Solution · Customer service
AI agents for customer service that know when to hand over.
Support teams answer the same questions all day: order status, returns, password resets, invoice copies. AI agents resolve those reliably, draft the harder replies with sources for your team and hand over to a person the moment confidence drops or money is involved. We design the thresholds, the escalation paths and the transparency your customers are entitled to.
Where agents pay off in customer service
Support is where AI agents deliver the most measurable value, and where a careless deployment does the most visible damage.
There are three levels of automation in service, with different risks. Triage classifies and routes. Agent assist drafts replies with sources for a human to send. Resolution means the agent solves a defined category end to end, including actions in your systems such as looking up a shipment or issuing a return label. Most teams should climb them in that order.
Why now: models read messy, multilingual customer messages well, helpdesk platforms expose the interfaces agents need, and customers accept AI for simple matters as long as it is fast, honest and easy to escape. The EU AI Act adds transparency duties for AI systems that interact with people, so pretending to be human is no longer an option.
We start with triage and assist because they relieve the team within weeks at low risk, then automate categories with high volume and clear rules. Money, strong emotions and ambiguity stay with people.
- Typical entry pointTriage and agent assist, then automated resolution per category
- SystemsHelpdesk or ticketing, CRM, shop and order systems, knowledge base, telephony
- Human checkpointLow confidence, refunds and credits, complaints, data-subject requests
- First pilotTypically 6–8 weeks
Use cases
Six workflows, from low-risk relief for the team to automated resolution with escalation built in.
Ticket triage and routing
The agent classifies every incoming email, chat and form by intent, urgency, language and customer tier, detects duplicates, routes to the right queue and adds a two-line summary. Team leads adjust the rules in plain language, and misroutes are measured weekly.
Agent assist with sources
For each ticket the agent drafts a reply from your knowledge base, order data and previous conversations, cites its sources and checks tone and policy. The human agent edits and sends; every edit is fed back. New team members are productive faster because the context arrives with the ticket.
Automated resolution for defined categories
Order status, return labels, address changes, password resets, invoice copies: the agent verifies the customer, executes the action through your shop, CRM or order system and confirms. The scope is a strict list. Anything outside it, and anything involving refunds or credits, goes to a person.
Email, chat and voice agents
Email and chat come first because they tolerate a second of thinking time. Voice agents for phone lines are emerging: latency, accents, identity verification and interruptions still need careful design. We plan them deliberately, starting with after-hours or narrow intents, always with a route to a human.
Knowledge base maintenance
The agent finds questions without a good article, drafts new articles from resolved tickets and flags content that is outdated after a product or policy change. Your knowledge manager approves every article. An internal knowledge assistant for the team is often the first step.
Quality monitoring and root-cause analysis
The agent analyses ticket data for recurring causes (a confusing invoice layout, a shipping partner, a product defect) and briefs product, operations and finance. It reviews samples of conversations for quality and, in multilingual support, answers in the customer's language with human review where you have no native speakers.
Worked example
An email support workflow with confidence thresholds
A typical scenario: an online retailer or a SaaS company receives a few hundred support emails a day, half of them about a handful of topics. This is how an agent handles the inbox, with people where it matters.
- Intake and context. An email arrives in the helpdesk. The agent identifies the customer by address or order number and pulls orders, previous tickets and the contract or subscription.
- Classification. It determines intent (“where is my order”, “cancel subscription”, “complaint”), urgency, sentiment and language, each with a confidence score.
- Automated path. If the category is in scope and confidence is above the threshold set for that category, the agent answers from live data such as the carrier's tracking status, applies your policy and sends. The message says it comes from an AI assistant and offers a one-click route to a person.
- Assisted path (human gate). If confidence is below the threshold, or the resolution needs an action with financial impact such as a refund or goodwill credit, the agent prepares draft, proposed action and sources, and a human agent approves or edits.
- Escalation path. Complaints, legal threats, data-subject requests and emotionally charged messages go to a person immediately with a summary. The customer gets an acknowledgement, not an automated resolution attempt.
- Second round. A customer reply runs through the same evaluation. A second unresolved round triggers escalation regardless of confidence, so nobody ends up in a loop.
- Logging and review. Every automated conversation is stored with reasoning and sources. Team leads review a sample each week and adjust thresholds per category.
- Metrics. A dashboard shows automation rate per category, first-contact resolution, handling time, escalation rate and satisfaction, separately for AI-handled and human-handled conversations.
Guardrails and risks in customer service
Transparency towards customers
The EU AI Act contains transparency duties for AI systems that interact with people, and being upfront is good service anyway. Every automated conversation identifies itself as AI and offers a way to a human. Check the current legal status with your counsel; our EU AI Act guide gives an overview.
No promises the agent cannot keep
The agent answers only from verified data and approved policy. It does not invent delivery dates, grant exceptions or interpret contracts. When it does not know, it says so and hands over instead of producing a confident guess.
Money needs a person
Refunds, credits, goodwill gestures and cancellations with financial consequences are prepared by the agent and approved by a human, with a second approver above defined amounts. The rules are yours; we make them executable.
Data protection
Customers are verified before order or account data is disclosed, prompts carry the minimum data needed, processing stays in the EU where required, and conversations are not used to train models without a legal basis. Your data-protection officer reviews the design before launch.
Measurement and drift
First-contact resolution, handling time, escalation rate and satisfaction are tracked continuously. Weekly samples catch quality drift, and every change to prompts, policies or models is tested against a fixed set of past tickets before it goes live.
How we start
Service assessment
We analyse your ticket data (categories, volumes, resolution paths), your systems and the state of the knowledge base, pick the first categories for assist and automation and record baseline metrics.
Design and governance
Escalation paths, confidence thresholds, disclosure texts, tone and refund rules are designed with team leads; data protection and, where relevant, works council questions are settled before the build.
Pilot
Agent assist for the whole team plus automated resolution for two or three categories. Weekly sample reviews, threshold tuning and a comparison against the baseline.
Scale and operate
More categories, further channels (chat, later voice), more languages, and managed operations that keep knowledge, prompts and integrations current. Roles such as AI supervisor and knowledge manager are defined along the way.
Frequently asked questions
Do we have to tell customers they are talking to an AI?
In general, yes. The EU AI Act sets transparency duties for AI systems that interact with people; obligations are being phased in, so check the current status with your legal team. Disclosure is also good practice: customers who know they are talking to an assistant forgive its limits more readily and use the route to a person when they need it.
What happens when the agent is unsure or the customer is angry?
It hands over. Low confidence, negative sentiment, complaint language, legal keywords and repeated unresolved rounds all trigger escalation to a person, with a summary so the customer does not have to start again. The agent never loops a customer through the same answer twice.
Can the agent issue refunds or credits?
It can prepare them, not execute them. Refunds, credits and goodwill gestures are proposed with reasoning and customer history, a human agent approves, and a second approver is required above defined amounts. Whether you ever allow autonomous small credits is your decision, taken later, with data.
How is our customers' data protected?
Identity is verified before account data is disclosed, prompts carry only the data a case needs, processing stays in the EU where required and retention follows your policy. Processor agreements with model and platform providers are part of the design, which your data-protection officer reviews before launch. The legal assessment stays with your counsel.
What happens to our support team?
Agents remove tasks, not usually roles. Repetitive tickets leave the queue; complex cases, escalations, quality review and knowledge management remain and grow. Some teams stop backfilling attrition, others move people into onboarding or sales support. We plan the change with team leads and, where one exists, the works council.
Which helpdesk systems do you work with?
Zendesk, Freshdesk, ServiceNow, Salesforce Service Cloud, HubSpot Service Hub and custom systems, connected through their APIs. We are vendor-neutral and evaluate the native AI features of your platform first; sometimes they cover triage and assist, and the custom work goes into resolution and integration.
Related pages
Enterprise knowledge assistants (RAG)
Assistants that answer from your SharePoint, Confluence, DMS, ERP and ticket history: permission-aware, with citations, evaluated, hosted in the EU.
Managed AI operations (AgentOps)
Monitoring, evaluation, model upgrades, cost control and incident handling for agents and LLM applications after go-live, with a monthly review and documentation your auditors can read.
AI agents for sales
Inbound qualification, account research, CRM updates from calls and emails, quote and tender drafting, and compliant outreach that reps approve before it leaves.
E-commerce & retail
Product content at scale, customer service for order status and returns, claims processing and catalogue data cleaning, built for peak season.
Let's find the first workflow worth automating.
A 30-minute intro call, no slides and no obligation. We listen, ask about your processes, and tell you honestly where AI agents would pay off and where they would not.