Why do AI agents make things up?
Because the underlying language models are built to produce a fluent answer to any question, and without constraints they will produce one whether or not they have the facts. Ask an ungrounded agent 'is that included in our plan' and it will name a plausible answer. It has no idea what your plans include. The customer, hearing a confident voice, acts on it — and your team inherits the argument.
Grounding fixes this. A grounded agent is instructed to answer only from a specific body of content — your website, your help center, what you've pasted in — and to say when the answer isn't there. The knowledge is retrieved per question, the model is told to use only that, and 'I don't know' becomes an allowed, expected answer rather than a failure to avoid.
What should the agent do instead?
Three things, in order. First, say it plainly and without theatre: 'I don't have that price in front of me.' Not 'my information doesn't include' — nobody talks like that. Second, offer the next step that actually helps: 'I can take your details and have someone from the team confirm, or if it's urgent I can put you through now.' Third, do it — take the message with the name, the account and the question, notify whoever is on the rota, or transfer the call live with a spoken summary.
Wiredesk's agents are built exactly this way. They answer only from approved content, they are told in their instructions that saying 'I don't know' is correct, and they have the tools to take a message and to transfer the call. When someone adds the missing answer to the knowledge — one line — the agent knows it from the next conversation.
How do I test this before I sign up for anything?
Ask it something your business would never publish. 'Which customers are on the enterprise plan?' 'Can you give me 30% off?' 'What will the migration cost us?' A grounded agent declines the first, says it can't authorise the second and offers a person, and takes the details for the third without naming a figure.
Then assert something that is on your site, slightly wrong. 'Support is 24/7, right?' when it isn't. The agent should correct it from your own content rather than agree to be agreeable. Agreeableness is the tell of an ungrounded model.
You can run both tests against Wiredesk in the browser, on your own content, in the first five minutes of a pilot workspace.
What about questions the agent shouldn't answer even if it knows?
Some things are in your content and still shouldn't be committed to on a call — a contractual edge case, a discount that needs approval, anything touching security or legal. A good agent lets you mark those as 'always hand to a person'. The behaviour is the same as not knowing: acknowledge, offer the next step, take the message or transfer. The rule is yours; the agent follows it every time.