AI use case
AI in customer service
Where AI genuinely helps in customer service — triage, drafting, knowledge retrieval and summarisation — with complexity, data requirements, risks, KPIs and ROI considerations.
Miguel Torres, Founder, Merjora · Updated 3 September 2026
In short
In customer service, AI creates the most reliable value in triage, retrieval and drafting — classifying and routing inbound contacts, finding the relevant policy or record, and drafting a reply an agent approves. Fully autonomous resolution works only for narrow, low-consequence intents with a clean knowledge base. The economics come from reduced handling time per contact and reduced escalation, not from removing the agent.
The workflow today
Inbound contacts arrive by email, form, chat or phone. An agent reads, classifies, looks up the customer record, decides an action, drafts a reply and logs the outcome.
Where AI helps
- Classifying intent and urgency, then routing to the right queue
- Retrieving the relevant policy, order or case history
- Drafting a first-pass reply grounded in the retrieved record
- Summarising long threads before escalation or handover
- Detecting sentiment or churn signals for prioritisation
Common use cases
- Email and form triage with automatic routing
- Agent-assist reply drafting with source citations
- Knowledge base retrieval inside the agent workspace
- Post-contact summary and CRM logging
- Self-service answers for a defined set of high-volume intents
Expected benefits
- Lower average handling time per contact
- More consistent responses across the team
- Faster onboarding for new agents
- Capacity to absorb volume growth without proportional hiring
Implementation complexity
Moderate. Triage and summarisation are straightforward. Reply drafting requires reliable retrieval from a maintained knowledge base, which is where most of the effort sits.
Data requirements
- Historical contacts with intent labels, or a taxonomy someone will label against
- A current, de-duplicated knowledge base or policy source
- Read access to the CRM or order system for grounding
Risks and controls
- Confidently wrong answers when retrieval fails — mitigate with citations and mandatory review for regulated topics
- Knowledge base decay silently degrading quality
- Agents accepting drafts without reading them, which moves errors downstream
- Personal data exposure if retrieval scope is not restricted per user
Example workflow
- Contact arrives and is classified by intent and urgency
- Relevant customer record and policy passages are retrieved
- A draft reply is produced with links to its sources
- The agent edits and sends; edits are logged
- Edit rate and handling time feed the weekly quality review
KPIs to track
- Average handling time
- First-contact resolution rate
- Draft acceptance / edit rate
- Escalation rate
- Cost per contact
ROI considerations
- Value is contacts per month × minutes saved × loaded hourly cost
- Count escalation reduction separately; it is often larger than the drafting saving
- Review time does not go to zero — model it permanently
- Deflection claims should be measured against a pre-launch baseline, not vendor benchmarks
Real-world example
Publicly documented industry example — not a Merjora client.
Aviva: AI in claims and service operations →Frequently asked questions
- Can AI fully replace customer service agents?
- Not in most businesses. Narrow, low-consequence intents can be resolved end to end; everything involving judgement, exceptions or regulated advice needs a human decision point.
- What is the fastest customer service AI project to deliver?
- Inbound triage. It uses text you already have, has a clear accuracy measure and fails safely, because a misrouted item is corrected by the receiving queue.
Could this apply to your business?
Merjora quantifies what this workflow costs you today and what changing it is realistically worth.
Discover your AI opportunitiesRelated reading
Editorial standard. Merjora publishes analysis, frameworks and publicly documented examples. We do not publish invented statistics, unattributed benchmarks or unverified customer stories.