We built an AI receptionist for law firms. Here’s what we learned
A few years ago, I needed legal help after an accident.
I called a law firm and spent a long time waiting on the line. At that moment, I was not comparing websites, reading every review, or studying each firm’s credentials. I simply wanted someone to answer.
That experience stayed with me.
For consumer-facing law firms, a missed call is often more than a missed conversation. It can mean losing a potential client to the next firm they contact. That is why we started building Lexidesk.
Lexidesk is an AI phone and chat intake platform designed specifically for law firms. It answers enquiries 24/7, asks the right qualification questions, books consultations, handles urgent transfers, and sends structured case summaries to the firm’s CRM.
One of the biggest lessons we learned is that answering every call is not enough. The difficult part is understanding how each firm evaluates a potential case.
A personal injury firm, an immigration practice, and a family law firm all have different qualification criteria, workflows, terminology, and escalation rules. The AI agent has to reflect how the firm actually works rather than follow a generic script.
We have already seen some encouraging results:
• King Law Offices increased qualified leads by 33.9% and reached a 100% answer rate.
• Modern Law reduced abandoned calls from 23% to around 1%, while conversion increased from 31% to 49%.
• IMD Solicitors increased qualified enquiries by 109% over six months.
We are still learning how much of the intake process law firms are comfortable automating and where people expect a human to step in.
I would love to hear from other founders building vertical AI products:
How do you decide which parts of a specialized workflow should be automated, and which should always remain human?
That’s exactly the kind of environment I had in mind.
I wouldn’t try to impose one universal standard across every firm. The starting point would be the firm-specific operating boundary you already define during deployment: qualification criteria, escalation rules, urgent-case triggers, prohibited behaviours, required handoff points and intended intake outcomes.
OpsWatch would treat those as the agreed assurance criteria, then independently sample production interactions against them.
For example, the review could test whether:
matters were qualified consistently with the firm’s own rules
urgent or high-risk conversations were escalated when required
the agent stayed inside the information-gathering boundary rather than drifting toward legal advice
human handoffs happened at the right point
summaries accurately reflected the underlying conversation
unexpected or ambiguous scenarios were handled safely rather than silently forced through the workflow
The important distinction is that Lexidesk would continue to operate and improve the system, while OpsWatch independently assesses whether the agreed production behaviour is actually occurring.
Sampling would also be risk-based rather than purely random — combining representative conversations with targeted reviews of escalations, edge cases, failures and higher-consequence interactions.
That separation becomes more valuable as intake volume grows, because the question changes from “did we configure the agent correctly?” to “can we independently demonstrate that it continues to behave within the agreed boundaries in production?”
A Lexidesk-style legal intake workflow would actually be a very natural OpsWatch pilot because the boundaries are already clearly defined.
Absolutely — I’d be happy to outline a small pilot.
Probably easiest to take the details off-thread. You can reach me directly at jason@mcgillintelligence.com.au and I can send through a simple proposed pilot structure covering scope, evidence required, timing and commercial terms.
Jason
McGill Intelligence / OpsWatch
Absolutely — I’d keep the first pilot deliberately small.
For a law-firm AI receptionist, we’d typically select one defined workflow and agree upfront on the behaviours and controls that matter most — for example intake accuracy, escalation triggers, sensitive or unusual enquiries, and whether the agent stays within its intended boundaries.
OpsWatch would then independently assess a risk-based sample of real production interactions, with extra attention on escalations, edge cases and higher-consequence conversations. The output would be a concise evidence-backed assurance report showing what performed as intended, what didn’t, and any areas worth tightening before volumes increase.
The idea is not to interfere with the implementation, but to give you and the law firm an independent view of how the agent is actually behaving in production.
Happy to outline a small pilot scope privately if useful.