Ask any e-commerce support lead what fills the queue and you will hear the same three answers: where is my order, how do I return this, and will this fit or work with that. These questions are repetitive, time-sensitive and spike exactly when the team is busiest, during sales events, holidays and delivery delays. They are also the questions customers increasingly expect to be answered instantly, in their own language, at any hour.
In 2026 AI agents no longer just answer these questions, they resolve them. They look up the order, create the return label, change the delivery address before dispatch and issue a refund within your limits. This guide covers what that takes, what it costs compared with the built-in agents of Gorgias, Zendesk, Intercom and HubSpot, which guardrails you need, and how to roll it out without refund leakage or angry customers.
The market in 2026: per-resolution pricing is the benchmark
AI in customer service has moved from experiment to standard practice. According to Salesforce research based on 3,075 respondents, 66% of service organisations now use AI agents, up from 39% in 2025, and 70% report measurable impact within the first 60 days. In Germany, Bitkom found that customer contact is the top area of AI use, cited by 72% of companies using AI.
Pricing has followed. Helpdesk and CRM vendors now charge per outcome, which gives every store a clear benchmark:
| Platform | Price | Notes |
|---|---|---|
| HubSpot Customer Agent | $0.50 per resolved conversation | Outcome pricing since April 2026 (source) |
| Gorgias AI Agent | $0.90 per AI interaction on annual plans, $1.00 on monthly plans | Built for Shopify-centric stores (pricing) |
| Intercom Fin | $0.99 per outcome | Minimum 50 outcomes a month on third-party helpdesks, 14-day trial (pricing) |
| Zendesk AI | $1.50 per automated resolution with commitment, $2.00 without | Copilot for human agents $50 per agent per month (comparison) |
| Salesforce Agentforce | $2 per conversation or $0.10 per action | Plus licences (pricing) |
| Crescendo (AI plus humans, outsourced) | $1.25–2.99 per resolution plus about $2,900 per month | 30-day launch guarantee (review) |
These prices set the ceiling for what a resolution is worth. They also show where a custom project makes sense: where the built-in agent cannot complete the task at all.
The ticket types that matter
| Ticket type | What the agent needs | Typical actions | Risk |
|---|---|---|---|
| Order status (WISMO) | Order system, carrier tracking | Verify customer, look up order, explain status, flag delays | Low |
| Address and order changes | Order system, warehouse status | Change address or items before dispatch, cancel unshipped orders | Medium |
| Returns and exchanges | Return policy, order history, returns portal | Check eligibility, create label, offer exchange or store credit, trigger refund | Medium to high |
| Product questions | Catalogue, specs, size charts, manuals | Answer with sources, recommend alternatives, check stock | Low, if grounded |
| Damaged or missing items | Order data, photos, carrier claims | Collect evidence, send a replacement or refund within limits | Medium |
| Subscriptions | Billing and subscription system | Pause, skip, change frequency, cancel | Medium |
Order status: the safest place to start
“Where is my order?” is usually the largest single category and the lowest-risk one. The agent verifies the customer with order number plus email or phone, retrieves the order and tracking events, and explains the status in plain language: “Your parcel left our warehouse on Tuesday and is with the carrier; delivery is expected Friday.” The step most stores skip is proactive messaging. When tracking shows a delay, the agent contacts the customer before they write in.
Returns and exchanges: where policy meets money
Returns are the category where automation saves the most handling time and carries the most risk. The agent needs your actual policy as rules, not as a paragraph in a help article: return window, excluded categories, condition requirements, who pays shipping, when to offer an exchange or store credit first. In the EU it must also respect the statutory 14-day right of withdrawal for distance purchases under the Consumer Rights Directive. A bot that refuses a legal return is a compliance problem, not just a bad experience.
Product questions: grounded or not at all
Pre-sales questions (“Is this jacket waterproof?”, “Will this charger work with my laptop?”) are where AI can lift conversion as well as cut costs. They are also where hallucinations do real damage. The agent must answer from your catalogue, specs and manuals, using retrieval with source references, and say “I don’t know, let me connect you” when the data is missing. An invented spec turns into a return, or worse.
Answering versus resolving
A built-in helpdesk agent is the right first step. It is cheap, quick to switch on and handles FAQs plus standard actions in common setups such as Shopify with Gorgias. Custom work pays off when the resolution path runs through systems the platform does not reach:
| Situation | Built-in agent is usually enough | Custom agent pays off |
|---|---|---|
| Single Shopify store, standard returns | Yes | Rarely |
| Orders also from marketplaces (Amazon, Bol, Zalando) | Partly | Yes, to unify order lookup |
| ERP or warehouse system as source of truth (e.g. stock, dispatch status) | No | Yes |
| Different return rules per country or product category | Partly | Yes |
| Five or more languages with local policies | Partly | Yes |
| Complex B2B orders, quotes, spare parts | No | Yes |
The logic is simple. Test the platform agent on your real tickets first. If it resolves the bulk of your WISMO and FAQ volume, keep it and invest elsewhere. Build only for the gap. Our guide to building a support agent that resolves tickets walks through that evaluation step by step.
Architecture and guardrails
A production setup usually looks like this:
- Helpdesk as the hub. Zendesk, Gorgias, Intercom or Freshdesk stays the place where every conversation lives, so human agents keep full context.
- The AI agent with a short list of tools: look up order, get tracking, check return eligibility, create return label, change address, issue refund up to a limit, hand over to a human.
- Identity verification before any personal data. Order number plus email or phone, or a logged-in session.
- Server-side policy checks. Every action is validated by your code against the order state and policy. The model asks, the system decides.
- Limits and approval queues. For example, refunds above a set amount, a third refund for the same customer, or orders flagged for fraud go to a person.
- Full logging. Every tool call, every decision, every message, ready for audits and dispute handling.
- Handoff with a summary. When a human takes over, they get a short summary of the case and the actions already taken.
Point 4 matters more than anything else. Prompt injection, where a customer message tries to trick the agent into doing something it should not, is not a solved problem. Even OpenAI has said it may never be fully eliminated. The answer is architectural: the agent should never hold more power than a junior support rep, and your systems enforce the rules. It is no coincidence that 49% of large companies in KPMG’s survey have explicitly defined scenarios where agents may not decide alone. Our article on prompt injection defense covers the patterns in depth.
Channels and languages
Customers write wherever they already are: email, site chat, Instagram DMs, WhatsApp and sometimes the phone. One agent behind all channels is easier to maintain than a separate bot for each. On WhatsApp, the rules changed in January 2026: general-purpose AI assistants are no longer allowed on the Business API, while a company’s own support bot for its own customers remains permitted. Our guide to AI chat funnels in WhatsApp, Telegram and Instagram covers the channel specifics.
For European stores, multilingual support is often the deciding factor. A Dutch shop selling into Germany and Belgium, or a German brand shipping across the EU, can answer in each customer’s language from a single knowledge base. Country-specific return rules and carriers still have to be modelled explicitly.
What resolved means and how to measure it
Outcome pricing creates a new source of disputes: what counts as resolved? HFS Research expects this definition to become the next battleground in service contracts. Write it down before launch. A practical definition: the customer’s issue was handled without human involvement and the ticket was not reopened within, say, seven days.
| Metric | Why it matters |
|---|---|
| Automated resolution rate by intent | Shows where the agent works and where it does not |
| Reopen rate | Catches resolutions that only looked resolved |
| CSAT for AI versus human conversations | Quality check from the customer’s side |
| First response time and time to resolution | The speed customers actually feel |
| Cost per resolution | Compare with platform prices and human handling cost |
| Refund leakage | Refunds the policy did not allow; should be close to zero |
| Escalation rate and reasons | Your backlog for improving the agent |
Quality is the main barrier: in LangChain’s 2026 survey 32% of teams named it as their top obstacle, with latency second at about 20%. Measure quality by intent from day one.
Compliance in short
- EU AI Act, Article 50. Since 2 August 2026, tell customers they are interacting with AI unless it is obvious (text).
- GDPR. Data processing agreements with your helpdesk, model and messaging vendors, minimal data in prompts, clear retention periods, EU hosting where possible, and a documented legal basis for every data flow.
- Consumer law. The agent must not deny statutory rights such as the EU withdrawal period or warranty claims.
- Reviews. Asking for reviews after delivery is fine, but generating them is not: the FTC bans fake reviews, including AI-generated ones.
A six-week rollout plan
- Ticket audit (week 1). Tag the last 1,000–3,000 tickets by intent. You will usually find that five or six intents make up most of the volume.
- Platform test (week 1–2). Switch on the helpdesk’s built-in agent in a test mode on those intents and measure what it resolves.
- Knowledge and policy clean-up (week 2). Turn the return policy into explicit rules, fix outdated help articles and consolidate product data.
- Build the gap (weeks 2–4). Custom actions and integrations for the intents the platform cannot resolve: ERP, warehouse, marketplaces, country rules.
- Shadow mode (week 4–5). The agent drafts and humans send. Compare its answers with the human ones on real tickets.
- Launch by intent (week 5–6). Start with order status, then address changes, then returns within a low refund limit. Review conversations weekly and raise limits only when the numbers hold.
Custom agent builds from agencies typically start around $10k and can exceed $50k depending on scope (market overview). For the pricing model behind your own support costs, see our article on outcome-based AI pricing.
Typical mistakes
- Launching on the help centre alone. A bot that only quotes FAQ articles deflects tickets without resolving them, and customers write again angrier.
- Policy in the prompt instead of the code. “Never refund more than €100” in a prompt is a suggestion. In your backend it is a rule.
- No identity check. Sharing order details or changing addresses on the strength of a name in a chat invites fraud.
- Hallucinated product specs. Without retrieval from real product data, the agent will eventually invent a size, material or compatibility.
- Hiding the human. A clear way out to a person keeps CSAT high and makes customers more willing to try the bot first.
- Counting deflection as success. If reopen rates climb, your resolution rate is an illusion.
- Forgetting peak season. Test with Black Friday volumes and delay scenarios before they happen.
- Ignoring local rules. Withdrawal rights, carriers and return addresses differ by country, and one global answer is wrong somewhere.
Conclusion
For online stores, AI support is no longer about a chatbot that recites the help centre. It is about agents that resolve order status, address changes, returns and product questions end to end, within limits your systems enforce. Start with the built-in agent of your helpdesk and measure it honestly against the $0.50–2.00 per resolution benchmarks. Invest in custom actions only where the platform cannot reach: your ERP, warehouse, marketplaces and country-specific rules. Define resolved, watch reopen rates and refund leakage, and expand intent by intent.
If you want an outside view of your ticket mix and what could be automated safely, book a free 15-minute audit. We will show you which intents the platform agent can handle and where a custom agent would pay off. Learn more about our AI agents and automation and how we connect them to your CRM and business systems.