AI strategy: Customer support

Prioritized AI adoption map. Date: 2026-06. Sources: GDPval, AEI, O*NET, and profession-automation-2026.md. ← Role hub

Context
Bots are taking over standard first-line requests as automation rises. Complex cases, retention, and empathy are becoming more valuable.

1. Automated handling (high priority)

No.	Goals	KPIs	Initiatives
1	Bots for common questions grounded in the knowledge base	% resolved by the bot · CSAT	Bitrix24/Zendesk AI · natural language understanding based on a trusted source
2	Classify and route tickets	routing time · accuracy	automatic sorting by topic and priority

2. Agent assistance (high priority)

No.	Goals	KPIs	Initiatives
3	Draft replies and ticket summaries	response time · quality	AI draft · fact check before sending

3. Retention (medium priority)

No.	Goals	KPIs	Initiatives
4	Detect churn risk in customer success	% of high-risk customers retained	Pylon and triggers · human works with the risk

4. The role (foundation)

No.	Goals	KPIs	Initiatives
5	Move people to complex and empathy-heavy work	share of high-stakes tasks · NPS	less standard work, more retention and conflict handling

Where to start: the first round

The tables are a map of what is possible, not your plan. Start with process scoring. Ask whether the task repeats, has enough volume, produces a checkable result, carries an affordable error cost, and has usable data. For most support teams, the first round looks like this:

No. 2, ticket classification and routing. It repeats for every request and has the greatest volume. Compare the assigned topic with the real message. An error costs a reroute, not a reply sent to the customer.
No. 3, draft replies and ticket summaries. The support agent sees the draft before sending, so review is built into the process.
No. 1, bots for common questions grounded in the knowledge base. Start only while the bot answers strictly from a trusted source instead of making up an answer, and audit it closely. Without those conditions, this becomes the first incident rather than the first round.

Not in the first round: No. 4, "churn detection," needs request history that contains a visible risk pattern. Without it, the model is guessing rather than predicting. No. 5, moving people to complex and empathy-heavy work, is the result of the first three initiatives, not a separate implementation task.

Keep these parts of customer support human

Emotionally difficult and hostile cases. Route them to a person immediately, without trying an automatic reply.
Unusual requests with no approved answer in the knowledge base. A person decides the exception and its wording.
A conversation with a customer who may leave. This needs a real conversation, not a template.
Customer personal data. Control and de-identify it, and do not send it to third-party services.
Responsibility when the bot is wrong. Escalate and investigate the case instead of silently correcting it.

For a similar path in practice, read the example of a support leader who worked through this exact split between bot and person: Sergey.

Review points and stop thresholds

Initiative	What a person checks	Stop threshold (example, replace with your own)
No. 2 routing	sample whether the topic and priority are correct	a misrouting rate above your threshold means returning disputed cases to manual sorting
No. 3 drafts and summaries	support agent checks facts and tone before sending	a claim in the draft without source support means fixing the prompt and manually checking the next batch
No. 1 bot for common questions	audit N conversations per day against the knowledge base	one answer outside the knowledge base, or a hallucination, means quarantining the bot until review and retraining are complete

Before you expand the flow, build a reference set of 20 cases: 20 real tickets with an approved answer and the correct route.

Two paths from here

You are a support agent or specialist: open the role hub, then use the Volume 1 workbook to move from your week to an audit and a 90-day plan.
You lead the support function: use process scoring, write the first agent contract, run the 90-minute workshop with the team, and complete the full path in the Volume 2 workbook.

Discipline
Unchecked answers damage trust. De-identify personal data. Keep difficult cases with people.

→ Role hub · audit · plan.

Sources: profession-automation-2026.md; book, Chapters 2, 3, and 7. Version: 2026-07-09. What changed: added "Where to start: the first round," "Keep these parts human," "Review points and stop thresholds," and "Two paths from here."