The CSR row spends 4 minutes of after-call work on every booking — 4 minutes of typing notes, updating the customer record, tagging the call type, flagging the dispatch handoff. At a 4-CSR floor handling 90 calls a shift, that is 18-20 hours per week of phone-row time burned on documentation that 95% of the next-touch staff never reads. AI call summaries — built into CallRail Conversation Intelligence, ServiceTitan call summaries, Avoca's post-call notes, and Housecall Pro AI Team — drop that 4 minutes to 15 seconds and produce a cleaner, more searchable, more handoff-ready customer record than the CSR typing manually ever did. The lift on after-call work (ACW) is 70-85%; the reclaimed time converts to additional booking capacity or seat reduction; the downstream effect on dispatcher confidence and tech preparedness is bigger than the ACW number suggests. This lesson is the build instruction for the AI call summary workflow at every level — what each vendor's product actually produces, how to configure the system prompt so the summary lands in the customer record cleanly, the verify discipline that catches the 5% of summaries with hallucinated facts, and the handoff cadence between CSR, dispatcher, and tech that converts the saved time into booked revenue.

The 4-Minute ACW Tax and What the Old Workflow Actually Cost

The pre-AI CSR documentation workflow lives in muscle memory at every shop that has not implemented call summaries. The call ends. The CSR has the headset on, the customer record open in ServiceTitan or Sera or HCP or Jobber, and 4 minutes of typing ahead before the next call can land. The typing covers: customer name, address, equipment, symptom, prior service history references, dispatch handoff context (urgency, access notes, preferred tech, gate codes, pet warnings), pricing context (was a dispatch fee discussed, was a membership pitched, was financing mentioned), and the tagging fields the FSM platform needs for routing. A skilled CSR types fast; a typical CSR types slow. The variance is 2-7 minutes per call. The shop's actual ACW averages around 4 minutes when you measure it honestly across 200 calls.

The 4 minutes look small. The annualized number is not. At a 4-CSR floor, 90 calls per shift, 21 business days per month, the ACW tax is roughly 25,200 minutes — 420 hours — per month. At a $19-$24/hour CSR loaded cost, that is $8K-$10K per month, $96K-$120K annualized. The shop pays it whether the documentation is good or not. And the documentation is usually not good — CSRs typing under volume pressure produce fragmented notes the next-touch staff cannot use, with key fields missed (gate codes, pet warnings, prior service context) because the CSR ran out of time before the next call landed.

The downstream cost is larger than the labor cost. A dispatcher reading a fragmented note misroutes a call; a tech arriving without prior service context redoes diagnostic work; a follow-up CSR fielding the next call has no record of what was promised. Each downstream miss is 15-45 minutes of recovery work and a small trust break with the customer. The 4-minute ACW tax is the visible cost; the downstream miss tax is invisibly larger. Owners who have lived through the transition consistently report that the bigger surprise is not the reclaimed CSR time but the dispatcher-and-tech handoff quality jump that lands in week two.

What the Four AI Summary Products Actually Produce

Four vendor products dominate the AI call summary surface for residential trades in 2026. Each one produces a slightly different artifact. Knowing which is which prevents the procurement mistake of buying the wrong summary engine for the FSM platform the shop runs.

CallRail Conversation Intelligence

CallRail's Conversation Intelligence module is the call-tracking-and-summary engine most independent trades shops already pay for — CallRail is the dominant call-tracking platform in residential home services, and the AI summary module typically adds $30-$50 per month per tracked number. The summary it produces is structured: a 3-5 sentence executive summary at the top, a key-facts list (customer name, address, callback number, equipment mentioned, symptom described, slot booked), a sentiment tag (positive, neutral, frustrated, angry), and a missed-opportunity flag (was a price quoted without a slot booked? was a service-area edge ignored? was a membership opportunity missed?). The CSR row gets the summary in the CallRail dashboard within 60-90 seconds of call end, and pushes it into the customer record in the FSM via copy-paste or API integration. CallRail's strength is depth — multi-year corpus of trades calls — and breadth across FSM platforms it integrates with (ServiceTitan, Sera, HCP, Jobber, FieldEdge). Its weakness is that the summary lives in CallRail's environment until pushed to the FSM, adding a handoff step.

ServiceTitan Call Summaries (Titan Intelligence)

ServiceTitan's in-platform call summary module is the play for shops on ServiceTitan as the FSM core. The summary lands directly in the customer record — no copy-paste, no API push, no second dashboard. The structured fields ServiceTitan exposes (customer details, equipment, prior service, membership tier, callback window) populate from the call audio against the existing customer record context, producing summaries that reference prior tickets and recognize the customer's equipment history. The advantage is integration depth: Titan Intelligence has access to the pricebook, the dispatch board, the membership module, and the customer history in real time, so the summary references all of them. The trade-off is platform lock — Titan Intelligence call summaries only run on ServiceTitan, and the shop pays for them through the ServiceTitan platform fee (roughly $400-$700 per seat per month for the platform that includes the AI module in 2026).

Avoca's Post-Call Notes

Avoca's post-call notes are the artifact produced after the Avoca AI receptionist handles an inbound call. The notes lean richer than CallRail's executive summary — Avoca's product team writes the system prompt around the trades-call surface and produces notes that include the rebuttal patterns used, the homeowner's stated concerns versus implicit concerns, and the recommended next-touch language for the human CSR follow-up call. The notes flow into ServiceTitan, Sera, HCP, or Jobber via API; the CSR row reads them in the customer record. Avoca's notes are the deepest summary on the market for the call-types Avoca handles (after-hours, overflow, the calls the AI took rather than the human). They do not cover calls the human CSR handled directly — those still rely on CallRail or the FSM's in-platform summary.

Housecall Pro AI Team

Housecall Pro's AI Team is the in-platform summary engine for HCP shops, launched broadly in 2026 across the $1M-$15M HVAC, plumbing, and electrical segment. Same in-software logic as ServiceTitan's Titan Intelligence: summaries land directly in the HCP customer record, reference equipment and prior service history, and trigger HCP's automation engine for downstream actions (on-my-way text, day-before confirmation, post-job review nudge). The 2026 HCP case studies show ACW dropping from 3-5 minutes to 12-18 seconds per call, with dispatcher handoff quality lift measurable in dispatch board re-routing reduction (5-8% fewer mid-day reshuffles driven by missing call-context).

The decision tree: ServiceTitan shop on ST native AI, use Titan Intelligence call summaries. HCP shop, use HCP AI Team. Jobber shop, use Jobber's in-platform call summary or CallRail layered on. Any shop running Avoca on after-hours and overflow uses Avoca's post-call notes for those call types specifically. Most shops end up running CallRail Conversation Intelligence layered on top of the FSM-native summary because CallRail's missed-opportunity flagging is the strongest secondary feature and the cost is low enough to layer without procurement headache.

What an AI Call Summary Actually Contains and the System Prompt That Shapes It

The summary is not a transcript. It is a structured artifact the next-touch staff reads in 8-12 seconds to absorb the call's context. The structure that works across vendors has six fields, and the system prompt the shop configures is what makes the summary land in those six fields consistently.

Field One: Executive Summary

Two to three sentences in shop voice that capture what the call was about, what got committed, and what the next-touch staff needs to know. Example: "Customer called about a 14-year-old Trane condenser making a knocking sound after Sunday's storm. Booked diagnostic Tuesday 8-10 a.m. with Marco; customer mentioned considering replacement this fall, flag for Comfort Advisor follow-up if diagnostic surfaces compressor issue." The system prompt enforces the three-sentence ceiling and the "what got committed" anchor — without those constraints, AI defaults to 6-8 sentences of narrative the next-touch staff skims and misses the commitment.

Field Two: Key Facts

Customer name (verified against caller ID), service address (geocoded against service area), callback number (verified by the AI receptionist or read back by the human CSR), equipment make/model/age, symptom in customer's own words, urgency tier, and dispatch slot booked (window and tech). The key facts are the dispatcher's primary read. A summary missing the equipment age or the slot window forces the dispatcher to back into the source — the call recording or the customer record — to find it. The system prompt makes the key-facts field non-optional: if a fact is missing, the field reads "not captured" rather than omitting the line, which surfaces the gap for human follow-up.

Field Three: Prior Context

Any references the call made to prior service, prior tickets, prior tech relationships, or prior promises. Example: "Customer referenced the Marin Park job from June 2024 — Marco was the tech, customer was satisfied with the install but felt the warranty conversation was rushed." This field is where the AI summary outperforms human CSR notes by the widest margin. A human CSR cannot remember 200 prior tickets; the AI summary pulls prior-context references in real time from the customer record (FSM-native) or surfaces them from the call audio (CallRail). The next-touch CSR reads the prior context and the conversation continuity holds across handoffs.

Field Four: Sentiment and Urgency

Sentiment tag (positive, neutral, frustrated, angry) and urgency tier (true emergency, same-day important, next-day acceptable, scheduled non-urgent). The sentiment tag drives downstream escalation: an angry sentiment on a booked call flags the dispatcher to prioritize the slot and the service manager to follow up after the visit. Urgency tier drives dispatch routing — true emergencies bypass the standard queue. The system prompt's job is to anchor sentiment to specific transcript markers (raised voice, profanity, repeated frustration phrases) rather than vibes; sentiment-by-vibes produces false positives and the floor stops trusting the tag.

Field Five: Missed-Opportunity Flag

CallRail's signature feature, now matched by Avoca and ServiceTitan: did the call miss a membership pitch, a financing mention, a service-area validation, a rebuttal opportunity? The flag does not punish the CSR — it surfaces the pattern for the 4 p.m. review huddle. A floor seeing 8 missed-membership flags per day from the same CSR knows where the coaching focus belongs that week. The system prompt defines the missed-opportunity set against the shop's actual offer mix (which membership tiers, which financing tiers, which warranty-extension options) rather than a vendor-generic checklist.

Field Six: Next-Touch Recommendation

The action item for the next person who touches the customer record. Example: "Comfort Advisor follow-up call Wednesday morning if Tuesday diagnostic surfaces compressor issue; include Section 25C federal tax credit walkthrough for heat-pump replacement option given equipment age and homeowner's stated replacement consideration." The next-touch recommendation converts the summary from passive documentation into active routing. The system prompt's job is to constrain the recommendation to the shop's actual handoff patterns (Comfort Advisor follow-up, dispatcher reroute, service manager escalation, marketing nurture handoff to Hatch) rather than generic suggestions.

The Verify Discipline for AI Summaries

AI call summaries hallucinate in roughly 3-7% of cases — wrong equipment age, wrong service address, wrong slot window, fabricated prior-service reference, mistranscribed callback number. The shop that pushes summaries to the customer record without verify discipline embeds the hallucinations in the customer record where they compound through future touches. The discipline that protects the customer record is the 30-second verify pass adapted from L1 Cardinal Rule discipline, applied to every AI-generated summary before it lands in the FSM.

The Five-Checkpoint Summary Verify

The five checkpoints map directly onto the Cardinal Rule structure. Numbers: equipment age, slot window, dispatch fee, any pricing referenced — all checked against source-of-truth (manufacturer plate for age, dispatch board for slot, pricebook for fees). Names: customer name spelled against caller ID, address geocoded against service area, equipment make/model against the customer record. Parts: any part numbers referenced (less common in CSR-row summaries than in tech-job notes) cross-referenced against supplier catalog. Warranty terms: any warranty language in the summary checked against the shop's warranty matrix. Financing and regulatory language: any financing reference checked against the actual lender portal output rather than AI-generated approximation. The pass takes 25-35 seconds when the CSR has it muscle-memorized, applied to every AI-summary before push to the FSM. The pass is faster than typing 4 minutes of notes; the verify discipline preserves the time savings without sacrificing the record integrity.

The Confidence Flag and the Low-Confidence Queue

All four products expose a confidence score (0-100); the vendor recommends auto-push above 85 and human-review-queue below. Tighten this: queue threshold 80, with the CSR-floor lead or service manager running a 25-second verify on each queued summary before push. The queue averages 6-12 summaries per day on a 4-CSR floor — 5 minutes of review time. The cost of the queue is trivial; the cost of skipping it is a hallucinated record the dispatcher reads as gospel and the tech arrives at the wrong house. The 4 p.m. review huddle absorbs the queue as its retrospective leg.

The Prior-Context Falsification Risk

The highest-frequency hallucination class is the prior-context field. AI confidently produces "customer referenced the Marin Park install from June 2024" when the customer made no such reference — the AI saw a Marin Park record in the customer history and inserted it into the summary even though it was never discussed on the call. The verify discipline catches this by reading the prior-context field against the actual call audio via the link-to-recording feature (CallRail, ServiceTitan, HCP all expose 15-second jump-to-timestamp). Without the audio-check, false prior-context compounds into the record and the next CSR builds the next conversation on a fabrication.

The Handoff Cadence Between CSR, Dispatcher, and Tech

The ACW reduction is the visible win. The handoff quality lift is the bigger win. The summary changes how three roles interact, and the shop that designs the handoff cadence deliberately captures the full value.

CSR to dispatcher. Pre-AI, the dispatcher reads a fragmented CSR note 6-12 minutes after the call ended. Post-AI, the dispatcher reads a structured 6-field summary within 60-90 seconds of call end. The routing decision is faster and better-informed; dispatch board re-routing reductions of 5-8% are the measurable downstream effect — fewer "I need to move Tuesday morning around because the dispatcher didn't have the gate code" rebalances.

Dispatcher to tech. Pre-AI, the tech read a 6-line CSR note and the dispatcher's "Marco, take this one" — minimal context, frequent on-arrival surprises (the customer is the elderly homeowner not the son; the equipment is in the attic not the basement; there's a dog in the yard). Post-AI, the tech reads a 6-field summary with key facts, prior context, sentiment, and next-touch recommendation pre-baked. On-arrival surprise rate falls 30-45% in published 2026 deployments; average time-on-stop drops 8-12 minutes.

Tech back to CSR for follow-up. The tech's job-notes (covered in L2 Ch4 Lesson 1) feed back into the summary chain. A return call from the same customer 2 weeks later lands at the CSR with the full call-and-job history pre-summarized: original call, dispatch, tech findings, repair-vs-replace pivot, financing conversation, follow-up commitment. The CSR reads in 12 seconds what previously required 4 minutes of opening tabs.

The handoff cadence is operating discipline, not technology. The summary lands in the FSM; the dispatcher reads it before routing; the tech reads it before rolling; the CSR reads it before the follow-up. Shops that train the cadence in week 1 capture the full ACW and handoff lift by week 4. Shops that deploy the tool and skip the cadence training capture only a fraction.

The Metrics and the 90-Day Rollout

The AI summary workflow moves five numbers, and the CSR-floor lead and service manager watch them on the dashboard.

ACW per call. Pre-AI baseline 4 minutes; target 15-30 seconds (verify pass plus push). 70-85% cut. The number drops fastest in weeks 1-2.

Customer-record completeness score. Percentage of records with all six summary fields populated. Pre-AI baseline 45-60% — most records have name, address, symptom but miss prior context, sentiment, next-touch. Post-AI target 92-96%, measurable from week 2.

Dispatch board re-routing rate. Frequency of mid-day re-routing driven by missing call context. Pre-AI baseline 12-18% of bookings re-routed; post-AI target 4-8%.

On-arrival surprise rate. Frequency of techs encountering at-door information that should have been in the booking note. Pre-AI baseline 15-25%; post-AI target 8-14%. The 30-45% reduction shows up as time-on-stop reduction.

Hallucination detection rate. Percentage of AI summaries flagged by verify discipline as containing factual errors. Healthy band 3-7%. Lower than 3% means verify is being skipped; higher than 10% means the system prompt needs tuning or the model needs vendor escalation.

The 90-day rollout sequence: week 0, configure the system prompts for the six-field summary structure, set the confidence threshold at 80, define the low-confidence-queue routing. Week 1, deploy to one CSR seat as pilot; CSR runs verify on every summary; floor lead reviews the low-confidence queue. Weeks 2-4, expand to full floor; ACW drops to target; verify discipline gets muscle-memorized. Weeks 5-8, dispatcher and tech read-cadence trained; on-arrival surprise rate drops; handoff lift becomes visible. Weeks 9-12, the customer-record completeness score lands at 92-96% and the workflow stabilizes. Total payback typically lands at month 2-3 — the labor cost reduction alone covers the AI summary product cost within 60-90 days at most shop sizes.

Key Takeaways

• The 4-minute ACW tax is the visible cost; the downstream miss tax is bigger. 4-CSR floor burns 420 hours per month on documentation. Fragmented notes cause dispatcher misroutes, on-arrival tech surprises, and follow-up CSR confusion that cost 15-45 minutes of recovery work per miss.
• Four AI summary products dominate residential trades in 2026: CallRail Conversation Intelligence (cross-platform, $30-$50/mo per number), ServiceTitan Call Summaries (Titan Intelligence in-platform), Avoca's post-call notes (deepest for AI-handled calls), Housecall Pro AI Team (in-HCP). Most shops end up layering CallRail on top of the FSM-native summary.
• The summary has six fields: executive summary (2-3 sentences in shop voice), key facts (customer/address/equipment/slot), prior context (references to prior service/tickets), sentiment and urgency (anchored to transcript markers, not vibes), missed-opportunity flag (membership/financing/service-area), next-touch recommendation (action item, not narrative).
• AI summaries hallucinate in 3-7% of cases. The 5-checkpoint verify pass (numbers, names, parts, warranty terms, financing language) catches them in 25-35 seconds. Skipping verify embeds hallucinations in the customer record where they compound through future touches.
• The highest-frequency hallucination is the prior-context field. AI inserts records from customer history that the call did not reference. The verify discipline reads prior-context against the call audio via the link-to-recording feature in 15 seconds.
• The handoff cadence is operating discipline, not technology. Dispatcher reads pre-route; tech reads pre-roll; CSR reads pre-follow-up. Skipping any read erodes the lift. The full handoff training lands in week 1; the lift compounds by week 4.
• Five metrics move: ACW per call (4 min → 15-30 sec, 70-85% cut), customer-record completeness (45-60% → 92-96%), dispatch re-routing rate (12-18% → 4-8%), on-arrival surprise rate (15-25% → 8-14%), hallucination detection rate (3-7% healthy band).
• The low-confidence queue threshold is 80. Summaries below 80 confidence route to the floor lead for 25-second review; the queue averages 6-12 per day on a 4-CSR floor and absorbs into the 4 p.m. review huddle. The cost of the queue is trivial; the cost of skipping it is a hallucinated customer record the dispatcher reads as gospel.
• Payback lands at month 2-3. Labor cost savings alone cover the AI summary product cost within 60-90 days at most shop sizes. The handoff quality lift is the bonus the ACW number does not capture.