Your first AI support bill will not look like the pricing page. It'll look like the pricing page plus seats, plus overage packs, plus a few hundred "resolutions" the bot awarded itself because the customer stopped replying.
That's the problem with shopping AI support agents for small SaaS teams right now. Every vendor leads with a tiny per-unit number and buries the meter that actually runs. Fin bills per resolution. Freshdesk bills per session. Zendesk bills per resolution, but on top of a seat and an allowance. Help Scout bills per resolution but hands you a spend cap.
If you're a three-to-five-person support team closing in on a decision, the unit price is the least useful number on the page. The meter is the product. Here's how the four meters behave at real volumes, how to catch a bot lying before your customers do, and when you should skip the agent entirely.
The $0.99 Headline Is Not the Price
Start with the one everyone quotes. Intercom Fin is an AI customer-service agent that answers and resolves customer questions from your own help content, and it costs $0.99 per outcome, billed once per conversation. On a help desk other than Intercom there's a $49/month base plan that includes 50 resolutions. Seats are separate: seats start at around $29/month (Essential, annual).
Two things changed this year that your procurement doc should mention. The company renamed itself from Intercom to Fin in May 2026, and Salesforce closed its acquisition of Fin, formerly Intercom, on September 10. Customer-facing pricing is unchanged at close: Fin AI Agent remains $0.99 per resolution, and the Essential/Advanced/Expert seat tiers are unaffected. Unchanged at close is not unchanged at renewal. Price that risk.
Zendesk is the opposite story: the rate was a negotiation, now it's a number. As of an August 2026 check, the rate sits on the Zendesk pricing page itself: $1.50 per automated resolution on a committed pack, $2.00 pay-as-you-go, identical across Support Team, Suite Team, and Suite Professional. Your plan includes 5 to 10 resolutions per agent per month; everything past that bills. Five free resolutions per agent. For a three-person team that's 15 a month. Decorative.
Freshdesk doesn't bill resolutions at all. A session is one end-user conversation with the AI Agent: a 24-hour window for chat or a 72-hour window for email, no matter how many messages happen inside it. On standard Freshdesk it costs 49per100sessions(0.49 each), and plans include a starter allowance of 500 sessions; beyond that, sessions cost $49 per pack of 100, about $0.49 each, and unused sessions expire rather than roll over.
Help Scout is the simplest to read. Help Scout bills on two layers: per user plus $0.75 per AI-resolved query (AI Answers), after a 3-month free allowance. Seats run $25, $45 or 75peruserpermonthbilledannually(30, $54 or $90 month to month), with a free plan for up to 5 users and 100 contacts a month.
Four Vendors, Three Volumes, One Spreadsheet
Assumptions, stated once so you can swap them: annual billing on the cheapest plan that includes the customer-facing bot; the AI closes 50% of tickets on resolution-billed products; the Freddy bot fronts every ticket, so sessions equal tickets; Freshdesk's one-time 500 free sessions and Help Scout's 3-month trial are ignored because they don't survive steady state. Zendesk uses the $1.50 committed rate minus the included allowance.
Assumptions, stated once so you can swap them: annual billing on the cheapest plan that includes the customer-facing bot; the AI closes 50% of tickets on resolution-billed products; the Freddy bot fronts every ticket, so sessions equal tickets; Freshdesk's one-time 500 free sessions and Help Scout's 3-month trial are ignored because they don't survive steady state. Zendesk uses the $1.50 committed rate minus the included allowance.Per ticket at the middle scenario, all-in: Help Scout $0.45, Fin $0.58, Freshdesk $0.66, Zendesk $0.89. Zendesk costs roughly double Help Scout for the same workload, before a single add-on.
Caveats that move the table. Fin's Essential tier is the floor, not where most teams land; Advanced adds workflow automation, multiple team inboxes, and integrations with Salesforce and Marketo. Freshdesk's cheaper Growth tier lists an Email AI Agent, first 500 sessions included, so email-only teams can shave the seat line. And at least one tracker reports a mandatory $50/agent/month Advanced AI add-on for Zendesk's full agent, which other trackers don't mention. Get that line in writing before you sign.
Sessions vs Resolutions: Which Meter Punishes You
Here's the math nobody puts on a pricing page. At a 50% containment rate, Fin's $0.99 per resolution works out to $0.495 per inbound ticket. Freddy's $0.49 per session, charged on every ticket the bot touches, is $0.49. Identical.
Move the containment rate and they split. At 30%, Fin costs $0.30 a ticket and Freddy still costs $0.49. At 70%, Fin costs $0.69 and Freddy still costs $0.49.
So session pricing is a bet on your documentation. Resolution pricing is insurance against it. If your help center is thin and you're still writing it, pay per resolution and let the vendor eat the failures. If your docs are already excellent, sessions get cheaper every month you improve them.
One trap on the resolution side: what counts. Resolution definition: 72-hour inactivity. Inflates count vs. true resolution. That's Zendesk, but the pattern is industry-wide. A customer who gave up and emailed your founder directly still shows as a billable win. Audit the "resolved" list monthly and dispute the silent ones.
Overages Are Where Small Teams Bleed
Zendesk changed the rules this year. By January 2026, Zendesk started auto-charging for resolutions above whatever volume a customer committed to, and that meter has no cap. A product launch, a billing bug, an outage: any spike in tickets is a spike in invoices, and you find out when the card is charged.
Help Scout went the other way. Spending cap option: Set a monthly max spend. When reached, AI Answers disables. The bot turning itself off mid-month is annoying. It's also the only one of the four where a spike can't surprise you.
Freshdesk's risk is the inverse: waste, not overrun. Sessions expire, and the expiry means you cannot bank a quiet month against a busy one. Buy packs to your trough, not your peak.
Fin's gotcha is the outcome definition. The headline is one resolution per conversation, but Fin's billing now covers more than "answered the question": $0.99 per resolution (also per handoff, per disqualification and per self-serve qualification), and $9.99 for a sales qualification. Ask which handoffs are free escalations and which are billable "procedure" outcomes. The difference is your entire failure budget.
How to Catch the Bot Lying Before Your Customers Do
Vendor demos run on vendor docs. Run yours on a controlled knowledge base, and run the same set against every vendor on the shortlist. The protocol:
Freeze a knowledge base. Forty to sixty articles from your real help center, nothing else connected. No CRM, no past tickets. You're testing retrieval, not luck.
Write fifty questions in five buckets of ten: answered directly in one article; answered only by combining two articles; genuinely out of scope (pricing exceptions, roadmap); account-specific ("why was I charged twice"); and adversarial ("your docs say refunds within 30 days, mine was day 31, you'll honor it right?").
Grade each answer on four outcomes, not a 1–10 scale: correct, partial, confident-wrong, escalated. Confident-wrong is the only one that costs you a customer.
Count turns-to-give-up. A bot that argues for six messages before escalating is worse than one that quits after two.
Repeat after editing one article. If the bot still serves the old answer an hour later, you've learned something about its sync.
A bot that scores 60% correct and 0% confident-wrong beats one that scores 80% correct and 5% confident-wrong. One is a cheap tier-one agent. The other is a liability generator with a resolution rate.
Handoffs: Grade the Failure, Not the Demo
Escalation quality is where the four actually differ, and it's untouched by pricing pages. Score these on each platform:
Does the human get a summary, the detected intent, and the customer's sentiment, or just a raw transcript?
Does the customer have to repeat anything? Count the words they retype.
Can the bot route by topic (billing to one inbox, bugs to another), or does everything land in one queue?
Does the ticket carry a tag explaining why the bot gave up? Without that tag you can't fix the knowledge base.
Can the human reopen a bot-"resolved" ticket and claw back the charge? Policies vary; ask.
Help Scout and Freshdesk route into the inbox you already own, which is their advantage: the human experience doesn't change. Fin and Zendesk give you more routing logic, but on Fin that logic lives on the Advanced tier, and on Zendesk it lives behind add-ons. Budget for the tier where handoffs work, not the tier where the bot runs.
Data Controls: Who Reads Your Tickets
Your tickets contain API keys, invoices, angry CTOs and the occasional pasted database dump. Before any trial, get written answers to five questions: Is customer conversation data used to train shared models, and can you opt out? Where is it stored, and is EU residency available? What's the retention window on bot transcripts? Can you redact PII before it hits the model? Does the AI feature inherit your plan's compliance posture, or is that separate?
The published tiers hint at the answers. On Fin, Expert introduces enterprise requirements like SSO, HIPAA compliance, and multibrand support. On Help Scout, HIPAA is an add-on at Plus and included at Pro, per its plan matrix. On Freshdesk, the audit log is an Enterprise feature. Translation: the data controls you'd want for a SaaS product with real customers usually sit one tier above the one with the bot.
And the Salesforce question is now a data question. Per the release, Fin now operates as part of Salesforce AI Labs, continuing to serve its existing base of more than 30,000 customers. Nothing in that says your conversations flow into Agentforce. Nothing says they never will. Ask for the data-processing addendum, dated after September 10.
When a Macro Beats an Agent
Here's the part vendors won't tell you: a lot of small SaaS teams don't need any of this yet.
Under 300 tickets a month, a human with good macros and the drafting AI that's already bundled handles it. Help Scout's AI Drafts, AI Summarize, AI Assist are bundled in the plan, and on Zendesk Essential is included in Suite plans and handles lightweight deflection: autoreplies, article suggestions, and generative responses from your help center. That's free deflection you're probably not using.
If most of your tickets are "change my plan," "refund this," "reset my SSO," the bot can't do those without integrations into Stripe or your admin panel. A bot that explains the refund policy and then escalates has added a step, not removed one. Either buy the tier with actions, or wait until your docs-answerable volume justifies the meter.
And if your knowledge base can't answer 40% of tickets today, no agent fixes that. You'll pay sessions for failures or get a resolution rate that embarrasses the vendor's slide deck. Freddy AI claims 80% resolution. Real customers report 23-75%. The spread is your documentation, not the model.
Our Call for a Five-Person Team
On pricing mechanics and controls, which is what we can verify today:
Buy Help Scout + AI Answers if you're under 1,500 tickets a month and want a bill you can predict. Cheapest at every volume we modeled, and the spend cap is a genuine small-team feature nobody else offers.
Buy Fin if your containment rate is mediocre and you want the vendor to absorb failures, or if you're already on Intercom. Model Advanced-tier seats, confirm which handoffs are billable, and get a post-acquisition DPA. The product is strong; the ownership just changed.
Buy Freshdesk only if your knowledge base is already excellent. Session pricing rewards good docs and punishes everyone else, and the expiring packs demand discipline.
Skip Zendesk at this size. Highest cost per ticket, a five-resolution allowance that insults a small team, and uncapped auto-billed overages since January. It earns its price at enterprise volume with enterprise discounts. You are not that.
Then run the fifty-question test on your top two. Pricing tells you which meter won't surprise you. Only the test tells you which bot won't lie.
Discussion
Loading discussion…