AI Call Summaries: Log Every Phone Enquiry in Your CRM

Route your business number through a phone system that records, transcribes and summarises calls (Quo's Business plan or Aircall Professional, for example), switch on its CRM integration so each call is logged against the matching contact, then add one extraction step that turns the summary into proper fields. Budget roughly $23-$50 per user a month plus an afternoon of setup.

The summary on its own isn't the win. Most integrations drop a paragraph onto the contact's timeline, and a paragraph can't be filtered, reported on or chased. Decide first which six to eight facts every enquiry must capture, and make the AI fill those fields. The other gap is calls taken on personal mobiles: they never reach the system at all, so no amount of AI will log them.

Follow me on Instagram@sagnikteaches

The enquiry fields to define before you change any settings

Start with a sheet of paper, not a pricing page. Write down what you need to know about a new enquiry to decide whether it's worth a visit and who should call back. Those items become CRM properties (custom fields), and the AI's job is to fill them from the conversation. Anything that isn't a field stays in the summary text, where a human can read it but a filter can't find it.

Connect on LinkedInSagnik Bhattacharya

Here is the list a four-person kitchen fitter might settle on. The examples are illustrative, but the logic carries across trades: each field either decides priority or saves a question on the callback.

Subscribe on YouTube@codingliquids

Two rules keep this list honest. Every field needs a fixed set of allowed values where possible (a dropdown rather than free text), because AI output is far easier to check against five options than against anything it likes. And every field needs a "not given" value, so the AI has somewhere to put silence instead of guessing.

Three routes from a ringing phone to a CRM record

There are really only three ways to get call content into a CRM, plus a fallback for teams on mobiles. The prices below are list prices checked in September 2026; confirm them on the vendor pages before you commit, because per-seat plans and minimums change.

Two plan details trip people up. Quo's entry Starter plan only summarises calls handled by its AI agent and lacks the CRM integrations, so the Business plan is the real starting point for this job (see the Quo pricing page). Aircall's Essentials plan includes transcription, but conversation summaries sit on Professional, and both need at least three licences (see the Aircall pricing page). A two-person firm on Aircall still pays for three.

Six setup stages, about four hours of hands-on time

The order matters. People usually connect the CRM first and discover a week later that half the calls came in on an old mobile number and never got recorded.

Stage 1: Put every enquiry call on one number (30-60 minutes)

List every number a customer might ring: the one on the website, the one on the vans, the one printed on old quotes, the owner's mobile. Either port them to the new system or forward them to it. Check that the provider offers numbers where you trade, and ask how long porting takes from your current carrier, because it is rarely same-day. Install the calling app on everyone's phone so a fitter on site answers the business line through the app, not their personal dialler.

Stage 2: Recording notice and retention (20 minutes)

Recording rules vary by jurisdiction. Some places need only one party's consent, others need everyone on the call to agree, and data-protection rules add duties about how long you keep audio. Play a short notice on every call and set a retention period. Wording that works on a greeting:

Thanks for calling. We record and summarise calls so we can follow up accurately. If you'd rather we didn't, just tell whoever answers.

Then decide how long audio stays. Keeping the transcript and summary on the CRM record while deleting the audio after 30 to 90 days is a common compromise; if you're unsure what's right for your business, ask your data-protection adviser.

Stage 3: Connect the CRM and switch on auto-logging (30 minutes)

In the phone system's integrations screen, connect the CRM, switch on automatic logging for calls, and choose what happens with numbers the CRM doesn't recognise. My default is to create a contact only when the call lasted longer than about 45 seconds, so wrong numbers and robocalls don't pollute the database. If the contact already has an open deal, Quo also puts the logged call on that deal's timeline, which is what you want for existing jobs.

Stage 4: Add the extraction step (60-90 minutes)

This is where a paragraph becomes fields. Quo exposes "call summary completed" and "call transcript completed" events to Zapier and to its own webhooks; Aircall and most other systems have equivalents. The automation reads the transcript, asks an AI step to fill your fields, finds the contact, updates it, and creates a deal only for genuine new enquiries. Use a filter before the AI step to skip calls under 45 seconds: Zapier doesn't charge tasks for filter steps.

The prompt does most of the work. Give it your allowed values and forbid guessing:

You fill CRM fields from a phone call transcript for a kitchen fitting firm.
Return JSON only, using exactly these keys and allowed values.

enquiry_type: "full kitchen" | "worktops only" | "repair" | "existing job" | "supplier or sales" | "unclear"
caller_role: "homeowner" | "landlord" | "builder" | "not given"
room_and_size: short text, or "not given"
budget_band: "under 10k" | "10-20k" | "over 20k" | "not given"
timescale: "within 3 months" | "3-6 months" | "researching" | "not given"
source: "recommendation" | "search" | "signage" | "repeat" | "not given"
next_action: one short sentence
callback_due: ISO date if a time was agreed, otherwise "not agreed"
evidence: for budget_band and timescale, quote the caller's exact words

Rules:
- Only use what the CALLER said. If the caller didn't state something, use "not given".
- Never convert a vague phrase into a number. "Not too expensive" is "not given".
- If two people were on the call, fill fields from the customer, not our staff.

Here is how that plays out on a short, illustrative excerpt from a showroom call:

Caller: We've just bought the house and the kitchen's a bit tired. It's a galley,
maybe three metres long? We were thinking something in the region of fifteen,
sixteen thousand, and ideally done before the baby's due in March.
Staff: Lovely. How did you hear about us?
Caller: My sister had her utility done by you last year.

And an illustrative output from the AI step:

{"enquiry_type": "full kitchen", "caller_role": "homeowner",
 "room_and_size": "galley, about 3m long", "budget_band": "10-20k",
 "timescale": "within 3 months", "source": "recommendation",
 "next_action": "Book a design visit before the end of the month",
 "callback_due": "not agreed",
 "evidence": "budget: 'fifteen, sixteen thousand'; timescale: 'before the baby's due in March'"}

Two things to fix in that output. "Within 3 months" is an inference from a date the caller gave, which is fine only if March really is within three months of the call date, so pass the call date into the prompt. And "next_action" invented a deadline ("before the end of the month") that nobody agreed. Tighten the rule: next_action describes the step, never a date unless one was agreed on the call.

Stage 5: Test with ten real calls (one week)

Don't judge the setup on test calls between colleagues; they're too clean. Run it for a week on real traffic and check ten records against their transcripts. Score each field right, wrong or missing. If one field is wrong more than twice, fix the prompt or the allowed values for that field before moving on.

Stage 6: Build the list people actually open (20 minutes)

A log is only useful if someone looks at it. Create a saved CRM view called something like "Enquiries: callback due" filtered on enquiry type, callback date and no completed follow-up. Pin it for whoever does callbacks. Pair it with a missed-call text-back so unanswered calls get an immediate reply and a record even when nobody picks up.

A four-person kitchen fitter, costed month by month

Take an illustrative firm: a showroom manager, a designer and two fitting crews. The showroom takes around 70 calls a week. Roughly 30 are new enquiries, 25 are existing customers mid-job, and the rest are suppliers, deliveries and sales calls. Before the change, enquiries went on a paper pad by the phone, and when the owner counted one month of pad pages against the phone bill, several calls a week had no note at all.

The costs, at list prices:

• Quo Business for three users (manager, designer, owner) billed annually: 3 × $23 = $69 a month. The crews keep their own phones for site work.
• Zapier Professional billed annually: $19.99 a month for 750 tasks.
• Setup: about four hours of the owner's time, plus a week of checking.

The quick sum people skip is the task count. Each enquiry that passes the filter uses an AI step (1, 3 or 5 tasks depending on the model tier chosen), a find-contact step, an update step and a create-deal step. At the cheapest tier that's 4 tasks; at the middle tier, 6. Thirty enquiries a week is about 130 a month, so 520 tasks at the cheap tier fits inside 750, while 780 tasks at the middle tier doesn't. Run the automation only on calls from numbers without an open deal, and existing-job calls stop consuming tasks at all.

On time: if writing up an enquiry properly takes three minutes, 130 enquiries is six and a half hours a month of note-taking removed. The larger gain is harder to count: enquiries that used to go unrecorded now have a record, a budget band and a callback date.

Before and after: one enquiry, two records

The difference is easiest to see side by side. Before, the pad by the showroom phone said:

Tues. Lady re kitchen, new house. Ring back. (number on caller ID?)

Nobody could tell from that whether it was a $4,000 worktop swap or a $20,000 refit, who took the call, or whether the callback happened. After the change, the CRM record for the same call holds the enquiry type (full kitchen), caller role (homeowner), room (galley, about 3m), budget band (10-20k), timescale (within 3 months), source (recommendation), a next action (book design visit) owned by the designer, the full summary, the transcript and a link to the recording. The daily callback view surfaces it the next morning without anyone remembering.

Where AI call summaries put the wrong facts in your CRM

The errors aren't random. They cluster in a few places, and each has a recognisable symptom.

• Shared numbers. A couple ring from the same landline, or a builder rings on behalf of three clients. The integration matches on the number, so the second enquiry lands on the first person's record. Symptom: a contact with two addresses in its summaries. Fix: when a summary names a different person from the contact, route it to a review list instead of updating.
• Misheard numbers. "Fifteen" and "fifty" sound alike on a bad line, and the summary will state the wrong one confidently. Symptom: a budget band that doesn't match the job type. The "evidence" field in the prompt above exists so the checker can see the caller's words in two seconds.
• Suppliers becoming leads. A worktop supplier rings to confirm a delivery, and the automation creates a new deal called "kitchen enquiry". Symptom: a pipeline full of deals with no budget and no timescale. Fix: create deals only when enquiry type is one of the three real lead types.
• Withheld numbers. With no number to match, some setups create a blank contact per call. Send these to a review list instead.
• Invented commitments. Summaries like to tidy things up. "Customer agreed to a quote visit on Thursday" when the caller actually said "Thursday might work, I'll check". Symptom: a no-show that wasn't really booked. Keep the "only if agreed" rule and train staff to confirm times out loud before hanging up, so the transcript contains a clear yes.
• Transferred calls. A call passed from showroom to designer can arrive as two call records with two summaries. Check how your system handles transfers during the test week.

If you're reviewing transcripts for quality as well as accuracy, the routine in reviewing AI call transcripts for quality and compliance pairs well with the spot-check below.

The same setup for a locksmith and a surveying firm

The machinery stays identical; the fields change, and so does which field matters most.

For a two-van locksmith, the phone is the business. Most calls are urgent, and the costly disputes come from prices quoted verbally at 11pm. The fields that earn their place are: lockout or planned job, property type, whether anyone vulnerable is locked out, the price or price range quoted on the call, and the arrival time promised. The price quoted is the one to protect, because it is the fact customers most often remember differently. With the transcript logged next to the job, a disagreement about "you said $90" becomes a two-minute check rather than an argument. Night calls that nobody answers can go to an after-hours AI phone agent, which logs into the same place.

For a six-person surveying firm, calls are fewer but longer, and the valuable detail is in the middle of the conversation: the survey level wanted, the property's approximate value and age, where the buyer is in the purchase, any deadline the buyer is working to, and who will give access. A useful extra field is "reason for urgency", because a buyer with a lender deadline next week and a buyer who is "just getting quotes" need very different callbacks. Surveyors here often prefer the voice-note route for calls taken while on site, dictating the key details straight after hanging up.

A fortnightly spot-check that keeps the log trustworthy

Once it runs, trust decays quietly unless someone measures it. Every two weeks, spend 20 minutes on four numbers. The targets are my suggested starting points, not industry benchmarks:

1. Coverage. Count calls in the phone system's log for the fortnight, then count logged calls in the CRM. Aim for 95% or more of calls over 45 seconds. A drop usually means someone has gone back to answering on a personal phone.
2. Field accuracy. Pick 20 enquiry records and check the eight fields against the transcripts. That's 160 checks. Under 90% right means a prompt fix is due; look at which field fails most.
3. Duplicates. Search for contacts created in the fortnight with the same surname or address. More than two or three means your matching rule needs work. The tutorial on cleaning up a messy CRM with AI covers merging what's already there.
4. Enquiries without a next action. Any genuine enquiry older than two working days with no completed follow-up is a lead going cold.

Record the four numbers in a simple table each fortnight. After two months you'll know whether the setup is steady, and you'll have the evidence to decide whether the next job is sales calls too, which is covered in updating your CRM automatically after sales calls.

Call summaries and your CRM: follow-up questions

Do I have to record calls to get an AI summary?

In practice, yes. The summary is written from a transcript, and the transcript comes from a recording, even if the audio is deleted afterwards. Most phone systems let you set how long recordings are kept, so you can keep the summary and transcript on the CRM record and delete the audio after a short period. Tell callers at the start that calls are recorded, and check the consent rules where you and your callers are.

Can I get call summaries if my team uses personal mobiles?

Not automatically. Calls made on a personal mobile never pass through a system that can record them. The options are to move enquiry calls onto a business calling app on those same phones, which most VoIP systems offer, or to use a voice-note habit: after each enquiry call, the person records a 30-second memo that an automation transcribes, structures and files. The second option is cheaper but depends on people doing it every time.

Which CRM should a small trade business use for this?

Use the CRM you already have if your phone system integrates with it. Quo's native integrations cover HubSpot, Salesforce and Pipedrive on its Business plan, and Aircall connects to a long list of CRMs. If your job-management software is where enquiries really live, check whether the phone system connects to it directly or through Zapier before changing anything else.

How accurate are AI call summaries?

Good at the gist, less reliable on numbers, names and addresses, which are exactly the fields a CRM needs. Accents, road noise and people talking over each other cause most errors. Treat budget figures, measurements and addresses as unconfirmed until someone reads them back, and run a fortnightly spot-check of 20 calls against their transcripts to see where your setup slips.

Further reads

• How to Set Up an AI Receptionist Without Losing Callers — Let an AI answer when nobody can, and log it too.
• What It Costs to Set Up an AI CRM for a Small Team — Price the CRM side properly before adding call logging.
• Are AI Meeting Note-Takers Safe for Client Calls? — The privacy questions to settle before recording customers.
• Can AI Answer My Business Phone? What It Can and Can't Handle — What an AI can and can't handle on your main line.
• What Is an AI CRM and Which Features Actually Save Time? — Which built-in CRM AI features are worth switching on.
• How Much Do Missed Calls Cost a Heating and Cooling Business? — A worked sum on what unlogged and missed calls cost.
• AI Phone Answering for Restaurants: What It Costs and Pays Back — Published prices for restaurant AI phone answering, the costs that never make the pricing page, and a payback calculation built from your own missed calls.
• AI Receptionist or Answering Service for a Booking-Based Business? — AI receptionists cost about $0.50-$2 a call; human answering services $7-$15. Sort your calls first, then use this rule to choose or combine them.
• How to Measure an AI Receptionist's Return in the First 90 Days — A 90-day plan to prove whether your AI receptionist pays: the baseline to record first, eight weekly numbers, and a worked example for a butcher's shop.
• How to Set Up an AI Phone Line for Takeaway Orders — Seven steps from counting your calls to a tested AI phone line that takes takeaway orders into your till, with the three routes compared.
• Can AI Answer the Phone and Take Restaurant Reservations? — How AI voice agents book, change and cancel tables, which calls they must pass to staff, and a twelve-call test before you trust one.
• Which Event Planning Tasks Should You Hand to AI First? — Twelve event planning tasks scored for AI, the five to hand over first with worked examples, and the ones to keep even though AI will attempt them.
• How Florists Can Use AI for Orders, Reviews and Social Posts — Worked examples for a flower shop: a rushed phone order turned into a docket, a substitution message, a year of reviews analysed, and accurate captions.
• What Is an AI Receptionist and How Does It Handle Bookings? — A plain-English look at what happens inside an AI receptionist during a booking call, and the three very different things vendors mean by 'handles bookings'.
• Cutting Parcel-Tracking Calls With AI Delivery Updates — How a small courier firm can find why customers ring, send updates at the moments that matter, and let AI handle exceptions and leftover questions.
• How to Write Sales Call Scripts With AI That Don't Sound Scripted — Swap word-for-word scripts for an AI-built call map: openers, questions and objection cards drawn from your own best calls, then rehearsed aloud with AI.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Quo pricing page and Quo support documentation (HubSpot integration, webhooks); Aircall pricing page; HubSpot Sales Hub pricing page and HubSpot Knowledge Base on call recording and transcription; Zapier task and AI step documentation.