AI should handle the parts of a sales conversation that are repetitive, time-sensitive and factual — first response, requirement gathering, scheduling and follow-up. It should hand over anything requiring professional judgement, emotional sensitivity, negotiation or trust-building. The dividing line is not complexity; it is whether a wrong answer would cause real harm.
Key takeaways
- Automate speed and structure. Keep judgement and relationship human.
- The right test is not "can AI do this?" but "what happens if it gets this wrong?"
- Four categories always escalate: professional advice, emotional situations, negotiation, and anything unusual.
- A 15–25% escalation rate is healthy. Near-zero escalation is a warning sign.
- Handover must be instant, obvious and carry the full conversation history.
The wrong question and the right one
Most businesses ask: "Can AI handle this conversation?" Increasingly, the answer is yes to almost everything — which makes it a useless question.
The useful question is: "What is the cost of it being wrong?"
| If the AI gets it wrong | Cost | Automate? |
|---|---|---|
| Books the appointment at the wrong time | Minor, easily fixed | Yes |
| States the clinic's opening hours incorrectly | Minor, fixable | Yes |
| Misjudges a budget band | Small, recoverable | Yes |
| Suggests a treatment | Serious, potentially harmful | Never |
| Quotes an unauthorised discount | Financial and contractual | Never |
| Mishandles a grieving or angry customer | Reputational, unrecoverable | Never |
| Interprets a legal or tax notice | Professional liability | Never |
This is the entire framework. Everything below is application.
What AI should always handle
1. The first response. Always. There is no scenario where a slow human reply beats an instant machine reply, because the alternative is silence. This is the core of speed to lead.
2. Requirement gathering. Budget bands, timelines, property types, goals, service needed. Structured, factual, repetitive — and better done by a machine, which never forgets to ask question three.
3. Standard factual questions. Hours, location, parking, payment methods, what to bring, whether you accept new clients.
4. Scheduling. Reading real availability, offering specific slots, confirming, rescheduling, reminding. Pure logistics.
5. Follow-up sequences. Timed, structured, consistent. Humans are unreliable at this not because they are lazy but because follow-up always loses to whatever is urgent today.
6. Reactivation and recall. Six-month dental recalls, lapsed gym members, annual filing reminders. Nobody does these manually at scale.
7. Review requests. Triggered by job completion, sent at the right moment, every time.
What must always be human
Professional advice
Clinical, legal, financial, structural. Not because AI cannot produce a plausible answer, but because a plausible wrong answer in these domains causes real harm and real liability. The AI's job is to recognise the category and hand over immediately.
Emotional situations
Complaints. Distress. Anger. Bereavement. Anxiety before a procedure. A person in an emotional state needs to feel heard by another person, and a machine that handles this smoothly is often worse than one that handles it clumsily — because smooth handling of genuine distress reads as indifference.
Negotiation
Discounts, scope changes, contract terms, payment plans. These require authority and judgement about a specific relationship. The AI may state approved ranges. It may not move within them.
Anything unusual
This is the most important and most often forgotten rule. If the conversation departs from expected patterns, the AI should hand over rather than improvise. Confident improvisation outside scope is the main way these systems fail publicly.
Designing the handover
The handover is where most implementations fail. Get these right:
Escalation must be instant and always available. "Talk to a person" must work at any point, in any phrasing, without a menu. If a customer has to ask twice, the system has failed.
Context must travel. The human picking up must see the full conversation. Making a customer repeat themselves is the single most damaging thing a handover can do — it converts a good automated experience into a bad one retroactively.
Set expectations honestly. "I'm connecting you with Priya, she'll message you within the hour" is far better than a silent transfer into a void.
Never claim to be human. If asked directly, the AI says what it is. Businesses fear this costs trust. It does the opposite — discovering later that a supposedly human agent was a bot destroys trust permanently.
Escalate proactively, not just on request. The AI should recognise anger, confusion, repetition and emergency signals and hand over without waiting to be asked.
The escalation rate tells you if it is working
| Escalation rate | Diagnosis |
|---|---|
| Under 5% | Warning. The AI is likely answering things it should hand over. |
| 15–25% | Healthy. Automation is handling routine work, humans handle judgement. |
| Over 40% | Scope is too narrow, or the AI is failing at basic tasks. |
Businesses instinctively celebrate a low escalation rate as efficiency. In practice it is usually the symptom of an AI operating beyond its competence — and the first sign appears in a public review, not in a dashboard.
A worked boundary, by industry
| Industry | AI handles | Human handles |
|---|---|---|
| Dental clinic | Booking, reminders, recalls, directions | Any clinical question, anxious patients |
| Gym | Trial booking, timings, follow-up | Injury history, membership disputes |
| Interior design | Qualification, scheduling, proposal follow-up | Design direction, taste, negotiation |
| Solar | Ownership, bill, roof, shading, timeline | Technical feasibility, subsidy specifics |
| Real estate | Requirement capture, site-visit booking | Negotiation, legal, recommendation |
| CA practice | Documents, status, deadlines | Any tax position, notices, planning |
The honest limitations
Be clear-eyed about these:
- AI is confident when it is wrong. This is the core risk, and it is why narrow scope matters more than capability.
- It does not know what it does not know. Guardrails must be explicit, not assumed.
- Regional language and code-switching are handled unevenly. Test with Hinglish and your local language before launch.
- Voice notes and images often need a human.
- Customers can tell, and mostly do not mind — provided the AI is useful, honest and easy to escape.
Frequently asked questions
What parts of sales should AI handle?
The repetitive, time-sensitive and factual parts: instant first response, requirement gathering, standard questions, scheduling, reminders and structured follow-up. These benefit from consistency and speed, which is exactly where software outperforms a busy human.
When should AI hand over to a human?
Whenever the conversation involves professional advice, emotional distress, negotiation, or anything outside expected patterns. It should also hand over immediately whenever the customer asks, in any phrasing, without requiring them to navigate a menu.
Should an AI assistant admit it is not human?
Yes. If asked directly it should say so plainly. Concealing it is a short-term gain and a long-term loss, because customers who discover the deception later lose trust in the business entirely, not just in the assistant.
What is a healthy escalation rate?
Roughly 15–25% of conversations reaching a human. A rate below 5% usually means the AI is answering questions it should be handing over, which is a risk indicator rather than a sign of efficiency.
Will AI lead qualification annoy customers?
Generally not, if it is fast, useful and honest. Customers dislike being ignored far more than they dislike talking to software. Problems arise when the AI is slow, pretends to be human, cannot be escaped, or attempts to answer questions beyond its scope.
Draw the line correctly
fizaki designs the boundary with you — what the AI handles, what escalates, and how the handover works — before anything goes live.
Related reading: what an AI receptionist is and how to set one up.
Stop losing leads to a phone nobody answers.
Book a free 20-minute audit — we'll show you exactly how many leads you're leaking and what they're worth.