For a decade, most support chatbots did one thing: deflect. They matched a question to a help-centre article, posted a link and hoped the customer would go away. Customers learned to type “human” as fast as possible.
In 2026 the bar has moved. Support platforms now bill per resolved conversation, and 66% of service organisations already use AI agents — up from 39% a year earlier, with 70% reporting measurable impact within 60 days (Salesforce). The question is no longer whether to use AI in support. It is whether your agent actually resolves tickets — checks the order, issues the refund, changes the address — and what each resolution costs you.
This guide explains the difference between an answer bot and a resolution agent, compares the pricing of Zendesk, Intercom, Gorgias, HubSpot and Salesforce, walks through the architecture and gives you a four-week pilot plan.
Answering vs resolving
A customer asking “Where is my order?” does not want an article about shipping times. They want the tracking status of order 48213 and, if it is late, a decision. That requires identity, data and permission to act.
| Capability | Answer bot | Resolution agent |
|---|---|---|
| Knowledge | Help-centre articles | Help centre, internal policies, macros and the customer’s own data |
| Actions | None; links to forms | Order lookup, refund, cancellation, address change, rescheduling, return labels |
| Identity | Anonymous | Verified customer (login, email plus order number, one-time code) |
| Escalation | “Please contact us” | Warm handover to a human with summary and context |
| Main metric | Deflection | Resolution rate, reopen rate, satisfaction, cost per resolution |
Deflection measures how many customers stopped asking. Resolution measures how many problems were solved. Only the second one saves money without costing you customers.
The 2026 platform landscape and prices
Outcome-based pricing has become the market norm, and it gives you a clear benchmark for any custom build.
| Platform | Pricing model (2026) | Notes |
|---|---|---|
| Intercom Fin | $0.99 per resolution | Works on top of other helpdesks with a minimum of 50 resolutions per month; 14-day trial (Intercom) |
| Zendesk AI agents | $1.50 per automated resolution with commitment, $2.00 without | Copilot for human agents at $50 per agent per month (Macha) |
| HubSpot Breeze Customer Agent | $0.50 per resolved conversation | Outcome pricing since 14 April 2026 (Resolve247) |
| Salesforce Agentforce | $2 per conversation or $0.10 per action | Flex Credits at $500 per 100,000 (Salesforce) |
| Gorgias | Per automated interaction | Strongest fit for Shopify e-commerce stores |
| Botpress | Plus $150 per month for 250 conversations, Team $750 for 1,500 | $0.50–0.65 per extra conversation (Botpress) |
| AI-native support outsourcers (e.g. Crescendo) | About $1.25–2.99 per resolution plus a monthly fee | AI plus human agents under one contract (Macha) |
What this means in practice: if an agent resolves 3,000 conversations a month at $0.99, the platform fee is about $2,970. A custom solution must either cost less in total or do something the platform agent cannot.
When the built-in agent is enough — and when it is not
Start with your helpdesk’s own agent if:
- all channels already run through one helpdesk;
- most automatable tickets can be answered from the help centre;
- the actions you need (order status, cancellations) are covered by native integrations, for example Shopify in Gorgias;
- you serve customers in one or two languages.
Invest in custom work if:
- resolutions require actions in systems the platform cannot reach — an ERP, a warehouse system, a billing platform or a custom back end;
- the policy is multi-step: return eligibility depends on product, purchase date, country and customer tier;
- you serve customers in several EU languages and need consistent quality in each;
- you run several brands or shops with different rules;
- data residency or data protection requirements rule out sending certain data to a given vendor;
- chat, email and voice must behave consistently.
Often the right answer is a hybrid: the platform agent stays the front end, and custom tools give it the actions it lacks.
Architecture of a resolution agent
A production support agent has six layers. Skipping any of them is how pilots fail.
1. Knowledge layer
The agent answers from approved sources only: help centre, internal policies, macros, product data. Use retrieval with citations so every answer can be traced, and give each source an owner who keeps it current. Outdated articles are the most common cause of wrong answers. The same principles apply as for an internal RAG knowledge assistant.
2. Identity and context
Before any action, the agent verifies who it is talking to — through the logged-in session, a one-time code or the combination of email and order number. It then loads context: open orders, plan, recent tickets, language. Without context, the agent asks questions the customer has already answered.
3. Action layer
Each action is a tool with a strict schema and limits: “look up order”, “create return label”, “refund up to €50”, “change delivery address before dispatch”. Everything above a limit goes to a human for approval. Actions must be idempotent (a retry never refunds twice) and every call is logged with input, output and result.
4. Policy and guardrails
The agent follows written rules: what it may promise, which topics it does not handle (legal threats, safety issues, data deletion requests) and how it phrases sensitive answers. A classic failure is an agent inventing discounts or conditions that do not exist; the fix is to give it the policy as data and forbid promises outside it.
5. Handover to humans
Triggers for handover: the customer asks for a person, the agent’s confidence is low, the topic is sensitive, an action fails, or the same question comes back twice. The human receives a summary, the verified identity and the steps already taken, so the customer never repeats themselves.
6. Observability and evaluation
Log every conversation and tool call, review a sample every week and keep a test set of real tickets with expected outcomes. Re-run it before changing prompts, models or knowledge. Our article on AI evals explains how to set this up.
AI can also help here: manual quality assurance in contact centres typically reviews only 1–2% of calls (AssemblyAI), while an automated reviewer can score every AI and human conversation against your checklist — correct policy, correct action, tone, handover — and surface the ones a team lead should read.
Define “resolved” before you launch
HFS Research expects the definition of a “resolved” ticket to become the next battleground in service contracts; it has to be written down and verifiable (HFS Research). If you pay per resolution, this definition is literally your invoice.
A workable definition: the conversation was closed without a human, the customer did not reopen it or contact you about the same issue within a set window (for example seven days), and satisfaction did not fall below an agreed threshold.
| Metric | What it tells you | Watch out for |
|---|---|---|
| Automated resolution rate | Share of conversations solved without a human | Easy to inflate by closing conversations too early |
| Reopen and repeat-contact rate | Whether “resolved” really meant solved | Customers switching channel after a bad answer |
| Satisfaction on AI conversations | Customer perception | Low response rates; read the comments |
| Handover rate and quality | Whether the agent knows its limits | Handovers without summary cost agent time |
| Cost per resolution | Platform fee, tokens, maintenance | Hidden costs of fixing wrong actions |
| First response and resolution time | Speed gains | Fast but wrong is worse than slow and right |
Which tickets to automate first
Not every frequent ticket is a good first candidate. The best ones combine high volume, clear rules and data the agent can reach through an API. A typical shortlist for an online shop or subscription business looks like this:
| Intent | Data needed | Action | Risk level |
|---|---|---|---|
| “Where is my order?” | Order and carrier tracking | Show status, open a carrier claim if overdue | Low |
| Return or exchange request | Order date, product, return policy by country | Check eligibility, create return label | Low to medium |
| Address change | Order status, warehouse cut-off | Change address before dispatch | Medium |
| Invoice copy or VAT correction | Billing system | Send invoice, flag correction for finance | Low |
| Subscription pause, downgrade or cancellation | Plan, contract term | Execute within policy, offer alternatives once | Medium |
| Refund | Payment, policy, amount | Refund under a limit, approval above | High: cap and log |
Start with the low-risk rows, prove the numbers and move down the table. Complaints about product quality, legal threats and anything involving personal safety stay with people.
A four-week pilot plan
- Preparation. Export the last three months of tickets and group them by intent. Pick three to five intents with high volume and clear rules — typically “where is my order”, returns, address changes, invoice copies and password or account issues. Record today’s numbers for each.
- Week 1: knowledge and rules. Clean up the relevant help-centre articles, write policies as explicit rules and connect read-only tools (order lookup, customer profile).
- Week 2: shadow mode. The agent drafts answers and proposed actions; humans review and send. Measure how often the draft was correct without edits.
- Week 3: live on a slice. Switch the agent on for one channel or one language, with write actions under limits. Review conversations daily.
- Week 4: evaluate and decide. Compare against the baseline: resolution rate, reopen rate, satisfaction, handling time of human agents and cost per resolution. Then expand, adjust or stop.
To calculate the business case, multiply the expected automated resolutions by your fully loaded cost per human-handled ticket and subtract platform fees, tokens and maintenance. For example, if 8,000 tickets a month come in, 45% fall into the pilot intents and the agent resolves 70% of those, you get about 2,520 resolutions per month — roughly $2,500 in fees at Intercom’s $0.99. The savings side depends on your own cost per ticket, which you should measure, not guess.
Common mistakes
- Launching on a messy help centre. The agent will faithfully repeat outdated information. Clean the top articles first.
- Measuring deflection instead of resolution. A customer who gives up is not a resolved ticket.
- Write actions without limits. Every refund, cancellation or change needs a cap and a log.
- No warm handover. If customers have to repeat everything to a human, satisfaction drops and the agent gets the blame.
- One language for a multilingual market. German, Dutch or French customers expect answers in their language and with local policies.
- Building on the wrong channel assumptions. Since 15 January 2026, WhatsApp Business no longer allows general-purpose AI assistants; company bots serving their own customers, such as support or booking, remain allowed (TechCrunch).
- Trusting demos. Test on your own tickets. The resolution rate on a vendor’s demo data says nothing about your customers.
Compliance and data protection
Since 2 August 2026, Article 50 of the EU AI Act requires that customers are informed they are talking to an AI unless it is obvious; most violations of the AI Act can be fined up to €15 million or 3% of global turnover, with the lower amount applying to SMEs (AI Act, Article 50; ActuIA). A single sentence at the start of the chat is enough: say that this is the company’s AI assistant and that a human is available.
Under the GDPR you need a data processing agreement with the helpdesk vendor, the model provider and any other processor, a list of subprocessors, a legal basis for transfers to the US and clear retention periods for transcripts. Mask payment data and avoid sending more personal data to the model than an answer requires.
Conclusion
The best support agents in 2026 are not chatty; they are useful. They verify the customer, look up the data, take the permitted action and hand over cleanly when they reach their limits. Start with the agent your helpdesk already offers, measure resolution rather than deflection, and invest in custom actions only where they move the numbers. If you sell online, our guide to AI for e-commerce support goes deeper into orders and returns.
Not sure whether your platform’s agent is enough or whether you need custom actions? Book a free 15-minute audit and we will review your ticket mix and systems with you. More on how we build and run support agents on our AI agents page.