How real companies put AI to work (and what you can copy)

Why look at cases, not just tools

There's no shortage of AI tools. What's missing is a clear picture of how they fit into real work: who uses them, on which task, with what data, who checks the output and how you measure it. That's where most companies get stuck.

These are big companies, but each one was picked because it teaches something different that translates to a midsize business. One shows the upside and the stumble in customer service. One shows the owner setting the rules. One shows the cost of poor communication. And two show the pattern that works best: AI supporting the people who already do the work, using the company's own data.

Klarna: efficiency in support, and the cost of overdoing it

What they did

Klarna, the Swedish payments company, put an AI assistant in its customer service chat. On February 27, 2024, the company announced that in its first month it handled 2.3 million conversations, two-thirds of all chats. Klarna said it was doing the work of 700 full-time agents, resolving issues in under 2 minutes instead of 11, matching human agents on satisfaction and cutting repeat contacts by 25%.

What went wrong

A little over a year later, CEO Sebastian Siemiatkowski admitted the problem in a Bloomberg interview, reported by CX Dive in May 2025: cost had weighed too heavily, and the result was lower quality support. Klarna started recruiting people for support again and said customers need to know a human will always be there if they want one. In June, he called human service a “VIP thing”: AI for the simple stuff, people for what matters.

AI didn't leave the building. In November 2025, on its Q3 call, Klarna said the assistant now did the work of 853 people and had saved $60 million. The same report notes that support and operations costs still rose year over year, and quotes a Forrester analyst saying the company “overpivoted.”

The lesson for your company

Automating simple support works. Measuring only cost is a trap. Before you put a bot on your website chat or text line, pick two metrics beyond response time: customer satisfaction and how often people come back with the same issue. And always keep the path to a real person visible.

Shopify: the owner sets the rules

What they did

On April 7, 2025, Shopify founder and CEO Tobi Lütke posted an internal memo that had started to leak. The core idea: “reflexive” AI use is now a baseline expectation for everyone, leadership included. Plus one practical rule: before asking for more headcount or resources, teams must show why they can't get it done with AI. He also gave teams a question to work through: what would this area look like if autonomous AI agents were already part of the team?

AI usage also became part of performance and peer reviews.

What drew criticism

The hiring rule landed badly with many people because it sounded like “AI instead of people” at a time of real job anxiety. TechCrunch noted Shopify had about 8,100 employees in 2024 and had cut 20% of staff the year before.

The lesson for your company

In a midsize company, the owner decides whether AI actually gets used. Without a clear rule, it stays at “use it if you feel like it.” You don't need to copy the tone, but you can copy the question: before hiring for a repetitive task, has anyone tested whether AI plus a better process can handle it? That's a decision, not a tool.

Duolingo: poor communication is expensive

What they did

In late April 2025, Duolingo CEO Luis von Ahn posted a memo on LinkedIn declaring the company “AI-first.” According to Fortune, it said Duolingo would gradually stop using contractors for work AI could handle, would weigh AI fluency in annual reviews, and would only hire if a team couldn't automate more of its work.

What went wrong

The backlash was strong, inside and outside the company, including in the comments on the app's social accounts. Weeks later, von Ahn wrote that he hadn't handled the announcement well, said he didn't see AI replacing employees, and announced workshops, advisory councils and dedicated time to experiment. In August he told the New York Times he hadn't given enough context and that the company had never laid off full-time employees.

The business kept going. In its Q2 2026 shareholder letter (August 2026), Duolingo reports 58.7 million daily active users, up 23% year over year, and points to AI features such as Video Call as part of its strategy. You can't credit that growth to AI alone, and the company cites other factors too.

The lesson for your company

In a 30-person company, “we're going to use AI” reaches the break room as “they're cutting jobs.” Say what will change, in which tasks, what happens to the people doing those tasks today, and how everyone will learn. Training and time to experiment come with the announcement, not after the blowup.

Morgan Stanley: AI on top of the firm's own knowledge

What they did

In March 2023, Morgan Stanley Wealth Management announced a partnership with OpenAI to build an assistant that finds and summarizes the firm's internal content: research, reports, procedures. It rolled out to all advisors in September 2023.

In June 2024 came Debrief: with client consent, the AI takes meeting notes, lists action items, drafts the follow-up email (which the advisor edits and sends at their discretion) and saves the record to the CRM. In the same release, the firm says 98% of advisor teams had adopted the assistant.

What to watch

The time-saved quote (“about half an hour per meeting”) comes from an advisor the firm chose to feature. And the design has clear limits: AI prepares, the person decides and sends. The firm keeps repeating that the client relationship stays human.

The lesson for your company

Adoption was high because the tool solved a daily pain for the people using it (finding information, writing up meetings) and worked on the company's own knowledge, not the open internet. You can do the same at a smaller scale: put proposals, template contracts, price lists and common answers in one place and let AI look things up there. That's what we call company memory in the Business AIOS.

Find out how much of your company still depends on you.Take the free index

iFood: coordinated agents, and AI for small businesses

What they did

iFood is Brazil's largest food delivery platform. At the Bloomberg Línea Summit in October 2025, CEO Diego Barreto said 90% of support interactions run through three AI agents: one for consumers, one for restaurants and a third that coordinates the others, like a conductor. He estimated AI already accounts for about 30% of operating profit (EBITDA) and said satisfaction in this model beats human-only support.

For restaurants, iFood launched the Cris agent in August 2025. It works over chat: the owner asks questions and gets orders, cancellations, reviews and promotion ideas, based on their own data and their neighborhood. Internally, in May 2026, the company said it had 14,000 AI agents supporting about 8,500 employees with analysis, data lookups and reports.

What to watch

The EBITDA and satisfaction figures are the CEO's statements, not audited data. And the goal of rolling Cris out to every restaurant by March 2026 was a forecast; I found no public confirmation that it happened.

The lesson for your company

Two takeaways. First, specialized agents with a coordinator work better than one bot that tries to do everything; we explain that setup in AI agent orchestration. Second, Cris shows what a small business owner actually wants: to text “how did this week go?” and get an answer with their own numbers. You don't need to be iFood to have that.

What the cases do not show

All the cases above come from announcements, interviews and reports by the companies themselves or by the press that follows them. That matters. A company tends to publicize what worked, in the format that suits it, and to measure the result by the criterion it chose. The numbers are true in the sense that they were published, but they were not audited by third parties in this article.

• We do not see the cases that did not work. Those who tried and gave up rarely write about it, and that is where most is learned.
• We do not see the full cost. Team time, rework and adjustments usually stay out of the announcement.
• We do not see what already existed. All these companies had data, processes and technical teams before they started.
• We do not see what changed afterwards. Some cases, like Klarna’s, show their own course correction, which is why they are worth more than those that only celebrate.

The useful reading, then, is one of pattern and not of promise: what repeats across companies in different sectors, and what you can test small before betting.

The common pattern across the cases

Size and industry aside, the cases that worked best follow the same sequence. The stumbles show up when a step gets skipped: Klarna measured only cost, Duolingo skipped the conversation with its people.

For examples of where this sequence usually starts in each area (sales, finance, operations), see use cases. And if the question is which pain point to tackle first, where to start with AI in your company shows how to diagnose before picking a tool.

What to copy on Monday

Next step

The cases show the tool is the easy part. The hard part is choosing the right pain point, organizing your own data and deciding who approves what. That depends on how much of the company still runs through you.

A good first step is to measure that. The Dependency Index (IDD) is a free, few-minute quiz that scores your company from 0 to 100 and shows where it depends on you the most. You'll find it on the contact page, along with the option to request a diagnostic.

Cesar Augusto Carlos

Founder, NoOne Systems

Has been building several businesses with digital marketing since 2008. Today he designs systems that bring data, processes and AI together so owners can see the whole company and decide what matters.