In short: The best-documented AI in business examples include customer support, where a study of 5,000 agents found 13.8% more issues resolved per hour, and routine content and coding assistance. Results vary, and some high-profile deployments have backfired. Ask whether a claim has a baseline, a comparison group and all costs included, since many published case studies come from vendors.
Case studies in this field are plentiful and uneven. Vendors publish stories about their best customers, and companies announce savings to please investors. The examples below include what’s independently studied and what went wrong, along with a note on how far to trust each one.
Customer support: the strongest evidence
What happened. Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied about 5,000 agents at a Fortune 500 software company. A generative AI tool suggested responses, which the agents could accept or ignore. It was rolled out gradually, allowing a comparison with agents who didn’t have it.
| Measure | Result |
|---|---|
| Issues resolved per hour | Up 13.8% |
| Time per chat | Down about 9% |
| Chats handled per hour | Up about 14% |
| Customer satisfaction | No significant change |
| Less experienced agents | Improved by about 35% |
| Top performers | Zero or slightly negative effect |
Why it’s credible. It’s an academic study with a comparison group, published through NBER. The data dates from 2020 to 2021, so current tools may do better.
The lesson. The tool appears to have spread good practice from top performers to newcomers. An agent with two months of experience performed like a six-month agent working without it. The agent stayed in the loop throughout.
Customer service automation: a cautionary case
What happened. Klarna froze hiring in 2023 and leaned heavily into AI. It said its chatbots handled two-thirds of customer service conversations in their first month, doing the work of 700 agents. Headcount fell from about 3,800 to about 2,000, which the company attributed to natural attrition. By May 2025, CEO Sebastian Siemiatkowski acknowledged that too much focus on cost had lowered quality, and Klarna began hiring human agents again, on a remote, contract basis.
How to read it. The initial figures came from the company itself. “Conversations handled” isn’t the same as problems solved or customers kept. Klarna’s experience doesn’t prove that automation fails. It shows that cost savings can hide quality losses, which may take a while to surface.
Chatbots and legal liability
What happened. A customer used Air Canada’s website chatbot and was given incorrect information about bereavement fares, then bought tickets at full price. In Moffatt v. Air Canada (2024), the British Columbia Civil Resolution Tribunal held the airline liable for negligent misrepresentation. It rejected the argument that the bot should be treated as separate, and said customers have no duty to cross-check one part of a website against another.
The lesson. It isn’t a binding precedent for courts, but the principle is simple: a business can’t easily disown what its tools tell customers. Limit bots to verified information, and log what they say.
Software development: gains are real but uneven
The favorable evidence. Developers report widespread use and time savings, and AI is good at repetitive code, tests and explaining unfamiliar code.
The counterexample. METR’s randomized trial in early 2025 had 16 experienced open-source developers carry out 246 tasks. With AI tools they took 19% longer, while believing they’d been 20% faster. The sample was small, the tools have since improved, and these developers knew their code unusually well.
The lesson. Measure it yourself. Self-reports and vendor numbers can deviate far from reality, and results depend heavily on experience and the task.
Everyday office work
Survey results. In the Federal Reserve’s survey of households, fielded in October 2025, one in four workers had used generative AI at work in the prior month, and 81% of those said it saved them time. In BCG’s 2026 survey of 11,749 employees, 42% of regular frontline users reported saving a full day or more each week.
How to read them. These are self-reported and can’t be directly translated into company profit. BCG’s survey also found that more than half of those workers didn’t redirect the saved time to more strategic work. Time saved for a person isn’t the same as money saved for a company.
Common uses and what to measure
Across industries, here are the uses businesses tend to pursue. The results column shows what to check, not a promised outcome.
| Industry | Use | What to measure |
|---|---|---|
| Retail and e-commerce | Product copy, search and recommendations, support | Conversion, return rate, cost per contact |
| Logistics | Demand forecasting, routing, document handling | Forecast error, fuel cost, stockouts |
| Finance | Fraud detection, document review, compliance checks | False positives, review hours, losses |
| Health care | Clinical note drafting, scheduling, image review | Clinician hours, accuracy, safety events |
| Legal and professional services | Contract summaries, first drafts, research | Billable time, error rate, client satisfaction |
| Manufacturing | Predictive maintenance, visual inspection | Downtime, defect rate |
| Marketing | Variations of ad copy, audience analysis, translation | Cost per acquisition, revenue per campaign |
Note that some of these, such as fraud detection and predictive maintenance, rely on older machine learning that long predates chatbots. They’re often better established than the newest generative uses.
How to judge a claim
When you read a story of AI success, ask these questions.
- Who published it? A vendor’s customer story isn’t independent evidence.
- What’s the baseline? “30% faster” than what, and measured how?
- Was there a comparison group? Without one, seasonal changes or new staff may explain the difference.
- Is it outcomes or activity? Tickets closed may rise while customers come back with the same problem.
- Were all costs counted? Include licenses, integration, data work, review time and any rework.
- How long did it run? A six-week pilot might not reveal quality decay.
- What’s the failure rate? You rarely hear about stalled attempts. MIT’s NANDA research reported 95% of organizations finding no measurable return, though that study has its own weaknesses.
- Does it apply to you? Results at a large software firm may not transfer to a small clinic.
What the examples have in common
The better-documented successes share a few features: a well-defined task, plenty of examples, a person checking the output, and a measure of what happened. The failures tend to have removed human oversight quickly, measured only cost, or treated customer-facing accuracy as somebody else’s problem.
To apply this in your own organization, start with our guide on how to implement AI in business. Our article on AI adoption covers why promising pilots stall, and small firms may prefer AI tools for small business. For the wider effect on work, see AI and the future of work.
Frequently asked questions
What is the most common use of AI in business?
Surveys point to writing, summarizing, customer service and data analysis as frequent uses, with marketing and software development close behind. Exact rankings vary by survey and sector. Older types of machine learning are also deeply embedded in fraud detection, recommendations and logistics, though people may not think of those as AI.
Can small businesses get the same results?
Sometimes, on a smaller scale. A small firm won’t have the volume that makes a large deployment pay, but it can move faster and start with cheap tools. Savings on drafting and admin can be significant for a tiny team. Large-company numbers don’t transfer directly, so measure your own results.
How can I calculate return on investment for AI?
Compare the value of measurable gains, such as hours saved and redeployed, higher sales or fewer errors, with total costs, including licenses, integration, training and human review. Use a baseline and, ideally, a comparison group. Count hours as savings only if they lead to useful work or reduced costs elsewhere.
