AI customer support tools: can a chatbot really resolve 87 percent of issues?

The 87% resolution claim is everywhere in vendor decks right now. Zendesk reports its AI agents autonomously resolve over 80% of customer interactions. Smaller deployments hit even higher numbers. But Comm100’s 2025-2026 benchmark data tells a different story: the industry average for full AI resolution without human involvement sits at 44.8%. That gap between vendor claims and real-world performance is exactly what this analysis is about.

AI customer support tools are being deployed at a scale that was unimaginable three years ago. Decagon, valued at $4.5 billion after tripling its valuation in 2025, signed 100+ enterprise deals last year alone. Zendesk’s CEO projects AI will handle 50% of digital customer service interactions within three years and 80% within five. The infrastructure investment is committed. The question isn’t whether chatbots will dominate support — they already are. The question is whether the resolution numbers companies are quoting actually reflect what customers experience.

Resolution rates are real, but they’re highly context-dependent. Deploying a chatbot expecting 87% resolution without understanding your industry, query mix, and deployment scope is how you get a Klarna situation — reduce headcount by 40%, then quietly rehire when quality tanks.

This analysis covers:

• Why resolution rate benchmarks vary from 38% to 97% across industries
• How deployment scope distorts the headline numbers
• Where the real ROI lives (and it’s not always in resolution rates)
• What the pricing model shift means for vendor accountability

Key Takeaways

• The industry-average AI chatbot resolution rate is 44.8%, not the 80%+ figures vendors typically cite, according to Comm100’s 2025-2026 benchmark data.
• Resolution rates vary dramatically by industry: Non-profits hit 97.7%, iGaming bottoms out at 38.1%, meaning your vertical matters more than any vendor benchmark.
• High resolution rates don’t directly predict customer satisfaction — iGaming scores 4.1/5 CSAT despite the lowest resolution rate in the dataset.
• Outcomes-based pricing (Zendesk charges $1.00 per resolved ticket) is shifting vendor incentives away from deflection and toward genuine resolution.
• Bot-to-human handoff satisfaction reached 92.6% in 2025, exceeding standalone chatbot CSAT of 49.3%, which means handoff quality is an underrated metric.

How We Got to Peak Chatbot Deployment

Early-generation chatbots — the ones answering “track my order” with a canned response tree — had resolution rates in the 20-30% range on a good day. The NLU was shallow, backend integrations were limited, and “resolution” often meant the customer stopped asking, not that their problem was solved.

The 2023-2025 period changed the architecture fundamentally. Large language models gave chatbots genuine intent detection regardless of how a question was phrased. Real-time backend integration let bots pull live purchase history, account status, and inventory data. Zendesk’s AI agents are now pre-trained on 18 billion real customer interactions, giving them a practical knowledge base that rule-based systems never had.

The business case sharpened fast. HelloSugar, a salon chain, automates 66% of customer queries and saves $14,000 per month — enough to double locations without adding headcount. Lush Cosmetics saves 5 minutes per ticket and 360 agent hours monthly. These aren’t pilot programs anymore. They’re operational infrastructure.

Scale introduced a new distortion, though. As AI handles a broader share of incoming queries — Comm100 data shows AI now handles 75.3% of all incoming chats, up from 73.8% the prior year — bots encounter more complex, edge-case queries they weren’t designed for. That’s partly why the overall resolution rate dropped from 45.8% to 44.8% despite the underlying technology improving. Wider scope, harder problems.

Resolution Rates Are Real — But Your Industry Determines Everything

The 87% figure isn’t fabricated. It’s just not universal. Comm100’s benchmark data shows an industry resolution range of 38.1% to 97.7%. Non-profits hit near-perfect numbers because their query types are narrow and predictable. iGaming sits at the bottom because users ask about account-specific disputes, payment failures, and fraud — exactly the kind of high-stakes, context-heavy queries where chatbots currently break down.

The CSAT column is worth pausing on. iGaming has the worst resolution rate and one of the best satisfaction scores. Education has a strong resolution rate and below-average satisfaction. Resolution volume doesn’t predict satisfaction — query complexity, agent quality after handoff, and how well the bot sets expectations matter more.

Team Size Shapes Your Deployment Strategy

Small support teams (1–5 agents) and large teams (26+ agents) run fundamentally different chatbot strategies, and their outcomes reflect it.

Small teams run narrow-and-deep bots: focused on a limited query set, optimized hard for those specific cases. They handle 54.3% of incoming queries but resolve 89.0% of what they touch. Large teams run broad-and-shallow: handle 67.5% of queries, resolve only 41.2%. The small team model produces a chatbot that looks like it’s hitting the 87% benchmark. The large team model produces the 44.8% average.

This explains why vendor case studies almost always feature smaller, specialized deployments. HelloSugar’s 66% automation rate is real — but HelloSugar is a salon chain with predictable, high-volume, low-complexity queries. Generalizing that to enterprise telecom support is how teams set bad expectations.

The Incentive Problem: Deflection vs. Resolution

AI systems optimize for whatever metric they’re assigned. Deploy a chatbot to cut ticket volume, and it will cut ticket volume — by deflecting queries, not resolving them. Nearly 1 in 5 consumers who used AI for customer service reported no benefit, a failure rate roughly 4x higher than AI applications generally, per the Qualtrics 2026 Customer Experience Trends Report.

Klarna is the clearest case study. They reduced headcount 40% using AI, then rehired human agents when quality degraded on complex tasks. Today their AI assistant handles the equivalent of 800 agents with satisfaction scores matching human agents — but they got there by pulling back, reassessing, and redeploying more carefully. The rebound wasn’t automatic. It required deliberate scoping, not just better models.

The industry’s structural response has been outcome-based pricing. Zendesk now charges $1.00 per automated resolution only when the customer, the business, and the system all confirm the issue was solved. Intercom charges $0.99 per resolution on top of its base fee. This model directly aligns vendor incentives with genuine resolution rather than deflection volume. That’s a meaningful shift — because when vendors only get paid for real outcomes, they stop selling deflection as a feature.

Where the Real ROI Lives: Handoff Quality and Agent Relief

Standalone chatbot satisfaction sits at 49.3% — up 9.1% year-over-year, the largest improvement across all tracked metrics. That’s progress, but it’s still below human agent scores. The underrated metric is handoff satisfaction, which reached 92.6% in 2025, up from 86.7% the year prior and exceeding the overall CSAT average of ~82%.

The mechanism is pre-transfer context collection. When a bot gathers intent, account details, and conversation history before transferring, the human agent doesn’t start cold. That alone eliminates a significant source of customer frustration — explaining the same problem twice. Get this step right, and your overall satisfaction numbers move even if your resolution rate stays flat.

Agent-side benefits are real too. Reducing burnout on repetitive, confrontational queries matters at scale. Human agents who handle escalated issues consistently — instead of processing “track my order” loops for eight hours — perform better and stay longer. That’s a retention and quality argument, not just a cost one.

Three Scenarios Worth Planning For

Scenario 1: E-commerce or SaaS deploying AI for the first time. Your query mix is probably 60-70% routine — order status, password resets, billing questions. A narrow-and-deep deployment targeting those specific flows will likely hit 70-80% resolution on the AI-handled subset. Deploy broadly and measure aggregate resolution instead, and you’ll land at the 44.8% average and declare failure. Scope deliberately, then expand.

Scenario 2: Enterprise with 25+ support agents considering broad AI deployment. Comm100’s data suggests your resolution rate will sit around 41%. That’s not a failure state — it means AI is handling 67% of volume and freeing agents for the 33% that actually needs them. Measure cost-per-resolution and agent utilization, not just resolution rate. The business case lives in labor hours saved, not percentage points on a dashboard.

Scenario 3: Leadership evaluating vendor claims. Ask vendors for resolution rate data from your industry vertical, not their best-performing case study. Confirm their pricing model — outcome-based pricing with tripartite confirmation (customer + business + system) is the current gold standard for accountability. If a vendor can’t produce industry-specific benchmark data, that’s a signal worth taking seriously.

Three trends worth watching over the next 12 months:

• Personal consumer AI agents negotiating directly with company chatbots — a concept already circulating among Zendesk and Sierra product teams. If this scales, it shifts the incentive dynamic entirely.
• Resolution rate transparency requirements — the FTC has signaled interest in how companies report AI deflection versus genuine resolution. Regulatory pressure could force standardized definitions.
• Handoff quality as a primary metric — as standalone CSAT plateaus, smart operators will compete on how well their bot-to-human transition works, not just on resolution volume.

What the Data Actually Tells You

A few conclusions hold up across the dataset:

• 87% resolution is achievable — but only in narrow, well-scoped deployments with predictable query types
• The industry average is 44.8%, and broader AI deployment tends to pull that number down, not up
• Resolution rate doesn’t equal satisfaction — handoff quality and query fit matter more than raw resolution volume
• Outcome-based pricing is the structural fix the industry needed, and it’s arriving faster than most operators expected

Over the next 6-12 months, expect the benchmark data to bifurcate further. Teams that deploy narrow-and-deep will keep pushing toward the high numbers. Broad deployments will hover around the 40-50% range. The vendors who survive the current consolidation wave will be the ones who can prove genuine resolution, not deflection.

The mindset shift worth making now: stop asking “what’s your resolution rate?” and start asking “how do you define a resolution, and what happens to the other 50%?” That second question is where the real differences between platforms show up.

Photo by Steve A Johnson on Unsplash