Tuning with transcripts, sentiment, outcome classes and containment rate; plan for the 10 simultaneous calls limit
Tune the AI with transcripts, sentiment and outcome classes, and plan for the 10 simultaneous calls limit.
Read the lesson
Goal: Run a weekly tuning loop that raises the share of calls the AI finishes alone. Plan for busy hours.
You need: A live agent with at least one week of calls, access to Analytics, and the call map.
Why this matters
No agent is right on day 1. Callers say things nobody planned for. Prices change. A new service starts. The agents that work best are tuned every week by someone who reads the calls.
Tuning pays twice. First, more calls end in a booking or an answer. Second, calls get shorter, and each minute costs 15 credits. A tuned agent books more and costs less.
Busy hours also need a plan. The AI takes up to 10 calls at the same time. An 11th caller hears ringing with no pickup. In a storm or a cold snap, that limit matters.
The numbers you watch
| Number | Where | What it tells you |
|---|---|---|
Success Rate | Analytics → Calls | Calls answered out of all calls |
| AI vs human split | Analytics → Calls | How many calls the AI took |
Disposition | Analytics → Calls | How calls were sorted and resolved |
Resolution Rate, Transfer Rate, Avg Handle Time, Top Topics | Call Analytics Performance view | How the AI calls ended and how long they took |
| Credits used | Billing → Usage history | What the voice line costs |
The Disposition labels come from the calls themselves. Read them in each client's account. Do not assume a fixed list.
Containment rate
Containment rate is the share of AI calls the AI finishes without a person. Work it out yourself each week:
Containment = (AI calls − transfers − follow-up tasks) ÷ AI calls
Transfers and tasks are not failures. Emergencies should transfer. But each one is work for a person. When containment rises, staff get time back.
Do not chase 100%. Emergencies on the owner's list must always reach a person. A healthy agent contains most routine calls and transfers every true emergency.
Sentiment
The live view shows sentiment alerts when a caller sounds upset. The Create tab in Workflows also has a Negative Sentiment Alert recipe. It can alert the office after a frustrated call. Read every call that triggered an alert. Sentiment often points to a gap in the flow, such as a question the agent cannot answer.
The weekly tuning loop
Set aside 45 minutes a week per client. Follow the same five steps.
- Pull 20 calls. Take every transfer, every call under 30 seconds, every sentiment alert, and random calls to make 20.
- Tag each problem. Use four tags: wrong fact, flow gap, action failed, caller confused.
- Fix the cause, not the call. A wrong fact goes to
Knowledge. A flow gap goes to the voice instructions. A failed action goes to the calendar orActionssettings. - Test and publish. Retest the fixed case in
Talk in browser. ThenPublish. - Log the change. Write the date, the problem, the fix and the number you expect to move.
The weekly tuning loop
- 1Pull 20 callsTransfers, short calls, alerts, random
- 2Tag each problemWrong fact, flow gap, action failed, caller confused
- 3Fix the causeNot the single call
- 4Test and publishTalk in browser, then Publish
- 5Log the changeDate, problem, fix, expected number
- 6Measure next week
How to read a transcript
Read each transcript with three questions in mind.
- Where did the call turn? Find the first moment the caller sounded confused or repeated themselves. The problem is usually just before it.
- What did the agent not know? Find each call-back promise on a question it should answer. That is a missing quick answer.
- What took the longest? Long stretches often come from stacked questions or long read-backs. Those are flow fixes.
Read the caller's words, not only the agent's. Callers use words the owner never uses, such as "the water thing" for a water heater. Add those words to the instructions so the agent understands them.
Short calls under 30 seconds need a special look. Some are wrong numbers. Others are callers who hung up on the greeting. If many hang up in the first 10 seconds, shorten the greeting.
Where each fix goes
- Tag the problem in the transcript
- Is it a wrong fact?YesFix it in KnowledgeNoCheck the next tag
- Is it a flow gap?YesFix the voice instructions in BasicsNoCheck the next tag
- Did an action fail?YesCheck the calendar or ActionsNoCaller confused: shorten that step
- Retest in Talk in browser, then Publish
When to bring in the owner
Some fixes need the owner. Never change these on your own:
- A price or a discount.
- The emergency list or the safety wording.
- Which actions the agent may take, such as cancelling.
- What the agent promises, such as "same day".
Collect these in a short list. Raise them on the monthly review call. Get a written yes before you change them.
Tuning by business size
- Owner-operator: Fewer calls, so read every AI call for the first month. Then read 10 a week.
- Fleet: Read 20 a week, plus every transfer. Share one tagged example with the dispatcher each week. They spot problems you miss.
- Multi-location: Tune each location's agent on its own. Copy a fix to the other locations only after you test it there.
Change only what the data shows. If you change ten things at once, you cannot tell which fix worked.
Plan for the 10-call limit
Estimate how many calls run at once in the busiest window. Use this rule of thumb:
Calls at once = calls in the window × average minutes ÷ minutes in the window
If the answer gets close to 10, plan for overflow. There is no overflow setting and no backup forwarding number. You have three options:
- Missed-call text-back. A workflow with
Call missed→Send SMStexts callers who did not get through. Test whether it fires for overflow callers in the client's account. - A named callback person. One staff member watches
Conversationsand calls back during peaks. - A higher limit. The business can pay ServiceAgent to raise it. Ask ServiceAgent for the price.
No-answer forwarding also helps. Staff answer first, so fewer calls reach the AI at once.
What happens to caller 11
- A caller reaches the AI number
- Is the AI already on 10 calls?YesCaller hears ringing with no pickupNoThe AI answersEnds here
- No overflow setting catches caller 11
- Plan: text-back, a callback person or a paid higher limit
Read the numbers together
Each number alone can mislead. Read them as a set.
Success Rateup and containment flat: more calls are answered, but people still finish them. Look at the task and transfer reasons.- Containment up and
Avg Handle Timeup: the agent finishes more calls but talks longer. Check for long read-backs. Transfer Rateup after a change: the change may have made the agent unsure. Read those calls first.- Credits up with calls flat: calls got longer. Find the step that grew.
Write one sentence each week on what the numbers say together. That sentence goes into the monthly owner report.
Plan credits for the season
Voice is the biggest credit cost. Check Usage history each month and plan ahead for busy seasons. Bought credits always roll over. Plan credits roll over only on yearly billing. A bigger pack bought ahead costs less per credit than many 500-credit Safe Packs.
Build it in ServiceAgent
- Open
Analytics→Calls. RecordSuccess Rateand the AI vs human split as your weekly baseline. - Open the Call Analytics
Performanceview. RecordTransfer RateandAvg Handle Time. - Count transfers and follow-up tasks for the week. Work out containment.
- In
Voice Agents, pull 20 transcripts and tag them. - Fix one cause per problem in
Knowledge,BasicsorActions. - Retest each fix with
Talk in browser, thenPublish. - In
Workflows, buildCall missed→Send SMSif the client has peak hours near the limit. - Check
Billing→Usage historyand compare credits used with the plan. - Check it works: next week, the tagged problems drop and containment rises.
Worked example: Northside Plumbing and Polar Air Services
Northside: containment and credits. In week 3, Northside's agent took 20 AI calls. It made 3 transfers and 4 follow-up tasks.
Containment = (20 − 3 − 4) ÷ 20 = 13 ÷ 20 = 65%.
The tagged calls show one gap. Callers asked if Northside serves Riverside, and the agent could not answer. Riverside was missing from the service area quick answer. You add it and publish. In week 4, tasks drop to 2. Containment = (20 − 3 − 2) ÷ 20 = 75%.
Containment at Northside, weeks 3 and 4
Now credits. Northside's AI line takes about 77 calls a month. Calls average 4 billed minutes.
- Monthly voice use: 77 × 4 × 15 = 4,620 credits.
- Core includes 2,000. Shortfall: 2,620 credits.
- With auto-reload adding 500 each time: 6 Safe Packs of 500 = 3,000 credits for 6 × $20 = $120.
- Bought ahead: one 2,500 pack ($80) plus one 500 pack ($20) = 3,000 credits for $100.
Two ways to buy Northside's shortfall
- 6 Safe Packs of 500
- 3,000 credits
- $120 a month
- One 2,500 pack and one 500 pack
- 3,000 credits
- $100 a month
Tuning also helps here. Cut average billed minutes from 4 to 3.5. The line then uses 77 × 0.5 × 15 = about 578 fewer credits a month.
Polar Air: the 10-call limit. Polar Air has about 20 techs. On a cold Monday morning, an example burst sends 30 calls to the AI in 15 minutes. Calls average 4 minutes.
Calls at once = 30 × 4 ÷ 15 = 8 on average.
Calls at once at Polar Air
The average is 8, so short spikes will pass 10. You build the missed-call text-back. You name the dispatcher as the callback person for cold mornings. You show the owner the option to pay for a higher limit.
Common mistakes
- Tuning by feel. Fix: tag 20 calls a week and fix what the tags show.
- Changing many things at once. Fix: one cause per fix, logged with a date.
- Chasing 100% containment. Fix: emergencies must still transfer.
- Ignoring the limit until a storm. Fix: estimate calls at once before the busy season.
- Letting Safe Packs pile up. Fix: buy a bigger pack ahead for known busy months.
Do this now
Build a weekly tuning log for a practice business. Add columns for date, calls read, tags, fix, and expected number. Work out containment for one sample week. Then estimate calls at once for its busiest 15 minutes.
Practice the conversation
- Operator Each week I read 20 of your AI's calls and fix what went wrong.
- Owner What kind of things?
- Operator Last week it didn't know you serve Riverside. Now it does.
- Owner How do I know it's getting better?
- Operator I track how many calls it finishes without your team. It went from 65% to 75%.
- Owner What happens on a crazy Monday in January?
- Operator It takes 10 calls at once. Caller 11 hears ringing.
- Owner That's bad.
- Operator So I set up a text-back and test that it fires. Your office calls those callers back.
- Owner Can we just get more lines?
- Operator You can pay ServiceAgent to raise the limit. I'll get the price.
Quick check
- On a cold Monday, the agent is on 10 calls. An 11th customer rings. What happens, and what do you plan?
- Ringing with no pickup; plan text-back, a named callback person, or a paid higher limit
- Voicemail, so the caller leaves a message
- A hold message until a line frees up
- An automatic forward to the `Transfer to your team` number
- An agent took 40 AI calls this week, with 6 transfers and 6 follow-up tasks. What is containment?
- 85%
- 70%
- 30%
- 15%
- Tags show 5 calls this week where the agent gave an old opening time. What do you fix?
- Rewrite the whole call flow
- Update the hours in `Knowledge`, retest, then `Publish`
- Transfer all calls about hours
- Turn off `Answer questions`
ServiceAgent UniversityLearn to run a service business on autopilot. Free courses, Blueprints and certification for Service Operators.