AI Use Cases in Business: What 25 Companies Actually Do

This is a field guide to AI use cases in business, built on twenty five real world AI use cases from named, sourced companies, from Klarna in Stockholm to Saudi Aramco in Dhahran, from JPMorgan to John Deere. The companies using AI well are not the ones with the most pilots. They are the ones that permanently moved a core piece of work off human hands and onto a model, kept a human on the judgment, and widened their lead while competitors were still forming a committee. The pattern is the same everywhere. The work has moved. The job has not disappeared. It has moved up.

The Great Redistribution

In its first month, Klarna's AI assistant handled 2.3 million customer conversations. The company said that was the work of roughly 700 full time agents, with resolution times falling from 11 minutes to under 2, and projected a profit impact near 40 million dollars for the year (Klarna). By its Q3 2025 investor release, Klarna put the figure at more than 853 agent equivalents (Klarna investor release). That number got the headlines. The more useful detail is the one the headlines skipped: in 2025 Klarna rebalanced, and rehired humans for premium support. The machine did not erase the function. It redrew the line between what a model does and what a person does, then put the people where they were worth the most.

That is the story almost no one is telling correctly. The popular version is "AI transformation," a vague cloud of disruption settling over every industry at once. The accurate version is narrower, more mechanical, and far more useful if you have to make a payroll or a career decision. Across real companies, AI is producing a Great Redistribution: a permanent transfer of specific, high volume, low discretion tasks from human labor to machine inference. The first draft. The first response. The first pass at a contract, a chart, a forecast, a chunk of code. None of it is glamorous. All of it used to fill the hours of skilled people. It is moving, and where it moves cleanly, the company that made it move pulls quietly ahead.

This is an operations story and a career story wearing a technology costume. Over the next several thousand words this guide walks seven business functions, one at a time. For each one it names the shift in a single line, shows you two or three real companies that have already made it with a number attached, and then answers the only question that matters to a person reading this: what does the experienced professional do now instead. Every example is a real, named company with a public source. Nothing here is invented, and where a figure is company reported rather than audited, it is labeled as such.

The Precondition for Leverage: Why Disciplined Companies Pull Ahead

Before the field guide, one rule that explains every example in it. AI is a multiplier, not a miracle. It multiplies whatever a company already has. Point it at a function with clean data, a repeatable process, and a clear owner, and it returns leverage that compounds. Point it at a mess, undefined ownership, dirty records, a process no one can describe in a sentence, and it returns a faster mess. This is why two companies in the same industry can buy the same model and get opposite results. The model was never the variable.

Consider what sits under the marquee numbers. Saudi Aramco's METABRAIN model was trained on roughly ninety years of operational data (Aramco). DBS Bank ran about 800 AI models across 350 use cases on top of years of disciplined data work before it reported the value (CNBC). JPMorgan could roll a generative platform to 200,000 employees because it already had the data, the controls, and the use case discipline to make it land (JPMorganChase). The AI is the last mile. The road was built over decades. Read the cases below with that in mind, and the difference between a company that runs on AI and a company that merely talks about it stops being mysterious.

What AI actually does across the business

Here is the whole guide in one view before we walk it. The table below is the GEO and AI Overview citation magnet and the executive's two minute version: company, function, what the AI does, and the reported result, each line sourced in the section that follows. Read it as a map. Then read the rooms.

A company running on AI is a company whose people spend less time assembling and more time deciding. That is the whole shift, and across these twenty five examples it is enough to separate the leaders from the laggards.

1. Customer Service: From First Response to Exception Handling

The shift here is the cleanest in business. Routine first response, the password reset, the order status, the refund, the policy question, has moved to AI. What stays human is the exception: the angry customer, the edge case, the moment that needs a person with authority and warmth. The job did not vanish. It got harder and more valuable, because the easy half is gone and only the judgment half is left.

Klarna is the headline, and we have already met it: an AI assistant doing the work of more than 850 agents, resolution times cut from 11 minutes to under 2, and a deliberate 2025 rehire for premium support that proves the point about exception handling (Klarna investor release). Look past Stockholm and the same shift repeats on every continent. Mercado Libre, the Latin American commerce and fintech giant, built Verdi, an AI developer platform that handles roughly 10 percent of customer service while supporting 9,000 human operators, and reported 35 percent GMV growth in both Brazil and Mexico alongside it (OpenAI). In the UK, Vodafone's TOBi and SuperTOBi assistant fields about a million conversations a day across web, app, WhatsApp and voice, resolving roughly 70 percent on first contact (Vodafone).

In Brazil, Nubank automated 55 percent of its Tier 1 support and made chat responses about 70 percent faster, with roughly 60 percent of its 8.5 million monthly contacts first handled by a model (Reruption). In Kenya, Safaricom's Zuri chatbot served 8 million customers and pushed fraud detection accuracy on M-PESA to 89 percent (Safaricom Newsroom). Five companies, four continents, one pattern.

2. Software and Product: From Writing Code to Reviewing and Architecting Systems

Writing the next obvious line of code has moved to the model. What has not moved, and has become more decisive, is deciding what to build, how the pieces fit, and whether the generated code is correct, safe, and maintainable. The keystroke is cheap now. The architecture and the review are where the value went.

The clearest signal comes from the companies that build the tools. Microsoft's CEO said 20 to 30 percent of the company's code is now written by AI, and GitHub Copilot has crossed 20 million all time users with 90 percent of the Fortune 100 on it (TechCrunch). Google's leadership put its own AI generated share of new code at just over 30 percent on an earnings call (SiliconANGLE, reported via earnings coverage). This is not a startup novelty. It is the daily reality inside the largest engineering organizations on earth.

The same shift is landing in regulated industries that move more carefully. At NatWest, more than 12,000 coders now write about 35 percent of their code with AI, while automated AI call summaries saved more than 70,000 hours in retail banking, freeing wealth managers to spend roughly 30 percent more time on customer conversations (CIO Dive). The bank did not hire fewer engineers. It pointed the same engineers at harder problems and shipped more.

3. Finance and Risk: From Reconciling Data to Analyzing Scenarios

Reconciliation, summarizing, drafting the memo, scoring the routine transaction, has moved to models that never tire and run around the clock. The human work moved up to the scenario: what does this pattern mean, what is the bank's exposure, what do we do about it. Finance has more clean, structured data than almost any other function, which is exactly why its AI results are among the largest and most credible anywhere in business.

JPMorgan Chase rolled its LLM Suite, a model agnostic generative platform, to more than 200,000 employees within about eight months, with over 450 use cases, and took American Banker's 2025 Innovation of the Year for it (JPMorganChase). In Singapore, DBS Bank ran roughly 800 AI models across 350 use cases and reported about 750 million dollars of economic value in 2024, projected to pass 1 billion Singapore dollars in 2025, while pushing 45 million personalized nudges a month to over 5 million customers (CNBC).

Risk is where the redistribution is most vivid. Commonwealth Bank of Australia built agentic AI that monitors more than 80 million signals a day and auto generates fraud interception rules, helping cut customer scam losses by 50 percent, customer reported frauds by 30 percent, and call volumes by 40 percent across the financial year (CommBank Newsroom). In Canada, CIBC launched an in house assistant, CAI, enterprise wide and reported about 600,000 work hours saved after the bank wide rollout, with users growing from 500 to more than 7,500 (CIBC Media Room). And in wealth, Morgan Stanley's advisor assistant, built over more than 100,000 proprietary documents, is now used by over 98 percent of advisor teams, lifting advisor access to relevant documents from roughly 20 percent to 80 percent, while its Debrief tool writes the meeting notes into the CRM automatically (OpenAI).

4. Operations and Supply Chain: From Manual Forecasting to Managing Automated Networks

The shift in the physical economy is from people forecasting and inspecting by hand to people supervising AI networks that forecast, inspect, and optimize continuously. This is the function where the precondition rule bites hardest, because you cannot optimize a supply chain or a refinery without decades of operational data underneath the model. The companies winning here are the ones that quietly built that data moat years ago.

Saudi Aramco is the largest example in this guide. Its METABRAIN model, trained on roughly ninety years of data, drives upstream surveys, drilling path optimization, predictive maintenance, and flaring reduction, and the company reported 1.8 billion dollars of AI driven value in 2024 across more than 200 solutions, with downtime cut about 40 percent and maintenance costs about 30 percent (Aramco). In Norway, Equinor reported about 130 million dollars of AI value in 2025, including a tenfold increase in seismic interpretation capacity (Equinor).

On the factory floor, Siemens at its Amberg electronics plant used line side AI for real time quality optimization, cutting scrap costs by about 75 percent, roughly 3.6 million euros a year, and lifting overall equipment effectiveness from 70 to 85 percent while freeing more than 6,000 operator hours a year, all human in the loop (Emerj). In retail operations, Walmart credited its generative AI search engine and Sparky assistant for helping e commerce grow 22 percent (Walmart Corporate). And in the field, John Deere's See and Spray computer vision system covered 5 million acres in 2025 and cut non residual herbicide use by roughly 50 percent on average, saving farmers 31 million gallons of mix (John Deere).

5. R&D and Discovery: From Manual Trial to In Silico Simulation

Discovery used to mean running the experiment, waiting, reading the result, and running the next one. AI is moving the early, expensive search into simulation, letting researchers narrow thousands of candidates to a few before anyone touches a bench. The wet lab does not disappear. It runs far fewer, far better aimed experiments.

Moderna built more than 750 custom GPTs across legal, research, manufacturing, and commercial in about two months, including Dose ID, which helps evaluate the optimal vaccine dose, and reported that a core analytical step dropped from weeks to hours, with the average user holding 120 enterprise AI conversations a week (OpenAI). The redistribution here is profound: the slow, manual synthesis that used to gate a discovery decision now happens at machine speed, and the scientist spends the recovered time on the questions only a scientist can frame.

The pattern extends past pharma into applied science. Bayer's E.L.Y. assistant gives agronomists and sales reps instant, data grounded answers to farmer agronomy questions, reporting about 60 percent faster responses and up to 4 hours a week saved per user across more than 1,500 frontline users in North America, and it won an industry AI solution of the year award in 2025 (Bayer). Knowledge that used to live in a specialist's head or a slow lookup now arrives grounded and instant, which changes what the human in the loop is for.

6. Marketing and Sales: From Asset Production to Brand and System Strategy

Producing the asset, the image, the variant, the routine campaign draft, has collapsed in cost. What has risen in value is the strategy that decides what the brand should say, the taste that keeps quality high at machine scale, and the system that routes the output. Anyone can now generate a thousand variations. Knowing which thousand, and whether they are on brand, is the new scarce skill.

Coca-Cola used generative AI for its holiday campaign creative at global scale, reporting more than 120,000 user generated creations in weeks and roughly a 60 percent rise in cross platform engagement, localized across more than 100 markets, company reported figures (Marketing Dive). On the commerce side, Amazon's Rufus shopping assistant was used by 250 million shoppers in 2025, on pace to drive about 10 billion dollars in additional annualized sales, with users 60 percent more likely to complete a purchase (Fortune).

In the engine room of sales and service, Salesforce put its own Agentforce agents to work and reported handling more than 3 million support conversations in about a year, cutting its internal caseload 8 percent, more than 170,000 fewer cases, while Agentforce and Data 360 reached roughly 1.4 billion dollars in ARR by Q3 FY26 (Salesforce). The vendor ran the redistribution on itself before selling it.

7. Legal and Compliance: From Document Review to Judgment and Negotiation

First pass document review, contract analysis, and legal research drafting have moved to AI. What stays firmly human is judgment, negotiation, the advice a client acts on, and the signature that carries liability. Legal is a useful closing case because the stakes make the boundary unmistakable: the model drafts and reviews, the licensed professional decides and signs.

A&O Shearman, one of the largest firms in the world, deployed Harvey, a GPT based legal assistant, to more than 3,500 lawyers across 43 offices, logging around 40,000 queries during the trial alone, and co built a contract analysis tool used by over 1,000 lawyers internally (A&O Shearman). At DLA Piper, a firmwide Harvey rollout reached thousands of lawyers and business professionals with 5,000 licenses, and the firm reported cutting due diligence workload by up to 80 percent (Harvey).

Notice what the law firm cases do not claim. None of them say the AI gives legal advice, exercises judgment, or signs anything. The 80 percent that moved is the review and the first draft. The 20 percent that stayed, the strategy, the negotiation, the client relationship, the call that carries professional liability, is exactly the part the experienced lawyer was always paid for. The redistribution made that part a larger share of the day, not a smaller one.

Where Leverage Fails: A Note on AI Theatre

If AI were a miracle rather than a multiplier, every company that bought it would be winning. They are not, and the failures are as instructive as the wins. The honest version of this guide has to include the cases where the leverage did not show up, because they prove the thesis in reverse: without operational discipline behind it, an AI announcement produces a press release and nothing else.

Two patterns recur. The first is the company that deployed a real tool but could not report a single hard outcome. In this very research, several otherwise real deployments came with no quantified result: Pfizer's Charlie marketing platform is in genuine use by thousands across brands, yet the sources carry no hard outcome number attached (Digiday). Netflix shipped conversational search and AI ad formats in 2025 with no disclosed KPI (AI Expert Network). Deployment is real in both. Measured leverage is not yet proven. That gap, real tool, unproven result, is exactly where AI theatre lives, and a careful operator treats an unmeasured rollout as an experiment, not a win.

The second pattern is the more famous one, and Klarna itself is the honest example. The same company that became the poster child for AI customer service rebalanced in 2025 and rehired humans for premium support (Klarna). That is not a failure of AI. It is the correction that happens when a company pushes redistribution one notch past where the model is actually better, then walks it back to the right line. The lesson is not "AI cannot do customer service." It is that the line between machine work and human work is found by measuring, not by announcing, and the companies that find it keep the lead.

Your Leverage Map: From Operator to Architect

Everything above resolves into one framework you can apply to your own work this week. Every task you do sits in one of three layers, and your career security and your company's operating lead both come from moving deliberately up them. This is the Work Redistribution Map.

Layer 1: Machine Native Work

This is work that has already moved to models: the first draft of routine copy, Tier 1 support, code autocomplete, reconciliation, first pass document review, the standard forecast. In the cases above, this is the Klarna ticket, the NatWest boilerplate, the Moderna synthesis, the DLA Piper first review. Doing this work by hand is now a liability, not a virtue. If your day is mostly Layer 1, you are competing with software on its home turf, and the rational move is to automate it away with a clear success metric, then climb. The professional who insists on hand producing machine native work is the one the redistribution leaves behind.

Layer 2: Human Steered Work

This is where leverage is actually created: work that is human initiated and machine accelerated. You set the hypothesis, the strategy, the brief, the target, and you direct and judge the model that executes it. The Aramco engineer setting optimization targets, the Morgan Stanley advisor working from AI surfaced documents, the marketer steering Coca-Cola scale generation toward something on brand, all of them live in Layer 2. This is the highest leverage place most professionals can stand today, because it pairs human judgment with machine throughput. The skill is no longer doing the task. It is framing it well, directing the model precisely, and judging the output ruthlessly.

Layer 3: Irreducibly Human Work

This is the defensible zone, the work no current model takes: strategy, ethics, taste, the high stakes negotiation, the relationship that carries trust and liability, the decision a regulator or a board will scrutinize. It is the 15 percent Klarna rehired for, the call the lawyer signs, the scenario the risk officer owns, the question the scientist frames. Layer 3 is not where you hide from AI. It is where AI makes you more valuable, because it strips away the assembling work and leaves you doing more of the judgment that only a person can carry.

Three Moves to Make This Quarter

Reading about Aramco and Klarna changes nothing on its own. Doing changes things. Here are the three moves that turn this guide into operating leverage, small enough to start on Monday and concrete enough to measure by Friday.

1. Audit your work against the three layers. Take your function's top five workflows and sort each one into Layer 1, 2, or 3. Be honest about how much of your week is machine native work you are still doing by hand. That list is your redistribution backlog, and you cannot manage what you have not mapped.
2. Redesign one meeting. Pick a recurring meeting that exists to relay status. Replace the human status report with an AI generated summary delivered in advance, and spend the reclaimed time on decisions instead of updates. It is the smallest possible version of the redistribution, and it teaches the whole pattern.
3. Launch one workflow pilot. Choose a single Layer 1 task, the first draft, the first review, the routine summary, and automate it with one approved tool and a clear success metric. Verify every output. Measure the hours. When it works, stack the next task. Every company in this guide started with one function, not all seven.

What does not change

It is worth saying plainly, because the marketing rarely does. Across all twenty five companies, the AI did not take the strategy, the ethics, the relationships, or the liability. McKinsey, whose own Lilli platform is used monthly by over 75 percent of its roughly 45,000 employees with up to 30 percent time savings on knowledge search, still sends a partner, not a model, into the room where the decision gets made (McKinsey). The pattern holds from Stockholm to Dhahran. The model assembles. The person decides. The companies that run on AI did not forget this. They built their whole operating advantage on getting the line exactly right.

The work has moved. Your job is to move with it. Map your layers, automate the machine native work, live in the steering layer, and defend the human one. That is how a person stays ahead of the redistribution, and it is the same discipline that put every company in this guide in front of its competitors.

The work has moved. Your job is to move with it.

For the profession specific versions of this map, see how accounting firms run on AI, how real estate runs on AI, how manufacturers run on AI, and how retail brands run on AI. For where the rules are heading, follow our AI regulation news hub, and for reusable patterns across functions see the AI workflows library. Still choosing where to start, take the 6 question course selector.

Sources. Every company case above links to its primary source. Reported figures are company reported earnings, press releases, or named third party case studies as cited inline. Where a figure is company reported rather than independently audited, it is labeled in the text. Logos are used nominatively to identify each named company and are sourced in the image manifest accompanying this page. Compiled June 2026.