AI Integration Case Studies: What They Show

Most AI integration case studies you’ll be shown were picked to win a sale. That doesn’t make them false. It makes them a sample of the best results, written by the people who sold the tool. If you’re deciding whether AI is worth building into your business, you need the wins, the surveys that cover everyone else and at least one public failure, side by side.

We build AI integrations through our AI integration service, so weigh that. Every case study, survey figure and court finding below was checked against the company’s own announcement, the report or paper itself, or the tribunal decision in October 2026. Company figures are self-reported by the company or its vendor. None of them is independent evidence.

Short answer: The published case studies are real, and some of the numbers are large. Klarna said its AI assistant handled two-thirds of its customer service chats in its first month. US Foods said its sellers saved over 32,000 hours in six months. But they’re self-reported by the company or its vendor, and they come from big companies. The broad surveys are more sober. In McKinsey’s 2026 survey, 37% of respondents say AI has contributed to their organization’s EBIT. MIT NANDA’s 2025 report says 95% of organizations are getting zero return from GenAI. The independent studies show real gains on specific tasks, such as 14% more customer support issues resolved per hour. And Air Canada was held liable for what its chatbot told a customer. Treat a case study as proof that something is possible. Then test it on your own data, against your own baseline, before you commit.

What the Published Outcomes Say

As of October 2026, from the companies’ own announcements, the reports and papers themselves, and the tribunal decision:

Sources: Klarna, US Foods on AWS, Moffatt v. Air Canada, NBER w31161, Noy and Zhang, McKinsey, MIT NANDA, Stanford Enterprise AI Playbook, Grant Thornton.

Why Most AI Case Studies Are Useless as Decision-Making Tools

A case study answers “can this work for someone?” You need the answer to “will this work for us?” Those are different questions, and most case studies are only built to answer the first.

Survivorship Bias: Only the Wins Get Published

Vendors publish the projects that worked. The failures rarely get written up, so you only see one end of the distribution.

Even careful research has this problem. Stanford’s Enterprise AI Playbook studied 51 deployments that had already delivered measurable value, and its authors say so: the data is mostly self-reported, and readers “should consider potential selection bias toward successful deployments” (source). The same report found that two thirds of those companies had significant failed attempts before they got to value. Those earlier failures don’t appear in the success story.

The surveys show what the success stories leave out. In McKinsey’s 2026 survey, 37% of respondents attribute at least some EBIT impact to AI, essentially unchanged from the year before. Only about 6% are what McKinsey calls high performers: they attribute 5% or more of EBIT to AI and describe its impact as “significant” (source).

Enterprise Results Don’t Transfer to SMB Contexts

Most published outcomes come from large companies. Klarna describes 150 million consumers (source). US Foods serves about 300,000 restaurants (source). Stanford’s 51 cases come from 41 organizations “representing over a million employees” (source). In McKinsey’s survey, 36% of respondents work for organizations with more than $1 billion in annual revenue (source).

Size changes the picture in both directions. McKinsey found 40% of respondents from large organizations report scaling AI agents, while the share from smaller organizations stayed flat at 22%. MIT NANDA found the opposite on speed: top-performing mid-market companies reported about 90 days from pilot to full implementation, against nine months or longer for enterprises. NANDA notes these figures come from interviews, not company reporting (source).

When you read a case study, check three things: the company’s size, the state of its data and who built the system. If the case study doesn’t tell you, it isn’t comparable to your situation, however close the use case sounds.

Real Outcomes From 2025–2026: What the Data Actually Shows

Leave the vendor libraries aside and look at the surveys and research.

Where AI Is Delivering Consistent Results (With Numbers)

The clearest gains are on specific, well-defined tasks. In a study of 5,179 customer support agents, access to a generative AI assistant raised issues resolved per hour by 14% on average. Novice and low-skilled workers gained 34%, and experienced, highly skilled workers saw minimal impact (source).

In an experiment with 444 college-educated professionals on writing tasks, ChatGPT cut the time taken by 10 minutes (37%) against a control group average of 27 minutes, and evaluator grades rose too (source).

Stanford’s playbook compares how much the AI does on its own. Agentic implementations, where AI completes multi-step tasks end to end, showed a 71% median productivity gain against 40% for high automation, but were only 20% of cases. Human-in-the-loop setups, where a person approves each output, showed a 22% median (source). Remember these are the successful deployments, reported by the people who ran them.

Grant Thornton’s 2026 survey found that companies with fully integrated AI were nearly four times more likely to report AI-driven revenue growth than those still piloting it, 58% against 15% (source). That’s a self-reported correlation, not proof that integration causes growth. It does fit what the other sources say: depth of integration matters more than which tool you pick.

Where AI Consistently Underperforms, and Why

The drop-off happens between pilot and production. MIT NANDA found 60% of organizations evaluated enterprise-grade AI tools, 20% reached pilot and 5% reached production. In its words, most fail due to “brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.” General chatbots like ChatGPT and Copilot were widely adopted, but the report says they “primarily enhance individual productivity, not P&L performance” (source).

McKinsey’s numbers show the same gap. 80% of respondents say AI has improved their individual productivity, while the share reporting an EBIT contribution stayed at 37% (source).

Stanford’s interviews point to why. 77% of the hardest challenges were “invisible”: change management, data quality and process redesign, not the technology (source). A demo shows what the model can do. It doesn’t show whether your process, data and people are ready for it.

The Governance Problem: 78% of Executives Can’t Pass an AI Audit

The headline is shorthand. Grant Thornton’s survey measured confidence, not audit results: 78% of senior leaders lack full confidence that their organization could pass an independent AI governance audit within 90 days. The same survey found 46% say AI underperforms because controls and compliance aren’t working, and only 52% of boards have set clear AI governance expectations (source).

For a smaller business, governance means plain things. Someone owns the output. Errors get logged. There’s a defined path when the AI is wrong, and a way to switch it off. The Air Canada case below shows what happens when nobody checks what a customer-facing system says.

Four Published Cases, Read for a Smaller Business

Published, checkable outcomes from small businesses are rare. So each of the four below is a public record you can open yourself, with the company’s size stated and the lesson framed for a smaller business.

Customer Service Automation: What “30% Workload Reduction” Actually Means

A headline like “30% workload reduction” means nothing until you know the denominator. As an illustration of our own: if the AI only handles routine queries, and those are half your support volume, a 30% saving on them is a 15% saving on total support hours.

Klarna’s announcement is the best-known example. One month after launch, it said its OpenAI-powered assistant had 2.3 million conversations, two-thirds of its customer service chats. It said the assistant did “the equivalent work of 700 full-time agents”, cut the time to resolve an errand from 11 minutes to under 2, and was “estimated to drive a $40 million USD in profit improvement” in 2024 (source). These are Klarna’s own figures, published after one month, and the profit number is an estimate.

The independent study of 5,179 support agents is a better guide to what an assistant does for a support team: 14% more issues resolved per hour on average, with most of the gain among newer staff (source). For a small support team, the honest question is which share of your tickets the AI can actually take, and who reviews the rest.

Operations Workflow: The Difference Between Productivity and Revenue

US Foods built an application on Amazon Bedrock, using Anthropic’s Claude and Amazon Textract, to prepare sales proposals from menus, spreadsheets and handwritten notes. AWS’s customer story says the company built a proof of concept in 6 weeks, piloted it in one area, then gave it to 3,300 sellers. Proposal time fell from 3–4 hours to 20 minutes, and sellers saved over 32,000 hours within 6 months (source).

Notice what’s measured. Hours saved is a productivity number. The story also says US Foods “increased its return on investment by 10 times”, without saying what that’s measured against, and it reports no revenue figure. The time saved could turn into more sales, but the story doesn’t show that.

For a smaller business, the lesson is the same at any size. Decide before launch what the freed time is for, and measure that, not just the hours.

Marketing and Content: Where Gains Are Real and Where They Plateau

The gain on drafting is real and measured. In the MIT experiment, professionals including marketers and grant writers finished 20–30 minute writing tasks, such as press releases and short reports, 37% faster with ChatGPT (source).

The same paper also notes that 68% of participants using ChatGPT submitted its first output without editing it. That fits a one-off writing task. It isn’t how you’d want brand content produced week after week. We haven’t found a published primary study that measures how content gains hold up over months, so treat any plateau claim, ours included, as a judgment, not data. Our view: keep a person on the brief, the angle and the final edit, and use AI for drafts.

A Failed Integration: What Went Wrong and What It Cost

In November 2022, after their grandmother died, Jake Moffatt used the chatbot on Air Canada’s website. It suggested they could apply for a bereavement fare retroactively. Air Canada’s policy didn’t allow that. When Moffatt claimed the difference, Air Canada argued the chatbot was responsible for its own actions. The tribunal called this “a remarkable submission” and said a chatbot “is still just a part of Air Canada’s website” (source).

The tribunal found negligent misrepresentation and ordered Air Canada to pay $812.02: $650.88 in damages, $36.14 in interest and $125 in fees. The money is small. What it shows matters more: you’re responsible for what your AI tells customers, just as you are for every other page on your site. Before any AI answers customers, check its answers against your actual policies, limit what it may promise and give it a clear route to a person.

How to Evaluate AI Case Studies Before Signing a Contract

Vendors will show you their best work. That’s normal sales. Your job is to test what they show you.

Five Questions That Reveal Whether a Case Study Is Credible

1. What was the business size and technical maturity? If they won’t say, the case study is decoration. A company with 3,300 sellers and a vice president for machine learning isn’t your reference class.
2. What did it cost in total, not just licensing? Ask about build, staff time, running old and new processes side by side, and maintenance.
3. How long before results were measurable, and how were they measured? MIT NANDA measured ROI six months after the pilot (source). Results quoted after a few weeks often describe usage, not outcomes.
4. What failed, and how was it fixed? Stanford found 61% of successful projects included at least one prior failure (source). If the answer is “nothing failed”, ask again.
5. Is there a named contact you can speak with? A reference call with a real client is the most useful check you can make.

Red Flags in Vendor-Supplied Success Stories

These patterns don’t mean a vendor is dishonest. They mean the evidence is weak:

• Percentages without a baseline. “50% faster” than what?
• Ratios without a denominator. “10 times ROI” with no cost or return figure behind it.
• Testimonials without a task. Anthropic’s small-business launch post quotes a founder: “What used to take me 120 hours now takes me five minutes” (source). It’s a sincere quote at a launch, but it doesn’t say which task, so you can’t plan on it.
• Activity instead of outcomes. Conversations handled or documents processed, with no cost, quality or revenue result.
• Early snapshots. Klarna’s figures were published after one month.
• The vendor measured its own result. Ask who collected the numbers.

Related Posts

• What AI ROI actually looks like for mid-market businesses: setting a baseline before you measure
• How to measure AI automation success beyond cost savings: choosing metrics that show real outcomes
• AI project failure patterns: what goes wrong after the case study
• What to ask an AI agency before signing anything: the questions to bring to a sales call

Frequently Asked Questions

What percentage of AI integrations actually deliver ROI?

There’s no single number, because each study defines success differently. In McKinsey’s 2026 survey, 37% of respondents attribute at least some EBIT impact to AI. MIT NANDA’s 2025 report says “95% of organizations are getting zero return” from GenAI, based on interviews, surveys and public initiatives. Both are about large organizations more than small ones. Productivity gains on specific tasks are better documented, for example 14% more support issues resolved per hour in one study.

How long does a typical AI integration take before results are visible?

It varies too much for a standard answer, so we don’t quote one. Stanford found similar use cases took weeks at one company and years at another, and put the difference down to the organization, not the technology. MIT NANDA’s top-performing mid-market companies reported about 90 days from pilot to full implementation. US Foods reported its savings over the first six months. Decide what you’ll measure before launch, so you know when results are real.

Are the AI case studies vendors show representative of typical outcomes?

No. They’re selected for good outcomes and usually come from large companies. Even Stanford’s research on successful deployments warns of selection bias toward successes. Use vendor case studies to see what’s possible, and the broad surveys to see what’s typical.

What’s the biggest reason AI projects fail for small businesses?

None of the studies here isolates small businesses, so be wary of anyone who claims a precise answer. Across them, the causes are organizational. Stanford found 77% of the hardest challenges were change management, data quality and process redesign. MIT NANDA points to brittle workflows and a poor fit with daily operations. Both are problems you can check before you build.

How do I know if my business is ready for AI integration?

Three signs: you can define the task the AI will handle and the output you expect, your input data is consistent, and someone owns the quality of the output. If the task is vague, the data is messy or nobody is accountable for reviewing results, fix that first.

What should I do if a vendor can’t share a comparable SMB case study?

Ask why, and listen to the answer. A credible one sounds like: “We don’t have an exact match, but here’s the closest, and here’s why it’s comparable.” A weak one is “all our clients are confidential, trust us.” If there’s no reference at your scale, propose a small pilot on your own data with a success threshold agreed in writing beforehand.