence revenue. Freshworks could gain session-pack demand from customers expanding Freddy deployments. For both, successful implementation must lead to continuing paid use rather than a short experiment.
The materiality remains uncertain. The cited tariffs and operating definitions do not quantify the incremental profit these particular workflows contribute to any of the three groups. No company-wide earnings forecast or share-price conclusion follows from the existence of a price list.
Who’s at risk?
Interpretation: HubSpot faces a renewal risk if its contractual resolution measure and customers’ own quality measures diverge materially. That possibility follows from the distinction between a defined event and customer value; it is not evidence that its users are dissatisfied. Comparable repeat-contact and satisfaction evidence would help assess the gap.
Salesforce faces an analogous commercial risk if customers need many chargeable steps to complete a modest task. Freshworks customers need to match session capacity to actual demand because purchased capacity can expire. If either arrangement produces weak customer economics, expansion could slow at subsequent purchasing decisions. None of the cited documents establishes that these outcomes are occurring.
All three also bear the delivery-cost question. More customer activity may require more resources before it produces durable renewals. The size of that risk cannot be calculated from public tariffs alone. Useful evidence would connect comparable paid activity, retention and delivery costs over several reporting periods.
Speculation: outcomes could improve the bargain
One possible path is that clearer outcome definitions make customer budgeting easier and reward suppliers for reliable service. Another is that customers demand different definitions, stronger quality checks or tighter spending limits. Neither scenario has an assigned probability here.
The next convincing evidence would pair a stable definition with independently checked completion, repeat contacts, human effort and the full customer bill. Useful software still has a cost. Following the work after the assistant stops replying is how that cost becomes a business assessment.
Terms and sources checked 20 October 2026. Financial analysis for information; no investment recommendation.
Photo by MART PRODUCTION on Pexels. Headsets beside laptops on a wooden office table.