Customer-support agents can resolve routine questions, collect information, check account or order status, perform limited actions, and escalate with a summary. Intercom Fin has the clearest per-resolution model, Freshworks offers competitive session economics, HubSpot connects service with CRM context, and Salesforce Agentforce suits complex Customer 360 deployments. Real cost includes the helpdesk, usage, human exceptions, implementation, and knowledge maintenance.
Pricing models are not directly comparable
| Platform | Common AI charging approach | Platform cost to add | Best fit | Cost risk |
|---|---|---|---|---|
| Intercom Fin | Around $0.99 per confirmed AI resolution under current public model | Intercom helpdesk seats or Fin for another supported helpdesk | Digital support seeking predictable outcome pricing | Resolution definition, channels, messages, and human seats |
| Zendesk AI agents | Automated-resolution or usage capacity plus Zendesk suite/agent seats | Zendesk plans and add-ons | Mature ticket operations and omnichannel service | Seat, AI, workforce, QA, and volume layers |
| Freshworks Freddy AI Agent | Session-based packages; public comparisons cite roughly $100/1,000 sessions | Freshdesk/Freshchat or eligible product | SMB and midmarket value | Supported channels and session definition |
| HubSpot Customer Agent | HubSpot Credits consumed by eligible conversations/actions | Service Hub and CRM configuration | Companies using HubSpot across revenue teams | Credits plus paid service seats and hub tier |
| Salesforce Agentforce | Flex Credits per action or conversation/license structures | Salesforce, Data Cloud and service architecture as required | Enterprise customer workflows | Many actions per case and implementation cost |
Public rates and packaging change. Request a written definition of resolution, session, conversation, action, credit, overage, abandoned contact, reopened case, and channel.
Choose one narrow starting job
Begin with the top ten contact reasons by volume and handling time. Good first candidates are order status, standard returns, account navigation, password or setup guidance, appointment policy, subscription features, and documented troubleshooting. Avoid launching first on fraud, safety, legal threats, medical advice, vulnerable customers, cancellations with complex retention rules, or large refunds.
For each intent, write the authoritative source, identity requirement, questions the agent must ask, actions it may perform, transaction cap, forbidden claims, and escalation condition. “Handle billing” is too broad. “Explain the latest invoice from billing data and open a dispute ticket; never change payment details or issue credit” is deployable.
Intercom Fin: transparent outcome pricing
Fin can answer from approved knowledge and take supported actions within Intercom or alongside other helpdesks under available integrations. Intercom defines a resolution through customer confirmation or behavior under its rules and charges around $0.99 per resolution in its widely published model. Current FAQs note that pricing differs when buying Fin with Intercom versus for another helpdesk, and some outbound/message allowances have separate charges.
The calculation is straightforward. If Fin resolves 3,000 conversations, the resolution line is roughly $2,970 before helpdesk seats, messaging, implementation, and tax. If 40% of 10,000 monthly conversations are suitable and Fin resolves 70% of those, the billable outcome is about 2,800 resolutions—not 7,000. The remainder still requires human capacity.
Test what happens when a customer reopens a conversation, changes topic, or silently leaves. Review whether the vendor classifies the result as resolved and whether the business agrees.
Our pick: Intercom Fin for the clearest SMB and midmarket cost-per-resolution pilot
Freshworks: competitive session economics
Freddy AI Agent works with Freshworks service products to answer questions and perform configured workflows. Public 2026 comparisons cite pricing around $100 per 1,000 sessions, or roughly $0.10 per session, but included amounts, package minimums, product eligibility, and channel support must be confirmed. Current comparisons also note that some Freddy AI Agent experiences rely on Freshchat-supported channels rather than all email or phone interactions.
A low session price does not equal a low resolution price. If only half of sessions are eligible and 60% resolve, 1,000 sessions produce 300 resolved outcomes. The apparent $0.10 session becomes about $0.33 per resolution before helpdesk seats and human work—still attractive, but different.
Freshworks suits a small or midsize team that wants helpdesk, chat, and agent functionality without enterprise implementation. Test its channels, commerce actions, language quality, analytics, and escalation in the exact product combination.
HubSpot Customer Agent: best CRM context for SMBs
HubSpot’s Customer Agent uses approved content and customer-platform context to answer or handle eligible support interactions and hand off to people. The shared CRM can expose marketing, sales, and service history subject to permissions, useful when a “support” question is really about onboarding, renewal, or an open deal.
Usage is tied to HubSpot Credits under current packaging. Public HubSpot examples and third-party comparisons often model a customer-agent conversation near a dollar, but credit rates and included allocations change. Calculate eligible conversations, actions within each conversation, paid Service Hub seats, required hub edition, onboarding, knowledge migration, channels, and extra credit packs.
HubSpot is attractive when the company already uses its CRM and Service Hub. It is an expensive reason to migrate if the only need is a standalone FAQ bot.
Salesforce Agentforce: enterprise action pricing
Agentforce can use Salesforce records, flows, knowledge, Data Cloud, and platform permissions to handle service and employee workflows. It supports Flex Credits tied to actions, conversation-based options, and employee-license structures under current packaging. Updating a record, summarizing a case, answering a product question, or invoking a custom flow can draw from the credit pool.
Model actions per successful case. A conversation may require identity verification, several knowledge queries, order lookup, address validation, case update, and confirmation. The credit bill must include unsuccessful loops and escalation.
Salesforce is strongest when the enterprise already has Service Cloud, a governed data model, flows, and implementation capability. Data Cloud, integration, consulting, testing, and operational ownership may dwarf the headline agent rate.
Build a knowledge base the agent can trust
Inventory the articles used by human agents. Assign each an owner, effective date, review date, audience, geography, product, and policy status. Remove duplicate or contradictory pages. Separate internal procedure from customer-visible guidance and enforce permissions.
Write articles around complete intents: eligibility, steps, exceptions, examples, and escalation. Store exact amounts, dates, and account status in structured systems rather than prose. Require source citations in internal review and link the customer to the applicable policy where appropriate.
Test the agent against obsolete content, two policies with similar names, a regional exception, and a document containing instructions aimed at the model. Retrieved content is data, not an authority to change system rules.
Connect actions with strict controls
Begin in answer-only mode. Then allow the agent to read order or subscription status. Next permit reversible actions such as creating a draft ticket or resending a standard confirmation. High-impact actions require authentication, policy checks, transaction limits, and human approval.
Expose narrow tools such as get_order_status and request_return_review, not unrestricted database or browser access. Validate every parameter server-side. Use idempotency so retries cannot issue duplicate credits or messages. Log the source, tool call, result, approval, and final customer response.
Escalate when confidence is low, sources conflict, identity fails, the customer asks for a person, sentiment indicates distress, or the request involves fraud, safety, privacy, legal issues, discrimination, regulated advice, or exceptions. The handoff must include transcript, attempted steps, verified identity state, and reason.
Measure quality and total cost
Create a test set of at least 100 real, anonymized conversations across common, difficult, and adversarial cases. Score correct resolution, grounded answer, policy compliance, action accuracy, escalation, language, tone, and customer effort. Re-run after any model, policy, integration, or prompt change.
For monthly cost, add helpdesk seats, AI resolutions/sessions/credits/actions, messaging channels, model or integration usage, knowledge labor, implementation, QA, monitoring, and human handling. Divide by correctly resolved contacts, not vendor-reported deflection.
Track resolution without reopening, wrong-answer rate, repeat contact within seven days, escalation quality, customer satisfaction, average human handling time after escalation, and cost per correct resolution. Compare against a control group.
Verdict
Intercom Fin provides the easiest outcome-based pilot and transparent per-resolution calculation. Zendesk is the safer choice for established Zendesk operations, Freshworks offers compelling session economics for supported channels, HubSpot wins when CRM context matters, and Salesforce fits complex enterprise actions.
Start with five low-risk intents, answer-only behavior, and a 100-case evaluation. Add write actions one at a time. An agent is economical only when it resolves accurately without causing repeat contacts, unsafe actions, or a larger human cleanup queue.