Customer service is one of the few areas where AI's value is showing up in hard numbers, and also where the deploy-vs-operationalise gap is clearest.
It's an obvious fit. Support is high-volume, text-heavy, and full of repetitive questions that have known answers sitting in a knowledge base somewhere. If any domain should produce clean AI wins, it's this one, and it does. The catch is that the headline figures get quoted as if deploying a chatbot is the whole job, when the numbers actually describe organisations that did a lot of unglamorous integration work to earn them.
Where the gains are real
- In a study of 5,179 customer support agents at one large software company, an AI assistant that suggested replies raised issues resolved per hour by about 14% on average, and by 34% for newer, lower-skilled agents, with little effect on the most experienced (Brynjolfsson, Li and Raymond, NBER, 2023).
- In Salesforce's 7th State of Service survey of 6,500 service professionals (April to June 2025), reps using AI said they spend 20% less time on routine cases. Respondents estimated that AI handles 30% of cases today and projected 50% by 2027.
What "half of cases handled by AI" actually means
This is the figure that gets stripped of context fastest, so it's worth unpacking. First, it's a survey estimate and a projection, not a measured resolution rate: service teams guessed that AI handles 30% of their cases now and expect 50% by 2027. Second, "handled by AI" does not mean a chatbot improvised brilliant answers. It mostly means the easy, well-bounded, repetitive queries (order status, password resets, store hours, return policy) go to automation, freeing humans for the genuinely hard ones. That's a real and valuable outcome. But it's also a measure of how good your knowledge base, your systems integration and your escalation logic are, not how clever the model is. A bot pointed at a thin or out-of-date knowledge base resolves nothing; it just frustrates people more efficiently.
The catch
Deploying a bot is easy. Wiring it into your knowledge, your systems and your escalation paths, so it actually deflects work instead of annoying customers, is the real project.
Plenty of support teams have a bot. Far fewer have done the work that turns it into deflected tickets rather than a customer's first obstacle. The difference is integration: connecting the model to live order data, account state and CRM history; defining exactly which query types it owns; and building a fast, clean handoff that carries full context to a human the moment it's out of its depth.
A short scenario
Two retailers deploy the same vendor chatbot, the same kind of split we cover in AI in retail & e-commerce. The first drops it on the homepage pointed at a generic FAQ. It answers "what are your hours?" and fumbles everything else; angry customers hammer "talk to a human," and the bot becomes a speed bump. The second wires the bot into order and shipping data, scopes it tightly to the handful of query types it can actually resolve, and routes anything ambiguous straight to an agent with the full conversation attached, the same human in the loop discipline that makes any assistive AI system trustworthy. Same technology, opposite outcomes: the second deflects real volume, the first generates complaints. The model was never the variable.
Why getting it wrong costs more than doing nothing
There's an asymmetry that makes this domain unforgiving. A support interaction that a human would have resolved fine, but a badly-scoped bot mangles, doesn't just fail to deflect a ticket; it actively manufactures a worse one. Now the customer is annoyed and still needs help, and the human who eventually picks it up inherits a frustrated person and a tangled conversation. So a poorly operationalised bot can have negative ROI: it raises handling time and lowers satisfaction at the same time. That's the opposite of the gains in the study and the survey above. The gap between those two outcomes is entirely a matter of execution.
This is also why the productivity numbers are best read as results from specific settings, not a guarantee. The 14% came from one company's assistant suggesting replies to human agents, and most of the gain went to newer agents. A bot that faces customers directly is a different system: feed it a stale knowledge base, or bolt on a handoff that drops the conversation history, and the same budget produces the opposite result. The gain is available; it is not automatic.
What this means for your team
- Fix the knowledge base first. The bot is only as good as what it can retrieve; clean, current content is the prerequisite, not an afterthought.
- Scope tightly. Give AI the clearly-bounded queries it can own, and let humans keep the rest. A confident "I'll connect you to someone" beats a confident wrong answer.
- Engineer the handoff. The escalation path, with full context carried across, is where most of the customer experience is won or lost.
- Measure deflection, not deployment. "We have a bot" is not a result; "tickets per resolution dropped X%" is the same discipline behind measuring AI ROI.
The winners treat AI as an assistant to agents and a first line for well-bounded queries, with fast, clean handoff to humans for everything else. If your support automation is deployed but not yet operationalised, that gap is closeable, and wiring support automation in properly is the kind of integration work our team does.
Sources
- Brynjolfsson, Li and Raymond: Generative AI at Work, NBER Working Paper 31161, 2023
- Salesforce: State of Service, 7th edition, November 2025
