AI agents for customer service that resolve, not just deflect.

AI agents for customer service are autonomous software agents that handle a support conversation end to end — understand the request, act in your systems (helpdesk, CRM, order, billing) and resolve it or hand off to a human with context. Unlike a scripted chatbot, they reason over your live data, so they can actually issue the refund or update the order instead of just pointing the customer at a policy page.

A chatbot follows a decision tree and returns canned replies; it deflects. An AI agent understands free-text intent, pulls live data and takes actions in your systems, then escalates the edge cases with full context. The chatbot answers "here's how returns work"; the agent processes the return.

Out of the box, the common ones: Zendesk, Intercom, Freshdesk, Gorgias, Salesforce Service Cloud and HubSpot, plus commerce and billing (Shopify, WooCommerce, Stripe) and channels like email, Slack and WhatsApp. Anything with an API or webhook can be connected — the agent joins your stack, you don't migrate to ours.

Resolution rate and CSAT. Deflection only tells you a human wasn't involved, not that the customer was helped — a bot can deflect a ticket and still leave the problem open. We instrument resolution and customer satisfaction from day one and treat deflection as a secondary signal.

It answers from your own knowledge base with retrieval (RAG) and citations, runs inside guardrails, and is told to escalate rather than guess when it's unsure. Before launch we evaluate it against hundreds of your real past tickets, and we start it in assist mode or on a slice of traffic with every action logged and reversible.

Two to four weeks for the first production agent on a single high-volume workflow. The first days are scoping and connecting data; the rest is grounding it, evaluating against your real tickets and launching behind a human before widening scope.

No — they absorb the repetitive tier-1 contacts so your team handles the conversations that genuinely need a person. The agent escalates anything outside its remit with the full transcript and a suggested action, so handoffs are warm. The honest framing is augmentation, not replacement: fewer repetitive tickets, the same humans on the hard cases.

Buy if you're on a standard stack with standard use-cases — a vanilla Shopify-and-Gorgias shop can get a long way with an off-the-shelf agent, and we'll tell you so. Build when you have custom systems, custom policies, or enough volume that per-resolution or per-seat fees start to hurt. We build against the stack you already run, so there's no vendor lock-in and no per-seat tax — your only running cost is the model tokens.

Two options per agent: a one-off setup from €1,500, or a fixed monthly fee — hosting, retraining and ongoing tuning included, with no per-ticket or per-seat surprises. Off-the-shelf platforms often charge per resolution or per conversation on top of seat licences, and they define what counts as a "resolution"; a custom build is scoped up front and your only variable cost is the model tokens, usually cents per conversation. Start with one agent on one workflow and expand once the resolution numbers hold.

Most of them: Zendesk AI agents, Intercom Fin, Salesforce Agentforce and Ada all escalate to human agents inside their own helpdesk. The difference is what travels with the handoff. A good one passes the full transcript, the customer record, what the agent already tried and a suggested next step, so the customer never repeats themselves. A custom agent does that into whichever helpdesk you already run.

Yes, through your systems' APIs and inside limits you set: refunds below a threshold, order changes before dispatch, plan upgrades and cancellations, password resets. Anything above the limit, outside policy or irreversible goes to a person with the case already prepared. Every action is logged with its reason, so it can be audited or reversed.

On rules you can read: the request is outside its workflows, a policy exception applies (a refund above the limit, a late return), its confidence in the answer is low, the customer is angry or asks for a person, or the topic is on the never-automate list. We tune those triggers on your past tickets before launch and review the handoffs every week after it.

That's how we'd do it. Pick one queue — billing questions or order status — and run the agent on it for a fixed period, measuring automated resolution rate and CSAT against the same queue handled by people. If the numbers hold, the next queue is added on the same build; nothing is re-implemented, because the knowledge base, tools and guardrails carry over.

Yes. A custom agent has no per-resolution or per-seat fee: the build is scoped up front and the running cost is model tokens, usually cents per conversation. Platforms such as Intercom Fin ($0.99 per outcome) or Zendesk ($1.50–$2.00 per automated resolution) charge per outcome instead — simpler at low volume, more expensive as volume grows.

Yes, and it should. Resolved tickets, with personal data stripped, go into the knowledge base so the agent answers the way your best people already do, and a few hundred of them become the eval set it is tested against before launch. New answers reach the knowledge base through a review step, not automatically from live conversations.

Because the agent says so. Under the EU AI Act (Article 50), a system that talks to people has to make clear it's an AI unless that's obvious from the context. We put it in the first message and keep a "talk to a person" option available in every channel.