Use human agents instead of AI voice agents when a call is high in value, complexity, risk or emotion. Keep AI for routine, low-risk, calm calls. This framework scores each call type on those four factors and gives you a decision matrix.
Score every call on four factors
Rate each call type from 1 (low) to 3 (high) on four factors. Score the call type, such as renewal reminders or disputed bills, not the individual customer.
| Factor | What it measures | A 3 looks like |
|---|---|---|
| Value | Lifetime value or deal size | A large account or a big quote |
| Complexity | Unknowns, number of steps, need for judgement | Site-specific variables, no fixed script |
| Risk | Regulatory, financial or reputational harm | Required disclosures, money disputes |
| Emotion | Anger, fear, grief, distrust | Hardship, a complaint, a customer who suspects a scam |
The rule: any 3 on complexity, risk or emotion goes to a human. Otherwise, add the four scores. A total of 4 to 6 goes to AI, and 7 to 9 goes to AI with a human check. Tune these bands in your pilot.
Every row still needs a human exit. A Gartner survey of 3,566 customers found that 87% say a company using GenAI must offer a way to reach a human. Even your lowest-scoring calls need a clear path to a person.
Treat your first scores as a hypothesis. Gartner's analysis of 432 AI service use cases found that only a quarter show positive ROI. Another 42% cannot be measured. Measure before you widen the AI share.
Use the decision matrix to route calls
Add up each call type's scores, apply the complexity, risk and emotion rule, then read the route off this table.
| Call profile | Route |
|---|---|
| Low value, simple, low risk, calm | AI |
| High value, simple, low risk, calm | AI with a human check |
| Complexity scores 3 | Human |
| Risk scores 3 | Human |
| Emotion scores 3 | Human |
| High value plus high emotion | Senior human |
Here is how common revenue calls land:
- Renewal reminders and post-install satisfaction surveys: AI. They are short and predictable, and operators want proactive service calls before a complaint arrives.
- KYC completion nudges: AI. Lenders and fintechs lose people who start identity checks and never finish, and emails alone are not fixing it.
- Lead follow-up and lapsed-customer win-back: AI with a human check. A person reviews transcripts and takes over warm replies.
- Payment reminders: AI. Reminders are not collections, and a happy customer pays. Score a collections call separately. It often carries a 3 on risk or emotion.
- Large-quote negotiation and disputed amounts: human. The first has value and complexity, and the second has risk and emotion.
Complexity hides inside "simple" calls. A waste-management operator we spoke with would not automate quote follow-up, because quotes depend on container size and site access. A call that looks like a routine follow-up can score a 3 on complexity.
High-value calls need skilled people because small behaviours decide the outcome. In our insurance sales research, top agents close at 65% and bottom agents at 0% under identical conditions. Across 6,209 sales conversations, top performers ask 37% more questions per call and score 68% higher on empathy and rapport. A script cannot easily copy those skills on a large account, so give your best humans the calls that score high on value.
Hand over to a human by design
Write the handover rules before you launch, because buyers ask about handover on their own. They want to know if it is polite, what happens to a disputed amount, and whether anyone chases a handover nobody picked up. They also fear AI that goes off script.
Set one trigger for each factor, plus one for every call:
| Factor | Handover trigger |
|---|---|
| Any call | The customer asks for a person |
| Value | The account carries a high-value flag |
| Complexity | Two or more failed answers in a row |
| Risk | A disputed amount, or a regulated or sensitive topic |
| Emotion | Anger or distress words, or a mention of hardship |
Then design the handover itself:
- Warm transfer: Send the human a short summary, so the customer never repeats their story.
- Polite wording: Have the AI say a person is joining and why.
- A queue someone chases: A missed handover should alert a named owner.
- A fallback: If nobody picks up, offer a callback time and log it.
Ask every vendor how each of these works. Do not assume a product has escalation rules because the demo mentions them.
Disclosure is part of the design. In the EU, Article 50 of the AI Act says any AI that talks to people, including automated phone systems, must say it is AI from the first interaction, and it has applied since 2 August 2026. In the US, AI-voice calls need consent under the TCPA after the FCC's 2024 ruling. Open every call with the disclosure, then offer the human route.
Test the split on a small pilot
Pilot one outbound call type that scores low on all four factors, on a small sample. Start with something like renewal reminders or post-install surveys. We advise starting with outbound, because much inbound volume exists only because companies never send proactive updates. Regulated buyers want a short, low-risk pilot, with payment tied to a measured lift.
Let AI call one group and keep a human control group on the same call type. Then compare:
- Handover rate: how often the AI passes the call to a person. A high rate means the call is more complex or emotional than you thought.
- Answered and reached rate: a call where only the AI spoke should not count as reached.
- Outcome rate: renewals confirmed, surveys finished, KYC steps completed.
- Complaint rate: the first number to move if a risk or emotion score was too low.
Score 100% of calls, human and AI, against the same weighted scorecard. A call can be "perfect" on steps and still leave the customer angry, so let one critical error fail the whole call. Check that automated scores match your human auditor.
Insight7 scores every call against a weighted scorecard your team owns, not a sample. Its voice campaigns are in beta. In beta, a campaign can include a test call to your own phone and a scorecard when it ends, so check the detail in the app before you rely on it. A dashboard lets a reviewer read each transcript, and call outcomes are pulled from the transcript.
Every call that breaks the matrix is a row to re-score.
Review the matrix every quarter
Re-score every call type each quarter, because a score that was right in March can be wrong by June. Also re-score at once on three triggers: a rule change, a new product, or a spike in complaints.
Move a call type from human to AI only when the evidence is steady:
- The handover rate is low and stable.
- QA scores hold for a full quarter.
- The scorecard shows no critical errors.
Move it back to a human when complaints rise, a rule changes or you launch a new product. A call type that passes all three checks has earned more AI share. One that fails any of them goes back a row.
Keep a cost check beside the scores. Set the AI cost per call type against what it costs you to acquire or serve that customer. Write the number down. Champions need a financial case they can carry upstairs.
The pressure to show proof is rising. A Gartner survey of service leaders found AI spending up 38% while overall service budgets grew 2%. Your quarterly review gives your finance lead the evidence for that spend.
Related reading: AI Voice Agent Test Call Checklist With 20 Things to Check Before Going Live and AI Phone Surveys vs Online Surveys Compared by Response Rate, Cost and Insight Depth.
Questions people also ask
When should a human agent take over from an AI voice agent?
Take over the moment a risk or emotion trigger fires, even mid-call. A direct request for a person is an instant trigger whatever the call scores. Waiting until the call ends turns a small problem into a complaint.
Can AI voice agents handle angry customers?
Mild irritation scores a 2 on emotion, so a routine call can stay with AI plus a human check if the handover is fast. Real anger, distress or a formal complaint scores a 3. The matrix sends every 3 to a person.
Which customer calls should I automate first with AI voice agents?
Pick one outbound call type with low scores on all four factors, high volume and a clear outcome, such as renewal reminders. Outbound is the safer start.
Do customers have to be told they are talking to an AI voice agent?
In the EU, yes, from the first interaction. The US requires consent for AI-voice calls, and more rules are spreading. Where no law applies, still answer truthfully if asked.
How do I hand off a call from an AI agent to a human?
Check that a person is free before the AI says it is transferring. Pass the reason and a short summary with the call. If nobody answers, offer a callback time.
Should high-value customers always get a human agent?
No. Value alone usually earns AI with a human check. Send the call to a human when high value combines with complexity, risk or emotion.
How do I measure whether AI voice agents are working better than humans?
Compare outcome rate, complaint rate and cost per successful outcome against a human control group on the same call type. Cost per call alone misleads. Have a human auditor check that automated scores agree.
Route by the highest score
Route each call type by its single highest factor, then check the split against real call data. A matrix is only as good as the calls you review after launch, and Insight7 can help you review them all. Which call types could let a customer hear a mistake?
