AI Chatbots for Business: Where They Pay Off and Where They're Oversold

“AI chatbot for business” is the kind of search that happens at 11pm when you're staring at your support inbox thinking "there has to be a better way." Or it comes up in a Monday meeting when someone says "should we get a chatbot?" and nobody in the room can explain what that means in practice.

If you're already running one and wondering whether you picked the right tool, or whether the whole category is doing what the vendor promised, this is for you too.

The market context#

Gartner forecast in 2022 that conversational AI deployments in contact centres would cut agent labour costs by $80 billion in 2026, with roughly one in ten agent interactions automated by then (Gartner). That deadline is now, so read the forecast as a direction of travel rather than a settled result. Worth saying plainly: there is no clean primary source for a single industry-wide chatbot ROI figure, so treat any "$X back for every $1" number with suspicion, including the ones vendors put on a slide.

The underlying driver is simple. Businesses handle large volumes of repetitive customer queries. A human costs a wage per conversation, and software costs a fraction of that at volume. That gap is real. It only turns into ROI when the chatbot is pointed at the right work.

The table looks decisive, and for the top rows it is. But those advantages only hold when the chatbot is doing the right work. Point it at the wrong problems and they evaporate fast.

Where chatbots earn their keep#

You probably already have a sense of this, but it's worth being precise about which categories consistently show returns.

Customer support and self-service. Order status. Returns. Shipping FAQs. Account inquiries. Product availability. High-volume, repetitive, factual questions with definitive answers. A well-configured chatbot handles them faster and cheaper than a human, and the customer often prefers the instant response to sitting in a queue.

The resolution rates are worth looking at. Service teams surveyed by Salesforce estimate that AI now handles 30% of their cases, a figure they expect to reach 50% by 2027 (State of Service, 7th Edition). On Salesforce's own help site, Agentforce resolved 84% of its first 500,000 conversations. By the time volume passed four million, Salesforce's own figure had settled at 70%. Still strong, and the drop is instructive: resolution rates fall as the easy questions run out. Customer deployments sit in the same range. Fisher & Paykel moved from 40% to 70% self-service resolution, and Zendesk reports Vagaro resolving 44% of requests via AI with 92% CSAT. The pattern holds across industries, not just ecommerce.

Worth being upfront: these resolution rates are vendor-reported too. The reason to trust the direction is the independent consumer data. When SurveyMonkey and Qualtrics ask customers directly, comfort with AI for routine transactional queries is consistently high and falls off a cliff as the stakes rise. The vendor numbers and the independent numbers agree on where the line sits.

One thing worth flagging: if your FAQ page isn't already cutting ticket volume, a chatbot layered on top of the same content won't magically fix that. The knowledge base underneath matters more than the interface on top. (We wrote about why most FAQ pages fail if you want to fix the foundation first.)

After-hours coverage. Your team works business hours. Your customers don't. It's 2am, someone in a different timezone needs their tracking link for a gift arriving tomorrow. Without a chatbot, that ticket sits in the queue until morning, and the customer has already left a one-star review by the time anyone sees it. With one, they get the answer in seconds.

Onboarding and documentation. If your product has a knowledge base, help docs, or setup guides, a chatbot that surfaces the right article from a natural language question is properly useful. The customer describes what they're trying to do in plain language and gets the right answer back, without clicking through an FAQ tree.

Common thread across all three: structured information, factual answers, high volume, low ambiguity.

Where chatbots are oversold#

If you've sat through a chatbot vendor demo recently, you'll recognise this tier. The use cases that look impressive on screen but rarely pay off in practice. The difference here is that the evidence gap is enormous: virtually all positive conversion data for sales chatbots comes from vendors measuring their own platforms, without independent verification or comparison groups.

Complex sales conversations. A chatbot can qualify a lead ("what's your company size?", "what problem are you solving?"). Fine for top-of-funnel. But the buyer side tells a harder story, and the most credible read on it comes from independent research rather than the vendors selling the bots.

Gartner predicts that by 2030, 75% of B2B buyers will prefer sales experiences that prioritise human interaction over AI (August 2025). Academic research backs this up: Chang (2022) in the Journal of Business Research found that AI is effective for providing information and making initial contacts, but human salespeople are superior in understanding needs and building relationships in later sales stages.

Relationship management. Some vendors pitch chatbots for customer success, account management, or retention conversations. These depend entirely on context, history, and emotional intelligence. Automating them signals to the customer that you don't value the relationship enough to have a person handle it. The Reddit/SurveyMonkey "Hidden B2B Journey" report (published March 2026, n=1,202 US business decision-makers) puts a number on it: AI chatbots are trusted by just 39% of B2B buyers during research, below vendor websites (55%), search engines (54%), and review sites (46%). Only 18% of respondents use chatbots at all during B2B research.

Anything requiring creative problem-solving. When a customer's situation doesn't fit your policy framework. When they need an exception, a workaround, or just someone who'll listen and make a call. Salesforce's Connected Customer report (n=16,585) quantifies the comfort gradient: 40% of customers are comfortable with AI scheduling an appointment, and that drops to 17% for making financial decisions on their behalf. The higher the stakes, the less tolerance there is for automation.

The mistake most businesses make (and if you're running a chatbot that isn't paying off, this might be why) is treating it as a universal tool. It's a specialist. Use it outside its lane and you don't just miss the ROI; you actively damage the experience. Klarna learned this publicly. After announcing in February 2024 that its AI assistant handled 2.3 million conversations doing the work of 700 agents, CEO Sebastian Siemiatkowski admitted by mid-2025 that quality had declined. They began rehiring human agents.

The complexity gradient (not a simple yes/no)#

The headline stat you'll see everywhere is that 79% of consumers prefer humans over AI agents (SurveyMonkey, December 2025, n=2,017). But that number obscures what's going on. The same study includes a breakdown by interaction type that tells a much more useful story.

59% of consumers are comfortable using AI to return an item, and 65% to have it order food and drinks. For pricing questions, 71% prefer a human and just 12% prefer AI. For financial or billing disputes, 85% prefer a human and 5% prefer AI. The preference for humans scales directly with complexity and stakes.

The Qualtrics 2026 Consumer Experience Trends Report (20,000+ consumers across 14 countries) adds the failure data: AI-powered customer service fails at four times the rate of other AI applications. Nearly one in five consumers who've used AI for customer service report getting no benefit at all. Metrigy's CXO 2025-26 study (n=503) found that 80.1% would still prefer a human even if assured their issue would be resolved by AI.

So why deploy a chatbot?

Because for simple, routine issues, plenty of customers actively prefer self-service. The preference for humans is real, but it's conditional on complexity. When the choice is between an immediate automated answer to a routine question and sitting in a queue, most people will take the bot. As long as it's correct, and they can reach a human when it isn't.

That second condition is where most implementations fall apart. Customers hate chatbots for a specific reason: most of them loop, misunderstand, and make it hard to reach a person when the automation fails.

The top use cases where consumers find AI agents helpful are routing to the right person (50.4%), shipping confirmations (49.6%), and scheduling (46.9%), according to Metrigy. Every one of them is support work. Sales doesn't appear on the list.

What your team thinks about AI#

HubSpot's research found that 80% of customer support specialists say AI and automation let them spend less time on manual work like data entry and scheduling, and 78% say the tools help them work more efficiently and focus on the more important parts of their role (HubSpot).

These numbers reframe the conversation. The chatbot's job is to take the repetitive work off your support team's plate so they can spend their time on conversations that need a human. The frustrated customer who needs empathy. The complex return that requires judgement. The account that's about to churn if someone doesn't pick up the phone.

Nobody on your support team signed up to copy-paste tracking links eight hours a day. Those repetitive queries are the ones most prone to burnout, mistakes, and slow responses. Automate them, and your team does better work on everything else. If you're scaling a support operation, that split between "routine" and "human-required" work becomes the whole strategy.

Guardrails aren't optional#

The case law is already building. Air Canada was held liable when its chatbot gave incorrect bereavement fare advice (Moffatt v. Air Canada, 2024 BCCRT 149). A Chevrolet dealership's ChatGPT-powered bot was manipulated into agreeing to sell a car for $1 (AI Incident Database, Incident 622). These aren't hypotheticals; they're precedents.

The Qualtrics 2026 Consumer Experience Trends Report found that 53% of consumers now cite data misuse as their top concern when companies use AI to automate interactions, up 8 points year over year. Your customers are paying attention to this even if your vendor isn't.

For a business deploying a chatbot, the non-negotiables look like this: the bot drafts every reply from a verified knowledge base, not from a general-purpose model improvising. Every proposed answer then has to clear a confidence threshold before it sends, and the same gate catches answers that would commit you to things your policies don't. Fail the threshold repeatedly and the customer gets a human, not a guess. And asking for a person works instantly. Not after three rounds of rephrasing.

Hay is built around exactly this. Verified knowledge base. A confidence gate on every reply. Hard escalation boundaries. It works alongside your existing helpdesk (Zendesk, Intercom, Gorgias, and others) rather than replacing it. The AI handles the routine queries from your own data. Everything else goes straight to your team.

How to evaluate whether it's right for your business#

Skip the feature comparisons for now. Three questions matter more.

What percentage of your support volume is repetitive and factual? Pull a month of tickets. Categorise them. If more than 30% are order status, shipping questions, return policies, or other factual queries with definitive answers, you've got a strong chatbot use case.

Do you have the data for the bot to answer from? A chatbot is only as good as its knowledge base. If you don't have documented answers to your most common questions, the chatbot will either hallucinate (dangerous) or punt everything to a human (pointless). Build the knowledge base first.

Can you measure success honestly? The metric that matters is resolution rate: did the customer's problem get solved? Deflection rate (did the bot prevent a human from getting involved?) is easier to measure, but optimising for it will cost you. Your customers will feel the difference. We covered which support KPIs predict churn separately; the short version is that resolution quality matters more than resolution speed.

If the answer to all three is yes, start small. One ticket category. 30-day pilot. Measure containment rate and CSAT against your human baseline. Expand from there.

AI chatbots for business are past the "should we?" phase for most companies. The direction is clear, and the economics work for routine, repetitive queries. But "should we?" and "how should we?" are different questions entirely. The businesses getting value are the ones using chatbots for support and self-service, with verified data and clean escalation paths, measuring resolution rather than deflection.

Hay starts at €50/month with 500 resolutions included. No credit card required, and you get a testing mode before anything goes live. Start a 30-day free trial and see what your support queue looks like when the routine queries handle themselves.