Summary: This is a long tour of more than forty real AI deployments across fourteen industries: banking, retail and e-commerce, software platforms, steel, manufacturing, energy, logistics, aviation, agriculture, healthcare, pharma, telecom, insurance and education. For each case you get what was built, how it works technically, the numbers the company or reporter gave, and a label that says how far to trust those numbers. The last third of the guide pulls out the engineering patterns that repeat across all of them and turns them into a learning path.
How to read this guide
Industry AI writing has a trust problem. A single deployment can appear in a company press release, a vendor's marketing blog, an analyst's summary, and a newspaper story, each with slightly different numbers. Some numbers come from a controlled A/B test. Others come from a sales deck. If you treat them the same way, you will learn the wrong lessons.
So every case below carries a source label.
- A · primary The company itself, a peer-reviewed paper, or a published experiment. Engineering blogs, research papers, official press releases and customer references.
- B · reported Journalism or trade press that interviewed the company or read its filings. Usually reliable on facts, less precise on method.
- C · aggregator Case-study aggregators and vendor blogs that summarize others. Useful for orientation. Their numbers often lack a traceable source, so treat them as claims to verify.
One naming warning before the cases. There are two different recommendation papers called GenRec. One is from JD.com (arXiv 2604.14878) and was covered earlier on this site. The other is from Netflix (arXiv 2608.10257) and appears below. Always check the arXiv number.
The shape of an enterprise AI system
Before the cases, a mental model. Almost every deployment below has the same four layers, even when the model is completely different.
- Data in. Sensors, documents, transactions, clicks, images, call logs. This layer is where most projects spend most of their time, cleaning, labeling and joining.
- Model. A classifier, a time-series model, a vision network, a large language model, an optimizer. Often the easiest layer to change.
- Control. Who may call the model, on which data, with what logging, and what a human must approve. This is the layer that decides whether a project passes the security review.
- Action and measurement. What changes in the world when the model speaks, and how anyone knows it was right.
Keep those four layers in mind. When a case below reports a big number, ask which layer produced it. Very often the model is ordinary and the win came from data or control.
Banking and financial services
HDFC Bank: EVA, an early banking chatbot B · reported
What it is. EVA, short for Electronic Virtual Assistant, is HDFC Bank's customer-service chatbot. It was built by Senseforth AI Research, a Bengaluru company, and runs on the bank's website, mobile site and a customer portal.
How it works. EVA is a natural-language-processing system built around intents. The report says it handles "50,000 plus semantic variations" across thousands of banking intents. In plain terms, the team mapped thousands of things a customer might want, such as fees, IFSC codes or how to apply for a product, and trained the system to recognize the many ways people phrase each one. It also learns from daily interactions.
Numbers reported. In its first six months EVA answered 2.7 million queries from more than 530,000 unique users across 1.2 million conversations, with accuracy above 85 percent and 99.9 percent uptime.
Lesson. EVA predates large language models. It shows what intent-based systems did well: narrow, high-volume questions with known answers. Modern LLM assistants widen the range of questions, but the intent catalog is still a useful discipline, because it tells you exactly what the bot is supposed to handle and lets you measure accuracy per intent.
JPMorgan Chase: LLM Suite, OmniAI and COiN C · aggregator
LLM Suite. According to Emerj's summary, LLM Suite is a proprietary generative AI platform launched in summer 2024. It is model-agnostic, integrating models from OpenAI and Anthropic, connects to internal databases and applications, and is updated about every eight weeks. More than 200,000 employees were onboarded within eight months, about two thirds of the workforce. A variant called Connect Coach supports Private Bank advisors.
OmniAI. An enterprise machine-learning platform that standardizes how models are built and deployed. It supports real-time transaction monitoring and anomaly detection for fraud.
COiN. Contract Intelligence is the older and more famous system. Best Practice AI's summary, citing the Independent and TechEmergence, says it analyzes about 12,000 commercial credit agreements a year, extracts around 150 attributes from them in seconds, and saved roughly 360,000 lawyer hours a year. It runs on an in-house cloud called Gaia.
Other numbers in the summary. More than 450 AI use cases in production with a target of 1,000, up to $1.5 billion of estimated annual value, about $250 million a year saved by AI fraud prediction, and a technology budget above $18 billion. These are aggregated claims; treat the dollar figures as directional.
How it works technically. The interesting design choice is model-agnosticism. Instead of betting on one model vendor, the bank built a platform layer that routes requests to different models and connects them to internal data. That is the same pattern as Uber's gateway below: put the control and the connections in a platform the bank owns, and treat models as swappable parts.
Lesson. COiN is extraction, not chat. The value came from turning long legal text into a structured table that a human can check. That remains the most reliable document-AI pattern.
Morgan Stanley: an assistant for financial advisors B · reported
What it is. AI @ Morgan Stanley Assistant gives financial advisors chat access to the firm's own research, investment strategies and analyst insight, drawn from a knowledge base of more than 100,000 internal documents. A companion tool, Debrief, writes meeting notes and draft emails.
How it works. This is retrieval-augmented generation. The documents are indexed, a question retrieves the relevant passages, and GPT-4 writes an answer with sources. The firm ran a 300-advisor pilot first and iterated on frontline feedback before the wide rollout.
Numbers reported. More than 98 percent of advisor teams use the assistant, and Debrief is reported to save about 30 minutes per client meeting.
Lesson. Adoption at that level is rare. The reasons reported are a narrow, trusted corpus (only compliance-vetted content), answers with sources, and a pilot that let the advisors shape the tool. Trust came from the corpus, not the model.
Klarna: an AI assistant, and a partial reversal B · reported
What happened. In its first month, Klarna's OpenAI-based assistant handled 2.3 million conversations, about two thirds of customer-service chats, the work Klarna said was equal to about 700 agents. It works in more than 35 languages across 23 markets and handles refunds, returns, payments, cancellations and disputes. Klarna estimated about $40 million of profit improvement for 2024.
The second half of the story. Later reporting describes Klarna moving back toward human agents after quality and satisfaction concerns, using AI for simple cases and people for the "moments that matter".
Lesson. Automation rate is not the same as resolution quality. A system that closes two thirds of chats can still frustrate customers on the third that matter most. Measure satisfaction and repeat contacts, not only deflection.
Stripe: a foundation model for payments B · reported
What it is. Stripe trained a transformer-based Payments Foundation Model on billions of transactions. Instead of separate hand-built models for each fraud pattern, one large model learns a representation of each payment that downstream tasks can use.
Numbers reported. Detection of card-testing attacks for large merchants improved from 59 percent to 97 percent, and card testing fell by 64 percent almost immediately. Each transaction is scored in under 100 milliseconds.
Lesson. This is the same idea as Netflix's GenRec below, applied to payments: replace many specialized models and features with one general model trained on a huge history, then adapt it. The hard engineering is latency. A fraud score that arrives after the payment is approved is useless.
Kinectify, Ally Bank and Manulife on Azure A · primary
Microsoft's customer roundup lists three more financial cases. Kinectify built an anti-money-laundering platform on Azure Cosmos DB, Azure AI services and Kubernetes, and reports detecting 43 percent more suspicious activity with 96 percent faster decisions. Ally Bank uses Azure AI for personalized recommendations. Manulife containerized document-processing applications on AKS to enable online document submission. These are vendor customer stories, so they are accurate about what was built and optimistic about outcomes.
Retail and e-commerce
Walmart: My Assistant for associates A · primary
What it is. My Assistant is a generative AI tool for Walmart's corporate associates. It helps write first drafts, summarizes long documents in seconds and gives thought starters. It is loaded with Walmart-specific knowledge.
Rollout. It went from vision to viable in about 60 days. About 50,000 US campus associates got it in 2023. In January 2024 it launched in eight more countries (Canada, Mexico, Chile, Costa Rica, El Salvador, Honduras, Guatemala and Nicaragua), with India and South Africa planned, and about 25,000 international associates expected to have access by the end of 2024. It supports native-language interaction.
Lesson. Internal assistants are the lowest-risk place to start. The users are employees, the data stays inside, and mistakes are caught by a person before they reach a customer. Walmart's framing, "people-led and tech-powered", is a statement about where the human stays in the loop.
Amazon: Rufus and Amazon Q code transformation B · reported
Rufus. A shopping assistant trained on Amazon's catalog, customer reviews, community Q&A and web information. It answers specific and vague questions, compares products and recommends items. Reporting says more than 250 million customers used it in 2025, with monthly active users up 149 percent year over year.
Amazon Q for Java upgrades. Andy Jassy said Amazon used its Q assistant to upgrade Java versions across its own codebase. An upgrade that used to take about 50 developer-days fell to a few hours, the company reported saving the equivalent of 4,500 developer-years, upgraded more than half of its production Java systems in six months, and estimated about $260 million in annualized efficiency gains. Seventy-nine percent of the auto-generated code reviews shipped without further changes.
Lesson. The Java case is one of the clearest returns on AI anywhere, because the task is well-defined, verifiable and boring. A compiler and a test suite tell you if the upgrade worked. Look for tasks like that in your own organization first.
Mercado Libre and Shopify C · aggregator
Mercado Libre uses AI search tuned to user behavior and automated assistants that, according to reporting, resolve nearly 90 percent of user interactions without a human. Shopify has embedded generative AI across its merchant platform: Shopify Magic for content, Sidekick as an in-product assistant, AI store building and search. Aggregator write-ups cite small merchants cutting listing time by 40 percent; those merchant numbers are anecdotal.
ASOS, CarMax and JATO on Azure A · primary
ASOS built a conversational virtual stylist on Azure AI Foundry. CarMax used Azure OpenAI to write customer-review summaries for 5,000 car pages in months, work Microsoft says would have taken about 11 years by hand. JATO Dynamics built Sales Link, which generates tailored content for dealerships from data across 55 car brands and reports saving customers 32 hours a month. The CarMax case is a good template: summarization of a large, fixed corpus where each output can be reviewed.
Zomato, Swiggy and Myntra B · reported
Zomato partnered with Microsoft Azure to use Azure OpenAI models inside a data-private setup, and launched Zomato AI, a food companion that reads preferences and dietary needs. Swiggy built neural search that accepts conversational queries instead of keywords, used generative AI to create food images, and is building in-house models to help restaurant partners with onboarding and payouts. Myntra, part of Flipkart, launched MyFashionGPT with Microsoft, which answers open requests like clothes for a beach holiday across categories.
Lesson. Neural search is the quiet win here. Moving from keyword match to meaning match lets a user describe a need instead of guessing the right word, and it usually lifts conversion more than a chatbot does.
Software platforms and agent infrastructure
Uber: an MCP gateway for AI agents A · primary
Problem. Teams across Uber had built hundreds of independent Model Context Protocol integrations, so discovery, ownership and security were fragmented. MCP standardizes how an AI client finds and calls tools; it does not decide which tools exist or who may use them.
Design. Uber built MCP Gateway as a central service with two halves.
- Control plane, the MCP Registry. A catalog of more than 800 MCP servers and more than 5,000 tools, the single source of truth for discovery and ownership.
- Data plane, the proxy gateway. Executes calls at runtime and translates between MCP and the protocols Uber's services already speak (HTTP, gRPC and TChannel), routing through Uber's Muttley service mesh, so backend services do not have to change.
Discovery that is safe by default. A component called AutoCrawler discovers APIs from Uber's interface-definition registry and from native MCP servers and generates tool definitions, but leaves them disabled. Owning teams must review and enable them. Discovery is automatic; exposure is explicit.
Security on the path. Policies enforce access at the level of individual tools, and personal data is redacted from responses before the model sees them. One defense decides what an agent may call; the other decides what it may read.
Fighting context bloat. Five thousand tool schemas would fill a model's context before the user typed anything. Uber added Omni MCP for incremental tool discovery, response projection (GraphQL-style field selection so only needed fields return), and Code Mode, a CLI integration that lets agents call tools without loading every schema.
Lesson. The model is not the platform. The platform is the registry, the proxy, the policies and the context budget. Any company with more than a handful of agent tools will rebuild something like this.
Netflix: GenRec, an LLM-backed ranker A · primary
Problem. Netflix's ranking stack relied on thousands of hand-engineered features and specialized architectures, which made every new business goal expensive to support.
Method. Two phases. Phase 1 adapts an open-source LLM to Netflix data so it understands members and titles. Phase 2 post-trains it for recommendation using verbalized user histories (watch history written as text), a catalog-aware scoring head, reward signals, and cost-optimized serving through prefill-only inference.
Why prefill-only matters. A ranker needs a score, not a paragraph. If one pass over the history produces the score, the expensive token-by-token decode loop disappears and the model behaves like a classifier at serving time. And because a member's history changes slowly, the stable prefix can be cached, so repeat requests only compute the new tail.
Results. Statistically significant improvement over the production baseline in large-scale A/B tests, about 1.6 percent offline lift in mean reciprocal rank, using substantially fewer Phase-2 training examples and input signals.
Lesson. Fewer hand-built inputs, comparable or better output. The work moves from feature engineering to evaluation and serving cost.
DocuSign, H&R Block, KPMG, Volvo and TomTom on Azure A · primary
Document and workflow automation fills most of Microsoft's roundup. DocuSign built Intelligent Agreement Management on Azure AI and Azure SQL Hyperscale for its 1.6 million customers. H&R Block classifies more than 30 million tax documents a year with Azure AI Foundry and Azure OpenAI. KPMG Australia built KymChat to extract data for audits. Volvo Group automated invoice and claims processing with Azure AI Document Intelligence, saving more than 10,000 manual hours, about 850 a month. TomTom cut in-car assistant response time from 12 seconds to 2.5 seconds.
Lesson. Look how many of these are document extraction. Unglamorous, measurable, and reviewable, which is exactly why they ship.
Google and DeepMind: cooling data centers A · primary
What it did. Neural networks trained on historical sensor data predicted how hot a data center would get over the next hour and recommended cooling settings so the site did not over-cool.
Result. Up to 40 percent less energy used for cooling, which Google said equals a 15 percent reduction in overall PUE overhead, and the lowest PUE that site had seen.
Lesson. This is one of the oldest and best-documented industrial AI wins, and its shape repeats in steel and energy below: predict a physical system a short time ahead, then act within safety limits set by engineers.
Steel
Tata Steel: 260+ algorithms and a remote operations center C · aggregator
What it built. According to AI Expert Network's case study, Tata Steel's digital program started in 2018 under CIO Jayanta Banerjee. It built more than 260 AI algorithms for real-time decisions, digital twins of production facilities, predictive maintenance, defect detection, blast-furnace optimization and raw-material procurement models, and an Integrated Remote Operations Center (iROC) in Jamshedpur for centralized plant control. It ran an analytics academy to train staff. Tata Steel plants are members of the World Economic Forum's Global Lighthouse Network.
Numbers claimed. About $1.4 billion in savings, first-time-right production above 90 percent, and a 1:10 cost-to-value ratio. These come from an aggregator without a cited filing; verify before quoting.
Technical heart. The program focused on what the company calls YET-QP: yield, energy, throughput, quality and productivity. A digital twin is a simulation of a furnace or mill fed with live sensor data, so engineers can test a setting virtually before changing the real plant.
Tata Steel Nederland with Siemens: Senseye at IJmuiden A · primary
Problem. A shortage of skilled maintenance staff and the high cost of downtime in the Coldmill Annealing Bay at IJmuiden.
Solution. Siemens took over monitoring of roughly 300 IT/OT systems with 24/7 remote operations, patching and vulnerability management, and deployed Senseye Predictive Maintenance, which analyzes shop-floor data with AI and includes a generative AI copilot that guides maintenance staff.
Outcomes. A 67 percent reduction in equipment breakdowns, several main and utility shutdowns prevented, an early-detected filter clog that would have released harmful fumes, support for NIS 2 cybersecurity compliance, and scaling to other factories within six months.
Lesson. Notice that cybersecurity and maintenance were sold together. Once plant equipment is connected so that AI can read it, the same connection is an attack path. Operational technology security is part of the AI project.
Tata Steel Shotton: bearing-wear study A · primary, peer reviewed
Study. Chen, van Hillegersberg, Topan, Smith and Roberts compared three models for predicting bearing wear on a Tata Steel production line at Shotton: partial least squares regression (PLSR), an artificial neural network, and a random forest.
Result. The simplest model, PLSR, worked best on the data available. The authors stress that success needed careful understanding of the industrial process, thorough cleaning of the data and a user-friendly tool for the shop floor.
Lesson. This is the most honest case in the guide. With small, noisy industrial data, a linear method beat a neural network. Start simple, and earn the complex model.
ArcelorMittal, Nucor, SSAB and Outokumpu C · vendor blog
A maintenance-software vendor's roundup lists several steelmakers. ArcelorMittal: a 30 percent drop in unplanned maintenance events at pilot sites and 7 to 21 days of warning before failures. Nucor: plant-level AI for electric-arc-furnace electrodes and ladle furnaces, with 20 percent lower electrode consumption. SSAB: AI tied to its HYBRIT hydrogen steel project, with 45 percent fewer emergency cooling repairs. Outokumpu: ANYmal inspection robots covering about 1,890 inspection points a week each and cutting hazardous exposure. It also credits Tata Steel IJmuiden with 25 percent lower blast-furnace maintenance cost. The page is vendor marketing and does not cite sources; it is listed for breadth, not proof. (It also does not cover JSW, SAIL or Emirates Steel, despite the link title.)
Manufacturing and food
Sachsenmilch with Siemens: predictive maintenance in a dairy A · primary
Setting. Sachsenmilch Leppersdorf processes about 4.7 million liters of milk a day, roughly 170 truckloads, and runs 24/7 with near-100 percent availability required.
Solution. New vibration sensors, the Siplus CMS 1200 condition-monitoring system, and Senseye Predictive Maintenance, with SAP Plant Maintenance integration planned. Siemens ran the pilot and then trained the plant team to operate it themselves.
Outcome. The system caught a pump near the end of its life before it failed, saving costs the company called "low six figures". Technical manager Roland Ziepel said the pilot paid off.
Lesson. One prevented failure can pay for a pilot. The handover, training the plant's own team, is what makes it last.
Siemens Industrial Copilot B · reported
Siemens pairs Senseye's anomaly detection with a generative copilot that explains alerts and guides technicians through repairs. The design separates two jobs: a time-series model decides something is wrong, and a language model helps a human decide what to do. (The Business Insider page on this was unavailable when checked; this description comes from the Siemens reference and press release above.)
BMW: computer-vision quality inspection B · reported
Steyr uses Inspekto's S70 autonomous machine vision for engine parts, aimed at cutting pseudo-errors, where a classic vision system rejects good parts. Dingolfing compares order data with a live image of the badge on each finished car, training on about 100 images per feature of good parts, dust and oil droplets. Regensburg runs a GenAI4Q pilot with Datagon AI that gives tailored inspection recommendations for about 1,400 vehicles a day.
Lesson. False rejects are expensive too. A vision model that is "accurate" but flags good parts creates a manual re-check loop. Measure false positives separately.
Oil, gas and energy
Shell: 100+ AI applications C · aggregator
Shell runs more than 100 AI applications from upstream to retail. Seismic interpretation with AI reportedly cut an exploration analysis timeline from nine months to under nine days. Reinforcement learning optimizes drilling. Machine-learning models watch equipment for predictive maintenance. Computer vision at service stations spots hazardous behavior such as smoking near pumps, and smart charging improves EV charging. The strategic point the summary makes is standardizing platforms and data so models can be reused across businesses.
Shell India B · reported
In an interview, Mansi Madan Tripathy, chair of Shell companies in India, described AI for predictive analytics across assets, supply-chain optimization, LNG operations where "real-time sensor data and AI models" tune equipment settings at the Hazira terminal, AI-assisted seismic interpretation, and a smart charging algorithm that lowers EV charging cost. She also pointed to a December 2025 collaboration between Shell and SLB on agentic AI tools and secure data for upstream operations. (The original S&P Global interview returned an access error; details come from the republished version.)
Aramco, ExxonMobil, Chevron, TotalEnergies, BP and Eni C · aggregator
Saudi Aramco reports about 500 AI use cases and roughly $4 billion of savings in 2024, including AI seismic analysis. ExxonMobil runs proprietary high-performance computing for 4D seismic imaging and has a data-center power pipeline above 2.7 GW. Chevron is partnering with GE Vernova and Engine No. 1 on gas power plants built next to data centers, starting in West Texas. TotalEnergies runs the Pangea III supercomputer for reservoir simulation; Eni runs HPC5. BP uses AI and robotics to raise recovery from existing fields.
Lesson. Energy companies are on both sides of AI: they use it, and they are building the power plants that feed AI data centers. (The ExxonMobil Notion page in the original list was empty when checked.)
Logistics
UPS: ORION route optimization C · aggregator
What it is. ORION (On-Road Integrated Optimization and Navigation) plans the order of stops for each driver. It combines operations research, the mathematics of optimization, with machine learning, and works on about 250 million address data points.
Timeline. Package Flow Technology in 2003, ORION development in the early 2010s, first deployment in 2012, full US rollout to 55,000 drivers in 2016, and Dynamic ORION, which re-routes during the day using live traffic, deployed to 97 percent of the ORION fleet by 2021. Reported investment about $250 million.
Numbers reported. About 100 million fewer miles a year, about 10 million gallons of fuel and 100,000 metric tons of CO2 saved annually, and $300 to $400 million of annual savings. The famous left-turn rule, avoiding turns across traffic, is part of the optimization. A driver makes about 160 stops a day.
Lesson. Not all industrial AI is deep learning. This is a search problem with millions of possible plans, where a good answer in seconds beats a perfect one in hours. (The second UPS link in the list resolved to a generic case-study index with no UPS content.)
DHL and Maersk: warehouse robots B · reported
DHL deployed 100 autonomous mobile robots from Locus Robotics to replace manual cart picking, and reports 50 percent better order accuracy and 60 percent shorter cycle time. Maersk Contract Logistics used Hai Robotics case-handling robots and reports 30 percent higher storage density.
Lesson. Robotics wins in warehouses come as much from layout and workflow redesign as from the robots' intelligence.
Aviation
Delta and Airbus Skywise; Lufthansa AVIATAR B · reported
Delta uses APEX, which collects engine data during flights and predicts problems. Reporting says maintenance-related cancellations fell from about 5,600 in 2010 to 55 in 2018. Delta joined Airbus's Skywise platform to build cross-fleet predictive maintenance, which is reported to predict impending failures with over 95 percent success. Lufthansa began predictive maintenance in 2017 and now sells its AVIATAR platform to other airlines.
Lesson. Aircraft are the ideal predictive-maintenance case: rich sensors, strict maintenance records and very expensive failures. The records are what make labeling possible.
Air India: a virtual assistant A · primary
Microsoft reports that Air India's assistant handles 97 percent of queries with full automation, and includes visa and passport scanning and baggage tracking. As with Klarna, ask what happens to the remaining 3 percent, which are usually the hardest and most emotional cases.
Agriculture
John Deere: See & Spray A · primary
What it is. Cameras across a sprayer boom look at the ground as the machine drives. A vision model separates weeds from crop in real time, and individual nozzles fire only on weeds instead of blanketing the whole field.
Numbers. Deere says the system saved about 8 million gallons of herbicide mix over more than 1 million acres in 2024, an average 59 percent saving on corn, soybean and cotton. An Iowa State University study found an average 76 percent product saving and about $15.70 per acre. Arkansas research found post-emergence herbicide use could be halved.
Lesson. This is edge AI with a hard real-time budget: the decision has to be made in the few milliseconds between the camera seeing a weed and the nozzle passing over it. The model must run on the machine, not in the cloud.
Healthcare
Ambient AI scribes in hospitals B · reported
What it is. An ambient scribe listens to a doctor-patient conversation, with consent, and drafts the clinical note. The doctor reviews and signs it.
Deployments. Kaiser Permanente made Abridge's ambient documentation tool available across 40 hospitals and more than 600 medical offices. In an earlier rollout, the Permanente Medical Group reported 3,442 physicians using a scribe across 303,000 patient visits in 10 weeks, with most saving about an hour a day. The American Hospital Association reports these tools now reach more than 250 health systems.
Evidence. A multi-site JAMA study found modest effects: about 13 minutes less total EHR time and 16 minutes less documentation time over its measurement window, plus about 0.49 more visits per week. Mass General Brigham reported a 21 percent drop in burnout prevalence after 84 days.
Lesson. The controlled studies show smaller effects than the press releases. Both can be true: a modest time saving still matters if it reduces burnout. Medigold Health, in Microsoft's roundup, reports a 58 percent rise in clinician retention from automating report writing, another sign that the main value is human, not throughput.
Pharma and drug discovery
Pfizer and Chai Discovery B · reported
Deal. Pfizer gets early access to Chai-3 and a custom model trained on Pfizer's own data. Financial terms were not disclosed.
Technology. Chai-2, released in 2025, designs antibodies de novo: given a target and an epitope, it generates the complementarity-determining regions, the loops that bind. Reported antibody hit rates were about 20 percent, against well under 0.1 percent for traditional screening. Chai-3 is said to double the success rate and handle multispecific designs.
Lesson. The win is not generating molecules; it is how many of them bind in the lab. Hit rate is the honest metric.
The pharma AI deal wave B · reported
| Company | Partner and date | What it covers |
|---|---|---|
| Pfizer | PostEra (Jan 2025, up to $610M); XtalPi (Jun 2025); Boltz (Jan 2026) | AI lab and ADC payload design; physics plus AI modeling; fine-tuning Boltz models on Pfizer data |
| Eli Lilly | NVIDIA (Oct 2025, >1,000 GPUs; Jan 2026, up to $1B over 5 years); Insilico; Superluminal; Profluent (Apr 2026, up to $2.25B) | AI supercomputer and closed-loop robotic lab; small molecules; GPCR drugs; custom recombinases |
| AstraZeneca | Modella AI (acquired Jan 2026); Owkin (May 2026); CSPC | Oncology foundation models; an AI scientist for decisions; oral drug discovery |
| Novo Nordisk | NVIDIA Gefion (Jun 2025); OpenAI (Apr 2026) | Custom models on a Danish supercomputer; AI across discovery, manufacturing and operations |
| GSK | Noetik (Jan 2026, $50M upfront); Helix | Virtual-cell models for cancer; genomic data for targets |
| Bristol Myers Squibb | Anthropic (May 2026) | Claude access for 30,000+ employees |
| Merck KGaA | Google Cloud (Apr 2026, up to $1B); Valo Health; Siemens | AI infrastructure and Gemini Enterprise; target identification; connected R&D and manufacturing |
| Roche/Genentech | PathAI (May 2026, $750M upfront); Orionis | Digital pathology; molecular glues |
Lesson. Most headline values are milestone-based: the partner only gets the big number if drugs succeed. Upfront payments are the real signal of conviction.
Telecom
Airtel: network-level spam detection B · reported
How it works. The model runs inside the network, not on the phone. It scores each caller or sender using usage patterns such as call and SMS frequency and call duration, and compares them with known spam behavior, then labels suspected spam in real time.
Scale reported. About 2.4 billion calls and 1.5 billion SMS analyzed per day, roughly 100 million potential spam calls and 3 million spam SMS flagged daily, and cumulatively billions of spam calls detected. Vodafone Idea and Jio have launched similar systems.
Lesson. Running at the network gives a view no single phone has: one number calling thousands of people a minute is obvious from the operator's side.
Insurance
Allianz and Zurich: claims B · reported
Allianz launched an Insurance Copilot for motor claims in Austria in 2024 and Project Nemo, an agentic claims system, in Australia in July 2025, aimed at the flood of simple claims after natural catastrophes. Zurich built ZurichIQ, a modular generative AI platform used in claims, underwriting, legal and employee tools.
Lesson. Catastrophe events are the perfect agent use case: many similar, low-complexity claims arriving at once. The agent takes the volume so human adjusters can focus on the complex ones.
Education
Duolingo Max A · primary
Duolingo used GPT-4 for two features. Explain my Answer tells a learner why an answer was wrong, something impossible to pre-write for every possible mistake. Roleplay lets learners practice conversations with an AI character, such as ordering coffee in Paris. Duolingo met OpenAI weekly and provided data to improve the model for these tasks.
Lesson. The best generative features do something that was impossible before, not something that was merely expensive.
IT services
Infosys: Topaz A · primary
Infosys's annual 20-F filing describes its enterprise AI strategy under the Topaz brand. The filing was too large to analyze for this guide, so no figures from it are quoted here. If you read it, look at three things: how AI revenue is defined, how many client programs are counted as AI, and what the risk-factor section says about AI reducing demand for traditional services. That last section is often the most honest part of a services company's filing.
The patterns that repeat
Read across forty-odd cases and a handful of patterns cover almost all of them. Each pattern has a typical model, a typical data problem and a typical failure.
| Pattern | Examples | Typical model | Hard part |
|---|---|---|---|
| Knowledge assistant (RAG) | Morgan Stanley, Walmart, LLM Suite, KymChat | LLM + retrieval | Corpus quality and access control |
| Customer-service agent | EVA, Klarna, Air India, Mercado Libre | Intent model or LLM | Quality on the hard minority of cases |
| Document extraction | COiN, H&R Block, Volvo, DocuSign | OCR + LLM or classifier | Grounding and human review |
| Predictive maintenance | Senseye, Sachsenmilch, Delta APEX, Tata Steel | Time-series anomaly model | Labels, sensor drift, false alarms |
| Vision inspection | BMW, See & Spray, Shell stations | CNN or vision transformer | False rejects and edge latency |
| Optimization | UPS ORION, Google cooling | OR solver + prediction | Constraints and safety limits |
| Foundation model on own data | Netflix GenRec, Stripe, Chai | Large transformer, adapted | Evaluation and serving cost |
| Agent platform | Uber MCP Gateway, Allianz Nemo | LLM + tools | Permissions and context budget |
| Code transformation | Amazon Q Java | Code LLM | Verification by tests |
Verifiable tasks win first
The cleanest returns, Amazon's Java upgrades, Volvo's invoices, CarMax's summaries, Google's cooling, all share one trait: someone or something can check the output cheaply. A compiler, an accountant, a thermometer. Open-ended tasks with no checker, like general customer conversations, produce the most impressive demos and the most reversals.
Simple models earn the right to be complex
The peer-reviewed Tata Steel study found linear regression beat a neural network on real bearing data. Delta's maintenance and UPS's routing rely on decades of classical engineering. Large models win when there is a lot of clean data (Netflix, Stripe) or language involved. Elsewhere, start simple.
Control is the product in an enterprise
Uber's gateway, JPMorgan's model-agnostic platform, Morgan Stanley's vetted corpus and Siemens's bundled cybersecurity are all control layers. The model is replaceable; the permissions, logging and data boundaries are what the company actually builds and keeps.
Cost shapes the architecture
Netflix serves with prefill-only inference. Uber trims tool schemas and responses. Stripe scores in under 100 ms. See & Spray runs on the machine. In each case, the serving budget decided the design more than the benchmark did.
The human result is often the real result
Scribes reduce burnout more than they raise visit counts. Medigold reports higher clinician retention. Klarna learned that customers wanted humans for important moments. A project that only measures throughput will miss the outcome that keeps it funded.
A learning path
- Build a small RAG assistant over a corpus you trust. Index a few hundred documents, retrieve with citations, and measure whether the right passage is in the top five. This teaches the Morgan Stanley pattern.
- Build an MCP server with two tools and a policy. Allow one tool, block the other, and log every call. Then add response projection. This teaches the Uber pattern.
- Run predictive maintenance on a public bearing dataset. Compare PLSR, random forest and a small neural network, exactly as the Tata Steel study did, and set a threshold from a cost table rather than accuracy.
- Train a tiny scorer on user histories. Write histories as text, score items with one forward pass, and compare cost against generating an answer. This teaches the GenRec idea.
- Automate a verifiable code change. Pick a dependency upgrade in a real repo, let a model propose the change, and let the test suite decide. This teaches the Amazon Q lesson about checkers.
- Read one primary source per pattern. The Uber post, the Netflix paper, the Siemens reference, the Tata Steel paper, the Google cooling post. Compare each with an aggregator's summary of the same case and note what the summary lost.
Mistakes that repeat across industries
Trusting an unsourced number. Several aggregator figures in this guide have no traceable source. Quote them only with that caveat.
Measuring deflection instead of resolution. Klarna's story is the warning.
Ignoring false positives. BMW's pseudo-errors and maintenance false alarms both cost real money and trust.
Forgetting security when connecting machines. Tata Steel's Siemens project bundled cybersecurity for a reason.
Skipping the owner. Uber leaves new tools disabled until their owners approve them. Every AI system touches something someone owns.
Frequently asked questions
Which industry gets the most proven value from AI today?
On the evidence here, the strongest documented returns are in well-instrumented operations (maintenance, cooling, routing, agriculture) and in verifiable software and document tasks. Customer-facing chat has the biggest headlines and the most mixed outcomes.
Should a company build its own platform like Uber or JPMorgan?
Only once it has many teams and many tools. Before that, a managed service and a clear policy are enough. The platform pays off when fragmentation becomes the main problem.
Are LLMs replacing classical ML in industry?
In language-heavy and some ranking tasks, partly. In sensor and optimization tasks, mostly not. The two increasingly work together, as in Siemens's copilot sitting on top of an anomaly model.
How do I tell if a case-study number is real?
Look for the source behind it, the baseline it compares against and the time window. A number with all three is evidence. A number with none is marketing.
Glossary
| Term | Meaning |
|---|---|
| RAG | Retrieval-augmented generation: fetch relevant documents, then let a model answer using them. |
| MCP | Model Context Protocol, a standard for how AI clients discover and call tools. |
| Control plane / data plane | The configuration and catalog side of a system versus the live request path. |
| Digital twin | A live simulation of a physical asset, fed with real sensor data. |
| Predictive maintenance | Using sensor data to predict failures before they happen. |
| PLSR | Partial least squares regression, a linear method that handles many correlated inputs. |
| Prefill-only inference | Producing a score from one pass over the input, with no token-by-token generation. |
| Pseudo-error | A good part wrongly flagged as defective. |
| Operations research | Mathematical optimization of decisions such as routes and schedules. |
| Hit rate | In drug discovery, the share of designed molecules that actually bind in the lab. |
| PUE | Power usage effectiveness: total facility energy divided by IT energy. |
References
Grouped by source tier. All were opened for this guide unless marked otherwise.
A · primary
- Uber: Designing an MCP Gateway
- GenRec: An LLM-Backed Recommendation Ranker at Netflix (arXiv 2608.10257)
- Walmart: My Assistant expansion
- Siemens reference: Tata Steel Nederland
- Siemens press release: Sachsenmilch
- Chen et al. 2021, bearing wear prediction at Tata Steel
- Microsoft Azure customer stories
- John Deere: See & Spray savings
- Google: DeepMind cooling
- OpenAI: Duolingo
- OpenAI: Morgan Stanley
- Infosys annual report (20-F) (not analyzed, too large)
B · reported
- The Week: HDFC EVA
- Shell India interview (republished)
- Pharmaceutical Executive: Pfizer and Chai
- Pharma AI deals factbox
- AHA: health systems and ambient scribes
- Becker's: Kaiser ambient AI rollout
- GSMA: Airtel spam prevention
- Allianz: smarter claims management
- The Decoder: Amazon Q Java upgrades
- PYMNTS: Rufus usage
- Payments Journal: Stripe foundation model
- Supply Chain Dive: Delta and Skywise
- DHL and Locus Robotics
- BMW AI quality control
- Analytics India: Zomato and Azure
C · aggregator or vendor
- Emerj: AI at JPMorgan Chase
- Best Practice AI: JPMorgan COiN
- AI Expert Network: Tata Steel
- AI Expert Network: Shell
- Enki AI: AI and HPC in oil and gas
- Oxmaint: steel industry maintenance roundup
- NextBillion: UPS ORION
- Twig: Klarna automation
- AI Expert Network: Shopify
Links that could not be read
- S&P Global: Shell India interview (access denied; republished version used)
- ExxonMobil Notion page (empty when loaded)
- Business Insider NL: Siemens AI tools (server error)
- Best Practice AI: UPS (resolved to a generic index)