Skip to content

AI Customer Service Agents for B2B: Capabilities and Rollout

Business customers rarely use chat widgets. They email a shared inbox, quote a purchase order number, attach a delivery note and ask when the rest of the order will ship. Most of the answers already sit in the ERP, the help desk or the terms of delivery. An AI agent can gather that information, draft a reply and set up the case so that a person only has to check or decide.

This guide covers commercial customer service in B2B: order status, delivery dates, complaints, returns, invoice queries and the shared inbox. Technical field service, with fault diagnosis, spare parts and technician visits, needs different data and carries different risks. That is covered on the page about agents for technical service.

What is an AI customer service agent, and what is it used for?

An AI customer service agent is a language model that reads customer requests, calls your ERP, CRM and help desk through defined interfaces, and turns the results into replies or cases. Unlike a scripted chatbot, it does not follow a fixed decision tree. For each request it decides which information it needs and which tool to call.

B2B service differs from consumer support in a few ways that matter for design. The person writing is usually a buyer, a planner or a maintenance lead. Requests refer to framework agreements, partial deliveries and customer-specific terms. Volumes are lower than in retail, but each case often touches several documents. Typical tasks for an agent:

  • Order status and delivery dates: recognise the order number, query the order in the ERP, and put the confirmed date and shipping status into the reply.
  • Complaints: extract delivery note, item, quantity and fault description from the email, ask for anything missing, and prepare a complaint case.
  • Returns: check whether the return policy applies, collect the required details, and request or prepare a return authorisation.
  • Invoice queries: compare quote, order confirmation and invoice, explain the difference, or pass it to accounts.
  • Inbox triage: classify incoming emails by request type, customer and urgency, and route them to the right team.

Anthropic's guide to building a customer support agent describes support chat as a collection of separate tasks: answering questions, retrieving information and taking action on requests. Plan the same way. Do not set out to "automate customer service". Pick individual tasks you can build, test and approve one at a time.

How does an AI customer service agent work?

It has four parts: a language model that understands the request and writes the reply, a knowledge base with your rules and product data, tools that give it controlled access to the ERP, CRM and help desk, and handover rules that decide when it replies itself and when a person takes over.

The model. It reads the email, identifies the request and the documents involved, and writes the answer. For service work, reliable instruction following, consistent tool calls and a good grasp of business language matter more than creative writing.

The knowledge base. Terms and conditions, delivery terms, return rules, goodwill guidelines, product data sheets, recurring questions. The agent should answer from these documents, not from the model's general knowledge. How to build and maintain such a knowledge base is covered on the knowledge management agent page.

Tools. These are defined functions the agent may call, such as get_order(po_number), shipment_status(delivery_note) or create_case(customer, type, text). The model does not run them itself. It signals which tool to call with which arguments, and your application executes the call. That is how you control what the agent can read and what it can change.

Handover rules. They state which request types the agent may answer on its own, which it only prepares and which it forwards straight away. They also set tone, salutation and signature, plus statements the agent must never make, such as promising a credit note.

A typical flow for a status request:

  1. An email arrives in the shared inbox. The agent reads subject, body and attachments.
  2. It identifies the customer and the order number, then queries the order in the ERP.
  3. It checks the confirmed date, any partial deliveries and the shipping status.
  4. It writes a reply using the ERP data and saves it as a draft.
  5. A team member reviews and sends it, or, for approved request types, the agent sends it.

Which use cases exist in commercial customer service?

The use cases differ mainly in how much the agent is allowed to decide. Pure lookups such as order status suit automation, because the answer follows directly from system data. Complaints and returns need rules and often a human decision. Invoice queries involve money, so they belong in an approval step.

Lookups from the ERP. Order status, delivery dates, stock levels, tracking. The benefit shows quickly because the answer can be checked against a system. The risk is stale data: if the ERP shows a date that production moved last week, the agent passes on the wrong date.

Complaints. The agent collects what a complaint case needs, such as delivery note number, item, quantity, description and photos, and chases missing details before the case reaches quality assurance. Whether the complaint is accepted is a human decision.

Returns. The agent checks period and condition against the return policy and prepares the return. Exceptions, such as made-to-order parts or expired deadlines, go to the team.

Invoice queries. Discrepancies between quote, order confirmation and invoice are well suited to preparation. The agent lines up the documents and names the difference. It does not promise a credit or a price correction.

Inbox triage. Even without writing replies, an agent helps by classifying emails, spotting duplicates, matching senders to CRM accounts and flagging urgent cases. It is often a good first step, because no customer receives an AI-written message.

Where sales and field service begin. A request for a new quote or an upgrade is a sales matter. A sales agent or a quotation agent works with price lists, configurations and costing, whereas the customer service agent works with existing orders and documents. A reported machine breakdown belongs to technical service. The customer service agent should recognise these cases and route them, not answer them.

What do you need before you start?

You need three things: reliable data in the ERP and CRM that the agent can reach through interfaces, written rules for delivery, returns and goodwill, and a team that defines request types, rates answers and owns the handovers. If any of the three is missing, the project tends to stall.

Data and systems.

  • Read access to orders, shipments and invoices in the ERP, for example SAP S/4HANA, Microsoft Dynamics 365 Business Central or an industry-specific system, via an API or a database view.
  • Access to customer master data in the CRM, so the agent can match a sender to an account.
  • A shared inbox or help desk where drafts can be stored and cases created.
  • A knowledge base with current documents and a named owner.

Rules. In many mid-sized companies, goodwill and return rules exist only in the heads of experienced staff. For an agent they have to be written down: what counts as transport damage, when a credit is issued without a return, which accounts have special terms. That work pays off even without AI.

People. Someone has to choose the request types, rate answers during testing and decide when a request type may run automatically. In countries with works councils or similar employee bodies, such as Germany, involve them early, because the agent changes how the team works and produces data about case handling.

Data protection. Customer emails contain personal data. If you use an external model or platform provider, under the GDPR you will usually need a data processing agreement as set out in Article 28. If data must not leave your infrastructure, models can also run in your own data centre; the options are described on the sovereign AI page. A customer service agent can be operated in line with the GDPR, but that does not replace a legal review of your setup. Treat this section as orientation, not legal advice.

How do you roll out an AI customer service agent step by step?

Roll it out in stages. First pick request types and collect test cases, then run in draft mode with human approval, and only then switch on automatic replies for request types that proved stable in testing. Each stage has a measurable goal, and you move on only when it is met.

1. Analyse incoming requests. Take a sample from the shared inbox, for example one month of emails, and group it by request type. You will see which types are frequent and which data each answer needs.

2. Choose request types. Start with one to three types whose answer follows directly from system data, usually order status and delivery dates. Complaints and invoice queries come later.

3. Define test cases and scoring. Collect real, anonymised requests together with the correct answer. Anthropic's support guide recommends defining success criteria with the support team and lists query comprehension, response accuracy and correct escalation among the measures. Decide what pass rate makes a request type stable.

4. Build tools and permissions. Define the interfaces to ERP, CRM and inbox. Start with read-only access. Add write actions, such as creating a case, only once the answers are right.

5. Draft mode. The agent saves replies as drafts, and staff review, correct and send them. Every correction points to a missing rule or a data problem. Stay in this stage until scores are stable.

6. Automatic replies for approved types. Only request types that were stable in draft mode run automatically. Everything else stays in draft or is forwarded. Spot checks continue.

7. Operations. Decide who maintains the knowledge base, who runs spot checks and who signs off when rules change. A service agent is not a project with an end date. It is a process that needs an owner.

In our own operations, the lesson has been not to cut draft mode short. That is where the cases nobody wrote a rule for show up.

What are the limits and risks?

The main risks are wrong answers based on stale data, commitments nobody approved, and customers who do not know they are writing to an AI. On top of that, a model faced with an unclear request can produce an answer that sounds right but is wrong. Each risk can be reduced, none can be removed completely.

Stale or incomplete data. The agent is only as good as the ERP. Before go-live, check how current delivery dates and stock levels really are.

Invented answers. Language models can fill gaps with plausible text. Countermeasures: answers only from retrieved data and the knowledge base, a clear instruction to ask or forward when data is missing, and test cases built around exactly those situations.

Unauthorised commitments. Credits, discounts, delivery dates not in the system: rule them out explicitly in the instructions and test for them.

Automated decisions. If an agent rejects a complaint without human review, that may be a decision based solely on automated processing. GDPR Article 22 protects individuals from such decisions where they have legal effects or similarly significant effects. In B2B this mostly concerns sole traders and individual contacts. The simple rule: rejections and decisions about money are made by a person.

Transparency. Under Article 50(1) of the EU AI Act, AI systems intended to interact directly with natural persons must be designed so that those persons are informed, unless it is obvious. The obligation has applied since 2 August 2026, and according to law firm Goodwin the Digital Omnibus did not postpone it. A short notice in automatically sent replies is a simple way to meet it; check the details with your legal adviser, especially if you also serve customers outside the EU.

When this approach does not fit. If your inbox receives few but highly individual requests, if rules and terms are not written down anywhere, or if the ERP has no usable interface, an agent adds little. Fix data and rules first.

How do you get started?

Start by analysing the shared inbox and picking a single request type, usually order status. Give the agent read access to the ERP for that type and let it write drafts only for a few weeks. That shows, at low risk, whether your data, rules and team are ready for more.

A practical order for the first weeks:

  1. Pull a sample from the inbox and count requests by type.
  2. Choose the request type with the clearest data path.
  3. Collect fifty to a hundred real requests with correct answers as test cases.
  4. Set up read-only access to the ERP.
  5. Start draft mode and review corrections systematically.
  6. Then decide: automatic replies for this type, the next request type, or data clean-up first.

For an overview of agents for other departments and how they are deployed, see the AI agents page. If your team wants to work through the basics first, the glossary explains the terms and the guides section has further articles.

Frequently asked questions

Will an AI agent replace my customer service team?

No. It takes over routine lookups and prepares cases, while goodwill decisions, escalations and difficult conversations stay with people.

Do I have to tell customers they are talking to an AI?

In the EU, Article 50(1) of the AI Act has required this since 2 August 2026 for systems that interact directly with people, unless it is obvious. Other jurisdictions have their own rules, so take legal advice for your setup.

What is the minimum system access the agent needs?

Read access to orders, shipments and invoices in the ERP, access to the shared inbox or help desk, and a maintained knowledge base with terms, return rules and product data.

How is this different from a field service agent?

A customer service agent handles commercial requests such as status, invoices and returns. A field service agent works with fault reports, manuals and technician scheduling.

Sources

  1. EU AI Act, Article 50: Transparency Obligations for Providers and Deployers artificialintelligenceact.eu
  2. Goodwin: EU AI Act Transparency Obligations Are Now in Force (August 2026) goodwinlaw.com
  3. Anthropic: Customer support agent (Use-case guide) platform.claude.com
  4. GDPR Art. 22 – Automated individual decision-making dsgvo-gesetz.de
  5. GDPR Art. 28 – Processor dsgvo-gesetz.de