B2B customer service with AI: what an agent on WhatsApp can resolve and what it cannot

In a distribution business, customer queries do not arrive through one channel: they arrive through the rep's personal WhatsApp, through the company mobile, through three different mailboxes and through the phone of the admin who handles everything. They get answered when someone sees them, they get lost when that person is on leave, and nobody knows how many there were or how long they took.

In this article we look at what makes customer service different in B2B distribution, what kind of queries an AI agent can resolve unsupervised today, where it absolutely has to hand off, and how to measure whether it is working.

1. Why B2B service is a different problem

1.1 The volume is concentrated in five questions

Unlike consumer business, where queries are varied, in distribution the vast majority of incoming messages repeat:

• When is the order I placed arriving?
• How much does this product cost for me?
• Do you have stock of this?
• How much do I owe? Can you send me my account statement?
• An item was missing from my last order.

Five questions answered with data that is already in the ERP, and that today consume most of the time of a commercial team that should be selling.

1.2 The answer depends on who is asking

Here is the technical difference from a retail chatbot. In B2B there is not one price: there is the price on the list assigned to that customer, with their volume discount and their particular terms. The stock shown as available may be reserved. Credit has a limit. The same question has different answers depending on who asks it.

An agent that is not connected to the customer's data cannot answer anything useful, and that is why most generic chatbots fail in this context: they end up being an option menu that hands off to a human.

1.3 The real cost of dispersion

When customer service lives in each rep's personal WhatsApp, three things happen. The customer's history leaves with the person when they resign. No indicator of response or resolution time exists. And queries that arrive outside business hours — which in distribution are many, because the customer's buyer works nights or Saturdays — get answered the next day, or not at all.

2. What an agent can resolve today

2.1 Status queries and hard data

Order status, estimated delivery date, tracking number, account balance, recent invoices, current price for that customer, availability of a product. These are queries with a deterministic answer: the data is in the system and the agent's only job is to identify the customer, understand the question and fetch the right figure.

That is 60% to 70% of the volume, and it is where automation is safe, because no judgement is at stake.

2.2 Capturing and classifying what it cannot resolve

Everything else also gains from the agent, even when it does not resolve it. Every incoming message ends up classified by type (complaint, order, commercial query, collections, logistics), associated with the right customer and routed to the right person with the context already assembled.

The value here is not automation but traceability: nothing gets lost and everything has an owner.

2.3 Taking orders

When the message is an order rather than a query, the agent can interpret it, validate it against catalog and stock, and leave it ready to post. It is the same logic we develop in order intake with AI, operating over the same support channel.

2.4 Proactive outreach

The agent does not only respond: it also initiates. Letting a customer know an order shipped, reminding them of an invoice coming due or confirming a delivery are outbound messages that reduce inbound queries. A good share of support volume exists because the customer has no information, not because they have a problem. Automated collections is the clearest case of this.

3. What it should not resolve on its own

3.1 Complaints and credit notes

A missing item, a damaged product or a billing discrepancy carries economic and emotional consequences. The agent can take the complaint, ask for the photo, classify it and open it with all the information — but the resolution involves commercial judgement and should stay in a person's hands.

3.2 Negotiating prices and terms

An off-list discount, an extension of payment terms or a credit exception are commercial decisions. The agent reports what is currently in force; it never negotiates something new.

3.3 Customers in a sensitive situation

Accounts far past due, customers with an open dispute or key accounts in the middle of a renegotiation should have an explicit rule for immediate hand-off. The cost of an automated message landing badly in those cases is disproportionate.

3.4 When the agent is not sure

The correct design includes a confidence threshold. If the agent cannot identify the customer with certainty, if the question is ambiguous, or if the data it needs is not available, it hands off. An agent that invents a delivery date does more damage than one that does not answer.

4. Implementation and measurement

4.1 Channels and identifying the customer

The WhatsApp Business API and email cover most of the volume in distribution. The first technical problem to solve is identification: linking a phone number or a mailbox to the customer code in the ERP. Without that mapping resolved, nothing else works.

It is best to start from the base of contacts you already know, and to have an onboarding flow for the ones that do not appear.

4.2 Connecting to the data

The agent queries orders, invoices, balances, price lists and stock in real time. The golden rule is that it queries rather than replicates: a mirror database refreshed each night produces stale answers, which is exactly what creates the distrust that then takes years to reverse.

4.3 Governance and limits

Define explicitly which information the agent accesses, what it may say and what it may not. An agent with access to a product's margin is a security problem. One that can confirm deliveries that are not confirmed is a commercial problem.

4.4 KPIs

The indicators that matter from the first month:

• Volume by query type: the map of where the team's time goes.
• First response time and total resolution time.
• Automatic resolution rate: the percentage resolved without human intervention.
• Hand-off rate and the reason for it.
• Out-of-hours queries answered, which used to be zero.

The realistic goal in the first months is 50% to 70% automatic resolution of the total. Promising more than that is the fastest way to lose internal credibility.

4.5 How to start

One channel, one query type and one customer segment. The usual approach is to start with order status on WhatsApp for the most active customers, measure for two or three weeks, and only then add query types. Internal resistance from the commercial team drops a lot once they see the agent takes the repetitive questions off their plate, not the customers.

Frequently asked questions

Does the customer realize they are talking to an agent? Yes, and you should tell them. Transparency does not reduce satisfaction when the answer is fast and correct. What does reduce it is discovering a deception.

Is it useful for small distributors? It depends on message volume, not company size. If there are more than 30 or 40 repetitive queries a day, the return shows up quickly.

Does it replace the support team? It changes their work. Data queries get automated and the team is left with complaints, negotiation and relationships. In practice the team does not shrink, it gets reassigned.

What about phone support? It is the same problem with a different interface. We cover it in AI voice agents.

How nBlock answers every message instantly

nBlock's Customer Service block unifies WhatsApp and email on top of your ERP data:

• An AI agent that answers status, price, stock and balance queries with each customer's real information.
• Automatic classification of every incoming message, with the customer already identified.
• Hand-off to the rep with context, and follow-up until it closes.
• Live response, resolution and volume KPIs.

---

Want to see how many of your daily queries can resolve themselves? Book a demo.