AI & CX ยท 15 min read

AI in customer care: where to automate, where to hold back.

Six years of chatbots and three years of AI copilots across our support operations have convinced us: automating everywhere degrades NPS. Here is the right dose of AI in 2026.

By Hasina A., Head of Customer Care ยท IHAENG โ€ข Published March 28, 2026

In 2023, every CEO wanted "a GPT-4 chatbot". In 2024, we deployed the first agent copilots. In 2025, LLM voice bots hit the market. In 2026, the question is no longer "do we need AI?" but "where should we put it, and where should we absolutely not?".

Here is what we have learned running CX operations for 40+ clients across a wide spectrum of AI/human mixes.

Three usage zones, three risk levels

Zone 1: Back-office automation (green light)

Anything the customer doesn't see directly can โ€” and should โ€” be automated. Examples:

  • Automatic ticket categorization (90%+ accuracy on frequent topics)
  • Response suggestions for the agent (Intercom Fin, Zendesk AI)
  • Automatic summarization of long conversations
  • Sentiment detection to prioritize at-risk tickets
  • Voice-of-customer analysis of post-call transcripts
  • Draft email composition from a few keywords

In this zone, AI augments the agent โ€” it doesn't replace them. Measured gains at our clients: โˆ’25% on handle time, +18% on first contact resolution.

Zone 2: Self-service & standard answers (amber light)

Repetitive, non-sensitive questions can be automated on the front end, provided you:

  • Maintain an up-to-date knowledge base (otherwise the AI hallucinates or contradicts itself)
  • Offer a permanently visible "escape to human"
  • Log every interaction for QA
  • Measure the REAL deflection rate (not the vendor's bot stats)

Use cases that work: "where is my order?", "how do I reset my password?", "which payment methods do you accept?", "what are your opening hours?". Anything factual, low-risk, with a single answer.

Realistic deflection rate in 2026: 35-50% in e-commerce, 15-25% in high-complexity industries (finance, healthcare).

Zone 3: Emotional or complex interactions (red light)

Never automate without a human in the loop:

  • Emotional complaints (anger, disappointment, worry)
  • Sensitive financial questions (refunds, disputes)
  • Health, safety, and legal topics
  • VIP customers or strategic accounts
  • First contact with a new customer ("wow" effect vs. disappointment)

Automating these interactions saves $2 per ticket and burns $44 in LTV. We see it client after client.

Voice AI: hype vs. reality

LLM voice bots (Rasa, Retell, Vapi) are impressive in demos. In production, they are still too fragile for industries where errors are costly. Recurring issues in 2026:

  • Perceived latency of 400-700ms โ€” enough to make the interaction feel "off"
  • Struggles with accents, background noise, and interruptions
  • A tendency to invent precise information (prices, lead times)
  • Difficulty handling "off-script" calls without hallucinating

Our recommendation: in 2026, deploy only on tightly constrained cases (appointment reminders, automated NPS surveys, simple inbound qualification). Complex support stays human.

Hybrid: the model that works

The optimal setup at our clients:

  1. Layer 1: AI self-service for 30-50% of simple requests
  2. Layer 2: AI-augmented agent for 100% of human-handled interactions
  3. Layer 3: Dedicated expert with no AI for the 5-10% of sensitive/VIP cases
  4. Cross-cutting layer: back-office AI (categorization, analytics, summaries, dashboards)
"We thought we had to choose between 'all AI' and 'all human'. The real question is: which AI, on which part of the journey." โ€” Hasina A.

Four classic mistakes

  1. Deploying without a clean KB. AI amplifies documentation gaps. Clean the KB first, not after.
  2. Not measuring real deflection. Many "40% deflection" claims are actually 15% real deflection + 25% of customers who gave up in frustration.
  3. Removing the escape to human. A customer who feels trapped turns into a lifelong detractor.
  4. Forgetting continuous evaluation. LLMs drift, product catalogs evolve, prices change. Without monthly evaluation, quality quietly erodes.

Our benchmark metrics

A few internal benchmarks from our 2025-2026 AI/CX deployments:

  • Bot deflection (well-implemented text chatbot): 28-42% depending on industry
  • Post-bot satisfaction: must stay > 3.8/5 (below that, it's a hidden detractor)
  • Fallback rate: 15-25% of bot conversations end up escalated to a human
  • Handle time with AI copilot: โˆ’20% to โˆ’30% vs. an agent without a copilot
  • Cost per interaction: bot $0.16-0.38, agent $2.75-6.60, depending on stack and geography

Wrapping up

AI in customer care is neither the silver bullet marketed by solution vendors nor the threat denounced by purists. It's an amplifier โ€” one that makes a quality support operation even better, and a mediocre one even worse.

The real question in 2026 is no longer "should we add AI?" but "is our support foundation solid enough to benefit from it?". At IHAENG, we always start by auditing that foundation before recommending an AI rollout.

Want an AI-readiness audit? In 5 days, we measure your deployment potential and prioritize the workstreams.

HA
Hasina A.
Head of Customer Care IHAENG ยท Director of the Antananarivo hub
Get in touch