Key takeaways
- Judge AI call handling on outcomes along the whole chain, from call reached to instruction, not on calls answered alone.
- Build a baseline from phone system data for the same hours and weeks, before anything changes.
- Score a sample of calls for quality every week; most businesses using AI apply human checking to its outputs[1].
- Out-of-hours calls used to reach voicemail, so compare them with what voicemail captured, not with staffed hours.
- Report a short scorecard monthly, with the context that explains each movement.
Why AI call handling gets judged unfairly
Most firms judge a new phone tool in one of two ways, and both mislead. The first is anecdote: one partner hears about one bad call and the tool is finished. The second is a vanity number: 'it answered every call', which says nothing about whether the right people were reached or whether anyone instructed the firm. A fair judgement needs a small set of measures chosen before the change, collected the same way before and after, and read alongside a sample of real calls.
This matters because most organisations are still early. The ONS found only 10% of UK businesses with ten or more employees that use AI report using it extensively[2]. Whether a firm moves from a trial to everyday use usually depends on how well the first few months are measured.
Start with the outcome chain
A new enquiry passes through a chain of stages, and each stage can fail on its own. Measuring every link shows where the problem is, rather than producing one number nobody can act on. The phone is still where much of this happens: it is the most common way people communicate with their main adviser on a contentious issue, at 28%[3].
- Reached: the call connected to the firm rather than ringing out or hitting an engaged tone.
- Answered: a person or assistant spoke to the caller before they hung up.
- Understood: the matter type and urgency were identified correctly.
- Captured: name, contact details, other party and summary were recorded accurately.
- Routed: the enquiry reached the right team or person within the target time.
- Booked: a consultation was offered and accepted.
- Instructed: the caller became a client.
Speed through the chain counts. The Legal Services Consumer Panel found 69% of consumers took a week or less to choose a provider, and 42% shopped around before choosing[4]. A captured enquiry that waits three days for a call-back may already be lost.
The core metrics and how to calculate them
| Metric | Definition | Calculation | Data source |
|---|---|---|---|
| Answer rate | New-enquiry calls answered before the caller hangs up | Answered ÷ calls reached | Phone system call log |
| Correct routing | Enquiries sent to the right team first time | Correctly routed ÷ enquiries captured | Call notes reviewed by team leads |
| Capture completeness | Required fields recorded and accurate | Complete records ÷ enquiries captured | Sample review against recordings |
| Urgent escalation time | Minutes from call to a person accepting an urgent matter | Median across urgent calls | Escalation log |
| Call-back speed | Hours from capture to first human contact | Median, working hours only | Case management or CRM timestamps |
| Consultation rate | Enquiries that book an initial consultation | Booked ÷ in-scope enquiries | Diary and enquiry log |
| Instruction rate | Enquiries that become instructions | Instructed ÷ in-scope enquiries | Matter opening records |
| Caller feedback | Complaints or negative comments about the phone experience | Count per hundred calls | Complaints log and call notes |
Two choices make these numbers trustworthy. Count 'in-scope' enquiries separately from wrong numbers, sales calls and existing clients chasing updates, or the conversion figures will swing for reasons that have nothing to do with the tool. And use timestamps from systems, not recollection, for anything measured in minutes or hours.
Build a fair baseline
Pick the comparison period
Use at least four weeks immediately before the change, and note anything unusual in them: a staff absence, a marketing campaign, a bank holiday or a seasonal peak.
Export call data, don't estimate it
Pull calls reached, answered and abandoned from the phone system for the same hours the new tool will cover.
Count what voicemail captured
For out-of-hours and overflow calls, the honest baseline is the number of usable voicemails and call-backs made, not the number of calls.
Trace a sample to outcome
Follow a sample of baseline enquiries to booking and instruction so you have conversion rates to compare against.
Freeze the definitions
Write down what counts as in-scope, urgent and complete, and use the same definitions throughout the pilot.
Quality: listen as well as count
Numbers tell you where to look; recordings tell you what happened. Review a fixed sample each week, chosen at random rather than from complaints, and score each call on five checks: matter type identified correctly, no advice given, details accurate, right route and a tone the firm would be happy with. A call fails if any check fails.
This is normal practice, not distrust. DSIT's AI adoption research found that 84% of UK businesses using AI apply at least some human input or checking to AI outputs, and 67% apply significant checking[1]. Keep the reviewer the same for the baseline and pilot samples where possible, because scorers vary.
Review prompt for each sampled call: Would I be comfortable if this caller read the transcript aloud to the managing partner? If not, note why and what rule would have prevented it.
Listen to what callers are telling you
Complaints data shows what clients care about. Poor communication (24%) and delay or failure to progress (22%) were the most common complaint types brought to the Legal Ombudsman in 2025/26[5]. A phone tool that answers quickly but sends enquiries into a slow call-back queue improves the first number and does nothing for the second.
Callers also notice the channel. In the Consumer Panel's 2026 tracker, 17% of consumers would have preferred to speak to a legal professional more than they did, rising to 24% of those served by telephone and 25% of those served by email or online[6]. Track how often callers ask for a person and how quickly they get one.
Source: Legal Services Consumer Panel, 2026[6]
Show the data
| Item | Value |
|---|---|
| All consumers | 17% |
| Served by telephone | 24% |
| Served by email or online | 25% |
Comparing like with like
- Separate staffed hours from out-of-hours. During office hours the question is whether the tool helps busy staff; out of hours it is whether it beats voicemail.
- Separate new enquiries from existing clients. Existing clients chasing updates need a different route and a different measure.
- Watch for marketing changes. A new campaign or directory listing changes call volume and quality whatever answers the phone.
- Allow for the learning curve. Rules and wording change in the first weeks; compare the final fortnight of a pilot with the baseline as well as the whole period.
- Do not move the targets. If a measure was chosen before the pilot, report it even if it looks poor.
An example monthly scorecard
| Measure | Baseline, 4 weeks | Pilot, 4 weeks | Comment |
|---|---|---|---|
| In-scope new enquiries | 212 | 226 | Similar volume; no campaign running |
| Answer rate | 81% | 97% | Overflow and evening calls now answered |
| Correct routing | Not measured | 92% | Most misroutes were wills vs probate |
| Median call-back time | 19 hours | 4 working hours | Morning review of overnight notes |
| Consultation rate | 31% | 38% | Evening slots offered on the call |
| Calls failing quality review | Not measured | 3 of 80 sampled | Two gave timescales; wording changed |
Example: in this fictional scorecard the answer rate rises sharply, but the more useful movement is the call-back time, because that is what clients complain about. The three quality failures matter more than their number suggests; two of them involved the tool estimating how long a matter would take, so the rule was tightened and the test script extended. A partner reading this page should be able to see what improved, what went wrong and what was changed, in under five minutes.
Measuring the effect on staff time
Partners often ask whether the tool saved time, and it is easy to answer badly. 'Hours saved' worked out from call counts multiplied by an assumed call length overstates the gain, because reception staff rarely did nothing else while a call was on hold. Better measures are concrete and observable:
- Interruptions: how often fee earners were pulled out of work to take or return a new-enquiry call, counted for a sample week before and during the pilot.
- Admin per enquiry: minutes a support assistant spends turning a call into a usable record, timed on a small sample.
- Rework: enquiries where someone had to call the client back because details were missing or wrong.
- Out-of-hours load: messages and missed calls waiting at 9am on a Monday, before and after.
Ask the reception team as well. A short fortnightly check-in about which calls the tool handled well and which it made harder often explains a movement in the numbers faster than any report.
Reporting and decisions
Report monthly on one page: the chain metrics, the quality sample, the three most common reasons calls were escalated, and any change made to the rules. Agree in advance what result would justify continuing, what would justify extending the trial and what would end it. For how to set up the trial itself, see running a 30-day AI reception pilot; to judge results by fee earner as well as by channel, see measuring consultation conversion by fee earner.
Frequently asked questions
What is the most important metric for AI call handling in a law firm?
There is no single one. Answer rate shows whether callers are reached, but routing accuracy, call-back speed and the share of enquiries that book a consultation show whether the firm actually benefits.
How long should the baseline period be?
At least four weeks immediately before the change, covering the same hours the tool will cover, with any unusual events noted.
How many calls should we review for quality?
Enough to see patterns: a fixed random sample every week, scored by the same reviewer against the same five checks. Increase the sample in the first month or after any rule change.
How do we measure out-of-hours performance fairly?
Compare it with what happened before, which was usually voicemail. Count usable messages and call-backs in the baseline against captured, correctly routed enquiries in the pilot.
References
Department for Science, Innovation and Technology (GOV.UK) (2026). Artificial Intelligence Adoption Research (opens in a new tab) ↑
Office for National Statistics (2026). Artificial intelligence in UK businesses: 2023 to 2026 (opens in a new tab) ↑
Legal Services Board (2024). Legal Needs Survey 2023: Summary report (opens in a new tab) ↑
Legal Services Consumer Panel (2026). How consumers are choosing legal services (Tracker Survey 2026) (opens in a new tab) ↑
Legal Ombudsman (2026). Legal Ombudsman 2025/26 annual complaints data and insight (opens in a new tab) ↑
Legal Services Consumer Panel (2026). How consumers are using legal services (Tracker Survey 2026) (opens in a new tab) ↑