AI Receptionist for Law Firms: What to Buy, What to Build, and What Each Costs

Almost every law firm that calls me about this describes the problem as a phone problem. Calls come in, nobody picks up, work walks. The fix sounds obvious: put something on the phone that always picks up.

That framing produces a system nobody uses after month two. The phone was never the constraint. Qualification was.

The actual arithmetic

A firm with six offices takes a lot of calls. Some fraction are prospective clients with a real matter. A larger fraction are existing clients chasing an update, opposing counsel, process servers, vendors, and people who found the number on a directory and have a question the firm cannot help with.

The expensive failure is not the missed call. It is the paralegal who spends forty minutes with a caller who was never going to be a file, while a genuine matter sits in voicemail. Answering every call faster makes that worse, not better. You have simply increased the volume of unqualified contact reaching people who bill.

So the thing worth building is not a receptionist. It is a filter that runs before one.

What good intake actually collects

For most practices the qualifying set is short and boring, which is exactly why it automates well:

• Practice area. Is this something the firm does. A surprising share of inbound is not.
• Jurisdiction. Where did it happen, and are you licensed there.
• Timing. When did it happen. This is the limitation period question and it is the one that decides urgency.
• Conflict inputs. Names of the parties involved, collected so a real conflict check can run.
• Contact and consent. How to reach them, and permission to do so.

That is a five minute conversation a trained intake person has hundreds of times a year. It is repetitive, it follows a decision tree, and it is exactly the shape of work a language model handles well. It is also the work that most often gets skipped at 6pm on a Friday.

The three places it goes wrong

Clearing conflicts it should only be collecting for

A system that tells a caller "we can take your case" before a conflict check has run has created a problem for the firm, not solved one. The correct behaviour is to gather party names and say the firm will confirm. This is a configuration decision, and it is the first thing I check in any deployment I inherit.

Missing the clock

Limitation periods are the one piece of intake where being approximately right is worthless. If a caller describes something that happened close to a deadline, that call needs to reach a human today, not sit in a queue until Monday. Any system worth running has an explicit urgency path keyed to dates, and it is tested before launch.

Handling distress badly

Personal injury and family practices get calls from people in genuine crisis. A system that responds to that with a scripted question about practice area is worse than voicemail, because voicemail does not pretend to be listening. Escalation on distress signals is not a nice-to-have, it is the difference between a deployment that reflects well on a firm and one that ends up screenshotted.

The three options, honestly compared

There are three ways a firm solves this, and most of the writing about it is published by someone who sells one of them. Here is the version from someone who only gets paid for the third, and who tells most firms to pick the second.

A human answering service is the oldest answer and still the correct one for a small practice with low call volume. A person takes a message reliably. What they do not do is qualify, because they do not know your practice areas, your jurisdictions or your limitation periods, so everything still reaches you undifferentiated.

An off-the-shelf AI receptionist is where most firms should start, and several vendors will have you running on a legal-specific template inside a week. It answers, follows a qualification script, and drops the result somewhere: an email, a calendar booking, a spreadsheet. Pricing is a subscription plus usage, published openly, and small against the value of one recovered matter. The real advantage is that leaving is cheap, so a bad fit costs you a month rather than a project.

A custom build is not a better version of the second option. It is a different purchase, and for most firms it is the wrong one.

The threshold where building wins

There is one, and it is narrower than anyone selling builds will admit.

Buying is right while the output of intake is information a human then acts on. An email containing a qualified matter is fine. Somebody reads it and opens a file.

Building becomes right when the output has to be a record inside the system your firm already runs on: a matter created in Clio or Filevine, party names in the conflict fields, the incident date in the field your limitation reporting reads, the recording attached, and the right lawyer notified by the rules you actually use rather than the rules a template assumed. Off-the-shelf products stop at the boundary of their own database, because a product integrated that deeply with one firm's configuration has stopped being a product.

The honest test: if your intake currently ends with somebody re-typing details from an email into your case management system, that re-typing is the thing worth automating, and it is the thing you cannot buy. If nobody is re-typing anything, buy the subscription and stop reading.

How to read a proposal

The voice part is close to solved. Any vendor can give you a competent conversational agent with a Canadian number this month, and none of them are really competing on that any more.

So when you are quoted for a build, count the sentences. If the proposal is mostly about the model, the voice and the script, and light on where a qualified matter lands and who gets told, it is a proposal to give you a second inbox at project prices. You could have bought that on a subscription.

What to measure

Not call volume. Not average handling time. Two things:

• Matters captured outside office hours that converted to a file. This is the number the system exists to move.
• Unqualified contact reaching billable staff. If this has not fallen, the filter is not filtering.

A firm I work with runs intake this way across six offices in three provinces. The number that changed was not calls answered. It was how many people who bill by the hour spent that hour on someone who was never a case.

Is it worth doing

If your intake is one person who handles it well during business hours, and your practice area does not generate urgent after-hours contact, probably not yet. Fix the hours before you buy software.

If enquiries arrive at night, on weekends, and while everyone is in court, and if the cost of a missed matter is measured in five figures, then the arithmetic is straightforward and the only real question is how well it gets wired into what you already run.

And if the honest answer for you is the subscription, take the subscription. I would rather you spent a hundred dollars a month than a project budget on something a template already does.

If you are past that threshold, the scoping engagement that works out what your intake actually has to do, before anyone commits to a number, is the MVP Roadmap. More on how this and the rest of it fits a practice is on the page for law firms.