The Future of AI Customer Service: What Is Coming in 2027 and 2028

The next phase of AI customer service is not about better answers. It is about AI that takes actions — and Gartner projects that by 2029, agentic AI will autonomously resolve 80% of common customer service issues without human intervention, cutting operational costs by 30% (Gartner, 2025, forecast). The short answer for 2027 and 2028: AI shifts from answering questions to acting on them, voice gets usable for narrow call types, and the businesses that win are the ones who automate reliably, not fastest. This guide walks through the forecasts worth trusting, the hype worth ignoring, and the specific moves a small business should make now.

What is agentic AI in customer service, and can AI agents take actions instead of just answering?

Agentic AI is software that completes multi-step tasks on its own — looking up an order, processing a refund, rescheduling an appointment — rather than only returning text. This is the single biggest change between 2026 and 2028. The old chatbot says "I can help you cancel that, please contact our team." An AI agent actually performs the cancellation, updates the record, and confirms it back to the customer.

The distinction matters because most tools labeled "agentic" today are not. Menlo Ventures found that only 16% of enterprise AI deployments qualify as true agents — most are fixed-sequence workflows wearing an agent label (Menlo Ventures, 2025). Gartner is blunter: of the thousands of vendors claiming agentic capability, it estimates only about 130 are genuine, the rest engaging in what it calls "agent washing" (Gartner, 2025).

For a small business, the practical version of agentic AI is modest and useful: an AI agent that does not just say "your order is on its way" but pulls the tracking record, spots a delay, tells the customer before they ask, and offers a next step. That is the direction of travel — capture and route a lead, run a guided multi-step intake, or trigger an action — not a science-fiction autonomous worker. Omago, an AI agent platform that helps SMEs automate customer conversations across WhatsApp, Telegram, and web chat, leads with exactly this: agents that take actions, not just reply, with live integrations (Airtable confirmed by name).

How much of customer service will be automated by 2027 and 2028?

The credible projections cluster around 30% to 50% of cases resolved by AI in the next two years, climbing toward 80% by the end of the decade — but every one of these is a forecast, not a measured fact. Salesforce reports 30% of service cases were resolved by AI in 2025 and projects that rises to 50% by 2027 (Salesforce State of Service, 2025, forecast). Gartner's 80%-by-2029 number sits further out and assumes mature agentic deployments that most companies do not yet have (Gartner, 2025, forecast).

What you actually achieve depends on maturity, not the headline. Vendor-reported averages run high — Intercom's Fin agent averages a 66-67% resolution rate across more than 6,000 customers, with over 20% exceeding 80% (Intercom, 2025) — but independent case studies land lower, around 42-50% in early maturity (Intercom case studies, 2025). McKinsey frames the ceiling differently: generative AI could unlock up to 60% of addressable care volume and reduce human-serviced contacts by up to 50% (McKinsey, 2023; McKinsey, 2025).

Here are the forecasts shaping 2027-2028, all labeled as projections:

One forecast deserves special attention for SMEs. Gartner projects that by 2028, 70% of customer-service journeys will begin and end inside conversational assistants built into customers' mobile devices (Gartner, 2024, forecast). In plain terms, more customers will "ask their phone" before they ever reach your website — which makes structured, discoverable content and clean knowledge bases a competitive issue, not just a support one.

Will AI replace customer service jobs by 2028?

No — the data points firmly to augmentation, not replacement, at least through 2028. The most-cited replacement narrative collapses against Gartner's own survey: in a poll of 321 customer-service leaders in October 2025, just 20% reported reduced agent headcount due to AI (Gartner, 2026). The other 80% are redeploying people, not eliminating them.

Spending tells the same story. Gartner projects that by 2028, over 50% of customer-service organizations will double their technology spend without an equivalent reduction in talent (Gartner, 2026, forecast). Companies are buying AI on top of their teams, not instead of them. Forrester expects roughly 30% of enterprises to create parallel "AI management" functions — people whose job is to coach, tune, and unblock AI agents (Forrester, 2026, forecast).

There is a cautionary edge here too. Forrester's 2026 B2C predictions warn that "in 2026, a third of companies will harm experiences with frustrating AI self-service" as cost pressure pushes premature deployments (Forrester, 2026). The lesson for a small business: the role that disappears is the repetitive-FAQ role, and the role that grows is the person who owns the knowledge base and handles the hard, emotional, high-stakes conversations.

Is voice AI good enough for customer service in 2026, and how does it compare to messaging?

Voice AI is real and improving fast, but in 2026 it is materially harder than text — so the honest answer is "it depends on your call profile." The biggest barrier is quality: 72% of organizations cite performance quality as the top obstacle to deploying voice AI agents (Deepgram, 2025). Accuracy still degrades with accents, background noise, and emotionally charged calls, and every "can you repeat that?" cycle erodes trust.

The progress is genuine, though. Human conversation expects sub-300ms turn-taking; modern speech-to-speech models now hit 160-400ms end to end, versus 1,000-2,000ms for older cascaded pipelines, with sub-800ms as the production target (Hamming AI, 2025-2026). Momentum is visible on the supply side too — voice startups made up 22% of a recent Y Combinator class, and OpenAI cut realtime voice API pricing about 20% (a16z, 2025; OpenAI, 2025). But momentum is not maturity.

Use this to decide where voice fits versus messaging:

Watch one trap in voice marketing: "containment" (the call never reached a human) is routinely sold as if it were "resolution" (the problem got solved). A frustrated caller who hangs up is "contained" but not served. That same honesty gap shows up across the category, which is why the difference between containment and resolution metrics is worth understanding before you buy anything.

What is a good AI resolution rate, and how should an SME measure ROI?

A realistic resolution rate for a well-run SME deployment starts around 30-50% and climbs toward 65-80% only with a mature knowledge base and ongoing tuning — and the single biggest determinant of results is deployment maturity, not which vendor you pick. Chasing a vendor's 80% headline before you have done the groundwork is the fastest way to disappointment.

The named outcome benchmarks give you honest goalposts:

That Forrester ROI figure is directional only — it comes from a 2020 commissioned enterprise study, not an SME benchmark, so treat it as a ceiling, not a promise. The Gartner cost-per-contact gap is the cleaner number to plan around: at roughly $1.84 for self-service versus $13.50 for an assisted contact (Gartner), the economics favor automating the simple, repetitive volume and reserving human time for the rest.

Measure it properly with a short discipline:

1. Set a baseline first. Capture current first-response time, resolution rate, CSAT, ticket volume, and cost-per-contact before you deploy. Without it, ROI is unprovable.
2. Track resolution, not deflection. A customer who gives up still "deflects." Only resolution proves the problem was solved.
3. Compare AI CSAT to your own human CSAT, not industry averages — customers score AI roughly 5-10 points harder, so a small gap is normal.
4. Segment by intent. Authentication, order status, and refunds resolve far better than disputes or sentiment-heavy issues.
5. Set staged targets. Aim for around 70% bot CSAT in month one, 75-80% by month three, with weekly QA review.
6. Watch re-contact rate (customers returning within ~72 hours) as a guardrail on "resolved" claims.

If you are formalizing this, our 30-60-90 day KPI plan for AI agents lays the cadence out in detail.

Can you trust AI customer service, and is my business liable if it gives wrong information?

You can trust it when it is engineered for honesty — and yes, your business is fully liable for whatever your AI says. The core risk in 2027-2028 is not that AI is dumb; it is that AI can be confidently wrong, producing a fluent, authoritative answer that is simply false. Peer-reviewed research from Simhi et al. (Technion, Oxford, and Hebrew University, 2025) shows large language models can hallucinate with high certainty even when they hold the correct knowledge — they sound most confident exactly when they are wrong.

The legal reality is already settled. In Moffatt v. Air Canada (2024), a tribunal held the airline liable for negligent misrepresentation after its website chatbot invented a bereavement-fare policy, explicitly rejecting the defense that "the chatbot is a separate legal entity" and ordering C$812.02 in damages. The business owns what its AI says. Customers know it is risky, too: 84% of consumers believe human agents are more accurate than AI, only 8% prefer AI over humans, and 61% feel humans better understand their needs (SurveyMonkey, 2025).

The reassuring part is that hallucination is largely an engineering problem with known fixes. Grounded models can reach hallucination rates as low as 0.7-1.5% on summarization benchmarks, though even flagship reasoning models exceed 10% on harder real-world content (Vectara HHEM, 2025-2026) — which is exactly why grounding, abstention, and escalation matter. Here is the operator-level guardrail checklist:

• RAG grounding — force answers from your verified content, not the model's memory.
• Knowledge-base curation — remove stale and conflicting articles, which cuts grounded-but-wrong answers materially.
• "I don't know" behavior — let the model abstain and escalate when context is missing instead of guessing.
• Confidence thresholds — above ~85% proceed; 70-85% proceed but flag; below ~70% escalate to a human.
• Escalation triggers — explicit request, repeated failure by the second or third attempt, detected frustration, and high-risk intents like refunds or billing.
• Warm handoff with full context — pass the transcript so the customer never repeats themselves.
• Source citations — showing the source the AI used lifted CSAT 8-12% in one study, with no change to underlying accuracy.
• Transparency — tell customers they are talking to AI; disclosure builds trust, it does not destroy it.

This is the whole philosophy behind a messaging-first approach: an AI agent grounded in your knowledge base, with honest abstention and clean escalation, is far safer than an ungrounded bot optimized to always answer. If you want the deeper version, see how to keep AI customer service trustworthy with guardrails.

What should small businesses do now to prepare for 2027 and 2028?

Start with grounded messaging and web automation on well-defined intents today, then add action-taking deliberately — because the winners through 2028 are the ones who automate reliably, not the ones who automate first. Gartner's projection that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear value, and weak risk controls (Gartner, 2025, forecast), is a warning about rushing, not a reason to wait.

Here is the sequence that holds up against the data:

1. Deploy what is mature now — after-hours messaging, FAQ handling, lead capture, and simple guided flows. These are affordable and measurable today.
2. Build escalation and knowledge discipline first. Curated content and clean handoffs are what separate a 45% resolution rate from an 80% one.
3. Add action-taking deliberately. Begin with low-risk, reversible actions; never let high confidence authorize an irreversible action unsupervised.
4. Choose a platform that evolves, so your investment in knowledge bases and conversation flows carries forward as capabilities like voice and richer agentic actions mature.
5. Plan for "ask your phone" discovery. With Gartner projecting 70% of journeys starting in third-party mobile assistants by 2028, structured, accurate content is now a discoverability asset.

The honest framing matters most here. The next two years will not "fire your team," and they will not magically resolve everything. They reward businesses that treat AI as augmentation, ground it in real information, and measure resolution instead of believing the marketing. If you are still weighing whether to automate or hire first, when to automate versus hire covers that decision directly.

Frequently Asked Questions

Will AI replace customer service jobs by 2028?

No. In a Gartner survey of 321 customer-service leaders in October 2025, just 20% reported reduced agent headcount due to AI, and Gartner projects over 50% of organizations will double tech spend by 2028 without cutting talent (Gartner, 2026). The pattern is role evolution — AI handles routine volume, humans handle complex and emotional work, and a new group manages and tunes the AI.

What is agentic AI in customer service?

Agentic AI refers to systems that take multi-step actions — processing a refund, rescheduling an appointment, updating a record — rather than only generating a reply. It is the defining shift from 2026 to 2028. Be cautious, though: Menlo Ventures found only 16% of enterprise deployments qualify as true agents, and Gartner estimates only about 130 of the thousands of "agentic" vendors are genuine (Menlo Ventures, 2025; Gartner, 2025).

How much of customer service will be automated by 2027?

Salesforce reports 30% of cases were AI-resolved in 2025 and projects that rises to 50% by 2027 (Salesforce State of Service, 2025, forecast). For an individual SME, realistic early resolution rates run 42-50% and climb toward 65-80% only with a mature, well-curated knowledge base and ongoing tuning (Intercom, 2025).

Is voice AI good enough for customer service in 2026?

It depends on your call profile. Latency has dropped to a natural 160-400ms with modern speech-to-speech models (Hamming AI, 2025-2026), but 72% of organizations still cite performance quality as the top barrier to deploying voice AI (Deepgram, 2025). Voice fits high-volume, transactional, phone-first calls; messaging remains cheaper, easier to ground, and lower-risk for most SMEs.

Is my business liable if the AI chatbot gives wrong information?

Yes. In Moffatt v. Air Canada (2024), a tribunal held the airline liable for negligent misrepresentation after its chatbot invented a refund policy, rejecting the argument that the chatbot was a separate legal entity (BC Civil Resolution Tribunal, 2024). You own what your AI says — which is why grounding answers in verified content, citing sources, and escalating on low confidence are not optional.

Sources: Gartner (2024, 2025, 2026 customer service and agentic AI predictions), Salesforce State of Service (2025), Forrester (2026 B2C predictions; Total Economic Impact of IBM watsonx Assistant, 2020), McKinsey (2023, 2024, 2025), Menlo Ventures State of Generative AI in the Enterprise (2025), Deepgram State of Voice AI (2025), Hamming AI (2025-2026), a16z AI Voice Agents Update (2025), OpenAI (2025), Intercom (2025), Vectara HHEM Leaderboard (2025-2026), Simhi et al. (Technion/Oxford/Hebrew University, 2025), Moffatt v. Air Canada, BC Civil Resolution Tribunal (2024), SurveyMonkey (2025).