Should a professional services firm put a chatbot on its website?

Who the answer is about, and what happens when the visitor asks something the tool should not answer.

1. Subject. An advice chatbot answers questions about the visitor's own situation: their dispute, their tax position, their symptoms. A triage assistant answers questions about the firm: what it does, who it is for, what it charges, how work starts.
2. Source. An advice chatbot draws on the model's general knowledge. A triage assistant answers only from the firm's published pages and says when it cannot.
3. Refusal. An advice chatbot tries to help with everything. A triage assistant declines the substantive question and points to the relevant page or offers to hand off to a person.
4. Risk. An advice chatbot creates a duty of care the firm cannot supervise. A triage assistant repeats what the firm has already chosen to publish.

The first is a liability with a friendly face. The second is a receptionist that has read the website.

Because no partner can defend an answer they did not see, given to a person they have not met, about a matter they have not assessed.

A visitor who asks a website whether their dismissal was unfair, whether a transaction is taxable, or whether a symptom needs a referral is asking a question the firm would normally answer only after taking instructions. A model answering in the firm's name, on the firm's domain, produces advice the firm did not give and cannot stand behind. The regulatory and insurance position is the firm's problem, not the vendor's.

We will not build that product, and we say so on the service page for AI Operations. If a firm wants it anyway, we are not the right people.

Only what the firm has already published, plus a clear route to a human.

A defensible triage assistant answers from the firm's own pages: services, prices, sectors, how the first meeting works, what the firm will not take on. When a visitor asks for advice on their own situation, it says it cannot help with that and points to the page that describes how the firm would help, or offers to pass the visitor to a named person. It does not guess, it does not fill gaps from general knowledge, and it does not pretend to be a person.

The value is in the questions it absorbs: the tyre-kicking, the early due diligence, the visitor who wants to know the price before picking up the phone. Each of those would otherwise take a fee earner's time before any engagement exists.

A written position on what the assistant may say, where the conversation goes, and what is stored.

The firm should be able to publish, in plain language, what the assistant answers from, which provider processes the text and under which terms, whether anything is stored and for how long, and what the visitor should not type into it. Our own position for the assistant on this site is published at Where does what you type go?, and a client deployment gets the same page.

Building the assistant is AI Operations work, priced from £8,000 as a fixed project once the scope is agreed. Where the firm has not yet decided which tools and data it can defend at all, AI Governance and Assurance comes first.