Should a law firm put an AI agent on first contact?

This decision arrives as a technology question and is really a scope question. Nothing about an automated first contact is inherently unsafe or inherently valuable — what determines the outcome is the boundary drawn around what it is permitted to do. Firms that draw it narrowly tend to find the agent straightforwardly useful. Firms that let a vendor draw it get a system conducting substantive intake at the exact moment a prospective client relationship is being formed, which is the highest-risk place in the funnel to be improvising.

Start with what is genuinely gained. The enquiry arrives, someone or something responds immediately, the caller is acknowledged, the general nature of the matter and its urgency are captured, and a specific callback time is committed. That is real and it is most of the value, because the failure this replaces is not a slow lawyer — it is an enquiry that received nothing at all until the next business day, or never.

Now the boundary. A first-contact agent must not give advice, must not state or imply that the firm has agreed to act, and must not take detailed matter intake before a conflicts check has run. Each of those is a rule about what is said, not about who says it, and each is as binding on a person as on a system. The difference is that a person exercises judgement at the edges and a system does exactly what its script and its model allow — which is an argument for a narrower scope, and also an argument that the narrow scope will actually be adhered to, which is not always true of a rushed human intake.

The measurement problem is the reason this decision goes badly. The thing that would settle it is the share of enquiries that become signed matters, defined over every enquiry reaching the firm — including the ones never answered and the ones outside its practice areas. Most firms cannot produce that number, because their denominator starts at the point an enquiry entered the case management system, which is to say after intake has already filtered it. A firm that measures signing over enquiries-that-reached-intake will show an improvement from almost any change to intake, including a harmful one.

So the honest sequence is to fix the denominator first. Log every inbound enquiry at the point of arrival, across every channel, before any human or automated triage touches it. That single change makes the agent's effect visible and is worth doing whether or not the agent is ever bought. Without it, the firm will have a vendor's dashboard reporting on a population the vendor also defines.

One further property is worth being deliberate about: what happens when the agent does not know. The safe behaviour is to stop, say a person will follow up, and capture the details. The dangerous behaviour is to continue plausibly. That single design choice does more for the risk profile of an automated first contact than any amount of policy elsewhere, and it is a question that can be asked of a supplier before signing.