2026-08-13 · 9 min read
How to add live chat to your website in 10 minutes
Live chat installs in about ten minutes — one script tag or a CMS plugin. Here is the full setup: install route, knowledge base, office hours and handover.
AI agent vs chatbot in customer service: what actually changed
Almost every support vendor renamed its chatbot an AI agent. The label is free; the capability behind it is not. Here is the question that separates the two, and what to check before you sign.
Also available in Dutch: AI-agent of chatbot voor je klantenservice: wat is het verschil echt?
Somewhere between 2024 and 2026 the word “chatbot” went out of fashion and “AI agent” replaced it on nearly every support vendor's homepage — frequently without a single line of the underlying product changing. That makes the term useless as a buying signal and makes demos hard to compare. The distinction is still real. It just is not about how well the thing writes.
The dividing line is write access, not writing ability
Fluency stopped being a differentiator. Any product built on a current language model will answer a return-policy question in idiomatic English, well enough that you cannot tell vendors apart from a scripted demo. What differs is whether the system is wired into your order database, your payment processor and your subscription billing with permission to write to them — and what governs that permission.
| Customer says | A chatbot does | An AI agent does | What to verify |
|---|---|---|---|
| “Where is my order?” | Explains how to find tracking in the account area, or asks for an order number and repeats it back. | Looks up the order, reads the live carrier status, and answers with the actual delivery date. | Ask to see it answer for a real order during the trial, not a seeded demo record. |
| “I want to return this.” | Quotes the returns policy and links to the returns form. | Checks eligibility against the policy and the purchase date, then creates the return and sends the label. | What happens when the item falls outside the window — does it refuse, or escalate with context? |
| “Cancel my subscription.” | Explains where the cancel button lives. | Cancels it, or files the cancellation for approval, and confirms the effective date. | Whether cancellation is reversible, and who gets notified. |
| “You charged me twice.” | Apologises and creates a ticket. | Verifies the duplicate charge, then routes a refund through an approval step rather than issuing it silently. | Whether a refund can ever execute without identity proof. This is the one to press on. |
Why the distinction got urgent this year
Two things landed at once. Language models became reliable enough that vendors were willing to let them call functions against production systems, and the EU AI Act's Article 50 transparency obligations became applicable on 2 August 2026, which forced everyone to be explicit about what the AI is and what it does. An agent that can move money sits in a different risk category from one that can only produce sentences, and the paperwork now reflects that. We covered the disclosure side separately in what the AI Act requires your chatbot to disclose.
What an agent needs before it is allowed to act
Write access without preconditions is not an upgrade, it is an incident waiting for a trigger. Three things have to be in place first, and they are worth checking in this order, because each is useless without the one before it.
- Identity proof. The system must know it is talking to the account holder before it reads private data or changes anything — normally a one-time code to the email or phone already on the order. An agent that acts on a name typed into a chat box will act for anyone who types that name.
- A stated policy it can evaluate. “Refunds within 30 days if unopened” is a rule a machine can check. “Use your judgement” is not, and an agent handed the second will improvise the first.
- An approval path for anything irreversible. Not every action needs a human, but every action needs a decision about whether it needs a human — recorded once, in advance, rather than improvised per conversation.
The approval step is the product, not the friction
The instinct is to treat human approval as a temporary compromise, to be removed once the AI has earned trust. In practice the queue is what makes write access safe enough to grant at all: the agent proposes a specific action with its evidence attached, a human confirms or rejects, and the action then executes under an idempotency key so a retry cannot double-refund. That is the pattern behind every write action here — outlined on the features overview and in the AI-to-human handoff playbook.
The useful part is that approval is per action type, not global. Sending a tracking link needs no review. Refunding €400 does. Once a category has run clean for a few hundred cases you can promote it to automatic with evidence behind the decision, instead of guessing on day one.
Questions that separate an agent from a renamed chatbot
- Which specific write actions ship today, against which systems? Ask for the list, not the roadmap.
- What identity check runs before a private-data lookup, and can it be switched off?
- Which actions require approval, who configures that, and is the setting per action type?
- What does the audit trail record — the action, the evidence, the approver, the timestamp?
- What happens on a duplicate request: is there an idempotency key, or will a retry run the action twice?
- When the agent cannot act, does the conversation reach a human carrying full context, or does the customer start over?
- How is success counted, and does a reopened conversation reverse it? That answer sets your bill under per-resolution pricing — see what counts as an AI resolution.
When a chatbot is still the right answer
If most of your volume is questions rather than requests — how does this work, do you ship here, what is in this product — then answering well is the whole job, and write access adds risk without adding resolutions. Content-heavy support gains more from a well-structured knowledge base than from integrations; that is a help-centre writing problem first. Start by counting how many of last month's tickets ended with someone changing a record. If it is under a fifth, buy the answering quality and leave the actions for later. Both are priced on our pricing page.
What is the difference between an AI agent and a chatbot?
A chatbot generates a response; an AI agent changes the state of a connected system, such as issuing a refund, editing a delivery address or cancelling a subscription. Both may run on the same underlying language model and both may answer equally well, so the practical test is not conversation quality but whether the tool holds write access to your systems and what governs that access.
Is an AI agent safe enough to issue refunds automatically?
It can be, provided three conditions hold: the customer's identity is verified before anything is read or changed, the refund policy is expressed as a rule the system can evaluate rather than a judgement call, and irreversible actions above a threshold you set are routed through a human approval queue. Without all three, an automatic refund path can be triggered by anyone who knows an order number.
Do I need an AI agent, or is a chatbot enough for my business?
Count how many of your recent tickets ended with a human changing a record in another system, such as a refund, an address edit or a cancellation. If that share is small, most of your value comes from answering accurately, and a well-fed knowledge base will do more than integrations will. If a large share of tickets end in a system change, answering alone leaves the actual work sitting in the queue.
Does the EU AI Act apply to customer service AI agents?
The Act's Article 50 transparency obligations, applicable since 2 August 2026, require that people are told when they are interacting with an AI system unless that is obvious from the context. This applies to a support assistant whether or not it can take actions. The above is general information rather than legal advice, and the obligations attaching to any specific deployment should be assessed for your own situation.
Why did every vendor start calling their chatbot an AI agent?
The term signals autonomy and supports higher prices, and nothing prevents a vendor from applying it to a product that only produces text. Because the label is not defined by any standards body, treat it as marketing rather than a specification, and evaluate instead the concrete list of write actions a product ships today against the systems you actually run.
Keep reading
2026-08-13 · 10 min read
Live chat vs chatbot: which one does your team actually need?
Live chat is a person; a chatbot is software. The honest answer is that you do not have to choose — modern widgets run AI first and hand off to a human.
2026-08-09 · 10 min read
When and how to hand off from AI to a human agent: an operational playbook
Customers forgive a bot that does not know. They do not forgive one that will not let go. The triggers worth wiring, what has to travel with the conversation, and how to tell a good handoff from a bad one.
Resolve more tickets automatically.
See how honestly-measured AI resolutions cut your support load — start on the Free plan, no credit card, no sales call to get started.