HDATF
English
Start a conversation
Menu

AX case studies

Explore how companies put AI into practice. Each case covers the work problem, the technology and data used, the changed process and the reported result, newest first.

Evidence reviewed 121Pending review 31Industries 23Research cutoff

AX means changing how work is done by integrating AI into operations and services.

Results 121 (30 / 121)

2026. 09

  1. 2026-09-17Source publishedAX107In operation

    Logistics / Supply chain planning United States

    Microsoft

    Cloud supply chain planning and sourcing

    The work problem
    The Microsoft cloud supply chain team handles work in which demand planning, sourcing, fulfilment and transport are tangled together. Planners once spent five to seven days just tracing why a demand plan had changed. The company wrote that early efforts helped people finish familiar tasks faster but rarely transformed outcomes. Adding agents to a broken process, it said, still leaves a broken process, because speeding up one step just creates a longer queue at the next.
    Technology and data
    Supply chain experts and engineers worked side by side, first mapping the end to end workflows and simplifying them. They then created a single source of truth so that every agent reasoned from the same data. On that foundation they deployed more than 100 purpose built agents across planning, sourcing, fulfilment and logistics. The agents investigate shifts in demand and model capacity, and they compare transport options across air, land and sea on cost, timing and carbon impact. Only within the permissions and approval thresholds defined by people do the agents help planners update or cancel purchase orders.
    Results
    Microsoft said cycle time fell by up to 75 percent in selected workflows. A footnote states that across 5 monthly planning cycles measured between April 2026 and August 2026, average cycle time declined from approximately 10 to less than 2.5 business days. It also states that across more than 20 demand plan investigations each month, the average time to produce a human validated explanation fell from five to seven days to less than a few hours, with some completed in less than 20 minutes. The same footnote says that as of September 2026 more than 111 agents had been deployed across cloud supply chain workflows, and that these figures come from Microsoft internal analysis of work led by a cross functional team of more than 150 people between September 2025 and August 2026. Microsoft wrote that this makes it possible to analyse changes as planning cycles unfold, model more scenarios, build better contingency plans and identify risks earlier, and said the results are specific to these workflows and measurement periods.
    Limits and open questions
    The source does not say which models or data sources were used, or what each deployed agent does. The figures come from Microsoft internal analysis, and the measurement scope is limited to the 5 planning cycles between April 2026 and August 2026 and the more than 20 demand plan investigations each month named in the footnote, so results outside that scope and any external verification remain unconfirmed.

    Sources

  2. 2026-09-17Source publishedAX103In operation

    Insurance / Insurance brokerage Germany

    MRH Trowe

    Permission-aware meeting minutes and employee AI operations

    The work problem
    At MRH Trowe, separate employee experiments with generative AI risked fragmenting tools and exposing sensitive client and insurance data outside governance. Basic chat was insufficient for answers grounded in internal information and multistep work. The company needed an environment where employees could use AI themselves while access and costs remained centrally governed.
    Technology and data
    The company connected LibreChat with Strands Agents and Amazon Bedrock AgentCore. Its first production agent takes an employee request, finds a Microsoft Teams meeting in the calendar, retrieves the transcript and drafts minutes with the date, participants, agenda and action items. The employee identity authenticated by Microsoft Entra ID is passed server-side so the agent can access only that employee's calendar and transcript. Agents, models and data are processed in the Frankfurt Region with session isolation.
    Results
    AWS reported that approximately 400 employees received access in the first month of production, with infrastructure and token costs of about $14 per seat per month. It described replacing manual post-call minute preparation with a short request. The company established a shared environment where employees can build and use work agents while usage and costs are tracked.
    Limits and open questions
    The adoption month, meeting-minute accuracy and working time saved were not disclosed. The roughly 40% additional infrastructure cost reduction described in the source is a potential future optimization, not an achieved result.

    Sources

  3. 2026-09-16Source publishedAX105Limited operation

    Manufacturing / Automaker Spain

    SEAT

    Manufacturability analysis for body-in-white sheet metal components

    The work problem
    SEAT's body development followed design, simulation and finally manufacturing validation. While work waited in the specialist simulation queue, geometry and tooling decisions became fixed, making late discoveries costly to revise. Designers needed to check manufacturability while geometry could still change.
    Technology and data
    SEAT connected CAD software, Altair simulation tools and its material database through Synera. When a designer uploads a component and a natural-language question, a supervisor calls specialist agents in sequence for tasks including geometry extraction and forming simulation. Established rule-based workflows and deterministic simulations perform engineering calculations. LLMs delegate work and explain simulation outputs against SEAT's engineering methodology. A reporting agent flags problem areas, while engineers decide whether to revise the CAD or rerun the analysis.
    Results
    Synera said the first benchmark demonstrated at CDFAM Barcelona completed a manufacturability assessment in about thirty minutes, compared with four weeks previously. It said live component analyses began before the April 2026 event and that adoption was expanding across the body-in-white program at publication. Designers receive results during design rather than waiting in the specialist queue, while engineers retain final judgment.
    Limits and open questions
    The four-week versus thirty-minute comparison is the first demonstration benchmark, not an average across all components. The source does not give the exact month of the first analysis, the number of components covered or a reduction in production defects.

    Sources

  4. 2026-09-15Source publishedAX104In operation

    Consumer goods / Household products United States

    Procter & Gamble

    Visual product inspection on high-speed production lines

    The work problem
    On P&G's consumer goods lines, materials stretch, wrinkle and overlap as they move at high speed. Low-contrast defects and complex packaging patterns are also difficult to inspect. Traditional rule-based vision systems required extensive reconfiguration whenever materials, packaging designs or production environments changed.
    Technology and data
    P&G uses Visual Inspection Cockpit, combining its deep learning models with Siemens Industrial Edge and industrial PCs powered by NVIDIA GPUs. Live camera images inspect products passing through the line, with results processed close to the equipment. The results connect to production control to trigger alerts or remove defective products. Plant engineers directly configure, train and update inspection models with the Visual Inspection Engineering Tool. Accumulated quality data supports analysis of process changes.
    Results
    Siemens said P&G is expanding the system across its global manufacturing operations and that scrap rates fell by 10 to 20 percent depending on the product. It said reusable Industrial Edge applications and established integration processes enable new deployments to be commissioned five to ten times faster than traditional bespoke vision systems. Production teams can continuously monitor quality while maintaining line speed.
    Limits and open questions
    The number of plants, adoption start month and measurement period were not disclosed. The source also gives no product-level comparison conditions for scrap reduction or false-positive rates.

    Sources

  5. 2026-09-08Source publishedAX106In operation

    Finance / Banking China

    China Merchants Bank

    Shared operation of accelerators for AI training and inference

    The work problem
    As China Merchants Bank extended AI across more financial use cases, it had to run large model training, fine tuning and online inference together on nearly 10,000 heterogeneous accelerator cards. Adding more cards would not solve the problem. A training job can only start useful work once all the required workers and cards are ready, so it needs stable capacity, while online inference has to scale up and down quickly with request volume. Whenever several tenants fine tuned the same base model, a separate copy of that base model was loaded for each of them.
    Technology and data
    The bank built a single control plane on Kubernetes while keeping the execution paths for training and inference separate. Kueue first decides whether a training job can enter the cluster based on queues and quotas. The Kubernetes scheduler and HAMi then assign accelerator capacity, and Fluid speeds up access to datasets and model weights. Twinkle, developed in house, runs training on Ray, and vLLM or SGLang serves online inference. On the inference side Prometheus collects request rate, queue depth and latency, and KEDA uses those signals to add or remove replicas. Platform architects decided that training and inference needed separate admission policies.
    Results
    The bank said it brought 99 percent of its accelerator compute resources under this framework. Average accelerator compute utilization rose from 35 percent to more than 60 percent, and under comparable model and service conditions the inference cost of processing one million tokens fell by more than 60 percent. With Twinkle, five tenants fine tuning with LoRA share one base model instance. Base model replicas drop from five to one, accelerator resource usage falls by 80 percent and the number of tenants that can train at the same time increases fivefold. The bank wrote that it uses five tenants by default in production and has validated eight.
    Limits and open questions
    The source does not say when the platform went into operation or over what period each figure was measured. The numbers come from the bank platform team comparing the same internal measurement method before and after the redesign.

    Sources

  6. 2026-09-02Source publishedAX102In operation

    Healthcare / Hospital Netherlands

    OLVG

    Writing clinical notes from pediatric consultations

    The work problem
    A pediatrician at OLVG hospital in Amsterdam used to write notes in the evenings and on weekends, risking forgetting parts of the conversation. Entering notes on a computer during visits could also interrupt conversation with patients. The hospital needed a way to reduce documentation work while letting clinicians focus on the conversation.
    Technology and data
    The hospital used Epic's Chart with Art for consultations in Dutch. AI documented the conversation in the background and queued up orders to help clinicians move to their next tasks. OLVG clinicians worked with Epic developers to refine its use in Dutch. The source includes a physician's account of the system capturing the essence of a consultation despite noise and an interpreter.
    Results
    Epic reported a 31% reduction in note-writing time per appointment for OLVG pediatricians. It also reported a 27% reduction in overall clinician time in the chart per appointment, from almost 17 minutes to just over 12 minutes. A hospital representative said early observations suggested somewhat lower mental and physical strain and frustration among clinicians using the tool.
    Limits and open questions
    The source does not state the adoption month, measurement period or number of participating clinicians. Final approval procedures for notes and orders and clinical accuracy figures are also not established.

    Sources

2026. 08

  1. 2026-08-27Source publishedAX111In operation

    Education / Online learning services United States

    Khan Academy

    AI tutoring for students and practice question authoring for teachers

    The work problem
    Khan Academy said early use of Khanmigo, its AI tutor, had varied and that some chats helped students move forward more than others. The organization said only around 15% of students who have access to Khanmigo engage with it, and wrote that it still has more work to do. Khanmigo could previously respond only in text, so it could not show students what a hard maths or science concept looks like. On the teacher side, teachers could use Khanmigo to generate practice questions but had no way to curate or refine what was sent to students.
    Technology and data
    Through a Google.org fellowship, six Google engineers worked alongside Khan Academy's product and engineering organization for six months starting in March 2026, and that work put Gemini models into Khanmigo. Khanmigo now detects the moment a visual could help a student and uses Gemini to build interactive material such as charts, geometric shapes and parallel coordinate plots. When a student drags a vertex or a line segment, Khanmigo recognizes where it was placed and adjusts the lesson to match. The Google engineers first defined the metrics for what makes a high quality diagram, then evaluated the visuals for relevance, helpfulness and clarity so they appear only when students benefit most, and built a feedback widget into Khanmigo. Gemini was also wired directly into the Practice My Knowledge tool for teachers. A teacher can generate multiple choice questions in seconds and upload their own material for Gemini to reference. Questions go to students only after the teacher reviews, edits or rejects them, and the teacher decides whether the practice is scored and receives reporting on the results. Khan Academy stated the principle that Gemini drafts the materials while teachers have the final say on what students see.
    Results
    In its April 2026 material Khan Academy said it was seeing an average of 269,000 interactions with Khanmigo on weekdays and more than 108 million interactions in total since the 2023 launch. The same material said only around 15% of students with access to Khanmigo engage with it. For the new features, the organization said it is seeing preliminary signs of increased student engagement and improvement in next item correctness. Next item correctness measures whether a student correctly answers the very next problem on the same skill in the same session without help from Khanmigo, and Khan Academy described it as a measure of independent learning transfer rather than performance with AI assistance. In its material of 27 August 2026 Google said the tools had moved from early pilots into real classrooms and that districts using Khanmigo can use the new features immediately. Teachers and administrators told Khan Academy the changes let them pitch question difficulty at the right level for their students.
    Limits and open questions
    The increase in student engagement and the improvement in next item correctness are described only as preliminary signs, with no measured values in the sources. The sources also do not give the number of districts or students using the new features, nor how often teachers edit or reject the generated questions.

    Sources

  2. 2026-08-19Source publishedAX113In operation

    Gaming / Online gaming platform United States

    Roblox

    Spotting early signs of child endangerment and ordering chats for review

    The work problem
    Roblox carries chat that many people share, and inside it people attempt to put children at risk. The difficulty is that such conversations are not explicit at the start. They only reach other detection systems once the wording becomes clear, and by then the situation may already have moved on. The company wrote that safety is a shared responsibility that no company can solve alone.
    Technology and data
    Roblox Sentinel is a model designed to detect early signals of potential child endangerment. It was built on contrastive learning and trained on patterns of both benign and eventually harmful conversations. New conversations are compared against those patterns so that subtle warning signs are flagged for review before they can escalate. Sentinel does not act on accounts by itself. It lets human reviewers prioritize the chats most likely to require attention. Reviewers then act on offending accounts and report problematic users to the appropriate authorities. Version 2 raised the number of score combining functions from two to six, each suited to a different type of data. By removing the need to recompute evaluation data for every setting, a sweep across 324 configurations finishes in less than three minutes. The same work used to take nearly an hour.
    Results
    The company said that for the 12 months ended August 7, 2026, nearly 70% of the cases it detected were due to Sentinel's early detection, and that it frequently caught them earlier than other methods. After the search over configurations improved, ROC-AUC, a standard measure of ranking quality, rose from 0.894 with the default settings to 0.996 with the best configuration identified. The company said versions of these models already run on Roblox, detecting attempts to solicit or share personal information, flagging early signs of child endangerment, and moderating voice chat in real time. In the same material it announced that three safety models including Sentinel, plus an evaluation dataset, were released to the ROOST Model Community.
    Limits and open questions
    The 70% figure is the share of detected cases that Sentinel caught first. The source does not say how many child endangerment cases were missed altogether. The ROC-AUC values of 0.894 and 0.996 are evaluation numbers used to pick model settings, not operating results. The source also does not give a false positive count, the size of the human review team, or the month Sentinel was first switched on.

    Sources

2026. 07

  1. 2026-07-22Source publishedAX082In operation

    Information technology / Enterprise software United States

    Salesforce

    Employee IT enquiries and handling lost devices

    The work problem
    TechForce, Salesforce's internal IT support organisation, handles IT enquiries from 75,000 employees. Reports of lost and stolen devices took more than an hour each for people to review and process. Another obstacle was that employees wrote device names in their own natural language. Some wrote phone, some iPhone, some Pixel.
    Technology and data
    Salesforce first ran TechForce Agent, built with Agentforce, internally. It started with the simplest task, answering knowledge questions. When lost and stolen reports were handed over, the wording of a conversation could not be tied precisely to the company asset database, creating the risk of locking the wrong device or sending a replacement to the wrong address. As a fix, a deterministic matching layer was built that cross-checks against serial numbers and device records, and this approach was later shipped in the product as the Agentforce Graph feature. The team uses Agentforce Observability to look at what employees asked, what the agent returned and where the conversation broke down, and uses that for the next round of improvement. The bar for widening to all employees is 70 percent accuracy.
    Results
    Salesforce said the share of enquiries the agent finished without human involvement rose to the low 30s straight away. Over 18 months of refinement, customer satisfaction for TechForce Agent rose from 76 percent to 90 percent, with a year-end target of 95 percent. It explained that lost and stolen handling, which used to take more than an hour, now finishes in about 15 minutes.
    Limits and open questions
    The source does not disclose the volume TechForce Agent handles or its cost effect. The 95 percent satisfaction is a target, not an achieved result.

    Sources

  2. 2026-07-08Source publishedAX032In operation

    Construction and real estate / Cement Switzerland

    Holcim

    Improving production processes and equipment operation

    The work problem
    When equipment in a cement plant stops suddenly, the whole production plan is thrown off. When demand forecasts miss, inventory and distribution cost rise together. Holcim said it would address both production and logistics with AI.
    Technology and data
    Plants of Tomorrow is an Industry 4.0 programme that bundles technologies from AI to robotics to modernise plants worldwide. The equipment-side tool is Holcim M-Predict. It watches equipment performance in real time with smart sensors and machine learning. The logistics and sales tool is Holcim Foresight. It pulls in thousands of external data points such as weather, GDP forecasts and market trends to forecast sales and inventory and to optimise vehicle and network utilisation. In Spain it runs Holcim+, an AI assistant that takes orders over a messenger.
    Results
    Holcim said its AI models predict equipment failure and the point at which maintenance is needed up to 90 days in advance. It explained that it applied AI at more than 100 plants. It said it is running 38 large-scale AI projects across four areas: production, logistics, commercial and administration. Holcim Foresight is in use in more than 30 markets, with a plan to extend to all markets by the end of 2026. In Spain it said 89 percent of customers order through the AI assistant.
    Limits and open questions
    Delivering a recurring EBIT effect of CHF 200 M a year from AI by 2028 is a target, not a result. Failure prediction accuracy and per-plant effects are not in this source.

    Sources

2026. 06

  1. 2026-06-25Source publishedAX025Pilot

    Manufacturing / Automaker Germany

    BMW

    Humanoid robots on the production floor

    The work problem
    Work that is monotonous, ergonomically demanding or safety critical remains on BMW Group production floors. Placing sheet metal parts for the welding process in the body shop is one such example. This work demands both speed and accuracy while also straining the body.
    Technology and data
    BMW Group, together with Figure AI, deployed the humanoid robot Figure 02 at the Spartanburg plant in the United States to position sheet metal parts for welding in the body shop. As the next step it said it will deploy the successor model Figure 03 for sequencing work in logistics. In sequencing, the robot picks parts that arrive jumbled in a large container and arranges them on a sequencing trolley. The trolley is moved to a set collection point, an automated tow tractor or a Smart Transport Robot carries it to the assembly position, and assembly staff receive the parts in order. Figure 03 has soft components for safety, wireless charging, speech-to-speech voice capability, and hands with tactile sensors and cameras in the palm.
    Results
    BMW Group said that at Spartanburg in 2025, Figure 02 supported the production of more than 30,000 BMW X3 vehicles over ten months. The company explained that this collaboration confirmed humanoid robots can safely carry out precise, repeated work under real production conditions. Figure AI founder Brett Adcock said eleven months of running Figure 02 proved that humanoids are no longer a laboratory experiment.
    Limits and open questions
    The sequencing work with Figure 03 is only now starting. The source has no throughput or quality figures for sequencing work.

    Sources

  2. 2026-06-03Source publishedAX023Pilot

    Mining / Copper Australia

    BHP

    Searching for substances for copper leaching

    The work problem
    Copper leaching is a complex process. BHP explained that finding better ways to recover copper from ore had relied on slow manual trial and error. Molecules that could improve leaching number in the millions, counting both those already known and those not yet discovered. Testing them one by one in a laboratory would take years, or even decades.
    Technology and data
    BHP used Microsoft Discovery together with Microsoft and the computational chemistry company Prescience Insilico. It is a platform that accelerates scientific research and development with high performance computing and specialised AI agents. The team swept more than half a million molecules that might extract copper more efficiently. It ran tens of thousands of quantum chemistry calculations and simulations to narrow the candidates. The models were matched to the real conditions of ore at BHP copper sites. The narrowed molecules are tested by scientists themselves in an Australian laboratory.
    Results
    BHP said it narrowed an almost limitless candidate pool down to a small set of options that could one day be used at global copper sites. Candidates are tested against its own orebodies and operating constraints. It explained that this means solving for what will actually work. Beyond speed and efficiency, it aims at lowering toxicity and reducing environmental impact alongside improving recovery and cutting cost. Vice President of Innovation Jessica Farrell said this collaboration can contribute to long-term copper growth targets.
    Limits and open questions
    It is still at the laboratory testing stage. There is no result in the source showing that copper recovery rose in commercial production. Which molecules were selected was not disclosed either.

    Sources

2026. 05

  1. 2026-05-28Source publishedAX096Adoption announced

    Manufacturing / Aerospace Multiple countries

    Airbus

    Expanding AI for technical documentation and design work

    The work problem
    Airbus wants to widen AI use across commercial aircraft, helicopters, defence and space. However, this is a field with high confidentiality that includes military work, so security and sovereignty requirements are strict. Labour-intensive work such as producing technical documentation also remains as it was.
    Technology and data
    Airbus signed a partnership agreement with Mistral AI. The agreement secures a licence to the whole Mistral AI product range, so models can be deployed where needed, whether on in-house infrastructure or a trusted cloud. The collaboration areas were organised into four branches. In industrial operations it covers automating the production of technical documentation for commercial aircraft and helicopters. In engineering and design it uses AI based simulation, such as optimising aircraft parts, to pull development cycles forward and support engineers in development, testing and certification. In next-generation products it explores automatic object recognition with edge AI loaded onto aircraft and spacecraft, and in defence it supports cyber investigation and coding assistance through on-premises deployment.
    Results
    Airbus said some of the collaboration areas it set out have already started. The company explained that the agreement gives it access to Mistral AI's key researchers and influence over the AI product roadmap. The Airbus vice president for digital said this is the stage of building the foundation that will underpin current and future products and services.
    Limits and open questions
    Which tasks have entered formal operation, and how many users and what results there are, is not in the announcement. The names of the Mistral models chosen were not disclosed either.

    Sources

2026. 04

  1. 2026-04-23Source publishedAX093Limited operation

    Manufacturing / Steel South Korea

    POSCO

    Anomaly alerts for steelmaking equipment

    The work problem
    Because of the high heat and the complex equipment layout at a steelworks, there were subtle signs of trouble that are hard to confirm by eye. Inside a reheating furnace the temperature is very high, so it was hard for operators to go close and see whether a valve was working properly. A small oil leak from a tank was also hard for an operator to notice straight away.
    Technology and data
    Field engineers in the EIC technology department at Pohang Works developed the predictive maintenance logic themselves. For the tuyeres on the No. 3 blast furnace body, they built logic that analyses electromagnetic flow meter and temperature data in real time and raises an alarm immediately if abnormal values persist for a set period. In the reheating furnace, the opening of a valve is compared with the amount of gas actually flowing and shown as a graph. If the slope of the graph differs from usual, they immediately recognise whether a part inside the valve has failed or air is short, and repair it. Oil tanks are monitored automatically around the clock with shift register technology. If the oil volume falls abnormally against the previous day, the system detects it at once and an engineer goes out for a field check.
    Results
    Pohang Works said that in May of last year, as of the announcement, it found potential failure causes early through two sensor alarms alone and carried out maintenance. In reheating furnace valve monitoring it was credited with achieving both safety and work efficiency. The EIC technology department staff who led the development said the core was turning complex field data into graphs and an alarm scheme anyone can understand. The company said it plans to keep extending the smart maintenance logic it developed in house to other equipment and spread a preventive maintenance culture.
    Limits and open questions
    The company presents this as AI based, but the logic described in the source is explained as persistence of abnormal values, comparison of gas volume against valve opening, and detection of a fall against the previous day. What role a trained model played is not confirmed.

    Sources

  2. 2026-04-22Source publishedAX095Limited operation

    Food and beverage / Coffee supply chain Multiple countries

    JDE Peet’s and Airbus

    Mapping coffee farms and forest risk from satellite imagery

    The work problem
    Attempts to stop deforestation in the coffee supply chain were scattered company by company. Looking only at an individual supply chain cannot protect the whole coffee growing region. There was also no public map that broadly captured which land is coffee farmland.
    Technology and data
    Airbus Defence and Space was selected as the technology partner of the Coffee Canopy Partnership. This collaboration is led by JDE Peet’s together with Louis Dreyfus Company, Sucden, Neumann Kaffee Gruppe, Touton, Sucafina and Tchibo, with support from the United Kingdom FCDO and the UN Food and Agriculture Organization. Airbus uses imagery from the 50cm resolution Pléiades satellites and the 30cm Pléiades Neo satellites. To this it adds AI models trained on large volumes of data to identify and observe coffee farms from space. The maps produced are used by governments, local communities and the coffee industry to find deforestation risk, support landscape restoration and protect smallholder livelihoods.
    Results
    Airbus said it applied the model in several East African countries. The area covered is more than 1.2 million square kilometres, including shade-grown areas and agroforestry zones. The company explained that early results showed precision and scalability that could set a new standard in agricultural observation.
    Limits and open questions
    Covering all of the world's coffee origins by 2027 is a goal, and it depends on further joint investment from the industry and institutions. The announcement has no figures on how much deforestation actually fell.

    Sources

  3. 2026-04-17Source publishedAX109In operation

    Retail / Jewellery retail Multiple countries

    Chow Tai Fook

    Work agents for store associates and managers

    The work problem
    Chow Tai Fook said large scale digital transformation had long been exceptionally difficult in gold and jewellery retail, because the products are highly non standardised and the service model rests on craftsmanship, customer trust and decades of frontline expertise. An extensive physical store network spanning multiple markets made it harder still to spread one way of working everywhere at once. Store associates had to switch between inventory platforms, pricing rules, real time gold prices and complex compliance manuals, and every switch risked pausing a customer's fleeting buying momentum.
    Technology and data
    Working with Microsoft, the company first standardised on Microsoft 365 E5, then used Microsoft Purview to control who may see which information and how sensitive data flows, and scaled AI on top of that. For stores it built AI Fook, a super agent ecosystem developed on Microsoft Foundry. When an associate asks a question in everyday language through Microsoft Teams, they immediately receive product craftsmanship stories, real time inventory insights and personalised styling recommendations without breaking the flow of the conversation. For managers, the AI Insights Platform answers natural language questions by combining consumer behaviour, mall dynamics and volatile gold prices in real time. Approvals run on two tracks. Routine transactions are cleared within seconds by AI, while complex cases trigger an automated risk scan and are decided by a person. In design, Azure OpenAI generates three dimensional concepts so that master artisans can concentrate on artistic judgement.
    Results
    Chow Tai Fook said it has deployed more than 400 customised AI agents, used by more than 24,000 employees. It said there are millions of AI interactions every month and that efficiency in core business processes has improved by more than 70 percent. The company said frontline associates rehearse different customer scenarios through immersive AI roleplay, and that the best selling behaviour captured by in store devices is refined into standard procedures used across the company. It said that combining this digital rehearsal with execution in the store has driven sales conversion improvements of up to 57 percent.
    Limits and open questions
    The source does not say when this system went into operation, nor over what period and against what baseline the 70 percent efficiency gain and the 57 percent conversion improvement were measured. It also does not describe which workflow each of the roughly 400 agents handles.

    Sources

  4. 2026-04-14Source publishedC036In operation

    Delivery and mobility / Ride-hailing and delivery United States

    Uber

    Improving food delivery search

    The work problem
    On Uber Eats, search is the main path to an order. Matching on characters alone breaks down on synonyms, misspellings, abbreviations, queries that mix languages, and words with several meanings. The source wrote that lexical methods see strings, not meaning. A structure was therefore needed that turns queries and documents into vectors and retrieves by meaning.
    Technology and data
    The design is a two tower structure that encodes queries and documents separately. Query embeddings are produced in real time by an online service, while document embeddings are prepared in advance by scheduled batch jobs. The backbone is a Qwen model fine tuned on internal Uber Eats data, and one model covers every vertical and market. Training uses an MRL based infoNCE loss, so a single model emits embeddings at several sizes. The index is an HNSW graph in Apache Lucene Plus holding both float32 and int8 vectors. Conditions such as area, city, document type and fulfillment type filter first, and approximate nearest neighbour search runs after that. Retraining and reindexing happen every two weeks, swapping blue and green at the column level inside one index rather than keeping separate indexes. Before deployment a run must pass a document count comparison, a byte for byte match on columns that should not change, and a recall comparison from replaying real queries. In operation, sampled requests check that the query model matches the model identifier on the index column, and the model deployment rolls back automatically if mismatches persist.
    Results
    The source said that lowering the shard level candidate count from 1,200 to about 200 cut latency by 34% and saved 17% of CPU with almost no effect on recall. int7 scalar quantization cut latency by more than 50% against fp32 while holding recall above 0.95. Cutting embeddings down to 256 dimensions kept quality loss under 0.3% for English and Spanish and reduced storage by about 50%. The company said the structure powers restaurant, grocery and retail search together, and that scheduled refreshes run without disrupting live traffic.
    Limits and open questions
    The source carries no results from an experiment that split real users. It does not report business measures such as order conversion or click changes. The figures given are offline evaluations and infrastructure performance measurements. The source also does not say when adoption began.

    Sources

2026. 03

  1. 2026-03-26In use fromAX041In operation

    Retail / Food retail France

    Carrefour

    Conversational grocery shopping

    The work problem
    Since 2023 Carrefour had run a conversational chatbot called Hopla on its own e-commerce site. It was a tool that built a basket to fit dietary constraints or a budget. However, customers could only use it once they came inside a Carrefour screen.
    Technology and data
    In 2025 Hopla+ replaced Hopla. Hopla+ looks at a customer's purchase history and suggests a basket. From the day the press release came out, customers could reach Carrefour products and services without leaving ChatGPT. Users talk with ChatGPT to get recipe ideas, check stock of products on shelves, build a basket to suit their needs and choose a delivery method. Order confirmation and payment happen on Carrefour.fr. Carrefour is found in the app list on the ChatGPT home screen. In early 2026 the company said it supports Google's Universal Commerce Protocol.
    Results
    Carrefour announced that it is the first European retailer to offer grocery shopping in ChatGPT. It said the goal of this launch is to reach the tens of millions of ChatGPT users in France. Emmanuel Grenier, who leads Group Supply Chain, E-commerce, Data and digital transformation, said it gives ChatGPT users a new way to shop and a simple, smooth purchase journey.
    Limits and open questions
    The scope stated in the source is French users and payment on Carrefour.fr. There is no statement that the same function is available in other countries. User numbers and sales results have not yet been disclosed either.

    Sources

  2. 2026-03-10Source publishedAX053In operation

    Energy / Oil and gas Saudi Arabia

    Aramco

    A portfolio of AI based operational improvements

    The work problem
    Aramco runs large facilities spanning exploration, drilling, production and refining. In a business like this it is hard to tell whether newly introduced technology actually paid off. The company decided to measure separately the value its technology created once deployed into operations, and it calls this indicator Technology Realized Value, or TRV.
    Technology and data
    TRV is calculated as capital or operating cost avoided, or as revenue increased by deployed technology, expressed as an increment against a business baseline. The indicator is verified by an external body. A representative case using AI is the Aramco Intelligent Live Earth Model. This model optimises well placement and widens reservoir contact in geosteering work. PE.AI is an AI based platform that generates synthetic well log data. The company explained that it also greatly increased its AI computing capacity to run real-time advisory systems and advanced analytics.
    Results
    Aramco disclosed that it produced 5.3 billion dollars of TRV in 2025. Of that, 2.6 billion dollars came from AI based solutions and 2.7 billion dollars from non-AI technology. It said cumulative TRV since 2023 is 11.3 billion dollars. It said the Aramco Intelligent Live Earth Model increased reservoir contact by about 10 percent and cut average drilling time per well by two days. It said AI half-precision computing capacity exceeded 570 PetaFLOPs, ten times the previous year.
    Limits and open questions
    TRV is a value the company calculates and an external body verifies, and it is different from accounting net profit. Reading the whole 5.3 billion dollars as an AI result would be wrong.

    Sources

  3. 2026-03-10Source publishedAX112In operation

    Healthcare / Health system United States

    Summit Health

    Preparing and submitting medication prior authorization requests

    The work problem
    Epic's material said prior authorization had historically been a time consuming process for both patients and clinicians. To prescribe a drug, staff must fill in the paperwork a payer asks for, send it, and wait for a reply. The work breaks the flow of care and leaves clinicians with administrative load. Epic said it built a setup in which Art, its AI for clinicians, works together with Penny, its AI for revenue cycle and operations, to cut through that complexity.
    Technology and data
    Art covers the clinician side of the work and Penny covers the revenue cycle and operations side. The two move together to gather what a prior authorization needs and to draft the responses that go on the submission. Staff review that draft and either send it as it stands or edit it first. Summit Health applied this setup to medication prior authorization submissions. Epic said Penny will go on to complete coding sessions and appeal submissions on its own, inside the guardrails each organization sets.
    Results
    Epic said that at Summit Health, Penny cut medication prior authorization submission time by 42%. It said 92% of the AI generated responses are accepted without edits. At the organizations using Penny most actively, coding related denials fell by more than 20% and denial appeal letters were created 23% faster. Epic summarized the effect as patients getting answers faster and clinicians facing less administrative friction.
    Limits and open questions
    The source does not say when Summit Health turned the feature on, how far the covered drugs and payers reach, or how many submissions were handled. The 42% and 92% figures are the ones Epic presented. The drop of more than 20% in coding denials and the 23% faster appeal letters are pooled across the organizations using Penny most actively, so they cannot be read as results for Summit Health alone.

    Sources

  4. 2026-03-10Source publishedAX110In operation

    Healthcare / Health system United States

    The Christ Hospital

    Finding and following up lung nodules buried in radiology reports

    The work problem
    Epic's material said lung cancer is the most common cause of cancer related deaths worldwide and that in 2020 only 16% of lung cancers in the United States were detected at a localized stage. Lung nodules often show up by chance on scans ordered for something else, such as trauma or cardiac symptoms, and radiologists record these incidental findings in the free text of their reports. Because the finding sits inside a long narrative, it is hard to tell which patients need to enter a monitoring program, and follow up was inconsistent. A study cited in the same material found that 36% of patients with stage IV lung cancer had nodules referenced in their radiology reports without receiving appropriate follow up. Joy Oh, chief information and digital transformation officer at The Christ Hospital, said the earlier approach depended on radiologists entering particular wording in particular fields, and that findings were missed because each radiologist writes differently.
    Technology and data
    After a radiologist finalizes a study, the impression and narrative are sent through Epic's AI extracted findings model. The model pulls mentions of lung nodules and lung masses out of the report and returns them in a daily report. That report carries the surrounding context of the finding, the radiologist's follow up recommendation such as additional scans, tissue sampling or a referral, and the due date the radiologist recorded. Three nurse navigators read the extracted findings and recommendations, check that they match the radiologist's impression, and then open the patient's chart to judge whether the recommendation is appropriate. Once a navigator verifies a finding, it becomes discrete data in the chart. Navigators use that data to see current and potential enrollment in the lung nodule program, unreviewed results, upcoming appointments and their own outstanding tasks on a single dashboard. Patients can be seen in the lung nodule clinic without a referral, so most are seen within three days of appearing on the daily report. Nurse practitioner Ashley Campbell can order additional diagnostic tests or hand the patient to one of ten pulmonologists who see patients on a rotating basis. When the team first turned the feature on, they ran the previous 90 days of reports through the model and surfaced more than 2,000 nodules that had not been documented as discrete data. Later the team found that nodules on scans ordered by oncologists or pulmonologists already had follow up in place, so they stopped sending those reports through the model and concentrated staff on scans ordered by primary care and the emergency department, where patients were more likely to fall through the cracks.
    Results
    The Christ Hospital said the number of lung nodules it tracks increased sixfold and that it diagnosed 23 lung cancers in the first six months of the program, matching its 2023 total in half the time. It said 64 additional patients began cancer treatment in the first eight months. Ashley Campbell said the feature has helped the team diagnose close to 70% of lung cancers at stage I or II, far above the national average of around 50%. Marcus Romanello, the chief medical officer, said the work is affecting mortality. In its March 2026 material Epic put the hospital's early detection rate for lung cancer at 69% against a national average of 46%. In the summary of the same material Epic said navigators followed up on almost 5,000 additional lung nodules, which led to 116 additional cancers being detected across the hospital's network.
    Limits and open questions
    The sources do not say how accurately the model extracts nodule mentions or how many extractions were wrong. They also do not give the month the program started. The early detection figures differ between sources: the June 2025 material gives close to 70% against a national average of around 50%, while the March 2026 material gives 69% against 46%.

    Sources

  5. 2026-03-09Source publishedAX055In operation

    Finance / Corporate banking Singapore

    DBS

    Corporate customer support

    The work problem
    DBS runs a channel that takes enquiries from corporate and small and medium enterprise customers. The bank set itself the goal of becoming a bank that uses AI at scale. Doing so required a common standard so that several teams could build and expand generative AI applications responsibly.
    Technology and data
    The foundation is ADA, the bank-wide data analytics platform. This platform provides a data usage environment with controls and security. On top of it the bank laid a generative AI framework. Reusable components, control mechanisms, workflow capabilities and prompt writing standards were set as the standard. On this basis it built the corporate banking chatbot DBS Joy. DBS Joy responds around the clock to corporate and small and medium enterprise customers.
    Results
    DBS said DBS Joy provides round-the-clock support to corporate and small and medium enterprise customers. In the same letter the bank said it applied more than 2,000 AI models to more than 430 use cases through ADA. It wrote that the economic value created by data analytics and AI activity in 2025 was about 1 billion Singapore dollars. That amount is the total of the bank's whole activity, not the result of one chatbot.
    Limits and open questions
    The source describes DBS Joy in a single sentence. Answer accuracy, volume handled and the share passed to human agents are absent.

    Sources

  6. 2026-03-09Source publishedAX056In operation

    Finance / Banking Singapore

    DBS

    Adverse news review

    The work problem
    DBS runs a screening operation that looks at adverse news relating to customers. The bank picked this work out as a target for raising productivity with generative AI. Before that, it needed a standard to tidy up development practices that differed from team to team.
    Technology and data
    The bank first built a generative AI framework. Reusable components, control mechanisms, workflow capabilities and prompt writing standards were set as the standard. On this basis it built an adverse news screening tool with generative AI. The data is handled by ADA, the bank-wide data analytics platform. The bank said this platform cut code deployment time by 25 percent and reduced the model deployment cycle to under 10 weeks.
    Results
    DBS said the tool raises productivity in adverse news screening. It is listed as a generative AI case in operation in the CIO letter covering 2025 results. The same letter said more than 2,000 AI models were applied to more than 430 use cases through ADA. Performance figures isolating the screening work were not given.
    Limits and open questions
    The source mentions this tool in a single sentence. Screening accuracy, handling time and the share reviewed again by people are absent.

    Sources

2026. 02

  1. 2026-02-26Source publishedAX061In operation

    Telecommunications / Fixed and mobile telecom Germany

    Deutsche Telekom

    Customer enquiries and installation booking

    The work problem
    Telecom call centres receive a large volume of calls that land on agents. Enquiries with no existing script, or enquiries worded unclearly and ungrammatically, were hard to handle with automated responses.
    Technology and data
    Deutsche Telekom added generative AI to its Frag Magenta chatbot. It now handles enquiries with no set script, and enquiries that are unclearly worded or ungrammatical. It is used on both the voice and the chat channel. On the agent side the company is rolling out AI tools bundling automatic pre-call briefing, an agent chatbot called askT, and call summaries. In Germany a voice bot also takes fibre installation bookings.
    Results
    Deutsche Telekom said that in 2025 Frag Magenta's voice and chat handled 3.4 million calls instead of agents. That is 900,000 more than the cumulative figure through the third quarter. It wrote that in Germany the customer chatbot moved fully to an LLM basis, recording a 55 percent resolution rate, and service contacts fell by 10 percent. In the fourth quarter a voice bot took 40 percent of fibre installation bookings. As of December more than 1,400 agents use AI tools.
    Limits and open questions
    The 3.4 million is the number of calls that did not go to an agent, not a figure indicating resolution quality or customer satisfaction.

    Sources

  2. 2026-02-26Source publishedAX045In operation

    Travel and hospitality / Hotels United Kingdom

    IHG

    Hotel revenue and room demand management

    The work problem
    IHG took the view that it has to raise performance in three places: guest experience, guest acquisition and hotel performance. Hotel performance depends on how accurately room demand is read and priced. This report does not separately describe how revenue management worked before adoption.
    Technology and data
    Data science, machine learning and forecasting tools were put into the revenue management system. The system produces advanced analytics results and recommendations. IHG gave this as one example of AI use across the company. As another example in the same vein it gave improvements to the IHG One Rewards mobile app. The report does not disclose the system name, the kinds of data connected, or how hotel operators take up the recommendations.
    Results
    IHG said this technology deepens brand loyalty and sharpens competitive advantage. Under the hotel owners section it wrote that it widened the scope of technology applied in order to capture demand, raise revenue and optimise operations. Under the guests section it wrote that AI based technology raised the guest experience. How much the revenue management system contributed to revenue per room was not separately disclosed.
    Limits and open questions
    There are no figures isolating the AI contribution. The date of adoption and the number of hotels covered are also not in this document.

    Sources

  3. 2026-02-19Source publishedAX002In operation

    Manufacturing / Semiconductors South Korea

    SK hynix

    Virtual process metrology and anomaly cause analysis

    The work problem
    Semiconductor processes go through metrology that measures the result to confirm the process ran correctly. When the metrology rate is low, it was hard to tell apart performance differences between tools that ran the same process. The overlay prediction model in the photolithography process also concentrated only on raising accuracy, so it was hard to explain how each input variable affected the result.
    Technology and data
    SK hynix uses Panoptes VM, an AI based virtual metrology solution, together with the industrial AI startup Gauss Labs. It has been in use on the production floor since 2022. Virtual metrology is a way of predicting process results that were not directly measured. Using operating data from Panoptes VM, the two companies built a tool-to-tool performance matching framework and jointly propose a way to detect tool anomalies early and diagnose the cause quickly. On the overlay side they added GRACE, a gradient descent based framework. GRACE quantifies the effect of input variables on the predicted result and provides indicators that explain model performance. SK hynix has applied GRACE on the production floor since the year before this announcement, using it to spot process anomalies in time and find the cause.
    Results
    The two companies are presenting two papers at SPIE Advanced Lithography + Patterning 2026, one in an oral presentation and one in a poster session. The papers carry joint research results based on data from operating Panoptes VM on the actual SK hynix production floor. They said the data demonstrates that performance differences between tools, which were hard to analyse because of the low metrology rate, can be identified effectively. Gauss Labs Chief Executive Officer Kim Young-han said the result came from continuing to find use cases on the manufacturing floor even after the solution was adopted.
    Limits and open questions
    The source has no figures on yield or metrology cost improvement. Indicators showing the scope and effect of GRACE on the floor were also not disclosed.

    Sources

  4. 2026-02-17Source publishedAX059In operation

    Pharmaceuticals and biotech / Pharmaceuticals France

    Sanofi

    Supporting management and research decisions

    The work problem
    At Sanofi, data in research and development, manufacturing and supply, and commercial was divided by department. Decisions that cut across departments, such as clinical trial design or portfolio judgement, needed material gathered in one place.
    Technology and data
    Sanofi introduced the internal app Plai together with Aily Labs GmbH. Plai gathers internal data from every department, makes it queryable in real time and shows the company's activity as a whole on one screen. It uses AI to produce timely insights and personalised what-if development scenarios that support judgement. Clinical operations teams use Plai during trial recruitment to check whether participant composition is sufficiently diverse. The portfolio strategy team uses Plai's forecasts and recommendations when deciding which therapies to focus on. Manufacturing and supply teams also use Plai for inventory planning.
    Results
    Sanofi said on its official page that 20,000 employees use plai every day. In its 2025 Form 20-F it wrote that more than 22,000 employees use Plai for everyday decisions. The company explained that in manufacturing and supply, probability based planning predicts 80 percent of inventory disruptions and links 65 percent of risks to a root cause.
    Limits and open questions
    The user numbers in the two sources differ, at 20,000 and 22,000. The official page carries no publication date, so which point in time it refers to cannot be confirmed.

    Sources

  5. 2026-02-12Source publishedAX097In operation

    Energy / Power and gas utility United States

    Duke Energy

    Detecting impersonation adverts aimed at customers

    The work problem
    Scams aimed at Duke Energy customers continued. The company said it received thousands of scam reports across its service territory in 2025. Reports rose especially in February, when customers looking for bill assistance cluster. Scammers approached customers with fake adverts and impersonated support numbers.
    Technology and data
    Duke Energy uses artificial intelligence to sweep websites, social networks and paid adverts. The purpose is to find social engineering threats visible in public material. Fake adverts and impersonated support numbers are the main targets of detection. Items found are reported to the relevant platform with a takedown request. The company also runs its own scam reporting tool so customers can flag suspicious cases directly. It advises customers to slow down if a payment demand is urgent or threatening, and to verify only through the official number or website printed on their bill.
    Results
    As a recent takedown example the company cited a paid search advert that impersonated Duke Energy and carried a fake customer service number. It also said it took part in the tenth Utility Scam Awareness Day in November together with Piedmont Natural Gas, the Better Business Bureau and other utility companies. Jessica Bishop, Senior Vice President of Customer Service Operations, said that as scammers grow more sophisticated, the company is protecting customers proactively. The company also stated that it does not accept prepaid cards or cryptocurrency payments and does not threaten immediate disconnection.
    Limits and open questions
    The model used for detection, the false positive rate and the number of takedown requests are not in the source. Nor does it say at which stage people review and approve.

    Sources

  6. 2026-02-12Source publishedAX049In operation

    Agriculture, forestry and fisheries / Forest management Finland

    Stora Enso

    Forest management and biodiversity prediction

    The work problem
    Forest management has to balance timber production with biodiversity conservation. It was hard to know in advance what effect a given management action would have on rare species. A basis for deciding which conservation measures to take first among many was also needed.
    Technology and data
    Stora Enso turned its entire Nordic forest into a digital twin and uses it for data analysis. With this digital forest twin it simulates forest management actions and predicts the effect on rare species. A joint project with the International Union for Conservation of Nature (IUCN) that began in early 2024 concluded in 2025. In that project it ran a proof of concept attaching the STAR biodiversity metric developed by IUCN to Stora Enso's forest planning system. This framework decides which conservation measures to take first across the forest landscape.
    Results
    In its 2025 annual report Stora Enso said it moved towards better integration of biodiversity and timber production using AI and other technologies. The company explained that this framework addresses the most urgent threats first and in the most effective way, while minimising negative impact on timber production. The results were prepared with journal publication as the goal and presented at the 2025 IUCN World Conservation Congress.
    Limits and open questions
    Linking the STAR metric is the result of a proof of concept stage. The source has no figures such as prediction accuracy or area covered.

    Sources

  7. 2026-02-03Source publishedAX092Limited operation

    Manufacturing / Steel South Korea

    POSCO

    Controlling the iron ore sintering process

    The work problem
    Sintering is the key process that turns powdered iron ore into lumps suitable for charging into a blast furnace. Operating the sintering machine was work that site operators had to handle as simple repetition. Pohang Works set itself the task of going beyond automating individual equipment and weaving the control technologies into one.
    Technology and data
    Pohang Works put an AI based smart control system into the sintering process. Using sensors, it finely controls the charge line where raw material builds up. It automatically controls trolley speed so that the exhaust gas temperature follows the set point. The technology was developed jointly by the ironmaking department at Pohang Works and the Process DX Research Institute at the technical research laboratories. Kim Young-hyun, the deputy general manager in the ironmaking department who led the development, said he wanted to turn the know-how of Pohang Works' skilled operators into logic with AI and create a standard by which anyone can produce the best quality.
    Results
    Pohang Works said this improvement recorded an operating rate of 99 percent and a hit rate of 97 percent, sharply reducing variation in operation. It explained that a stable supply of high-quality sinter also contributed to stable furnace conditions in the downstream blast furnace. The model verified at No. 3 sintering is now being extended to other processes such as No. 2 and No. 4 sintering. It also said that automating sintering machine operation created an environment where site operators move out of simple repetitive work and concentrate on higher value work.
    Limits and open questions
    The source does not disclose the definition of the hit rate or its basis for comparison. It also does not give the operating rate before the improvement, so the size of the rise cannot be known.

    Sources

Download all data (JSON)