World Economic Forum: The AI First Operating System
Source: World Economic Forum in collaboration with Kearney, The AI First Operating System: A Blueprint for Operating and Business Model Innovation, June 2026
Key takeaways
AI first is fundamentally different from AI enabled. An AI enabled company applies AI to tasks within existing workflows. An AI first company systematically redesigns workflows, roles and decision rights so AI becomes a strategic lever for creating and delivering value at scale.
Most organisations are still layering AI onto existing structures. Despite more than $250 billion of estimated global AI investment in 2025, only 25% of companies say AI is having a transformative effect and 84% have not redesigned jobs around AI capabilities.
The World Economic Forum identifies five building blocks of an AI first enterprise: the intelligence engine, adaptive AI technology stack, operations redesign, human AI teaming and new value creation. These are designed to operate as an integrated system rather than a linear transformation roadmap.
Intelligence becomes a compounding asset. AI first companies build an intelligence engine where customer data, operational context, business objectives and real world performance continuously feed back into the system, creating speed, scale and scope.
Context becomes part of the technology architecture. The report argues that context should be dynamic rather than static, with agents assembling live information from authoritative enterprise sources and using ontologies, MCP registries, APIs and other interfaces to access the right information for each task.
Operations need to be redesigned rather than simply automated. The ambition is not AI applied to selected workflows, but workflows that are digitised, codified and made legible to an intelligence engine so the system can learn from them, improve them and increasingly run them.
Intelligence should be allocated like capital. Leaders should concentrate AI on workflows where scale, repetition, workflow friction and cognitive complexity create the greatest potential operating leverage rather than allowing isolated use cases to proliferate across the enterprise.
The boundary between human and AI work must be designed explicitly. AI may automate high volume repeatable activity, augment complex decisions and monitor continuously, while people remain responsible for judgement, accountability and the highest stakes decisions.
Human AI teaming changes the talent model. The report expects growing value from people who combine broad systems thinking and adaptability with deep domain expertise, while AI increasingly absorbs the executional middle layer.
The organisational model becomes more federated. A central intelligence team can provide shared infrastructure, models and governance, while business units own where AI is applied, how initiatives are funded and the resulting business outcomes.
Trust and visibility are operating requirements. As agentic workflows become less deterministic, organisations need real time execution visibility, traceability, intervention capability and mechanisms for monitoring how workflows evolve.
The end state is not more AI tools. It is an enterprise in which intelligence becomes embedded in how the organisation thinks, decides, executes, learns and creates value.
Executive summary
The World Economic Forum’s AI First Operating System, developed in collaboration with Kearney, examines how organisations change when artificial intelligence becomes part of the fundamental design of the enterprise rather than another technology layered onto existing work.
The research draws on engagement with more than 50 advanced AI first enterprises and thinkers, alongside insights from more than 150 executives and experts gathered through interviews, workshops and consultations. The report argues that a new class of organisation is emerging that designs operations around intelligence, uses data as a strategic resource and embeds AI directly into processes, decisions and execution.
The distinction between AI enabled and AI first is important. An AI enabled organisation uses AI to automate or augment tasks inside existing workflows. Performance improves, but the operating model remains broadly intact. An AI first organisation goes further by systematically redesigning workflows, roles and decision rights around AI. If the AI systems disappeared, the organisation’s core workflows and delivery model would no longer operate in the same way.
The Forum argues that most companies remain closer to the first model. Despite significant investment, AI is often deployed as a layer over existing processes. This produces local productivity gains but leaves the architecture of the organisation largely unchanged.
The alternative is built around five connected components: an intelligence engine, an adaptive AI technology stack, operations redesign, human AI teaming and new value creation.
At the centre sits the intelligence engine, a self reinforcing system that learns from every interaction and transaction. Around it sits a modular technology architecture capable of connecting data, models, context and orchestration while allowing models and tools to change over time. Operations are then redesigned so workflows become legible to the intelligence engine, human and AI responsibilities are deliberately allocated, and every execution cycle creates signals that improve future performance.
The workforce also changes. The report argues that AI first organisations increasingly need people who can frame problems, apply judgement, orchestrate AI systems and bring deep domain expertise. AI absorbs more of the technical and executional middle layer, making both broad transferable capabilities and deep expertise more valuable.
The organisational structure changes alongside the work. Rather than relying solely on centralised AI teams, the Forum describes a federated model in which a CEO led central intelligence capability provides common infrastructure, standards and governance while business units retain responsibility for execution and financial value.
The report therefore presents AI first transformation as a redesign of the enterprise operating system rather than a programme for deploying more tools.
Key stats and quotable claims
The Forum’s AI First Enterprises workstream brought together more than 50 advanced enterprises and thinkers.
The work also drew on insights from more than 150 executives and experts.
Estimated global AI investment exceeded $250 billion in 2025.
Only 25% of companies say AI is having a transformative effect on their organisation.
84% of companies have not redesigned jobs around AI capabilities.
One commercial insurance workflow cited by the report moved from 28 days to 2.8 hours.
Some AI first teams are achieving 15 times greater productivity with AI tools than without them.
Gamma crossed $100 million ARR with a team of roughly 50 people by late 2025.
Gamma users had created more than 400 million assets, with more than one million pieces of content generated daily.
Gamma’s inference related gross margin increased from approximately 31% to 77% within six months.
Osmo reduced a fragrance development process that previously took six months to 60 seconds.
Stripe’s payments foundation model identified more than 95% of card testing attacks in real time, a 22% improvement over previous methods.
ServiceNow evaluates agents against defined task completion thresholds typically between 80% and 95% before deployment.
Waymo has used more than 20 billion virtual miles to test real world scenarios before deployment in new cities.
More than 53,000 Cognizant employees across 40 countries participated in an AI assisted development initiative.
Cognizant’s 1C platform connects hundreds of enterprise applications and AI agents for 350,000 employees.
The report says the platform has helped reduce support tickets by 50% and increase operational efficiency by 50%.
Contextere reduced a 47 minute factory floor information gathering process to near instant context assembly and cut troubleshooting time by up to 80%.
AI first production teams are typically fewer than ten people, organised around a product, workflow or customer problem.
The report notes that getting an AI prototype into reliable production can require around ten times the capabilities needed to build the prototype itself.
Overall summary
The report begins with an analogy to earlier general purpose technologies. Electricity initially produced limited productivity improvement when factories simply replaced steam engines with electric motors while keeping the same layouts and processes. The larger productivity gains appeared when businesses redesigned the factory around what electricity made possible.
The Forum argues that AI is at a similar point. Organisations can add AI to existing processes and gain incremental efficiency, or they can redesign the enterprise around intelligence itself.
That creates three broad categories.
AI enabled organisations use AI inside existing tasks and workflows. If the AI tools disappeared, the organisation would continue operating broadly as before.
AI first organisations redesign workflows, roles, decision rights and technology architecture around AI. If the AI systems disappeared, the operating model itself would break down.
AI native organisations go further still. They are new businesses whose products, services and competitive advantage fundamentally depend on AI systems.
The Forum’s blueprint focuses on how organisations make the transition towards the second model.
The central idea is that intelligence becomes a reusable and compounding enterprise capability. Rather than building a separate model or integration for every individual use case, organisations create shared intelligence, context and orchestration layers that become more capable as more workflows use them.
Operations are redesigned around the same principle. The organisation does not scatter AI evenly across every activity. Leaders identify a small number of influential workflows, determine where intelligence can create the largest performance improvement and redesign those workflows end to end.
At the task level, the question becomes whether AI should act, assist or stay out. Different tasks may require different models, different levels of autonomy and different levels of human involvement.
The organisation then needs infrastructure capable of coordinating those decisions at scale. The report places particular emphasis on ontology, live context, orchestration, permissions, evaluation and operational visibility.
Human AI teaming completes the operating model. Employees increasingly contribute through judgement, problem framing, domain expertise, collaboration and orchestration while AI takes on more execution. Organisations must therefore redesign both skills and structures around the new distribution of work.
Deep dive
From AI enabled to AI first
Artificial intelligence is the latest class of general purpose technology, like electricity, computers and the internet. Historically, these technologies had a limited impact when incumbent organizations applied them within existing systems. Real breakthroughs came when pioneers rebuilt their businesses around the technology itself.
In electricity, early manufacturing adopters replaced steam engines with electric motors while retaining the same layouts and production processes, resulting in no productivity gains but reducing energy costs by 20–60%. The breakthrough came when pioneering factories redesigned their entire blueprints around electricity. These leaders reconfigured workflows, distributed power to individual workstations and rebuilt production systems, unlocking step changes in operating models and laying the foundation for the modern factory.
The speed and breadth of adoption are what make AI different. In October 2025, Wharton estimated that 82% of decision makers use AI weekly, up from 37% in 2023. Democratized access, reasoning and agency, multimodal intelligence and parallel execution are expanding what organisations can build around the technology.
For today’s market, intelligence is changing what kinds of organizations can be built. AI first organizations are applying this lesson from the outset, designing operations, decision making and business models around intelligence.
Three enterprise archetypes
The Forum distinguishes between three types of organisation.
AI enabled enterprises apply AI to automate or augment discrete tasks within existing workflows, improving performance without materially redesigning the operating model. AI is applied at the task level, talent models focus on domain operators experimenting with AI tools, and the technology stack largely integrates third party AI tools into existing systems.
The litmus test is straightforward: if the AI tools were removed, would the workflows and organizational structure collapse? If the answer is no, the enterprise is AI enabled.
AI first enterprises systematically redesign workflows, roles and decision rights so AI becomes the strategic lever for creating and delivering value at industrial scale. Core operating workflows are redesigned around AI, technology stacks are structured to embed AI into execution and decision making, and tasks are systematically allocated between humans and AI.
The corresponding test is whether the business could still operate in the same way if its AI systems were removed. If the answer is no, the organisation is AI first.
AI native organisations are created with AI as a core production capability. Their products, services and competitive advantage are fundamentally dependent on AI systems. Here the relevant question becomes whether the value proposition itself would still exist without AI.
Why AI enabled approaches limit the opportunity
Despite AI investment estimated at over $250 billion globally in 2025, many organizations are seeing only incremental gains. A global survey found that only 25% say AI is having a transformative effect on their company. Meanwhile, 84% of companies have not redesigned jobs around AI capabilities.
Like electricity, the primary reason is structural. In many cases, AI is deployed as a layer on top of existing processes and embedded narrowly in tools. This approach improves efficiency at the margins but does not unlock the technology’s full potential.
In contrast, AI first performance gains are early but striking. Commercial insurance workflows have moved from 28 days to 2.8 hours. Some companies have reached $100 million in annual recurring revenue in months rather than the four to eight years previously associated with the milestone. A 2.5 day continuous autonomous inference loop identified the first viable drug candidate for a previously intractable HIV target, while some teams are achieving 15 times more productivity with AI tools than without them.
The Forum argues that innovations in operating and business model logic sit behind these outcomes rather than model capability alone.
Five building blocks of the AI first operating system
Five building blocks define the fundamentals of how AI first enterprises deliver value. They should not be understood as a linear roadmap but as an integrated system. Progress in one area reinforces the others.
The first is the intelligence engine, where intelligence sits at the core as a compounding source of speed, scale and scope.
The second is the adaptive AI technology stack, a modular architecture designed for scale, flexibility and control.
The third is operations redesign, which determines where intelligence should be distributed and how workflows need to change around it.
The fourth is human AI teaming, where roles, teams and organisational structures support continuous collaboration between people and AI systems.
The fifth is new value creation, which determines how new intelligence capabilities are positioned in products, services and markets.
Together, these building blocks form an emerging blueprint for an enterprise designed around intelligence.
Block 1: Intelligence engine
Intelligence as a compounding enterprise capability
When intelligence sits at the core of an enterprise, it becomes a durable source of speed, scale and scope.
AI first enterprises build intelligence engines as an operating system around which the firm thinks, runs and creates value. By strategically embedding intelligence into the foundational business logic, these organizations ensure that operational data, proprietary intellectual property and sources of advantage compound.
At the centre is an intelligence engine, a self reinforcing and data driven flywheel that embeds organizational knowledge, learns from every interaction and transaction and grows smarter with use. It connects customer data, operational context, business objectives and real world performance into continuous feedback loops that learn, improve and scale with each cycle.
The Forum describes three connected loops: speed, scale and scope.
The speed loop
The objective of the speed loop is to run more experiments, generate and test hypotheses and validate ideas before committing resources. This lowers the cost of experimentation while increasing the pace of learning. Failures narrow the search space while successes become building blocks for the next cycle.
Over time, this creates a system where proprietary learning compounds.
The report describes examples including a drug discovery company moving from candidate hypothesis to validated simulation in days rather than months, a software company shipping functional prototypes from plain language briefs in seconds, and a robotics company testing physical behaviours in simulation before real world deployment.
The underlying advantage is not simply speed. The system continually improves its ability to learn.
Context rich data provides the learning base, defined business outcomes determine what the system optimises for, and autonomous inference allows AI increasingly to generate, test, evaluate, adapt and refine outputs against defined objectives.
Context needs to connect individual and company objectives
The report uses Workera as an example of objective led context building.
Workera connects to systems including Workday and ingests role data, resumes and strategic priorities. It then builds two layers of context: individual context around a user’s skills and experience, and company context around the capabilities the organisation needs to develop.
The system uses this context to assess skills, identify gaps and recommend targeted learning pathways. Context is considered useful when it improves the precision of the assessment, makes recommendations relevant to the individual’s role and aligns skill development with the organisation’s objectives.
The example illustrates a broader point in the report: context is not simply more data. It is the information required to make the system’s output relevant to the specific person, task and organisational objective.
Production requires evaluation rather than unlimited autonomy
Learning acceleration also requires evaluation, safety and governance.
The report uses ServiceNow as an example. ServiceNow built an AI agent evaluation framework with quantitative thresholds that agents must pass before they go live. Agents are scored against defined task completion rates, typically between 80% and 95%, with allowable variance based on the objective.
When outputs fall outside defined bounds, a human in the loop reviewer becomes involved at predefined checkpoints. Once agents are operating in production, statistical evaluation becomes essential because the volume of agentic decisions can be approximately 1,000 times greater than in preproduction and exhaustive human review is no longer feasible.
Autonomy therefore depends on measurable acceptance thresholds rather than an assumption that humans should either review everything or nothing.
The scale loop
The scale loop operationalizes intelligence as a multi use platform, applying the same models and workflows to increase output across adjacent functions.
This requires connecting AI systems directly to business metrics, integrating them into live workflows and ensuring they operate reliably under real world conditions. Once deployed, each use strengthens the system, improving its efficiency and quality of output.
The data feeding the system shifts from training data towards live business performance. Economic and operational alignment links system performance directly to business metrics and embeds AI into how work is executed. Production grade autonomy then enables the systems to operate reliably within defined boundaries.
The final step is a multi use platform where capabilities are reused rather than rebuilt for every use case. This reduces the marginal cost of growth and allows intelligence to expand into adjacent work.
Connecting AI performance to business outcomes
Claryo provides one example of linking intelligence directly to business economics.
Its spatial generative AI platform processes roughly 50 terabytes of visual data per warehouse each month to identify work stoppages, backlogs and bottlenecks in real time.
Instead of simply showing operational activity, the system links those observations to daily P&L and square foot economics. If pallets sit idle in an express lane, for example, the system can identify the delay and quantify the financial loss.
The important change is that system performance becomes observable in the same terms the business uses to measure value.
Human judgement can sit inside an AI decisioning layer
Taktile provides another example in regulated financial services.
High stakes decisions often sit inside workflows including onboarding, credit underwriting, anti money laundering and fraud prevention. Taktile’s customers deploy agents against time consuming work that had previously been difficult to automate, while rules, agents and human judgement operate together in a common decisioning layer.
The purpose is not to eliminate human decision making. It is to make complex decision systems faster, more observable and more governable while reducing unnecessary manual work.
The scope loop
The scope loop expands what the organization can do with intelligence. Organisations can reuse capabilities across workflows, share them among teams and recompose them into new products, services and markets.
A capability developed for one purpose can therefore create value elsewhere. A legal AI system can extend from contract analysis into tax and transaction work. A shared intelligence layer can become a foundation for fraud detection, personalization and risk decisions across the enterprise.
Over time, organizations can move towards capabilities that make previously inaccessible scientific, technical and commercial opportunities feasible.
At that point, the constraint begins to shift away from whether intelligence can be built and towards where the organisation chooses to apply it.
Block 2: Adaptive AI technology stack
The technology stack has to evolve with the intelligence
When enterprises design modular technology stacks, the intelligence engine can evolve as fast as the technology beneath it.
An AI first technology stack connects data, models, context and orchestration so intelligence can be used across workflows, improved through feedback and adapted as models and tools change.
Importantly, the report does not argue that AI first organisations need to replace every existing system of record. CRM, ERP and other core systems can remain the source of truth while AI sits across them, orchestrating workflows, improving decisions and changing how information is accessed.
The architecture should provide interoperability, identity and authority, open interfaces and the ability to modernise as models, tools and infrastructure change.
Turn data into fuel for the intelligence engine
The intelligence engine only compounds if the data feeding it gets better with every cycle.
AI first enterprises integrate feedback, training and evaluation into a single system so performance can improve over time rather than depend on periodic fine tuning.
Every instance of use becomes a signal. Whether an output is accepted, overridden or flagged, the information can flow back through the stack. Repeated across the system, these signals improve models, refine orchestration and make the overall system more capable.
The report distinguishes operational data from structured learning data and argues that business context still needs to be encoded even where models can work with unstructured information.
Synthetic data can also help with rare edge cases or scenarios where sufficient labelled production data is unavailable.
Own the control layers
By owning the control layers around the model, a modular architecture can absorb technology changes without continuously rebuilding workflows or giving up critical architectural control.
A customer support organisation running thousands of AI assisted interactions across chat, email and internal handoffs should not need to redesign its workflow every time the underlying model changes.
The Forum identifies three important practices.
First, own orchestration, not just models. The orchestration layer routes work based on performance, cost and policy, giving the enterprise control over how intelligence is deployed as capabilities change.
Second, manage the gateways between data and models. Models need context from systems including CRM, HR and finance. Connectors such as MCP can provide access through a common layer rather than relying on bespoke integrations for every system.
Third, build APIs around proprietary data. Keeping proprietary information behind internal APIs creates a stable interface even as models or vendors change while preserving controls including permissions, audit logs and caching.
Compose a model agnostic portfolio
The report argues against relying on one model for every business task.
Instead, organisations should treat models as a portfolio and route work based on performance, cost and reliability.
Lower cost models may handle repetitive production work. Frontier models may be more appropriate for high judgement synthesis or complex reasoning. Tuned or domain specific models may matter where proprietary knowledge or sensitive intellectual property creates a genuine advantage.
The objective is not to deploy the most capable model everywhere. It is to use the appropriate model for the task while maintaining enough evaluation capability to switch confidently as model economics and performance evolve.
Make context dynamic, not static
One of the strongest sections of the report is its treatment of context.
The objective is to assemble context in real time. Instead of relying on static copies of data, the system should pull live information from the authoritative source, preserve its authority and tailor what is retrieved to the task at hand.
A field technician should not rely on a prompt loaded with outdated manuals and policies. The system should retrieve the current equipment history, latest service bulletin and site specific operating context at the moment the information is needed.
The report recommends keeping context modular and centrally discoverable rather than copying everything into a single store. Agents should retrieve only what is relevant to the task.
It also argues that source authority needs to travel with the context. If multiple sources conflict, the agent needs to understand which source takes precedence.
Finally, context should be assembled just in time. A context router retrieves what is needed for the current task, while corrections from employees feed back into the system.
Indeed and the modular operating layer
Indeed provides a case study of this architecture.
Indeed built a centralized agent platform based on the Model Context Protocol that modularized how models, data and applications connect across the enterprise. The architecture virtualises underlying services and allows secure access across internal and external products.
The platform also includes automated risk guardrails, evaluation systems that track adoption, accuracy and productivity, and the ability to change models based on cost, performance and latency.
The goal is for every interaction to feed back into the system, improving orchestration and allowing capabilities to expand without disrupting the underlying operation.
Block 3: Operations redesign
Workflows become systems that learn
When intelligence allocation shapes how operations are designed, workflows stop being processes to execute and become systems that learn.
The ambition for an AI first enterprise is not AI applied to selected workflows. It is an organization where every workflow has been digitized, codified and made legible to the intelligence engine, so that the engine can learn from it, improve it and ultimately run it.
Operations and intelligence become inseparable. Every decision, process and action feeds a system that continuously becomes smarter, faster and more capable.
The enterprise does not simply allocate AI to the business. Eventually, the business itself runs on AI.
Most organisations are still some distance from that state. Many are running disconnected pilots that demonstrate local value but do not compound because they are not connected to a shared intelligence layer or designed for scale.
The bridge between pilots and AI first operations is to begin treating intelligence like capital.
Intelligence should be allocated deliberately
Rather than allowing AI use cases to proliferate without coordination, AI first leaders identify the most influential workflows, redesign them end to end around intelligence capabilities and ensure every execution cycle feeds signals back into the engine.
Over time, the operating model is rebuilt workflow by workflow.
The Forum describes three levels of intelligence allocation: deciding where to allocate intelligence, determining how to redesign work, and building the infrastructure required for intelligence native operations.
The organisation begins at the level of the business outcome, concentrates intelligence on a small number of important workflows and then allocates the right mix of human and AI capability to the underlying tasks.
Gamma: continuously reallocating intelligence
Gamma illustrates the economics of deliberate allocation.
By late 2025, the AI native presentation company had crossed $100 million in annual recurring revenue with a team of roughly 50 people. Users had created more than 400 million assets and were generating over one million pieces of content every day.
The company treated each part of content creation as an allocation decision: which tasks should use frontier models, which should use cheaper specialised models, which should rely on orchestration rather than fine tuning, and where human judgement remained important.
Gamma’s edge came from continuously reassigning work to the lowest cost, highest quality combination of model, system design and human input.
Six months after launch, inference related gross margin increased from approximately 31% to approximately 77%.
Each improvement made the next execution cycle faster and cheaper.
Where to allocate intelligence first
AI first organisations do not try to digitise everything at once.
They start with workflows where redesign around intelligence can produce the largest step change in performance.
The Forum identifies three useful characteristics.
The first is scale and repetition: continuous, high volume processes where small improvements compound.
The second is workflow friction: processes slowed by handoffs, multi step coordination or manual review.
The third is cognitive complexity: decision heavy or information intensive activity where AI can improve judgement and speed.
This is why areas such as customer support, procurement, underwriting, claims and troubleshooting often become early priorities.
They are selected not because they are easy, but because redesign can generate compounding returns.
Redesign work rather than layering AI on top
Once the right workflows are identified, the work is to redesign them, not to layer AI on top of existing processes.
AI first leaders start with two questions: what can be reliably handled by AI and what must remain with people?
AI may take responsibility for high volume repeatable tasks, assist with complex decision making where speed and consistency matter and support long term monitoring and exception detection.
Humans remain responsible for judgement, accountability and the decisions carrying the highest stakes or requiring contextual knowledge AI cannot yet replicate.
The boundary between the two is not fixed. It changes as the intelligence engine improves.
Levels of human agency
The report provides a five level human agency scale.
At one end, AI completes the task entirely without human involvement.
At the next level, AI drives task completion with minimal human oversight and requests human input only where needed.
In the middle, people and AI work together as equal partners.
At the fourth level, the human leads execution while AI provides assistance.
At the fifth, the task remains fully dependent on human involvement.
The important implication is that human involvement does not need to be binary.
Organisations can intentionally allocate different levels of agency at different stages of the workflow.
Match the model to the task
The same principle applies to models.
Not every task requires the most powerful or expensive system. AI first organisations route work based on the actual requirements of each activity and balance accuracy, cost and latency.
High stakes decisions in regulated environments may require extremely high accuracy and human review. High volume repeatable activity can often use smaller, cheaper models. Complex reasoning may justify frontier models even where they are more expensive.
The goal is not to use the most capable model everywhere. It is to use the right model for each task and continuously improve that allocation as the system learns.
Building intelligence native operations
The hardest part of the journey is not identifying the right workflows or redesigning individual tasks.
It is building the operational infrastructure that allows intelligence to coordinate across workflows, improve continuously and run reliably at scale.
This is what separates organisations that redesign a few workflows from those that rebuild operations around intelligence.
The Forum identifies three requirements: making the business legible to the intelligence engine, redesigning workflows end to end and designing for trust and operational visibility.
Making the business legible
AI first enterprises digitize and codify how the business works: its workflows, decisions, rules, handoffs and actions, so intelligence can reason about them, act on them and improve them over time.
The report places ontology at the centre of this architecture.
Ontology defines how the business operates, how data is used, how decisions are made and what actions can be taken.
Without it, handoffs become inconsistent, decisions are difficult to trace and the intelligence engine cannot operate reliably across workflows.
With it, the system can reason and act consistently, while every additional workflow strengthens the common infrastructure.
The report says ontology can codify three layers.
Data answers what the current state of the business is.
Logic defines how the business should reason about that state through relationships, rules, constraints and decision criteria.
Actions define what can happen next, including allowed activities, routing, escalation paths and workflow triggers.
Embedding business context and rules in this structure moves the organisation from prompt driven automation towards structured and verifiable workflows.
Redesign workflows end to end
AI first enterprises redesign end to end workflows around desired outcomes.
That means resequencing tasks, redefining handoffs, embedding verification and assigning explicit responsibilities to humans and AI.
The first tactic is resequencing work. Leaders work backwards from the outcome and assign the appropriate human or AI system to each stage.
The second is reallocating roles. Handoff points frequently create delay, duplication and inconsistent decisions. AI can take on routing, low stakes reviews and coordination where appropriate.
The third is embedding verification. The workflow itself defines who or what can act at each stage and preserves critical quality and compliance checkpoints, with clear escalation when work returns to a human.
Operational visibility becomes essential
As workflows become AI driven and continuously evolve, visibility is essential to maintain control.
Unlike traditional systems, agentic workflows do not always follow fixed sequences. They adapt in real time. Without visibility, organisations lose the ability to understand performance, intervene when needed and improve the system.
The Forum identifies four characteristics of strong operational visibility: real time execution monitoring across agents, humans and systems; traceability of inputs, decisions and outcomes; the ability to pause, override or redirect active workflows; and visibility into how workflows evolve as agents learn.
This means evaluation and governance need to operate inside the runtime environment rather than exist only as policy.
Cognizant: operational visibility at scale
Cognizant provides an example of visibility at enterprise scale.
More than 53,000 employees across 40 countries experimented with AI assisted development tools in sandboxed environments. This gave the company visibility into how AI was being used, where compute was consumed, which security risks appeared and where AI was generating value.
The organisation initially provided broad lightweight token access and then directed additional compute towards ideas showing the strongest promise.
The same principle underpins Cognizant’s 1C multi agent platform, which connects hundreds of enterprise applications and agents through a single searchable interface for 350,000 employees.
The report says this has helped reduce fragmentation, cut support tickets by 50% and increase operational efficiency by 50%.
Operational leverage becomes the economic objective
AI first enterprises treat intelligence like capital and allocate it across the organisation.
When intelligence is embedded across workflows and the engine operates reliably at scale, organisations can get more from the same operating base by increasing throughput, compressing cycle times, widening scope and improving quality at the same time.
But higher intelligence does not automatically mean lower cost. Scaling AI introduces variable costs from compute, model access, software licences and specialised AI roles.
The economic objective is therefore to ensure that the performance gains from intelligence native operations exceed the cost of delivering the intelligence.
The longer term ambition goes further. Once every workflow is digitised and codified, execution continuously feeds signals into the intelligence engine and the system repeatedly improves what the organisation can do, the distinction between AI in workflows and the intelligence engine running operations starts to disappear.
Intelligence becomes how the business operates.
Block 4: Human AI teaming
Human contribution changes as intelligence scales
When structure follows collaboration, human AI teams can achieve outcomes that neither could design, decide or deliver alone.
The report argues that as intelligence becomes embedded across workflows, talent becomes an increasingly important differentiator: who organisations hire, how people are developed and how the organisation directs intelligence towards outcomes.
The objective is to enable people to orchestrate, supervise and scale AI capabilities into measurable results.
If executed well, human AI teaming can remove mundane work and move human contribution towards judgement, creativity and orchestration.
The limiting factor then becomes how well human AI teams are structured to collaborate, govern and compound performance.
The Forum identifies three parts of an AI first talent advantage: growing human AI talent, organising human AI teams and evolving the enterprise itself for AI first execution.
Growing human AI talent
AI first organisations are already recruiting new profiles including design engineers, forward deployment engineers, evaluation specialists, AI researchers and safety engineers.
Not every organisation needs every new job title. The broader signal is that capability requirements are changing as work changes.
The report identifies a combination of breadth and depth.
Breadth enables employees to frame problems, adapt to new tools and use systems thinking to direct AI across domains.
Depth provides domain expertise and judgement that AI cannot easily substitute.
Between those two layers sits the executional middle layer, which AI increasingly absorbs. As a result, broad transferable capabilities and deep specialised knowledge both become more important.
The report’s AI first T shaped talent model combines continuous learning, AI fluency, task decomposition, context integration and orchestration with systems thinking, problem framing, decision making, collaboration and deep domain expertise.
Skill life cycles become continuous
In a traditional skill lifecycle, a person learns a capability, applies it and eventually retires it when the work changes.
AI disrupts that linear model.
First, tools evolve faster than skills can stabilise. Prompting habits, orchestration patterns and tool specific techniques can become outdated in months or sometimes overnight.
Second, deep domain expertise is anchored in context. When the tools, roles and work surrounding an expert change, the expertise itself needs to be refreshed.
The report therefore recommends codifying individual expertise into reusable AI capabilities, embedding organisational knowledge into shared AI systems and actively managing skill atrophy.
As AI performs more execution, human skills can weaken through disuse. Organisations may only realise how valuable tacit knowledge was when an AI system fails.
The report highlights problem framing, judgement, systems thinking, exception handling and evaluation as capabilities organisations should keep active even where AI can perform much of the visible work.
Its conclusion is particularly important: skills can be retrained; judgement is harder to restore.
Organising for human AI teams
AI first leaders recognise that production grade AI requires new team structures and norms.
The most effective teams are described as small, fast moving units combining technical, domain and operational capability around a single product, workflow or customer problem.
A prototype may require around 1,000 skills to build, but taking it reliably into production can require ten times that.
Moving from prototype to production therefore depends less on model capability alone and more on how teams are organised around it.
The report describes cross functional teams typically containing fewer than ten people and operating with relatively flat hierarchies.
The intelligence engine sits at the centre while people operate at the edge, using context and real time information to navigate uncertainty, apply judgement and translate intelligence into business outcomes.
Three emerging team types
The report identifies three increasingly important team archetypes.
Pilot and production teams move AI from experimentation into scalable products and services. Their key principle is that production requirements are incorporated from the beginning rather than being handed off after the pilot.
Customer embedded teams combine technical and domain capability directly around customer problems, allowing requirements to evolve through live feedback rather than slow functional handoffs.
AI safety teams embed evaluation, legal, compliance and responsible AI capability into the system from the start.
Across all three, the broader design principle is continuous collaboration rather than sequential handoffs.
AI first team norms
The Forum also describes several team norms.
AI should be attempted as a first tool for relevant problems, but outputs should always be evaluated against known evidence, sound business judgement and common sense.
Humans set direction while AI can determine much of how the work gets done.
Measures of success include AI adoption for relevant work, productivity improvement, reductions in cross team friction and stronger quality control.
The report also emphasises that the multiplier effect disappears without genuine domain experts capable of guiding AI output.
Quality standards remain human responsibilities. AI output should improve on prior performance, but human craft and judgement remain important for differentiation, safety and quality.
Evolving the organisation for AI first execution
Designing how intelligence moves through the enterprise becomes a new source of organisational advantage.
As organisations make context, work and decisions more visible and machine readable, they can move beyond structures designed primarily to route information through managerial layers.
Shared intelligence layers, distributed ownership and autonomous operations allow organisations to coordinate more quickly across silos, push decisions closer to value creation and scale AI without losing control.
The report is explicit that the CEO must own the AI strategy.
Redesigning work across the enterprise cannot be delegated solely to a central technical function or treated as a side initiative.
The federated operating model
The report argues that the most effective operating model for an AI first enterprise is federated.
A CEO led central intelligence capability provides shared tooling, infrastructure and governance while business units retain ownership of budgets, execution and decisions about where AI is applied.
The centre decides what infrastructure the enterprise runs, including approved models, data platforms, evaluation tooling and safety guardrails.
Business units decide where to apply AI, fund their own initiatives and remain accountable for the resulting business value.
This separation is designed to increase both speed and accountability.
If the central team owns the entire roadmap, demand accumulates centrally and the teams closest to customers lose control of prioritisation.
Embedded AI leadership within individual business units can carry shared standards into local execution while feeding lessons back to the enterprise centre.
Rakuten and distributed execution
Rakuten’s AI & Data Division provides an example of this model.
The division reports to the CEO and operates across more than 70 business units.
The central capability acts as an internal AI platform provider, creating shared models, search platforms, recommendation systems, personalization capability, customer service tools and multilingual agents.
Embedded AI leaders inside business units maintain links to both the local business leadership and the group AI organisation.
Business units can also prioritise and fund their own AI initiatives through their P&Ls, keeping investment connected to operational outcomes.
The result is shared infrastructure without removing local accountability for value.
Intelligence as a coordination backbone
The report then pushes the organisational argument further.
The intelligence engine can become the company’s coordination backbone when tribal knowledge, documents and operating capabilities are connected through a shared context layer.
Live signals from work can be captured, interpreted and acted upon through the system rather than repeatedly moving through teams and management layers.
This shifts the organisation away from heavily layered structures and towards more outcome oriented roles.
Individual contributors can access the context and operational assets required to execute directly, while the system takes on more of the coordination burden.
Teams can then operate with greater autonomy, pooled intelligence and faster decision cycles.
Block 5: New value creation
The intelligence engine extends into the market
The intelligence engine does not stop at the boundary of internal operations.
How an organisation positions intelligence in the market, what it offers, how customers experience it and how they engage with it, determines what information flows back into the engine.
AI first companies therefore design their customer facing layer not only to acquire and retain customers, but also to generate feedback that improves the system over time.
The fundamentals of a strong value proposition still apply.
AI novelty may encourage initial trial, but repeatable value determines retention. Customers still ask whether the product solves their problem, whether they can trust it, whether the value justifies the price and why they should choose it rather than an alternative.
The Forum divides new value creation into product design, trust and market positioning.
Design intelligence around the user rather than the interface
AI first product design intentionally packages intelligence so people can use it naturally and confidently in the flow of work.
This means moving beyond a generic chat interface and thinking about the user’s actual objective, environment and journey.
The goal is not to maximise AI use. It is to apply intelligence where it improves the outcome.
Greater AI capability can otherwise produce noise, friction and low quality generic output rather than value.
The report recommends matching the modality to the task, providing starting points rather than blank prompts, revealing complexity progressively, incorporating personalised retrieval and feedback, supporting collaborative team environments and adapting interaction patterns to cultural context.
Contextere: intelligence in the flow of frontline work
Contextere illustrates the principle.
The industrial AI company redesigned factory floor troubleshooting around how frontline workers already operate, using a voice first interface rather than requiring workers to move into a different workflow.
The system replaced a 47 minute information gathering process with near instant context assembly and reduced troubleshooting time by up to 80%.
The lesson is that product performance can depend as much on how intelligence is delivered into the work as on the underlying model.
Trust is part of the product
For AI first organisations, trust is not a layer added after launch. It is part of how the product is packaged from the beginning.
The report identifies four principles.
The first is to protect identity, credentials and control, particularly where AI acts on someone’s behalf.
The second is to match accuracy to what is at stake. Higher consequence decisions require higher accuracy, more oversight and more human involvement.
The third is to make confidence and boundaries visible, so users understand what the system knows, where uncertainty remains and what it is designed to do.
The fourth is to make the system verifiable and auditable, allowing users to understand how outputs were produced and what evidence supports them.
These principles become particularly important as agents move from providing information towards taking actions.
Positioning intelligence in the market
Every AI first company faces the same strategic question: how should intelligence be positioned so customers adopt it, trust it and return to it?
The answer depends on where intelligence sits in the offer.
The report identifies five archetypes.
AI as a feature makes an existing product smarter.
AI as the product means the model or application itself is what the customer purchases.
AI as process and platform changes workflows and coordination.
AI as infrastructure powers execution layers, rails and protocols underneath other experiences.
AI as invisible creates a new user experience where customers may not think about the underlying intelligence at all.
The position determines what customers pay for, where value accrues and what type of company is being built.
Agentic commerce changes how customers discover and buy
AI first companies also need to prepare for markets in which agents increasingly participate in discovery, evaluation and transactions on behalf of users.
The report describes five levels of maturity.
At the first level, the agent completes routine transaction steps after the user has decided.
At the second, the user describes the need and the agent converts it into search and options.
At the third, the agent remembers preferences and constraints.
At the fourth, the agent evaluates and selects within user defined guardrails.
At the fifth, the system anticipates needs and acts before explicit demand.
Most of the market currently sits closer to the second level, but the report expects the trajectory towards greater delegation and autonomy to continue.
The strategic question is therefore not simply whether a company sells an AI product. It is how its products, services and commercial infrastructure operate in a world where software agents increasingly participate in the customer decision itself.
Designing the AI first business model
The Forum concludes by extending the traditional Business Model Canvas for a world in which intelligence is abundant and programmable.
The original building blocks remain relevant: value propositions, customers, operations, resources, partners, revenue streams and costs still define a business.
What changes is the set of questions leaders need to ask.
For key activities, which work should be automated, augmented or left human led?
For key resources, what proprietary context, data and feedback become more valuable with every month of operation?
For the value proposition, what becomes possible only because intelligence is embedded?
For channels, where should agents interact directly with customers and at what level of autonomy?
For customer relationships, what new experiences can intelligence create?
For cost structures, how does intelligence change marginal economics?
For revenue streams, which new pricing models or sources of revenue become possible?
For key partners, what parts of the stack should be proprietary, commoditised, bought or accessed through partners?
The result is not a separate AI strategy sitting beside the business model. It is a way of asking how AI changes each component of the business model itself.
Conclusion
The emergence of AI first enterprises is one of the most consequential organisational shifts of today, and it remains at an early stage. AI first and AI native organisations are demonstrating new forms of speed, scale and operational leverage, but it is not yet clear which operating models will prove most resilient over time.
The question for leaders is therefore not simply whether to become AI first, but how fast and where to start.
The Forum argues that organisations treating the current period solely as an experimentation phase risk falling structurally behind competitors whose systems are already learning and compounding.
At the same time, the report does not assume that every existing operating model should be immediately replaced. One practical approach is to run parallel operating models, developing AI first workflows, teams and systems alongside existing ones and comparing their performance.
This allows companies to determine where intelligence creates genuine advantage and where more traditional methods still work effectively.
The larger objective is to build the organisational capacity to learn and adapt continuously as AI capability itself changes.
Key takeaways
AI first transformation begins when companies stop treating AI as another tool inside existing processes and begin redesigning the operating system of the enterprise around intelligence.
The intelligence engine becomes a compounding source of speed, scale and scope. A modular technology stack connects models to live data, context and orchestration while preserving enterprise control as vendors and capabilities change.
Operations are redesigned workflow by workflow. Intelligence is allocated towards the highest value areas, tasks are deliberately divided between humans and AI, verification is embedded into the workflow and ontology makes the organisation’s data, logic and possible actions legible to the system.
Human AI teaming changes the workforce model as well. AI absorbs more routine execution while human value moves towards problem framing, domain expertise, judgement, creativity and orchestration. Organisations must actively manage skill atrophy and redesign teams around continuous human AI collaboration.
The organisational model becomes increasingly federated, combining central infrastructure and governance with local ownership of execution and business value.
Trust, runtime visibility and human intervention remain essential as autonomy increases. Organisations need to know what agents are doing, which context they are using, how confident they are, when a person should intervene and what evidence supports their decisions.
The ultimate shift is therefore not from human work to AI work. It is from an enterprise where AI sits alongside the operating model to one where intelligence becomes part of the operating model itself.
