Quick Answer: An AI support agent answers customers and completes support tasks on its own, then hands the hard cases to a human. When you buy one, you are really buying three layers: the autonomous agent, the handoff, and agent assist for your human reps. Judge vendors on verified resolution rate, CSAT on AI-handled tickets, reopen rate and cost per resolution, and ask each one exactly how it counts a "resolution".
Every helpdesk vendor now sells an AI support agent, and almost every one of them quotes a resolution rate. The numbers are hard to compare, because the vendors don't count the same thing. One counts a customer who goes quiet for 24 hours. Another runs an LLM check over the transcript. A third charges per conversation whether anything got solved or not.
This is the buyer-side guide. It splits the purchase into the three layers that matter (the autonomous agent, the handoff to a human, and the agent-assist layer for your reps), compares how eight vendors define and price their work as of November 2026, and ends with the metrics and a 30-day pilot plan. If you're still building the business case and the use cases, start with our guide to AI agents for customer support, which shows how teams resolve 80% of tickets automatically and the roadmap to get there.
We build cowork.ink, AI coworkers that each get their own computer and do back-office work for teams. We don't sell a support widget, so none of the vendors below is a competitor of ours.
What is an AI support agent?
An AI support agent is software that talks to your customers and finishes support tasks without a human: it answers the question, looks up the order, issues the refund within policy, and escalates when it can't. It differs from the old decision-tree bot because it reasons over your knowledge base and connected systems instead of following a fixed script. Our explainer on how AI agents differ from chatbots covers that shift in detail.
For a buyer, the useful frame is that the product has three layers, and vendors are strong in different ones:
| Layer | Who it serves | What it does | What to measure |
|---|---|---|---|
| Autonomous AI support agent | The customer | Answers and acts end to end on routine requests | Verified resolution rate, reopen rate, cost per resolution |
| Handoff | Customer and human rep | Decides when to escalate and passes the context along | Time to human, repeat questions after handoff, CSAT on escalated tickets |
| Agent assist | Your human reps | Suggests replies, summarizes, surfaces knowledge, checks quality | Handle time, ramp time for new reps, QA scores |
Most AI support agent software sold today bundles the first two. Agent assist is often a separate SKU with per-seat pricing, so it can quietly double the bill or be missing entirely.
Resolution vs. deflection: why vendor numbers don't match
Deflection measures whether a customer avoided a human; resolution measures whether the customer's problem got solved. They are different numbers. A customer who gets a useless answer, gives up and calls your phone line was deflected from chat but not resolved.
"Deflection" is the older metric from the chatbot and self-service era. Vendors selling AI support agents have moved to "resolution" and increasingly price on it, but each one defines it differently. Decagon's own pricing post admits the problem plainly: "You'll get into situations where a user is upset, leaves, and that gets counted as a resolution."
How vendors define a resolution (November 2026)
The definitions below come from each vendor's own pricing or help pages, checked on November 13, 2026.
| Vendor | Billing unit | When it counts |
|---|---|---|
| Intercom Fin | Outcome | The customer confirms the answer worked, or leaves without asking for more help; 24 hours of silence after Fin's last answer is an "assumed resolution". A Procedure that ends in a configured handoff also counts as an outcome. Charged once per conversation. |
| Zendesk AI agents | Automated resolution | Since May 18, 2026, three tiers: assisted escalation (a human finished), contained resolution (the customer didn't ask for more help) and verified resolution (passed an LLM check). The first two don't count against the resolution allowance. |
| Gorgias AI Agent | Automated interaction | A request "fully resolved without a human agent"; if an agent replies within 72 hours, it is billed as a standard helpdesk ticket instead. |
| Freshdesk Freddy AI Agent | Session | Any unique interaction with the AI agent, resolved or not; on email, a 72-hour window from the first message counts as one session. |
| Decagon, Ada | Mostly conversations | Both say most customers choose per-conversation pricing, with a per-resolution option. Neither publishes its rates. |
Intercom's Fin outcomes and pricing article also lists what it never bills: escalations when a customer asks for a human, failed Procedures and conversations the customer abandons after a clarifying question.
Zendesk's automated resolution tiers are the most explicit about the gap. A "contained" conversation is one where the customer simply didn't ask for more. A "verified" one is where a model read the transcript and agreed the request was actually resolved. The distance between those two numbers is roughly the distance between deflection and resolution.
Any definition that counts a quiet customer as resolved will flatter the AI. Ask what share of reported resolutions are assumed (silence) versus confirmed (the customer said so, or a check passed), and ask for the reopen rate on both.
Three questions to ask every vendor
- What exactly triggers a billable resolution? Get the written definition, including the silence window and whether a later human reply reverses it.
- What happens when the customer comes back? If the same customer returns on the same issue within a week, is the original resolution reversed, and do you get the charge back?
- Can I export the raw conversations behind the number? You want to audit a sample yourself, not trust the dashboard.
Handoff design: when an AI support agent should escalate
A good handoff happens early, carries the full context and never makes the customer repeat themselves. The handoff is where AI support agents most often lose customers, and it's the layer buyers test least in demos.
When to escalate
Set escalation triggers before launch, not after the first complaint:
- The customer asks for a person. Always honor it, on the first request.
- Sentiment turns. Repeated frustration, profanity or threats to cancel.
- The action is outside policy. A refund above the limit, an exception, anything legal or medical.
- Low confidence or no source. The agent can't ground its answer in your knowledge base.
- A loop. The same question asked twice, or the conversation passes a set number of turns without progress.
- High-value accounts. Some teams route named enterprise customers to a human from the first message.
Pricing shapes this incentive. Intercom says it doesn't charge for escalations triggered by Fin's default behavior, such as a customer asking for a human. Sierra's pricing post says escalated conversations are, "in most cases," not charged. That's the right incentive: the vendor gets nothing for keeping a customer who should have reached a person.
What context to pass to the human
| Pass along | Why it matters |
|---|---|
| A three-line summary | The rep reads the problem in ten seconds, not the full transcript |
| What the AI already tried | Stops the rep repeating steps the customer already rejected |
| Customer identity and account data | Plan, order, recent tickets, verified or not |
| Why it escalated | Policy limit, sentiment, low confidence; it tells the rep what kind of problem this is |
| Suggested next step | A draft reply or action the rep can accept or change |
The same principles apply when one agent passes work to another; our guide to AI agent handoffs between agents covers the state-passing side. For where to put a human checkpoint by design rather than as a fallback, see human-in-the-loop AI agents.
AI agent assist: the layer for your human reps
AI agent assist is AI that works beside your human support reps: it drafts suggested replies, summarizes long threads, surfaces the right knowledge and checks quality, while the human decides what goes out. It's the least glamorous of the three layers and the one with the strongest independent evidence behind it.
The best-known study is "Generative AI at Work" by Erik Brynjolfsson, Danielle Li and Lindsey Raymond. In the NBER working paper, access to an AI assistant raised issues resolved per hour by 14% on average across 5,179 customer support agents, with a 34% gain for novice and low-skilled agents and minimal impact on the most experienced. The authors also report better customer sentiment and higher retention among agents.
What agent assist actually does
- Suggested replies. A draft grounded in your help center and past answers, edited and sent by the rep.
- Summaries. A recap when a ticket arrives from the AI agent, from another shift or after a long thread, and a wrap-up note when it closes.
- Knowledge surfacing. The relevant article, macro or policy pulled up next to the ticket, ideally with a link to the source.
- QA and coaching. Checks for tone, missed steps or policy errors, either live or after the fact.
- Live guidance on calls. Real-time transcription and next-step prompts for voice teams.
AI agent assist pricing
Agent assist is usually priced per seat, separately from the AI agent's per-resolution or per-conversation fee. Published rates as of November 2026:
| Product | Price (as listed) | Notes |
|---|---|---|
| Intercom Copilot | $29 per agent/month billed annually, $35 monthly | Each agent also gets 10 free Copilot conversations a month |
| Zendesk Copilot | $50 per agent/month billed yearly | Add-on for Professional plans and up |
| Freshdesk Freddy AI Copilot | $29 per agent/month | Add-on on Pro and Enterprise |
| Salesforce Service Rep Assistant | Part of Agentforce for Service | Replies, summaries and knowledge-based guided plans |
| Sierra Live Assist, Decagon Assist | Not published | Sold alongside each vendor's AI agent |
Which comes first: the AI support agent or agent assist?
If your knowledge base is thin, start with agent assist. A rep catches every bad draft, so you find the gaps in your content without customers seeing them. If you have high volume of repetitive, well-documented requests and a clean knowledge base, the autonomous agent pays back faster. Either way, the AI agent knowledge base behind both layers is the same, and most failures trace back to it.
AI support agent software: the vendor landscape (November 2026)
The eight best-known AI support agent vendors split into two camps: helpdesk platforms that add an AI agent (Intercom, Zendesk, Salesforce, Gorgias, Freshdesk) and AI-agent specialists that sit on top of a helpdesk (Sierra, Decagon, Ada). Prices change often; everything below was checked on vendor pricing, product and help pages on November 13, 2026. "Not listed" means we couldn't find it on the vendor's own site, not that it doesn't exist.
| Vendor | Pricing model | Published rate | Channels (vendor-stated) | Agent assist |
|---|---|---|---|---|
| Intercom Fin | Per outcome (resolution or Procedure handoff) | From $0.99 per outcome | Email, live chat, phone and more | Yes: Copilot, $29/agent/mo (annual) |
| Zendesk AI agents | Per automated resolution, allowance included in plan | 5–10 per agent/mo included; then $1.50 committed or $2.00 pay-as-you-go | Messaging, email, voice (needs Zendesk voice) | Yes: Copilot, $50/agent/mo (yearly) |
| Salesforce Agentforce | Per conversation, Flex Credits per action, or per-user add-ons | $2 per conversation; $500 per 100,000 credits (20 credits per action); $125/user/mo add-ons | Voice billed as Agentforce Voice actions; others per setup | Yes: Service Rep Assistant |
| Sierra | Outcome-based; per-conversation for routing | Not published | Digital and voice | Yes: Live Assist |
| Decagon | Per conversation or per resolution | Not published | Chat, email, voice | Yes: Decagon Assist |
| Ada | Mostly per conversation; per resolution for some enterprises | Not published | Voice, email, chat, WhatsApp, SMS, social, in-app | Not listed |
| Gorgias AI Agent | Per automated interaction | $0.90 (annual) or $1.00 (monthly); $1.50 overage | Helpdesk channels; AI Agent coverage not itemized | Not listed on pricing page |
| Freshdesk Freddy AI Agent | Per session | First 500 sessions included once; then $49 per 100 | Email, web chat, WhatsApp, social | Yes: Freddy AI Copilot, $29/agent/mo |
A few notes that the table can't hold:
- Sierra and Decagon describe their pricing models in posts from December 2024 and publish no rate card. Sierra says it gets paid "only when we complete a task for you". Decagon says most customers choose per-conversation pricing.
- Ada says on its platform page that most customers pay per conversation and that resolution-based pricing is for "enterprises with specific needs".
- Gorgias counts each automated interaction as a helpdesk ticket as well, so the AI fee sits on top of your ticket volume.
- Salesforce sells three different meters. If you're on Salesforce already, our Salesforce Agentforce review walks through which one fits.
We haven't given these products ratings. We haven't run all eight on the same ticket set, and a score without a shared test would be invented.
AI support agent pricing: what each model rewards
Each pricing model rewards the vendor for something different, and that shapes how the AI behaves in edge cases. Choose the model that matches what you can measure.
| Pricing model | Who uses it | What the vendor is paid for | Watch out for |
|---|---|---|---|
| Per resolution / outcome | Intercom, Zendesk, Gorgias, Sierra | The definition of "resolved" | Assumed resolutions from silence; bills rise as the AI improves |
| Per conversation / session | Decagon, Ada, Freddy, Agentforce | Every conversation, solved or not | Paying for failures; low resolution rates hide in the bill |
| Per action (credits) | Agentforce Flex Credits | Every step the agent takes | Long, multi-step conversations cost more than you budgeted |
| Per seat | Agent-assist add-ons, Agentforce user add-ons | Each human rep with access | Paying for seats that never open the assistant |
Ada makes the case against resolution pricing in a July 2025 post on its blog: since "resolution" isn't standardized, and better performance means more billable resolutions, the bill rises as the AI improves. Per-conversation pricing has the opposite flaw: the vendor is paid the same for a failure as for a fix. Neither is wrong. Under per-conversation pricing you have to measure resolution yourself; under per-resolution pricing you have to audit the vendor's definition.
How to evaluate an AI support agent vendor: the metrics
Judge an AI support agent on four numbers you can audit: verified resolution rate, CSAT on AI-handled tickets, reopen rate and cost per resolution. Ask the vendor how it calculates each one, and calculate them yourself from exported conversations.
| Metric | Definition | Ask the vendor | Common trap |
|---|---|---|---|
| Resolution rate | AI-resolved conversations ÷ all conversations the AI handled | Confirmed vs. assumed split; what is excluded from the denominator | Excluding conversations the AI "didn't recognize" inflates the rate |
| CSAT on AI-handled tickets | Satisfaction score for conversations that never reached a human, reported separately | Survey response rate on AI conversations | Blending AI and human CSAT hides a weak AI score |
| Reopen rate | Share of AI-resolved conversations where the customer returns on the same issue within a set window (seven days is a sensible start) | Do reopens reverse the billed resolution? | No window defined, so reopens are never counted |
| Cost per resolution | Total AI spend (fees, seats, setup and upkeep time) ÷ verified resolutions | All fees, including platform minimums and add-ons | Dividing by all conversations, not resolved ones |
Two more are worth tracking. Escalation quality: the share of handoffs where the human didn't have to re-ask anything. Agent-assist adoption: the share of suggested replies reps actually send. If reps ignore 80% of drafts, the seat price is buying nothing.
Here's a worked example with round numbers, for illustration only. If a vendor charges $1 per conversation and the AI verifiably resolves half of them, your cost per resolution is $2, not $1. Compare that number across vendors, not the sticker price.
You can't judge an AI support agent without knowing your current cost per ticket, CSAT and reopen rate for the same request types. Pull two to four weeks of baseline before the pilot starts.
A 30-day pilot plan for an AI support agent
Run the pilot on a small, well-documented slice of traffic, with definitions agreed in writing and a weekly human audit. This is a vendor evaluation, not a rollout; the full rollout roadmap is in our customer support agents guide.
Pick two or three intents on one channel
Order status, password resets or a common billing question on chat or email. Choose requests your knowledge base already answers well.
Agree the definitions in writing
Resolution, reopen window, what counts as an escalation and how the vendor bills each. Put them in the pilot agreement, not a slide.
Split the traffic
Send a fixed share of those intents to the AI and keep the rest with humans as a control group, so you compare like with like.
Read the transcripts every week
Audit a random sample of AI-resolved and AI-escalated conversations. Note wrong answers, missed escalations and handoffs where the human had to start over.
Decide on four numbers
Verified resolution rate, CSAT on AI-handled tickets, reopen rate and cost per resolution, each against the control group. If the vendor can't give you the raw data, that is your answer.
If you run agent assist in the same pilot, give it to half the reps on the human side and compare handle time and QA scores with the other half.
Where AI support agents still fall short
AI support agents struggle with anything that needs judgment, missing information or empathy under pressure. Plan for these limits instead of discovering them in production:
- Thin or contradictory knowledge. The agent answers confidently from an outdated article. Fix the content before blaming the model.
- Exceptions. Policy says no, but the customer is a five-year account with a real grievance. That's a human call.
- Multi-system problems. A billing error caused by a sync failure between two tools needs someone who can investigate, not just answer.
- Voice. Phone adds latency, accents and interruptions. If you're evaluating voice, read our guide to AI voice agents before you sign.
None of this argues against buying. It argues for keeping the handoff fast and the humans well equipped, which is what the second and third layers are for.
The Back Office of Support: Where cowork.ink Fits
An AI support agent covers the front line. Behind it, support teams still have a pile of recurring work that no helpdesk widget does: the weekly trend report, the account research before a hard escalation, the follow-ups that slip.
That's the work cowork.ink is built for. cowork.ink is not a chatbot and doesn't talk to your customers. It gives you AI coworkers, each with its own computer (a real browser, files, an inbox and a code runtime), that sign into your tools and hand you finished work in Slack or Telegram. For a support team, that looks like:
- The weekly ticket-trend report. Every Monday at 8:00, it opens your helpdesk in its browser, exports last week's tickets, groups them by topic, compares them with the week before and posts the summary with the PDF in your support channel.
- Account research before a hard escalation. Before a lead takes the call, it reads the customer's ticket history, plan and recent activity and writes a one-page brief.
- Follow-up chasing. It checks which tickets are waiting on another team and posts the list, with owners, where your team will see it.
Every run leaves a transcript of the steps it took, so you can check its work and correct it, and corrections stick. You can set up as many coworkers as you need, in whatever roles you need, in about three minutes each. It's $25/month billed yearly, or $29 monthly, covering every coworker, its machine and the models, with a 15-day money-back guarantee.
Let the AI support agent take the front line. Hire an AI coworker at cowork.ink for the work behind it.