AI in customer service: what to automate, what to keep human

AI in customer service pays off first on routine requests with clear policies, such as checking an order’s status or resetting a password. It also cuts the busywork around each ticket, like drafting replies and after-call notes. Keep people on anything that needs judgment or carries real risk, and give them AI tools that make them faster.

Why support is a good first place for AI

Support is one of the largest operating expenses at many companies, and one of the easiest to measure. In Salesforce’s State of Service survey, 95% of service decision makers at organizations using AI report cost and time savings, and 83% plan to increase their AI investment over the next year^[1]. Zendesk’s CX Trends 2025 report found that 90% of the companies it calls “CX Trendsetters” see a positive return from AI tools for their agents^[2].

Fig. 01

What service decision makers say about AI

• Report cost & time savings95
• Say gen AI improves service92
• Plan to increase AI investment83

I’d sort the savings into two kinds.

The first kind comes from requests that never need a person. An AI front door, in chat or on the phone, resolves simple requests end to end and passes the rest to an agent with the context attached. Each request it resolves well protects your margin without hurting the customer’s experience.

The second kind comes from making agents faster on everything else. An agent co-pilot drafts replies and pulls up the right help article while keeping a running summary of the conversation. That takes minutes off each case. Background automation handles the system updates and follow-ups that would otherwise eat into an agent’s day.

How much to read into the surveys

The direction is clear enough: leaders expect to increase AI investment over the next three years^[3], and most already see a return^[1][2]. But adoption is uneven, and definitions differ. “AI agents handle half of all interactions” can mean a scripted phone menu (IVR) at one company and fully autonomous workflows at another. Benchmarks tell you which way things are heading. They’re too loosely defined to set targets for your own team.

Fig. 02

AI investment plans in three surveys

• Increase AI spend (McKinsey, next 3 yrs)92
• Increase AI spend (Salesforce, next year)83
• Explore/pilot conversational GenAI (Gartner, 2025)85

A rough business case

Say your support team handles 1 million contacts a year, and AI saves two minutes on each one through better triage and faster wrap-up. That’s about 33,000 agent hours. At the median U.S. customer service wage of $20.59 an hour^[4], it comes to roughly $686,000 a year in wages, before benefits and overhead. Revenue from better retention, or from add-ons sold during support conversations, comes on top of that.

Most of that saving happens on contacts a person still handles, so I’d plan AI as a change to how your whole support operation works, with a chatbot as one part of it. Work on throughput first (how many contacts your agents can close), since that’s what lowers cost per contact. Quality comes next, measured in customer satisfaction (CSAT) and first-contact resolution. Revenue comes last, when support conversations start preventing cancellations and selling add-ons.

Who, or what, handles each request

Here’s the flow I’d start from:

1. A customer request comes in

2. AI triage

   Works out what the customer wants and confirms who they are

3. Simple, within policy

   1. Self-service

      Resolved end to end in chat or on the phone

   Needs judgment

   1. An agent with an AI co-pilot

      The AI drafts replies and finds the right article

   2. Complex or high risk

      A specialist

      With the authority to resolve it
4. Learning after every conversation

   A summary, plus updates to the help content

Think of it as a ladder: each request climbs only as high as it needs to. Whichever rung it ends on, the conversation finishes with a summary and any updates to your help content, so the next answer is better.

The pressure to get this right is growing: contact volumes are rising, and leaders still expect significant efficiency gains from AI over the next few years^[5].

Guardrails: clean data and human oversight

Start with the data. Good answers depend on clean, current policies and product information, including what each customer is entitled to. Early on, put work into how your help content is organized and how well the AI retrieves it, because a co-pilot is only as good as what it can safely cite. If you sell online, the same product data feeds AI shopping agents, so the cleanup helps there too.

Keep a person in charge of anything ambiguous or high-risk, and of conversations where the customer is upset. Even the operators most bullish on AI acknowledge it needs oversight, especially in sensitive workflows and when verifying a customer’s identity^[5].

Salesforce’s 4,000 support roles, in context

In 2025, Salesforce’s CEO said AI agents had let the company cut its support staff from about 9,000 people to about 5,000, with customer satisfaction scores about the same as before^[7]. That’s a striking result from a company with deep data plumbing and its own AI-first platform, which most support teams don’t have. Take it as a sign of what’s possible, and expect smaller first wins: triaging billing questions, say, or summarizing long email threads so agents can get straight to the fix.

Where to start

Pick “boringly valuable” use cases. Password resets, order status, warranty lookups and simple returns have clear policies and plenty of data, so they’re safe to automate and quick to prove.

Build an agent co-pilot before a fully autonomous agent. Making your own people faster and more accurate carries little political risk and pays back well. Once the co-pilot’s answers are reliably right, point the same capabilities at customers.

Don’t skip change management. The technology is the easy part; changing workflows and incentives is harder. Train your agents and consider new roles, such as an “AI supervisor.” Leaders who underinvest here see rollouts stall, even when the models are impressive^[3][5].

Which numbers to track

If the AI is working, it shows up in numbers you already report:

• Average handle time (AHT): co-pilot summaries and suggested actions take seconds off each interaction.
• First-contact resolution (FCR) and repeat contact rate: better triage and knowledge retrieval make it more likely the issue is fixed the first time.
• Cost per contact: automation cuts the number of touches and escalations each issue needs.
• CSAT and Net Promoter Score (NPS): less effort and faster answers show up in how customers rate you, sometimes right away.
• Revenue influenced by service: count “saves” (customers talked out of canceling) and “attaches” (add-ons sold during a support conversation), so support isn’t seen purely as a cost center.

Protecting customer data and trust

Customers want their information kept safe and their problem solved. Protect personal data (PII) and limit what each role can access. Add guardrail prompts that keep the AI within your policies.

Make red-teaming a habit: have someone deliberately push the AI toward risky territory, like financial advice or a high-dollar refund, and set a clear escalation path for anything that looks wrong. Label automated conversations clearly and make it easy to reach a person. Log every action the AI takes, so you can audit it later.

A 90-day rollout plan

• Days 1–30: Find the three request types (intents) with the most volume and friction. Choose one channel, chat or voice, and launch a co-pilot for your busiest agent queue. Decide which single metric you’ll judge it by, plus a few supporting ones.
• Days 31–60: Finish the security review and run targeted red-team tests. Connect the AI to the systems it needs, like order management and billing. Where it’s safe, turn co-pilot suggestions that have proven reliable into automated actions.
• Days 61–90: Add a second high-volume request type. Set up an AI steering committee and write a one-page policy for how AI is used in support. Then send leadership a quarterly report that ties the automation to dollars saved and to churn.

The test I’d apply to each request type: if the policy is clear and a wrong answer is cheap to fix, it’s a candidate for automation. If not, it belongs with a person, and the AI’s job is to make that person faster.

Sources