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 259Pending review 24Industries 23Research cutoff

AX means changing how work is done by integrating AI into operations and services. Compiled from public sources. These are not results from ATF Works customers.

Results 259 (30 / 259)

2026. 11

  1. 2026-11-10Source publishedAX243In operation

    Retail / Food retail United States

    DeCicco & Sons: Detecting shelf out-of-stocks and prioritizing replenishment work

    The work problem
    As DeCicco & Sons expanded its stores, it needed to maintain the availability of products customers wanted and its standard of personal service. The retailer wanted an accurate picture of actual store conditions and shelf information that multiple business systems could share. The aim was to let employees focus on resolving identified issues instead of walking around searching for them.
    Technology and data
    Simbe's autonomous shelf-scanning robot Tally travels through stores several times a day to capture shelf conditions. Product information, aisle images and product locations are connected through APIs to existing systems for store operations, online product availability and supplier decisions. Rather than adding another standalone tool, the information enters existing workflows. Employees use it to prioritize corrective work and carry out actions such as replenishment.
    Results
    Simbe reported a 35% reduction in out-of-stocks and up to a 28% reduction in the average time products remained out of stock during the first 60 days at pilot locations. On that basis, DeCicco & Sons expanded the platform across its New York and Connecticut stores. Executives and a store manager at the retailer said faster visibility into actual shelf conditions helped teams correct issues sooner and spend more time serving shoppers and supporting department managers.
    Limits and open questions
    The availability and time improvements are results from the first 60 days at pilot stores, not chainwide averages after expansion. The source does not disclose revenue growth or a payback period.

    Sources

  2. 2026-11-10Source publishedAX239In operation

    Manufacturing / Automaker United States

    Ford Motor Company: Natural-language queries of transport operations data to support supply-chain decisions

    The work problem
    Ford's North American transport organization must make decisions as supply-chain conditions change quickly. Producing reports separately or waiting for analysis slows the delivery of information to the people who need it. Terms such as on-time arrival also depend on customer contracts and transport modes, making a general language model's answers insufficient for judging operating figures.
    Technology and data
    Mo, used by Ford, is an AI analyst that interprets natural-language questions using project44's logistics data structure and customer-specific business definitions. project44 says it connects transport contracts, modes, actual milestones and estimated arrival times as business context. Snowflake Cortex Analyst translates the interpreted questions into SQL to retrieve figures from stored structured data. Answers and recommended next steps go to transport staff; the source does not say AI independently approves Ford's actual transport actions.
    Results
    Doug Cantriel, Ford's Head of North America Transportation, said Mo puts accurate, data-grounded answers directly in employees' hands without exchanging reports or waiting for analysis. This is a change in information access confirmed by a Ford representative. project44 also provided aggregate customer usage and separate internal testing figures, but did not identify them as Ford-specific savings or accuracy results.
    Limits and open questions
    Ford-specific cost savings and response-speed improvements were not disclosed. Other customers' savings and the supplier's internal test results in the source are not Ford's outcomes.

    Sources

  3. 2026-11-10Source publishedAX240In operation

    Chemicals and materials / Basic chemicals Nigeria

    Indorama Eleme Fertilizer & Chemicals Limited: Analyzing operating anomalies and supporting optimization decisions at ammonia and olefins plants

    The work problem
    Indorama's ammonia and olefins plants in Nigeria must operate safely while monitoring energy consumption, throughput and equipment availability together. Engineers need analysis of key operating parameters to identify production constraints and emerging anomalies, and must understand the technical basis for any action.
    Technology and data
    KBR said it deployed KBR INSITE at these process units. The platform combines historical and real-time plant data with physics-informed AI, process-design expertise and site operating knowledge. Integrated with Applied Computing's Orbital AI, it continuously monitors asset performance and identifies optimization opportunities, process and equipment anomalies, and root causes. KBR says the ammonia model was designed to learn cause and effect across the whole plant rather than only individual data patterns. Site engineers make operating decisions using the technical explanations and traceable recommendations.
    Results
    Indorama Chief Technical Officer Deepu Sivdas said near-real-time analysis of key operating parameters identifies tangible opportunities to optimize energy consumption, increase throughput and detect operating constraints. He said combining physics-based models with data analytics explains the reasoning behind conclusions, giving engineers confidence to act on recommendations, and that practical benefits are already visible in day-to-day operations.
    Limits and open questions
    No numerical energy savings, throughput gains or availability improvements were disclosed. EBIC's separate contract in the same release targets operation by year-end and is not part of this operating case.

    Sources

  4. 2026-11-06Source publishedAX241In operation

    Retail / Food retail United Kingdom

    Waitrose: Optimizing grocery delivery plans using traffic, vehicle capacity and time-slot demand

    The work problem
    Waitrose needed to use its vehicles efficiently while delivering large volumes of groceries at customers' chosen times. Traffic, vehicle capacity, delivery windows, order demand and seasonal peaks interact, creating many daily planning decisions. Maintaining customer choice of delivery slots as volumes grew was a key challenge.
    Technology and data
    Satalia worked with Waitrose's operations teams to develop, test and refine an AI-powered delivery optimization platform around real operating conditions. It supports delivery-network planning and management by considering traffic, vehicle capacity, time windows and demand together. The teams simulated peak Christmas demand at the Coulsdon hub with 130 vehicles. The supplier said that at publication the platform operated across 170 stores, two Customer Fulfilment Centres and 1,400 vans, supporting approximately 160,000 home deliveries per week.
    Results
    Satalia reported a 10% reduction in operating costs, an 8% reduction in delivery mileage and a 12% increase in deliveries per route. It said customer delivery-slot availability increased by 4% despite rising volumes, while today-for-tomorrow slots at Coulsdon increased by 50%. Waitrose's Head of Change Delivery and Technology said the system enabled more precise and convenient delivery times while reducing environmental impact.
    Limits and open questions
    The source does not disclose the measurement period or a comparison group. Coulsdon's 50% improvement concerns next-day delivery slots at that site and must not be applied to the whole network.

    Sources

  5. 2026-11-05Source publishedAX225In operation

    Chemicals and materials / Basic chemicals Austria

    Borouge International: AI real-time optimization and operating-decision support for ethane crackers

    The work problem
    Borouge International needed to improve plant availability and productivity while reducing energy use and emissions and maintaining safety. As it expanded AI and digitalization from production to the supply chain, the company was building capabilities to analyze operating information quickly and support decisions. Real-time optimization of ethane crackers in the UAE is one workflow with confirmed deployment.
    Technology and data
    The company reports an already deployed AI-enabled real-time optimization system across three large ethane crackers and 20 furnaces in the UAE. It analyzes more than 2,500 operating parameters per minute to support production and energy decisions. The approach connects optimization expertise, AI and real-time operational intelligence. Expansion to North America and other regions remains planned. The Honeywell autonomous-operations proof of concept in the same release is a separate project; this record does not assume the optimization system controls equipment without human approval.
    Results
    The company says the deployed system supports faster data-driven decisions, higher productivity, lower energy consumption and reduced emissions. It does not disclose a measured improvement rate or savings attributable specifically to this system. The release's value-generation target for 2027 and potential gains from the autonomous-operations trial are not presented as achieved results of the operating optimization system.
    Limits and open questions
    Model architecture, individual input parameters and operator approval procedures were not established. International expansion and a fully autonomous control room are not included in the completed operating scope.

    Sources

  6. 2026-11-05Source publishedAX224In operation

    Construction and real estate / General contracting Japan

    Kajima: Autonomous construction-site rounds and collection of photos and point clouds for progress management

    The work problem
    Kajima needed to address construction labor shortages and reduce engineers' repetitive rounds, visual documentation and three-dimensional surveying. Structures and material layouts change as construction advances, making controls customized for fixed environments difficult to reuse. The company needed robots to collect routine site data while engineers concentrated on safety, quality and schedule judgments.
    Technology and data
    The company mounts a FieldAI autonomy unit containing cameras, LiDAR and an edge AI device on robots. Its FFM model processes images and point clouds on the device, combining physics-based reasoning and uncertainty assessment to choose travel directions. A robot explores an unfamiliar area and creates a map, then avoids obstacles while collecting photos and point clouds and updating the map. It returns to its charging station and uploads data for web access. Staff can specify a location and view live video, while people retain technical judgment and safety management.
    Results
    Kajima reports deployment at more than 10 Japanese construction sites in roughly a year following an enterprise service agreement in 2025, and more than 2,000 autonomous missions across robot types, tasks and sites. It describes continuous use on active sites with daily environmental changes, rather than restricted demonstrations. Automatically captured photos and point clouds make construction status and dimensions available for progress management.
    Limits and open questions
    The source gives no site-level staffing, time-saving or accident-reduction rates. It identifies the year of the agreement and expanded deployment but not its starting month, so publication date is used for ordering.

    Sources

  7. 2026-11-05Source publishedAX237In operation

    Travel and hospitality / Integrated travel services Multiple countries

    TUI: Threat modeling from digital service architecture diagrams against internal security standards

    The work problem
    Four TUI security engineers supported more than 100 digital product teams. Reviewing diagrams, identifying threats, mapping them to internal standards and documenting an assessment took 4 to 6 hours. Capacity of only 3 to 4 sessions a month created a backlog and made consistent assessment of every new service difficult.
    Technology and data
    Engineers upload diagrams and context. A LangGraph workflow running on AgentCore uses Claude Sonnet 4.5 through Amazon Bedrock to read components and data flows. Instead of loading every security standard, it selects relevant standards through document metadata, generates STRIDE threats, checks gaps and refines the results. Each result includes mitigations using internally approved tools and mapped controls. When an expert assessment is supplied, rating differences are highlighted for human review.
    Results
    TUI reports internal pre- and post-deployment measurements showing up to 70% less work per assessment and monthly capacity rising from 3 to 4 sessions to 10 to 12. As of October 2026, it reports 255 registered users, 86 threat models and 43 active creators. Teams can request assessments themselves and review evidence and recommendations rather than remaining in the previous review queue.
    Limits and open questions
    Time reductions and adoption are TUI-reported operational metrics. The source provides neither reduction in actual security incidents nor threat-detection false-positive rates. Automatic Jira ticket creation and integration with pre-deployment checks remain planned.

    Sources

  8. 2026-11-05Source publishedAX221In operation

    Information technology / Insurance data analytics Not classified

    Verisk: Connecting requirements, implementation, review and testing in insurance software delivery with stage-specific AI agents

    The work problem
    After adopting coding tools, Verisk engineers still handled requirement clarification, task decomposition, test planning and code review through manual handoffs. The story co-written by AWS and Verisk describes repeated clarification and manual work caused by incomplete acceptance criteria and context lost between stages. Software used in insurance also needs a record of who reviewed and approved a change, not just faster delivery.
    Technology and data
    Verisk built separate Requirements, Grooming, Developer, Code Review and QA agents on Amazon Bedrock and AgentCore. Inputs include Jira tickets and attachments, repository code, API specifications, internal documents and artifacts from earlier stages. MCP connects existing enterprise tools, and Jira state changes or pull requests trigger individual agents. The Developer Agent uses Claude Code on Bedrock. Product owners approve requirements, developers approve implementation and review decisions, and QA engineers approve test plans. Validated artifacts accumulate as context for the next stage.
    Results
    AWS and Verisk report that, for features developed through the complete workflow, test planning fell from 3 to 4 hours to under 10 minutes, and automation setup fell from 1 to 2 days to under 30 minutes. More than 90% of agent outputs were accepted by human reviewers with minimal modification. Cost per story point, a measure of work size, was 20% to 30% lower when labor and AWS compute were counted together. Weekly releases are an engineering-team target, not an achieved result.
    Limits and open questions
    The number of measured features, measurement period and composition of the comparison group were not disclosed. Equivalent effects across all Verisk engineering teams have not been established.

    Sources

  9. 2026-11-04Source publishedAX218In operation

    Manufacturing / Industrial equipment Not classified

    BACA Systems: AI answers to questions across fragmented business screens

    The work problem
    At BACA Systems, staff had to move between screens and systems and search manually to answer business questions. Andrew Russo, the company's VP of Business Systems, cited questions that previously would have taken 45 to 60 minutes of this work. The company began using Agentforce Coworker to make information easier to find in an environment containing 1,300 custom objects.
    Technology and data
    Employees ask Agentforce Coworker questions and review its responses. Salesforce explains that the tool uses existing Salesforce records, business metadata and connected data to coordinate relevant agents and workflows and return results and recommended actions for review. It operates within existing employee permissions, sharing rules and configured access controls. The actual BACA Systems workflow confirmed in the source is information retrieval through questions and answers; the new integrations announced in the same article are not all assumed to be deployed at this company.
    Results
    Salesforce said BACA Systems enabled Coworker in 20 minutes and began receiving answers to questions that previously required extensive manual searching. Russo said the tool answered questions that just the day before would have required 45 to 60 minutes of moving between screens. That figure describes the previous search workload, not a directly measured average handling time after adoption.
    Limits and open questions
    The source does not disclose response time after adoption, answer accuracy or the number of employee users. It also does not identify the specific questions or individual connected data systems.

    Sources

  10. 2026-11-04Source publishedAX230In operation

    Finance / Banking Australia

    Bankwest: Automating basic digital chat inquiries, handing off to colleagues and assisting response drafting

    The work problem
    As Bankwest became a digital bank, it needed both fast online responses and human explanations for complex financial questions. Asynchronous customers may return later, so colleagues must retain context across long conversations. They also need to explain financial terminology, including mortgage concepts, in language customers understand.
    Technology and data
    Bankwest uses Bee, powered by Microsoft technology, to answer basic questions such as password resets and locking stolen cards in natural language. When it cannot resolve an issue or a human is needed, Dynamics 365 Contact Centre passes conversation context and a summary to a colleague. The system summarizes long threads, and Copilot drafts plain-language responses that colleagues personalize. They adjust their tone using a live sentiment indicator, while bank teams review prior interactions to improve Bee's prompts and replies.
    Results
    Bankwest colleagues featured by Microsoft say contextual summaries and draft responses support faster, more consistent explanations. A home lending specialist says simplifying financial terminology has improved conversation quality. The article says digital chat represents a majority of the bank's customer interactions and describes positive customer and colleague feedback. These observations are not quantified effects attributable solely to AI.
    Limits and open questions
    No before-and-after measures of average handling time, automated resolution rate or sentiment accuracy are disclosed. The digital chat share cannot be treated as an outcome of Bee alone.

    Sources

  11. 2026-11-04Source publishedAX219In operation

    Finance / Banking United States

    Fulton Bank: Operating an AI coworker for employee information retrieval and work preparation

    The work problem
    Fulton Bank applied Agentforce Coworker to help employees find information, prepare for work and move tasks forward. The source does not quantify bottlenecks or detail the previous process. This is therefore classified as an operational application supporting employee access to information and work preparation, rather than automation of a specific banking review or transaction.
    Technology and data
    Employees receive business information and results through Coworker. Salesforce explains that the tool uses existing customer relationships, records, business metadata and connected data to coordinate relevant agents and workflows. Employees review results and recommended actions, and processing operates within existing permissions, sharing rules and access controls. The source presents this common product architecture alongside Fulton Bank's actual deployment scope, but does not identify the bank's individual data sources or approval procedures.
    Results
    Salesforce said Fulton Bank deployed more than 20 production use cases within weeks, supporting approximately 3,000 users. Brandon Bernard, the bank's Senior Vice President and Sales Technology Strategist, described substantial daily usage and real examples of employees finding information faster, preparing more effectively and moving work forward. The announcement describes deployment scope and user experience, not measured cost savings or outcomes in financial processing.
    Limits and open questions
    The source does not publish the list of workflows, daily active user count or time saved. The approximately 3,000 supported users should not be interpreted as daily users.

    Sources

  12. 2026-11-04Source publishedAX228In operation

    Information technology / Automotive software India

    KPIT Technologies: Monitoring public competitor developments and preparing monthly strategy reports

    The work problem
    KPIT's marketing organization handles market intelligence, investor communications and proposal support as well as campaigns. Gathering and synthesizing industry developments and market signals from multiple sources consumed time. The marketing leader interviewed by Microsoft said preserving enterprise security and controls mattered because the team also handles sensitive customer intellectual property.
    Technology and data
    KPIT built a competitive intelligence workflow with Microsoft 365 Copilot and Power Automate. Microsoft says it automatically monitors strategic developments and public announcements from 14 competitors and produces monthly reports with recommendations and threat assessments. Copilot draws on work context in the existing Outlook, Teams, SharePoint and PowerPoint environment. Marketers use the collected information and reports for business decisions and strategy discussions.
    Results
    KPIT's marketing leader says the team no longer has to begin research deliverables with a blank document and can find a starting point faster while improving quality. Competitor dashboards and automated reporting workflows are presented as actual changes in daily work. The qualitative result is a broader role for marketers with less specialized expertise, who can use the material to contribute a point of view.
    Limits and open questions
    No numerical measures of missed competitor developments, report accuracy, time saved or revenue impact are disclosed. The final report approval process is not described in detail.

    Sources

  13. 2026-11-04Source publishedAX229In operation

    Information technology / Automotive software India

    KPIT Technologies: Evaluating account opportunities and preparing sales discussions with automotive manufacturers

    The work problem
    KPIT's account teams need to connect external market information with internal customer knowledge before talking with automotive manufacturers. Time spent collecting material limits attention available for proactive proposals. The marketing leader featured by Microsoft describes the need to find competitor developments, technology trends and customer priorities across multiple sources.
    Technology and data
    KPIT uses a purpose-built Copilot agent to combine public evidence and internal knowledge into an account revenue readiness report. It ranks winnable opportunities, separates fact from inference and lays out account plans for 30, 60 and 90 days. Copilot also helps find patterns in a repository of more than 1,500 proposals produced each year. Account teams use this material to prepare discussions and proposals for automotive manufacturers.
    Results
    A KPIT marketer interviewed by Microsoft says linking external automotive manufacturer intelligence with the firm's institutional knowledge allows more time for proactive propositions rather than information gathering. The marketing leader says proposal analysis surfaces competitive patterns, emerging technology trends, customer priorities and future skill requirements. Actual use is described, but this does not establish a measured sales outcome.
    Limits and open questions
    The accuracy of opportunity rankings, actual contract conversion and a before-and-after comparison of report preparation time are not disclosed. No capability to independently finalize prices or contracts is established.

    Sources

  14. 2026-11-04Source publishedAX235In operation

    Telecommunications / Fixed and mobile telecom Malaysia

    U Mobile: Cross-system investigation of subscriber issues and preparation of corrective actions

    The work problem
    To resolve failed data add-on purchases or roaming issues, staff had to separately check interface logs, business system APIs containing eligibility information, and internal policy documents. If switching between systems failed to reveal the cause, escalation to backend IT prolonged the subscriber's wait.
    Technology and data
    U Mobile built AiDaH with AWS and Axrail and made it available through Microsoft Teams. Specialist agents on Amazon Bedrock AgentCore query subscriber, order and billing systems through MCP tools and attach supporting knowledge retrieval. Agents divide questions spanning multiple domains. An action that changes a live subscriber account executes only after a support agent reviews and confirms it. Rules handle simple requests, while models handle ambiguous ones, rather than using generative AI for every lookup.
    Results
    U Mobile says AiDaH is part of its L2 support standard procedures and covers more than 150 documented operational scenarios. Its internal measurements show monthly provisioning and network tickets falling 84.5%, general inquiries 81.6%, and billing and charging tickets 48.7% between January and September 2026. Instead of gathering findings separately, staff can review a consolidated diagnosis and proposed actions.
    Limits and open questions
    The article provides no control group or isolation of other operational changes behind the ticket reductions. End-to-end issue recovery without human involvement remains a future goal.

    Sources

  15. 2026-11-02Source publishedAX244In operation

    Insurance / Specialty insurance United States

    Ascot Group: Identifying first-notice-of-loss emails and creating insurance claim records

    The work problem
    Ascot's U.S. claims organization used a lean team to process first notices of loss in a dedicated mailbox. Documents for existing claims, status questions and other correspondence arrived at the same address. Staff had to distinguish genuine new loss notices, find the policy valid for the loss date and create a claim, which delayed intake.
    Technology and data
    Bevaya's FNOL AI agent identifies new loss notices in the mailbox and applies rules for each line of business. It queries policy systems to connect the right policy and date, then creates the claim in Guidewire. Ascot's claims organization worked with the supplier to define classification rules and system connections, and kept reviewing and improving results. Human review and traceable audit records remain part of the process; expert judgments such as coverage interpretation and payment decisions are not all delegated to AI.
    Results
    Bevaya's customer-story summary reports an 85% no-touch rate, about 925 claims automated per month, and a reduction in claim-opening time from 2 days to 20 minutes. Ascot's U.S. Chief Claims Officer explained that faster intake and assignment to the appropriate adjuster let customers begin claim handling sooner. Classification and summarization of supporting documents were described in the interview as the next expansion step.
    Limits and open questions
    The summary's success-rate figure differs from the interview remarks, so no success rate was adopted. The 85% figure is the basis in the customer-story summary and was not combined with a higher current rate mentioned during the presentation. The source does not explicitly state the year of the initial operating month, so publication date was used for ordering.

    Sources

2026. 10

  1. 2026-10-30Source publishedAX208In operation

    Construction and real estate / Homebuilding United States

    Lennar: Showing data quality and freshness evidence with internal data answers

    The work problem
    Lennar's divisional systems and definitions produced conflicting reports. Employees using its LenAI assistant still had to ask the data team whether answers were trustworthy.
    Technology and data
    LenAI answers natural-language questions with a trust score combining Ataccama data quality, freshness and pipeline success. It also checks source grounding and additional validation. Employees decide whether the evidence supports a business decision.
    Results
    In Ataccama's customer story, Lennar said visible evidence reduced hesitation and increased adoption. Users can judge data reliability without a separate conversation with the data team.
    Limits and open questions
    Measured answer accuracy and time savings are absent. Self-remediation agents remain under development.

    Sources

  2. 2026-10-30Source publishedAX216In operation

    Logistics / Third-party logistics United States

    Worldwide Express: Resolving freight invoice rate discrepancies and accessorial-charge exceptions

    The work problem
    Worldwide Express staff checked carrier rules, researched causes, contacted carriers and entered resolutions into its transportation management system whenever invoices did not match orders, delivery records or contracted rates. The AWS case says unresolved billing exceptions tied up $15 million in cash flow each month. Carrier-specific fee rules and dispute procedures made fixed-rule automation insufficient.
    Technology and data
    Working with AWS and LTM, the company connected Amazon Bedrock inference and agents running on Amazon EC2 to its SpeedShip transportation management system. Separate agents match invoices, examine contracted-rate differences and validate extra charges against delivery evidence. They read contract rates, shipment details and proof of delivery through existing APIs and retrieve each carrier's tariff rules and procedures. Staff initially approved recommendations individually in the existing interface. The source says bulk approval became the default after accuracy exceeded 99%. Rejections and human overrides feed improvements to prompts and knowledge resources.
    Results
    According to the AWS case, six months into operation the solution covered more than 40 carriers and agents resolved 10,000 exceptions per week with an error rate below 1%. The company reported time-to-cash improving by four weeks and annual cost savings exceeding $1 million. The source describes outstanding exceptions falling from $15 million to $4 million and $10 million in cash flow being released. Staff shifted from routine resolution to complex, higher-value disputes.
    Limits and open questions
    The source does not disclose comparison samples or how errors were measured. The narrative's $10 million cash-flow release and an executive's statement about accelerating an additional $10 million in customer billings each week use different bases and were not combined.

    Sources

  3. 2026-10-29Source publishedAX223Pilot

    Manufacturing / Automaker Germany

    BMW: Improving R&D incident-ticket quality and identifying historical incidents with the same root cause

    The work problem
    BMW's R&D incident workflow needed better incomplete ticket titles and descriptions, and retrieval of past incidents with the same root cause to avoid duplicate investigations. Internal German and English data carries proprietary categories and identifiers. General models given detailed instructions could invent identifiers or use the wrong output language. AWS and BMW selected this workflow to validate model customization and evaluation methods.
    Technology and data
    During a 12-week collaboration, the teams used ticket edit histories and root-cause analysis outcomes. Bedrock models generated training examples, while a separate judge filtered them for information preservation, structure and language compliance. Task-specific QLoRA adapters were trained on a small base model in SageMaker and served through one real-time endpoint with adapter selection. Tasks rewrite ticket titles and descriptions, ask for missing information and judge whether a new incident shares a root cause with given historical candidates. BMW controls use-case definitions, evaluation criteria, curated data and adapters.
    Results
    AWS and BMW report that, on held-out evaluation data, title quality rose from 3.74 to 4.71 on a five-point scale, and description quality rose from 2.84 to 3.87. The model also improved its ability to recognize when no candidate shared the root cause. These are experiments using one training configuration and fixed candidate lists, not measured field time savings or operating results for an end-to-end retrieval service.
    Limits and open questions
    The source explicitly describes single-seed, single-hyperparameter evaluations. End-to-end retrieval, continuous production deployment and actual engineer time savings were not established.

    Sources

  4. 2026-10-27Source publishedAX202In operation

    Telecommunications / Fixed and mobile telecom Australia

    Telstra: Reviewing billing and payment discrepancies between legacy and new platforms during customer service migrations

    The work problem
    Microsoft's October 27, 2026 article says Telstra needed to reconcile billing and payments across legacy and new platforms during customer migrations. The challenge was to identify inconsistencies and corrective actions while maintaining operational trust and governance.
    Technology and data
    Telstra partnered with Microsoft Frontier Company to deploy an AI-powered Migration Payment Agent. It uses information across the legacy and new systems to proactively identify inconsistencies, recommend corrections and create enriched cases for human review. AI analysis is combined with human validation; a recommendation does not itself establish automatic payment or billing changes.
    Results
    Microsoft reported support for more than 272,000 service migrations and recommendation accuracy of at least 95% in production. It described improvements to complex operations and customer outcomes alongside reusable AI capabilities. The migration count and recommendation accuracy are separate measures.
    Limits and open questions
    The accuracy denominator, measurement period, cost savings and reduction in manual work are undisclosed. The specific system integration for acting after human approval was not established.

    Sources

  5. 2026-10-26Source publishedAX211In operation

    Logistics / Third-party logistics United States

    C.H. Robinson: Calculating ocean and air freight price quotes and sending them to customers

    The work problem
    C.H. Robinson's ocean freight quoting involved manually checking origin and destination ports, terminal fees and inland routing together. Terminal fees change continually, and shippers sometimes leave the port unspecified, requiring staff in different time zones to gather external information. The company said an air quote previously took a day and an ocean quote two to three days.
    Technology and data
    The company built multiple AI agents and subagents to interpret shipper requests, obtain external information and decide routing. Agents check terminal websites and fee-notice emails. When a port is unspecified, they use pickup locations and company data to suggest ports, inland transport modes and consolidation centers. The company says AI augments experts while retaining human oversight and customer-specific attention.
    Results
    In its October 26, 2026 release, C.H. Robinson said AI had calculated and delivered over 80,000 quotes for more than 3,600 customers and cut ocean and air quote times to two to three minutes. The capability started with ocean less-than-container-load quotes for U.S. and Canadian customers and expanded to full containers, air cargo and markets including Mexico and Europe. The company says earlier information helps customers respond to shipping changes and disruptions.
    Limits and open questions
    The source does not disclose quote accuracy or the rate of staff corrections. Chinese customers are in a pilot, and expansion to some other regions remained a future plan at publication.

    Sources

  6. 2026-10-26Source publishedAX213In operation

    Logistics / Third-party logistics United States

    C.H. Robinson: Classifying U.S. Customs documents, extracting information and requesting missing documents

    The work problem
    C.H. Robinson's U.S. Customs team manually processed different kinds of documents from shippers, carriers and government agencies. New trade agreements, additional tariffs and stricter compliance rules made international customs clearance more complex and labor-intensive. The company described manual handling of over four dozen types of documents.
    Technology and data
    New AI agents and automations split incoming documents, classify them and attach them to individual shipments. An AI agent extracts required information, while the system identifies and requests missing documents. Automated invoices and information packets are sent at the end. This reduces repetitive document work while customs brokers continue handling critical validation and compliance steps.
    Results
    In its October 26, 2026 release, C.H. Robinson said time spent had so far fallen by 23% with this customs support. The company said brokers could focus on critical validation and compliance work. This figure concerns customs support, not processing-time results for ocean and air price quotes or ocean bookings.
    Limits and open questions
    The source does not disclose the measurement period or comparator for the 23% reduction, or document extraction accuracy. It does not mean the entire customs process finishes without people.

    Sources

  7. 2026-10-23Source publishedAX203Limited operation

    Food and beverage / Processed food Switzerland

    Nestlé: Answering consumer questions about product ingredients and use

    The work problem
    Nestlé needed fast, consistent answers across brands and languages. According to the Salesforce customer story, more than half of consumer service cases arrived through digital channels such as email, chat and WhatsApp. As product and ingredient information kept changing, the challenge was to scale service without building a separate AI for each brand or adding manual work.
    Technology and data
    Nestlé embedded an Agentforce consumer service agent in Service Cloud. Through a connection that does not duplicate data, Data 360 reads product ingredient information in Snowflake and also supplies brand website content and service knowledge. If the agent cannot answer confidently or a consumer asks for a person, it hands over the conversation history to a service representative. In the same workspace, the representative receives conversation summaries, relevant material and suggested answers to continue the interaction.
    Results
    Salesforce said the agent autonomously resolved 40% of consumer inquiries in Nestlé pilots and earned a satisfaction rating of 4.8 out of 5 for agent-led interactions. It said the first agent launched in eight weeks and expansion into each country also takes eight weeks from preparation through technical deployment. These are pilot results, not a resolution rate for all consumer inquiries worldwide.
    Limits and open questions
    The source does not disclose pilot sample size, measurement period or country-level results. Ingredient-answer accuracy and company-wide service cost savings are also absent.

    Sources

  8. 2026-10-22Source publishedAX197Limited operation

    Insurance / Insurance brokerage Hong Kong

    Sun Flower Insurance Brokers Limited: Comparing medical insurance plans and generating advisory recommendation reports

    The work problem
    For every customer enquiry, Sun Flower advisors had to search insurer documents and compare premiums, benefits, exclusions, eligibility and waiting periods. More insurers meant repeating the same checks, while consistently explaining options suited to each customer's needs was also difficult.
    Technology and data
    AccessOrange organized plan information in SharePoint and defined required customer inputs and recommendation rules. The internal Insurance Plan Advisor, built with Microsoft Copilot Studio, filters plans against customer information and generates a branded PDF comparison report. Advisors review the AI recommendations before advising customers. Delivery included user acceptance testing, bug fixes, accuracy checks and staff training.
    Results
    According to AccessOrange's case study, customer staff and overseas colleagues participated in testing and early usage. The customer confirmed that the plan comparison process for each enquiry could be at least half faster than manual work. This was customer feedback from testing and early usage, not an average processing time measured across company-wide operations or a result from automated underwriting.
    Limits and open questions
    User counts, measurement samples and comparison accuracy figures were not disclosed. Adding historical claims and medical case data remains a future plan.

    Sources

  9. 2026-10-22Source publishedAX205In operation

    Manufacturing / Elevators and escalators Germany

    TK Elevator: Pre-visit elevator maintenance guidance and post-job voice records

    The work problem
    TK Elevator technicians had to ask where an elevator was and locate handwritten service records and lengthy technical manuals on site. Sometimes they discovered missing tools late and returned to the vehicle. After the job, they wrote notes in an uncomfortable vehicle workspace. Sensor readings alone could not bring together diagnostic knowledge from manuals and previous technicians' experience.
    Technology and data
    The company built an agent app using Microsoft Foundry and Azure OpenAI on connected-equipment data, service history and technical documentation on Azure. On a phone or tablet, it briefs technicians before a visit about the elevator's location, issue and necessary tools. Afterward, a voice agent records the technician's account and asks follow-up questions for missing details. Technicians inspect the equipment and decide on action using their own judgment, while the new records become knowledge for later maintenance.
    Results
    Microsoft's on-site reporting describes actual use: a technician searches manuals and previous service history in the field and records work without typing. TK Elevator representatives explain that field feedback supports more precise troubleshooting and breakdown prevention. The story gives company scale of 25,000 technicians and 1.4 million maintained units worldwide, but does not state an adoption rate showing that all use the app.
    Limits and open questions
    App user numbers, adoption start month and numerical improvements in service time or breakdown rate are not disclosed. This is not autonomous elevator repair or replacement of the field technician's judgment.

    Sources

  10. 2026-10-21Source publishedAX210In operation

    Information technology / Technology consulting India

    Birlasoft: Extracting contractual obligations and tracking fulfillment during project delivery

    The work problem
    In the Birlasoft handover process described by Microsoft, delivery managers received contracts prepared by legal, finance, taxation and sales teams, read them and checked their meaning with stakeholders. Obligations were scattered across master agreements, statements of work, schedules and service commitments. Sequential manual reading risked missing connections between distant clauses or overlooking a commitment.
    Technology and data
    Birlasoft built Contract.ai to extract obligations and link them to their original sources. Azure Data Factory ingests documents, Data Lake stores them and Databricks supports analytics and reporting. A planning agent coordinates specialist agents responsible for extraction, output validation and inconsistency flags. Staff track obligations that are due, overdue or complete. Human oversight remains because the obligations have legal and financial implications.
    Results
    Microsoft's customer story says Contract.ai processes more than 1,000 statements of work annually, identifies an average of more than 10 obligations per project and delivers realized annual savings of approximately 4,000 person-hours. A Birlasoft programme manager said teams can focus on delivery rather than contract interpretation, with real-time visibility improving risk management and fulfillment of customer commitments.
    Limits and open questions
    The source does not disclose extraction accuracy, the omission rate or the comparison conditions used to calculate time savings.

    Sources

  11. 2026-10-21Source publishedAX198Pilot

    Manufacturing / Electronics Germany

    Siemens: Preparing PCB optical inspection with a digital twin and synthetic images

    The work problem
    When introducing a new electronics model or variant, optical inspection programs traditionally could only be prepared and trained after a physical prototype board became available. In high-mix production with frequent process changes, waiting for physical samples and manually preparing inspection burdened production changeovers.
    Technology and data
    Siemens uses its Electronics Factory Erlangen in Germany as the demonstration site. Process Preparation X validates and normalizes PCB designs, bills of materials and component information from Teamcenter into a digital twin of the assembled board. NVIDIA Cosmos generates synthetic images of a reference board and defective boards to calibrate inspection equipment before physical production. For production, the described workflow includes AI-supported secondary inspection to analyze whether a potential defect flagged by optical inspection is a real defect or a false call.
    Results
    Siemens said it was demonstrating this inspection preparation workflow using a real factory production environment and domain expertise. The aim is to reduce dependence on physical boards and manual work. The announced reductions of up to 85% in manual AOI programming effort and 40% in false calls, and more than 5% improvement in first-pass yield, are project targets, not achieved measurements.
    Limits and open questions
    Achieved processing time, false-call rates and yield figures from the demonstration were not disclosed. Final approval authority for inspection decisions and the full scope of commercial deployment were not established.

    Sources

  12. 2026-10-20Source publishedAX195In operation

    Information technology / Cloud managed services South Korea

    Bespin Global: Running 511 AI agents around the clock across the company after analyzing internal work

    The work problem
    According to a press release Bespin Global issued on October 20, 2026, the company, which builds AI services and solutions, follows a Self-Proof principle of proposing to customers only solutions it has first applied to and verified in its own work. The company devoted the last chapter of its 2026 industry AX customer casebook, published the same day, to the results of its own company-wide AX transformation.
    Technology and data
    The company analyzed 33,924 tasks on a monthly basis, selected more than 900 automation projects, and designed and verified 751 agents. Of these, 511 AI agents now run around the clock across the company. The press release said the cases in the casebook use the company's own AI platforms HelpNow AI Platform and HelpNow AI Foundry and its data integration solution HelpNow DataOps, applying RAG, agentic AI and multi-agent orchestration on multiple clouds such as AWS, Google Cloud and Naver Cloud.
    Results
    According to the press release, Bespin Global saved 36,800 hours a year and reduced human error by 99% with these agents.
    Limits and open questions
    The press release does not say which tasks the agents handle or how the time savings and human error were measured.

    Sources

  13. 2026-10-20Source publishedAX226In operation

    Consumer goods / Ceramics France

    Bleu Clémentine: Writing product descriptions, newsletters and social-media copy for a ceramics business

    The work problem
    Morgane, a ceramicist in Bordeaux, France, runs Bleu Clémentine and must write product introductions and customer updates alongside making tableware and home decor. In a small business, promotional writing competes with production for the owner's time. The purpose of this Google customer example is to reduce repetitive marketing-copy work and return time to the core craft, not to sell writing itself.
    Technology and data
    Morgane uses Gemini to write product descriptions, newsletters and attention-grabbing social-media hooks. Generative AI helps prepare promotional text needed by the ceramics business, and the saved time goes back into making products. The source does not describe specific input templates, review procedures or connected external apps. Newly announced branding-image and finance tools elsewhere in the article are therefore not attributed to this business.
    Results
    Google reports that this use saves Morgane two to three hours each day, which she reinvests in making tableware and home decor. The reported outcome concerns the owner's working time. The article does not provide increases in sales volume, revenue or advertising conversion rates.
    Limits and open questions
    The measurement period, calculation method and final-copy review process were not established. Existing writing use is kept separate from results of newly announced Gemini features.

    Sources

  14. 2026-10-20Source publishedAX227In operation

    Food and beverage / Restaurant and retail butcher shop United States

    Burger & Butcher: Analyzing local-search performance and supporting store updates and new-menu planning

    The work problem
    Burger & Butcher in New Jersey, USA, operates a restaurant and retail butcher shop at one location without a dedicated marketing team. Between operational duties, owner Josh needs to understand how local customers find the store, prepare neighborhood-event updates and plan new offerings. Running both the store and its marketing creates a need for help translating search information into copy and menu ideas.
    Technology and data
    Josh uses Gemini with a connected Google Business Profile. He tracks local search and Maps performance, prepares weekend updates tied to neighborhood events and analyzes regional flavor trends. These insights inform a new breakfast program and take-home butcher-counter offerings. The source describes an owner using the tools between shifts; it does not say AI automatically publishes account updates or makes menu-sales decisions.
    Results
    Google presents this as actual Gemini use by a single-location business without dedicated marketing staff for customer-discovery insights, local copy and product planning. The established workflow change is that the owner connects search performance and regional preferences to work between shifts. No numerical gains in visits, revenue or working time are disclosed.
    Limits and open questions
    The analysis period, accuracy and pre-publication review procedures were not established. Newly announced financial-analysis and image-generation features are not added to this store's adoption scope.

    Sources

  15. 2026-10-20Source publishedAX215In operation

    Consumer goods / Cosmetics France

    L’Oréal: Advisor reply drafts, intent checks and promotional codes

    The work problem
    L’Oréal advisors spent time matching brand voices, checking intent and navigating screens for promotional codes.
    Technology and data
    Agentforce in Service Cloud summarizes conversations and generates replies and codes. Data 360 connects case history, customer records and knowledge with SAP S/4HANA orders. Advisors review, edit and send drafts. AI checks intent before closure; brand instructions guide tone.
    Results
    Salesforce reports 64% faster handling without draft edits and about 40% with edits. Over 300 advisors use it daily across 25 North American brands. Europe remains planned.
    Limits and open questions
    64% is conditional, not an all-case average.

    Sources

Download all data (JSON)