Cockpit
    November 4, 2026 · 18 min

    AI Operating Model Study: What It Describes and How It Changes Work, Leadership and Organization

    AI Operating ModelOrganizationLeadershipWays of WorkingCase Studies

    What is an AI operating model?

    I’ve been asking myself that question for about a year. I’m still curious!

    I talk to people and try things myself. I haven’t seen every case firsthand, though. It’s hard to tell what’s real and what’s hype. Some of the claims on LinkedIn and Substack are wild. Which ones hold up?

    I asked people I know and trust. Their organizations are at different stages. Some have barely started using AI. So I researched the question in depth.

    This study brings together my experience with ways of working, leadership and organizational design, my conversations with others, and what I found in the research.

    I used AI to find data, reports and primary sources. I chose the examples, decided how to structure the study and drew my own conclusions.

    The examples

    I chose seven real cases from different areas of business:

    • Software development

    • Production planning and procurement

    • Customer service and operations

    • Case routing and administration

    • Supply chain and purchasing

    • B2B sales

    • Finance and healthcare administration

    I chose cases documented in annual reports, research studies or official records.

    The structure

    • What outcome did the organization want?

    • What does the end-to-end process look like?

    • How did the process work without AI?

    • How does it work with AI today?

    What I draw from the cases

    • What are the defining features of an AI operating model?

    • What do they mean in practice for ways of working, leadership and organization?

    This study helps me explain, with examples and evidence, what using AI means for an organization. All the sources are freely available online. I want to write pieces that readers can trust and use to find their bearings. They also show how I think and how I work.

    Enjoy the read!

    Seven cases from practice

    Spotting supplier risks early (BMW Group)

    What outcome does BMW want?

    BMW wants to hear about possible human rights violations and environmental harm early enough for its purchasing team to act. The team can then ask a supplier to remedy the problem and follow up to see whether it has taken the agreed action.

    What does the end-to-end process look like?

    A complaint, a site inspection or a media report can flag a problem. BMW checks which supplier is involved and whether the evidence supports the report. If purchasing confirms a breach, it agrees on corrective action with the supplier and monitors implementation along the supply chain.

    How did the process work without AI?

    BMW uses standard questionnaires, on-site audits and formal complaints channels. All three remain in place. To monitor the media without AI, staff had to read reports and painstakingly match the companies mentioned to BMW’s direct and indirect suppliers. They could not achieve comprehensive global coverage that way.

    How does the process work with AI?

    A web crawler continuously collects publicly available news and social media posts. The system searches sources in more than 50 languages for signals across more than 140 risk categories. Reports about upstream suppliers beyond the first tier (Tier N) can also trigger alerts. An alert flags a possible breach. The purchasing team alone determines whether a violation has occurred and decides and negotiates what happens next.

    Qualifying sales prospects in the self-pay market (Stanford report)

    What outcome does the sales team want?

    An aesthetic medicine sales team wants to identify practices worth visiting in person. It needs to answer two questions. Which practices in its territory offer the treatments in question? Which have enough patients and room to grow to justify a visit?

    What does the end-to-end process look like?

    The team first defines its sales territory. It brings together scattered information to build practice profiles and rank prospects. This part of the process ends when the field sales team receives the list. Whether a visit leads to a qualified sales conversation or a sale falls outside the process examined here.

    How did the process work without AI?

    In regulated markets where insurers reimburse treatment, claims data provide clear indicators of activity. In this entirely self-pay market, there were no central registries or claims data. Sales representatives had to visit practices on the off chance of finding a prospect, search websites by hand or follow up on referrals.

    How does the process work with AI?

    An AI system combines public information from practice websites, treatment listings, reviews and social media. It estimates treatment volumes and ranks the practices. Field sales representatives receive a list of prospects in priority order. The available data do not show how reliably that ranking predicts future revenue.

    Routing patent applications (European Patent Office)

    What outcome does the patent office want?

    A new patent application needs to reach examiners with expertise in the relevant technical field. Only then can the substantive search for prior art begin. Getting the initial assignment right affects both the time the examination takes and its quality.

    What does the end-to-end process look like?

    When an application arrives, it receives a preliminary technical classification. A classification code routes the electronic file to the responsible technical directorate. If the receiving department finds that the application falls outside its field, it consults colleagues and arranges a transfer. A specialist classification unit settles disputes.

    How did the process work without AI?

    An experienced member of staff would read through the title, abstract and patent claims and assign a preliminary classification by hand, following the official examination guidelines.

    How does the process work with AI?

    A trained system analyzes the full text of the application. It can suggest a classification or assign the file automatically. Examiners use the existing procedure to reroute files that reach the wrong department. The patent examiners retain legal and technical responsibility for examining the application.

    Routing passenger reports on board trains (Deutsche Bahn / Railmate)

    What outcome does Deutsche Bahn want?

    When passengers report broken air conditioning or a dirty seat, the service is most valuable if the crew can fix the problem during the journey. Reports of serious faults need to reach the train’s home maintenance depot without delay.

    What does the end-to-end process look like?

    Passengers can report a problem using a QR code at their seat, the ICE portal or the app. The system analyzes and categorizes the free-text message, then sends it to the train manager’s work phone or directly to the maintenance systems. The full process ends when someone fixes the problem on the train.

    How did the process work without AI?

    Passengers either went looking for a crew member or sent a written message to the central customer service team. Staff read those messages manually and replied days later. By then, the reply was of no help on the journey in question.

    How does the process work with AI?

    Railmate uses natural language processing to sort reports into more than 200 categories. Deutsche Bahn receives around 3.2 million pieces of feedback a year. Each month, about 3,000 urgent reports go straight to the crew’s mobile devices. These figures document the routing of reports. They do not show how many problems the crew resolved during the journey.

    Comparing technical bids (Audi / Tender Toucan)

    What outcome does Audi want?

    Before Audi buys complex production equipment, its planners need to check suppliers’ bids against the technical specifications. A reliable comparison shows, line by line, which requirements each supplier commits to meeting, where its bid deviates and what it leaves unanswered. The planners can then focus negotiations on those points.

    What does the end-to-end process look like?

    The team starts with the specifications and the suppliers’ bids. Planners first verify the list of requirements, then compare each bid against it. The process examined here ends when the technical specialists sign off on the comparison table. The commercial decision to award the contract is a separate step.

    How did the process work without AI?

    Planners searched the documents by hand. Specifications could run to 1,000 pages, with bids of more than 100 pages each, often from as many as 15 suppliers. A team could spend several weeks just comparing text.

    How does the process work with AI?

    The planning team uses Tender Toucan to generate a structured checklist from the specifications. The specialists first check and approve that list. Only then does the AI search the bids for supporting passages and display the cross-references. The specialists check each assessment against the original document. Audi reports time savings of up to 30 percent in the analysis phase.

    Developing a product with coding agents (Anthropic / Claude Cowork)

    What outcome did Anthropic want?

    Anthropic wanted people who had never used a terminal to be able to give Claude access to local folders, for example to turn a collection of notes into a report. The deliverable was a working preview of a desktop app.

    What did the end-to-end process look like?

    The starting point was an observation. People were using Claude Code, an agent built for developers, for tasks beyond its original purpose. The product team designed the interactions and user interface, set limits on file access and built the application. After an intensive final development phase, subscribers received the preview.

    How would the process work without AI?

    Software developers would design the architecture and components, write the code by hand, set up unit tests and fix bugs over successive development cycles, using existing frameworks.

    How did the process work with AI?

    The developers specified how the software should behave and had Claude Code edit and refactor code directly in the project directory. The developers set the architectural constraints, reviewed the code and ran tests. The widely reported “ten days of development” referred to the final implementation phase. A small core team could draw on components that already existed internally.

    Predicting which insurance appeals will succeed (WellSpan Health)

    What outcome does WellSpan want?

    When a health insurer denies coverage for a hospital stay, the hospital can appeal. WellSpan wants to identify cases with a reasonable chance of success so its clinical team can spend its limited time on appeals worth pursuing.

    What does the end-to-end process look like?

    After a denial, the hospital examines the medical rationale and decides whether to appeal. If it files an appeal, the payer’s response establishes the outcome. A crucial gap remains. For denials the hospital never appeals, no one can know what the outcome would have been.

    How did the process work without AI?

    Physician advisers reviewed records manually and drew on their own experience to decide which denials to challenge. Under time pressure, they could overlook appeals that might have recovered reimbursement for the hospital.

    How does the process work with AI?

    A predictive model trained on 2,473 historical cases estimates the likelihood that an appeal will succeed. The clinical team uses the score to prioritize cases, writes the justification and decides whether to file. After implementation, a higher proportion of denials were overturned. Regulations and internal policies changed during the same period, so the study cannot isolate the algorithm’s contribution to the improvement.

    What an AI operating model describes

    An AI operating model describes how an organization delivers an outcome when people and AI work together. It starts with the desired outcome and sets out which tasks people and systems perform, how their contributions fit together, and who can decide, make corrections or intervene. It covers the work itself, the technology it requires, who keeps it running and how the organization learns from actual outcomes.

    At its heart is the division of work that leads to an outcome someone takes responsibility for. The seven cases show why a description of an AI tool alone tells us too little. Someone has to investigate a risk alert. A passenger’s report has to reach the right person in time to act. A technical bid comparison has to stand up to expert scrutiny. The operating model explains how the organization turns what the AI produces into something people can use.

    Which outcome matters, and to whom?

    The starting point is to ask what outcome matters to the people involved. A system’s immediate output and the outcome of the entire process are two different things.

    At BMW, a media report is an initial lead. An investigation establishes whether a supplier has breached a requirement and what action is needed. At Deutsche Bahn, classifying a report correctly makes it possible for someone to respond. The passenger may still be sitting in a carriage with broken air conditioning. At Audi, the process under review ends with a technical comparison that specialists have verified. Awarding the contract is a further step.

    An AI operating model therefore identifies who receives the service, what outcome they need and where the process under consideration ends. That boundary determines which tasks belong to the process and how to measure success. Measuring the routing of passenger reports answers a different question from measuring how long passengers wait for a problem to be fixed.

    Which tasks do people and AI take on?

    AI plays different parts in the seven cases. BMW finds leads in public sources. In the sales case, the system combines information into practice profiles and a ranked contact list. The European Patent Office and Deutsche Bahn use AI to classify and route cases. At Audi and WellSpan, systems prepare the ground for expert judgment. The developers of Cowork say that Claude Code made changes to the code itself.

    These distinctions matter for the operating model. Finding information, making a recommendation and carrying out an action are different tasks. Each calls for different instructions, checks and opportunities to intervene.

    For each step, the organization needs to specify what the system contributes and what people do. Their work extends beyond carrying out the task and making the final decision. It includes specifying the request, supplying the right material, checking intermediate results and handling exceptions.

    Who can decide, approve and intervene?

    Dividing up the work also means assigning decision rights. An AI operating model specifies what a system may do on its own, which steps require human approval and who can correct a result or stop an action.

    At the European Patent Office, a department can reroute a file it received in error. At Audi, specialists first check the requirements list and later verify the bid comparisons. At WellSpan, the clinical team decides whether to pursue a case.

    Decision rights therefore need to be explicit at each step. Saying that a person remains responsible leaves practical questions unanswered. The person accountable must be able to intervene in time and have the information needed to exercise judgment.

    How are handoffs and exceptions handled?

    Often someone has to take over before an AI output can be used. At Deutsche Bahn, a report must reach the crew or maintenance staff in time and with enough detail to act. At BMW, an alert must become a case someone can investigate. At the patent office, there must be an agreed way to correct a misrouted file.

    The operating model therefore identifies who receives each output, what information they need and what happens if anything is unclear. It also covers cases the system cannot classify meaningfully and cases the intended recipient cannot handle.

    A handoff is fully defined only when people know how to send a case back as well as how to pass it on. Who flags missing information? Who sorts out a wrong assignment? Who takes a case that the existing rules do not cover?

    What makes the process reliable?

    Dividing up the work requires technical and organizational support. People need access to the data and original documents they rely on. Systems need limits on the tools they can use. It must be possible to trace the steps taken, and someone must be responsible for incidents and changes. The sources on decision rights, AI agents and governance in this study address these requirements.

    Shared rules have to work in the context of each process. Specialists can verify a result only if they can inspect the evidence behind it. Approval is effective only if it comes before the action. A right to intervene is of limited use if nobody knows who is responsible or how to reach them when something goes wrong.

    How are outcomes checked and processes improved?

    An operating model also needs feedback from the work itself. That feedback shows whether classifications, assessments and handoffs helped and where the process needs to change.

    At the patent office, corrections to misrouted files can reveal weaknesses in the process. WellSpan knows the outcome of the appeals it files. It cannot know what would have happened in cases it chose not to appeal. The sales case documents how the team built a qualified contact list, but gives no evidence of how well the ranking predicted which prospects would buy.

    An AI operating model therefore records which outcomes the organization can observe, which remain unknown and who evaluates the evidence. The findings may call for changes to the AI model, the categories it uses, review steps, handoffs or responsibilities. What matters is that people can turn operational experience into specific changes.

    What the seven cases mean for ways of working, leadership and organization

    Ways of working

    Define the task, check the output and take it forward

    When AI takes on part of a task, the mix of work people do changes. At Audi, the tool produces a requirements list and a comparison of bids. Specialists check and supplement both. At Anthropic, the developers describe specifying the behavior they wanted and directing Claude Code as it implemented the changes.

    People now do more than produce the output themselves. They also define the task precisely, check the underlying material and judge whether the result is fit for use. Producing less by hand does not automatically reduce the amount of work that requires expertise.

    Work plans therefore need to make explicit provision for giving instructions, reviewing outputs and following up. Staff need the right documents, the necessary expertise and enough time. Adding a review step on paper achieves little unless people can carry it out under actual working conditions.

    Put checks where errors can travel downstream

    Audi illustrates how checks at different stages serve different purposes. If the initial checklist misses a requirement, the team may overlook it in every supplier’s bid. Verifying the checklist therefore serves a different purpose from checking the assessment of a single bid.

    This does not mean every result needs several reviews. What matters is where an error can arise and how far its effects can spread. Those are the points where people need a suitable check or a way to correct the error.

    Match new leads to the capacity to act on them

    Media monitoring helps BMW identify additional risks. In the sales case, the team discovers practices missing from its earlier list of qualified prospects. Both outputs create work. Someone must investigate the concern or verify the prospect and make contact.

    Teams need a way to select and pursue the leads they generate. Which cases come first? What information does the team need to act? What happens when more leads arrive than people can handle? A higher number of leads achieves little if the team cannot follow up on the ones that matter.

    Leadership

    Set decision authority according to the consequences of error

    The cases differ in how easily people can spot and correct mistakes. A patent file sent to the wrong department can be rerouted through an established procedure. At WellSpan, the outcome of an appeal that nobody files can never be known. An unduly low score may therefore go unnoticed.

    Leaders need to account for those differences when deciding how much discretion to allow. They need to consider the consequences of an error, how easily someone can detect it and whether it can be corrected. Those judgments determine where approval, spot checks and the right to intervene are needed.

    For example, WellSpan could ask clinicians to review a sample of low-scoring cases to see whether any had a stronger chance of success than the scores suggested. This is a design option. The published case does not report using it.

    Assess throughput, quality and impact separately

    The number of cases processed measures system activity. It tells us little about the quality of the work or its benefit to the recipient. Deutsche Bahn’s feedback volumes do not show how many problems were resolved during a journey. The reported duration of Cowork’s final development phase does not establish how much time the team saved compared with a similar project without coding agents.

    Evaluating a process therefore calls for three separate questions:

    • How much work gets processed?

    • How accurate and useful are the outputs?

    • What improves for the person who ultimately receives the service?

    The right measures depend on where the process ends. For a bid comparison, turnaround time and the quality of the technical assessment may be appropriate. For passenger service as a whole, it also matters whether and when someone fixes the problem.

    Consider changes in workload and capacity together

    More alerts, preliminary assessments and automatically generated outputs change the work that follows. Leaders therefore need to assess both the system’s performance and the capacity available for review, decisions and action.

    The seven cases do not support a general conclusion about whether organizations need more or fewer staff. They do show that AI shifts tasks or makes it possible to take on work that previously went undone. Each organization has to examine where a particular process frees up capacity and where it needs more.

    Organization

    Assign responsibility for the end-to-end outcome

    A passenger report connects a digital reporting channel and an automated routing decision to the work of the train crew or maintenance staff. An improvement in classification achieves little if the report arrives too late or lacks the information the recipient needs.

    Someone needs clear responsibility for the entire process, including how the technical system, the people doing the work and the feedback fit together. This need not require a new organizational unit. What matters is that someone can review the whole process and initiate changes across departmental boundaries.

    Distinguish shared infrastructure from domain responsibility

    Several applications can share access management, logging and technical testing. The judgments people make require expertise specific to each field. Production planners assess whether a bid meets a technical requirement. At WellSpan, clinicians judge the grounds for an appeal and its prospects of success.

    This suggests a way to divide responsibilities. Technical infrastructure can be shared, while specialists in each area retain responsibility for judgments in their field. The practical question is which services should be provided consistently across the organization and which decisions need to stay close to the people accountable for the outcome.

    Give people the authority to act on feedback

    Corrections and subsequent outcomes can reveal where a process needs improvement. Acting on that feedback requires clarity about three responsibilities. Who collects and evaluates it? Who can initiate changes? Who checks the proposed changes and takes responsibility for putting them into practice?

    The patent office’s procedure for rerouting files gives it a starting point for gathering feedback. WellSpan has the results of the appeals it filed. When interpreting those results, it matters which cases have a known outcome.

    For an organization to learn, people need an agreed way to move from an observation to an assessment and then to a change in rules, systems or responsibilities. Simply having the data does not establish that process.

    Conclusion · How AI changes leadership and organization at the point of delivery

    The seven cases offer no single organizational chart for AI. They do reveal a pattern. Value emerges when people connect AI outputs with expert review, decisions and the work needed to put those decisions into practice. An AI operating model must therefore describe the full path from a lead, a report or a request to an outcome someone is accountable for, including handoffs, exceptions and feedback.

    Leaders need to assign decision authority according to the consequences of an error, how easily it can be detected and whether it can be corrected. Human oversight is especially valuable where errors are hard to reverse, where regulation applies or where reputations are at stake. Leaders also need to assess throughput, quality and actual impact separately. A high volume of processed cases does not establish that the service or its outcome has improved.

    For the organization, responsibility must extend from the start of the work to its outcome. Data access, logging, security and operations can be provided centrally. Expert judgment and accountability for the outcome belong close to the people who know the field.

    A central part of leadership is therefore to shape the handoffs. People need to know who clarifies a request, checks intermediate results, decides what to do with exceptions and makes sure work reaches the next person. For people to trust the process, they need to know how to intervene, have access to the original evidence, have enough time to review it and know where to report an error.

    Ultimately, an organization can scale its use of AI only if the entire process has the capacity to cope. More alerts, suggestions and automatically generated drafts create additional work. Without the capacity to check, clarify and act, greater visibility delivers little value. An organization learns when it routinely turns experience from day-to-day work into changes to its rules, systems, responsibilities or processes.

    References

    Supply chain, sales and patent administration

    BMW Group’s procedures, the anonymized sales case and the European Patent Office’s examination guidelines.

    01 · Our due diligence obligations

    BMW Group · March 2026 · Description of procedures (in German) · View source

    Sections 1–4. BMW describes its media monitoring, complaints procedures, corrective measures and how it checks whether those measures work.

    02 · BMW Group Report 2025

    BMW AG · March 2026 · Group report (in German) · View source

    The sustainability statement explains how purchasing assesses suppliers and follows up on agreed measures.

    03 · The Enterprise AI Playbook: Lessons from 51 Successful Deployments

    E. Pereira, A. W. Graylin and E. Brynjolfsson · Stanford Digital Economy Lab · April 2026 · Interview-based report · View source

    Chapter 7, p. 63. The paragraph on sales in aesthetic medicine describes the anonymized case. The researchers selected successful deployments for the study.

    04 · Guidelines for Examination, B–V, 2. Preclassification

    European Patent Office · 2026 edition · Official guidelines, English version · View source

    This section describes the initial technical classification and how it is recorded in the electronic file.

    05 · Guidelines for Examination, B–V, 2.1. Incorrect preclassification

    European Patent Office · 2026 edition · Official guidelines, English version · View source

    This section sets out how to transfer an application to another search division and how to handle cases where the classification is in doubt.

    06 · AI use at the EPO

    European Patent Office · Official account of AI use · Accessed November 4, 2026 · View source

    The office describes AI-assisted file routing and the professional responsibilities of its staff.

    Passenger service, production planning and software development

    Reports from Deutsche Bahn and Audi, and sources on the development and release of Cowork.

    07 · “DB Mobilität erleben” 2024: Innovations for a Strong Rail System

    Deutsche Bahn AG · April 2024 · Media briefing (in German) · View source

    “Railmate,” p. 3. This section gives the number of categories, annual feedback volume, reports forwarded to train crews and an example involving the temperature on board.

    08 · Annual Report 2025

    Deutsche Bahn AG · March 2026 · Group report (in German) · View source

    Sustainability statement, “Consumers and end-users” (ESRS S4), p. 237. Deutsche Bahn describes the reporting channels, the use of Railmate and how reports reach the departments concerned.

    09 · How and where does Audi use artificial intelligence?

    Audi AG · August 7, 2025 · Company article (in German) · View source

    The section on Tender Toucan gives the document lengths, the sequence of steps in the comparison and the planning team’s review of the requirements list.

    10 · Artificial intelligence: Audi increases efficiency in production

    Audi AG · May 28, 2025 · Press release (in German) · View source

    Audi describes how specialists verify the results and estimates the potential time savings in bid analysis. Both Audi sources are company publications.

    11 · Claude Release Notes

    Anthropic · January 12, 2026 entry · Release notes · View source

    The entry for the first Cowork preview documents its release date, platform, availability and access to local files.

    12 · Anthropic’s Felix Rieseberg: Claude Cowork, Mythos, and the SaaS Extinction

    Felix Rieseberg in conversation with Matt Turck · The MAD Podcast · April 10, 2026 · Interview · View source

    From 11:19, Rieseberg puts the ten days leading up to release in context and describes the earlier work and the demand the team had observed.

    13 · Anthropic’s viral new work tool wrote itself

    Megan Morrone · Axios · January 13, 2026 · Report featuring the developers’ account · View source

    Boris Cherny and Felix Rieseberg describe how they used Claude Code. The account of who did which tasks comes from the people involved.

    Healthcare administration and the operating model

    The WellSpan study and the four sources used to interpret the cases.

    14 · Transforming appeal decisions: machine learning triage for hospital admission denials

    Timothy Owolabi · JAMIA Open 8(1), ooaf016 · February 25, 2025 · Implementation study · View source

    The “Materials and methods” and “Discussion” sections document the 2,473 cases and explain the limitations of the before-and-after comparison.

    15 · The AI-First Operating System: A Blueprint for Operating and Business Model Innovation

    World Economic Forum and Kearney · June 23, 2026 · White paper · View source

    This proposed design connects workflows, technical foundations and the division of work between people and AI.

    16 · Designing Decision Rights for AI

    I. M. Sebastian, P. Weill, T. Haskamp and J. vom Brocke · MIT Center for Information Systems Research · June 18, 2026 · Research briefing · View source

    The authors relate decision rights to uncertainty and the potential consequences of errors. Their research draws on 30 interviews with 27 executives from nine organizations.

    17 · Understanding and using AI agents: AI agents in the enterprise

    M. Kintz et al. · Fraunhofer HNFIZ, IAO and IAIS · February 2026 · Technical guide (in German) · View source

    The chapters on AI agents and tools explain how an agent can take actions and respond to the results.

    18 · Updated Model AI Governance Framework for Agentic AI

    Infocomm Media Development Authority · Singapore · May 20, 2026 · Fact sheet on the updated framework · View source

    The guide and its examples explain limits on access, human approval and responsibility for day-to-day operations.

    Editorial Note

    I used AI to help me research and write this study and to translate it into English. I selected the seven cases, framed the questions and drew the conclusions. My interpretation draws on my experience with ways of working, leadership and organizational design.

    The seven cases illustrate how organizations use AI in specific settings. The annotated references explain what each source documents and where the evidence is limited.

    © 2026 Bianca J. Schulz – 77Applied. All rights reserved.

    Keep reading: