The first big fight for life after the smartphone

Apple’s new lawsuit against OpenAI is not just about whether a few former employees walked out with trade secrets. It is one of the first visible battles over who controls computing in a world where AI agents, not apps, sit between people and the digital systems they use every day.

For more than a decade, the smartphone has been the organising centre of personal computing. We live inside it. We tap icons, switch between apps, juggle notifications and authenticate transactions. But every major technology era has ended the same way: not with a formal announcement, but with leading companies quietly stepping beyond the boundaries that once defined them. Mainframes did not vanish when the PC arrived; they simply stopped defining what “computing” meant. Newspapers, retailers and travel agents did not close the day the web launched; they lost their central position over time. The smartphone, too, is unlikely to disappear overnight. What is at risk is its role as the place where digital interactions begin.

Seen in that light, Apple vs OpenAI is not just a legal case. It is a test of what comes after the app-and-screen era.

Apple’s complaint, filed in the Northern District of California on July 10, 2026, accuses two former employees and OpenAI-related entities of misappropriating confidential information. According to public reporting, Apple claims these individuals accessed internal systems without permission, downloaded sensitive materials, discussed unreleased projects in hiring conversations and were encouraged to bring physical parts or design artefacts to interviews. These are allegations, not proven facts, and they will be tested in court. But the breadth of what Apple describes is revealing. The company does not frame its secrets as a single component or one new product. It talks about entire systems: hardware architecture, manufacturing processes, specialised equipment, power and thermal management, supplier relationships and the finely tuned coordination required to ship new devices at global scale.

That scope tells us what Apple thinks is truly at stake: not just a design, but an operating model for building physical computing platforms that can support the next era.

The deeper question: who owns the agent?

The legal arguments will revolve around contracts, access logs and trade-secret law. Beneath that, the strategic question is simpler and far more consequential:

Who will own the AI agent that sits between human intention and digital action?

In the smartphone era, your intent flows like this:

• You decide what you want to do

• You pick up your phone

• You choose an app

• You follow that app’s steps to get to an outcome

A simple request such as, “Move my trip to Tuesday, keep the customer meetings, find a comparable hotel and tell everyone affected,” is scattered across multiple apps and services. You open the airline app, adjust the booking, find a hotel site, update your calendar, message participants and approve payments. You are the orchestration layer. In the agentic era, that structure flips. You express an outcome; the agent orchestrates the rest. It interprets your goal, checks your constraints, chooses tools, talks to systems and executes the workflow inside your permissions.

The phone is still present, but no longer necessarily in charge. It becomes one of several channels through which an agent sees the world and takes action. Owning that agent—the default layer through which people express intentions and let things happen—could turn out to be more powerful than owning any single app or even any single device.

On paper, Apple and OpenAI looked like natural partners. Apple controlled the hardware, operating system, identity, security model and customer relationship. OpenAI brought frontier models and a rapidly evolving AI stack. Apple could keep the user experience coherent while quietly invoking OpenAI’s intelligence when needed. That arrangement worked in the first wave of generative AI, when models mostly answered questions and stayed behind familiar interfaces. But it was never likely to remain stable once AI began to act on a user’s behalf.

A model provider does not want to live forever behind someone else’s operating system. It eventually wants direct access to context, permissions to act, and a path to transactions and ongoing relationships. A device company, in turn, cannot be comfortable if a third party’s intelligence becomes more important than the device itself. The tension is tolerable when AI is a feature. It becomes existential when AI is the new interface.

OpenAI’s hardware push changes the game

That is why OpenAI’s move into hardware matters so much. The company is working with former Apple design chief Jony Ive and other experienced hardware leaders on new consumer devices, including a smart speaker and related products designed around native AI capabilities rather than traditional apps.These projects are not side experiments. They are an effort to control the post-smartphone interface. 

The smartphone is app-centric. Users select an application, learn its interface and click through a series of steps. A truly AI-native device could be intent-centric. Users state what should happen; the system chooses tools, services and data sources, and quietly completes the work.

In that world:

• The model provides reasoning.

• The device provides presence.

• The operating system provides authority.

• Identity systems provide permission.

• Applications and services provide execution.

• Payments provide economic agency.

That stack makes the AI agent tangible. It lives in a device, has a continuous sense of context and can act within well-defined guardrails. The company that ties these layers together does not just ship a better assistant. It becomes a gatekeeper for how people interact with the digital and physical world around them. Apple and OpenAI are racing toward the same centre from opposite directions. Apple starts with devices and is pushing “up” into intelligence. OpenAI starts with intelligence and is pushing “down” into devices and operating environments. Their paths were always going to collide.

The post-smartphone world: survival vs centrality

The important question is not whether OpenAI can out-design Apple and produce a “better” phone. The question is whether the smartphone itself remains the organising centre of everyday computing. History suggests that mature technologies rarely vanish. Mainframes, radio and television all survived later waves. They just stopped defining their categories. A more useful metaphor is the transition from horses to automobiles. Early cars were expensive, unreliable and poorly supported; roads, fuel stations and repair networks took years to develop. For a long time, horses and cars coexisted. What changed was not the existence of horses, but their role in the transportation system.

The same pattern could play out with smartphones.

• The phone remains an excellent camera, secure identity device, communications hub and high-quality display.

• It may live in our pockets for many years.

• Yet it may gradually stop being the place where digital journeys begin.

Instead of asking, “Which app should I open?” people may increasingly state what they want to happen. The agent will decide which apps, merchants, services, devices and payment systems to use. The phone will still participate, but more as secure infrastructure beneath the agent than as the main stage. In strategic terms, the phone survives physically but recedes from the foreground of the digital economy.

For Apple, that is far more threatening than another smartphone competitor. The iPhone’s power does not come only from hardware sales. It comes from being the front door: controlling identity, app discovery, payments, notifications, sensors and access to services. If an external agent becomes the primary interface through which users express needs and delegate action, Apple risks being pushed one layer down the stack.

The device would still be in your hand. The control point would move above it.

Why Apple’s execution machine matters

Apple’s complaint underscores something the industry often overlooks: the difference between invention and execution. A small team can build a compelling prototype of a new device or agent. Turning that into millions of reliable, secure products is a different discipline entirely. It demands deep expertise in batteries, thermals, materials, acoustics, tolerances, logistics, failure rates and supplier ecosystems.

A chef can invent an extraordinary dish in a single kitchen. Building a global chain that reproduces that experience every day requires a different operating system: sourcing, training, equipment, quality control, distribution and culture. Apple’s advantage has never been just in the “recipe” of a single product. It lies in the “kitchen” it has built: the supply chains, factories, processes and disciplines that let it deliver integrated devices and services at planetary scale.

For any AI company moving into hardware, the journey from intelligence to physical execution is long. A powerful model is not a finished product. A captivating demo is not a manufacturing system. A conversational interface is not a device that can run 24/7 in the messy real world. That is why Apple’s institutional knowledge is so sensitive—and why, if Apple’s allegations are proven, it would see any leakage as an existential threat rather than a narrow IP problem.

The physical agentic stack

In earlier discussions of Agentic AI, the industry has drawn a line between conversational systems and operational systems. Generative AI makes intelligence accessible through language. Agentic AI makes that intelligence capable of pursuing goals, invoking tools, handling multi-step tasks and completing work within defined boundaries.

What Apple vs OpenAI highlights is that this “agentic” concept is now extending into physical computing.

For an agent to be truly useful in daily life, it needs:

• Identity, so it can prove who it is acting for

• Permissions and policies, so it knows what it is allowed to do

• Memory and context, so it can act consistently over time

• Tools and integrations, so it can actually change the world, not just describe it

• Hardware sensors, processors, displays and connectivity, so it can see, hear, speak and respond in real environments

In other words, the agentic stack is not only digital. It is physical.

The players who can integrate models, devices, identity, payments and services into a coherent, trustworthy system will not just ship gadgets. They will shape the norms and economics of the post-smartphone era.

Talent, boundaries and trust

The case also surfaces a long-standing tension in Silicon Valley: where professional experience ends and proprietary knowledge begins. The region has always thrived on mobility. Engineers move between companies, founders leave incumbents to start new ventures, and knowledge spreads through networks. But the AI race is tightening the screws. When an AI company wants to build hardware, it naturally targets people who have already done batteries, cameras, supply chains and factories. Likewise, when a device company doubles down on AI, it wants model researchers, inference engineers and agent architects.

The line is crossed when hiring is not just about skills, but about extracting internal methods, supplier details, confidential roadmaps or proprietary processes. Apple’s complaint describes recruitment conversations in which candidates allegedly discussed sensitive project details and were asked to bring physical components or design artefacts.

Regardless of how this specific case is decided, it signals the need for stronger governance:

• Interviewers must be trained to stop candidates from disclosing confidential information.

• Technical assessments should emphasise problem-solving over showcasing proprietary work.

• New hires should receive explicit guidance on what knowledge they can and cannot use.

• In sensitive areas, companies may need clean-room documentation to prove independent development.

As agents gain more autonomy, the provenance of models, data and even hardware design will become a central trust issue, not a back-office compliance task.

Beyond intelligence: trust as architecture

Agentic systems carry a different risk profile from chatbots. A chatbot can be wrong; an agent can take the wrong action.

It can move money, reschedule a surgery, cancel a flight, notify a client or expose sensitive data. Once intelligence is allowed to act, trust stops being a marketing message and becomes part of the system architecture.

That architecture includes:

• Identity and authentication

• Fine-grained permissions and policy controls

• Audit logs and explainability

• Human escalation paths and revocation mechanisms

• Clear data boundaries and minimisation practices

Apple has built much of its brand on secure devices, privacy and tight integration. OpenAI’s strengths lie in model capability and rapid product iteration with broad adoption. Both approaches now meet at the same point: this is about winning the trust this entity to be your default agent

A larger platform war opening up

Zooming out, the Apple–OpenAI conflict is a preview of broader shifts.

• Partnerships will become more unstable, as companies cooperate at one layer (say, devices and models) while competing at another (interfaces and agents).

• Economic power will move from attention to intention: from capturing clicks in apps to interpreting goals and executing on behalf of users.

• The most defensible positions will sit at the intersections: between models and devices, devices and identity, identity and payments, agents and services.

Consumer markets are only one front. Inside enterprises, the same race is underway among cloud providers, SaaS platforms, model companies and services firms, each trying to become the orchestration layer through which work flows. Whoever owns business logic and the experience layer for agents will wield outsized influence over the “agentic enterprise.”

In this sense, Apple’s lawsuit is not just a reaction to a specific set of employees or a single partnership. It is one of the first shots in a larger platform war, where the real prize is control of the interface between human intention and machine action in a post-smartphone era. The smartphone will likely remain in our pockets. But the battle now is over what sits on top of it—and who gets to choreograph the agents that increasingly run our digital lives.

The Halo Is Changing: AI, Accenture, and the Reinvention of IT Services

There are moments in every industry when the numbers matter, but the narrative matters even more. Accenture’s recent market reaction feels like one of those moments. This is still a formidable company with global scale, trusted boardroom access, deep enterprise relationships, delivery discipline, and a long history of navigating technology transitions. It has not suddenly become weak, nor has it forgotten how to serve clients. The sharper point is that investors are now asking whether the old IT services value equation deserves the same valuation halo in an AI-native world.

For decades, the global services industry carried a powerful halo. Scale was seen as strength. Headcount was seen as capability. Global delivery centers were seen as industrial muscle. Certifications, offshore leverage, utilization, partner tiers, and pyramid structures were interpreted as signs of maturity. Large transformation programs were seen as proof of customer intimacy. The ability to mobilize thousands of people across geographies became the defining proof that large services firms were indispensable to enterprise technology. That halo was not artificial. The industry earned it by helping enterprises move from mainframes to ERP, from ERP to cloud, from fragmented systems to digital platforms, and from manual operations to automated workflows.

But AI is changing the lens through which the industry is being judged. That does not mean services companies are becoming irrelevant. In fact, the opposite may eventually prove true. As enterprises move from AI experimentation to AI at scale, services companies may become even more central because they sit at the intersection of technology, process, data, governance, compliance, change management, and execution. What is changing is not the relevance of services. What is changing is the measure of relevance.

This is where Phil Rosenzweig’s idea of the Halo Effect becomes useful. When companies perform well, observers describe their leadership as visionary, their culture as disciplined, their people as energized, and their strategy as bold. When performance turns, the same organization is suddenly described as complacent, overexposed, bureaucratic, or strategically confused. The underlying company may not have changed in a day; the interpretation has. Performance contaminates judgment. During success, everything is covered in gold dust. During pressure, everything is seen through smoke.

The services industry must avoid both extremes. The old halo said large services companies would always win because they had reach, relationships, talent, process maturity, and trusted delivery. The new reverse halo says they are vulnerable because AI will automate the work on which they depend. Both views are incomplete. IT services are not dead. Enterprises are not suddenly simple. Legacy estates remain complex. Data is still fragmented. Business processes are full of exceptions. Cyber risk is increasing. Regulatory scrutiny is rising. SaaS, cloud, data, and AI ecosystems are becoming more interconnected, not less. What is fading is not services; what is fading is the comfort of undifferentiated, labor-heavy, effort-priced services.

The old IT services equation was built around effort: people multiplied by billing rates, utilization, and duration. The industry perfected this model. It industrialized talent sourcing, training, global delivery, quality processes, program management, and account mining. It became a remarkable execution engine for the enterprise world. For every major technology wave—cloud, SaaS, digital, cybersecurity, data, analytics—the answer usually involved more migration, more integration, more support, and therefore more services effort. AI is different because it does not simply create new work; it questions the unit of work itself.

If a code migration that once required 200 people can be done by 40 people plus AI agents, who captures the saved value? If testing, documentation, support, configuration, process mapping, and knowledge management can be compressed dramatically, does the services firm protect the old effort model or lead the new productivity model? The right answer is clear: services firms must lead the productivity shift. Firms that help clients compress work, improve quality, reduce risk, and accelerate transformation will not lose relevance. They will move closer to the center of enterprise reinvention.

One analogy may help. The traditional services model was like a large factory floor. When demand increased, you added more stations, more workers, more supervisors, more shifts, and more throughput. Scale mattered because throughput depended on coordinated human effort. The AI-native services model looks more like a power plant. The question is not how many people are standing on the floor; the question is how much energy the system generates, how efficiently it converts input into output, how reliably it runs, and how safely it operates under stress. Services companies cannot simply sprinkle AI tools across the old factory and call it reinvention. They have to redesign the machinery itself: pricing, delivery methods, workforce architecture, incentives, IP creation, governance, and client value measurement.

This is why Vishal Sikka’s warning deserves attention. His argument that services-led and software companies may need radical reinvention, even potentially outside the glare of public markets, should not be read only as a call to go private. It is better understood as a call for strategic freedom. Public companies are often rewarded for quarterly stability: predictable margins, utilization, cash conversion, dividends, buybacks, and incremental growth. But AI reinvention may require bold moves that create short-term discomfort—cannibalizing legacy revenue, shrinking effort-heavy work, investing heavily in AI-native platforms, changing incentives, retraining at scale, acquiring capability rather than capacity, and accepting temporary margin pressure. The larger question is whether services companies can act with the independence and speed required to disrupt themselves before clients or competitors do it for them.

The positive interpretation is that services companies are not starting from weakness. They have enormous advantages if they choose to use them differently. They understand enterprise complexity. They know how large companies actually operate. They understand messy legacy estates, regulatory constraints, process exceptions, security concerns, data fragmentation, and organizational resistance. They have relationships with CIOs, CFOs, CHROs, business heads, and transformation leaders. They know how to run programs across geographies, functions, and platforms. In an AI world, where the hardest work will be moving from pilot to production, these advantages can become even more valuable.

The Netflix analogy is relevant here. Netflix did not survive because it loved DVDs. It survived because it was willing to attack the DVD model while that model was still profitable. It moved from mail-order DVDs to streaming, then from streaming distribution to original content, and then into global entertainment infrastructure. The hardest time to reinvent is not when the old model is dead; it is when the old model is still producing cash. That is where many services companies are today. The traditional model is not dead. It still produces revenue, employs millions, serves clients, and creates value. But the next model must be built before the old one visibly declines.

The services industry also needs to avoid the classic mistake of judging performance only in absolute terms. A company can improve and still fall behind. Revenue can grow and still disappoint. Margins can remain healthy and still be repriced. Internal dashboards can show thousands of people trained on AI, hundreds of pilots launched, dozens of partnerships announced, and many internal productivity programs underway. But these are absolute measures. The real question is relative: are services companies moving faster than client expectations, faster than competitors, and faster than the rate at which AI is commoditizing traditional effort?

This is where the measurement system must change. In the old model, the industry measured headcount, utilization, pyramid ratio, billability, certifications, bookings, revenue, margin, and delivery efficiency. Those metrics still matter, but they are no longer sufficient. The new measures must include AI-led productivity, cycle-time compression, defect reduction, autonomous resolution, reusable IP adoption, revenue from AI-native offerings, client outcomes achieved, work eliminated through automation, agent governance maturity, and value delivered per employee. The industry does not need to abandon discipline. It needs a new discipline.

Another analogy helps here: the difference between a barometer and a thermostat. Many companies use AI metrics like a barometer. They report the weather: how many people were trained, how many pilots were launched, how many tools were deployed, how many partnerships were signed. That is useful, but it does not change the room. A thermostat changes the temperature. Services firms now need thermostat metrics: how much work was actually compressed, how much quality improved, how much speed increased, how much cost was avoided, how much revenue shifted to differentiated offerings, and how much measurable value was created for the client. The industry does not need more AI weather reports. It needs proof that the operating temperature has changed.

The future value equation for services companies will not be people multiplied by rates. It will be intelligence multiplied by trust. The winning firms will combine domain depth, data capability, AI-native delivery, reusable IP, governance, ecosystem orchestration, and change leadership. In banking, life sciences, manufacturing, insurance, telecom, retail, healthcare, and the public sector, AI cannot be deployed safely with generic coding skills alone. It requires business context, regulatory understanding, process redesign, exception handling, auditability, security, human oversight, and measurable outcomes. AI without domain depth is like a brilliant intern: fast, impressive, and occasionally dangerous. Enterprises will still need guides, architects, integrators, and accountable transformation partners. The opportunity is for services firms to become those partners at much higher levels of value.

This is why the next phase could put services companies at centerstage. The first phase of generative AI was model fascination. The second phase was pilot proliferation. The third phase will be enterprise adoption at scale. That third phase is where services companies matter most. Models alone do not transform enterprises. Copilots alone do not redesign operating models. Agents alone do not resolve data quality, security, compliance, integration, and change management. The real enterprise AI challenge is not creating impressive demos. It is making AI work reliably, safely, repeatedly, and measurably inside complex organizations. That is a services problem as much as a technology problem.

The center of gravity will therefore move from “AI tools” to “AI operating models.” This is where services firms can lead. They can help clients redesign processes around agents, modernize data foundations, integrate AI into enterprise applications, build control towers for agent governance, establish human-in-the-loop models, measure productivity, manage risk, retrain workforces, and convert scattered pilots into scalable transformation programs. This is not lower-value work. It is higher-value work. But it must be priced, sold, staffed, and measured differently.

There is also a human dimension to this transition. The IT services industry employs millions of people. Behind phrases like automation, productivity, compression, and AI leverage are careers, families, aspirations, and identities. Reinvention cannot simply be a financial exercise. Developers must become AI-augmented engineers. Testers must become quality architects. Business analysts must become process intelligence specialists. Project managers must become transformation orchestrators. Support teams must become automation supervisors. Domain experts must become agent trainers, model evaluators, and governance owners. The worst version of the future is AI as a blunt cost-cutting weapon. The best version is AI as a way to raise the quality, leverage, and dignity of work.

The final lesson from the Halo Effect is humility. Markets overpraise during good times and overpunish during bad times. A single stock reaction does not define the destiny of a company or an industry. Accenture and other leading services firms have reinvented many times before, and it would be unwise to underestimate their ability to do so again. But it would be equally unwise to dismiss the signal. AI is forcing the services industry to justify its value in a world where intelligence is abundant, automation is accelerating, and clients are under pressure to do more with less.

The old halo was built on scale. The new halo will be built on intelligent scale. The old model sold effort. The new model must sell outcomes. The old model celebrated utilization. The new model must celebrate work compression. The old model priced people. The new model must price value. The old model implemented systems. The new model must redesign work. The old model was powered by labor arbitrage. The new model will be powered by agentic leverage, domain depth, data readiness, governance, and trust.

This is not the end of IT services. It may be the beginning of a more important chapter. Services companies can move to centerstage in the enterprise AI era because they are uniquely positioned to translate AI potential into enterprise reality. But to do so, they must change what they measure, what they reward, what they sell, how they deliver, and how they define value. The halo has not disappeared. It is changing. The companies that understand this will not merely survive the AI transition. They will define it.

Labels: GenAI, Halo Effect, IT Services

Mastering Token Economics: How Agentic AI Reshapes Enterprise Opex and Strategic Control - Part 2

Part 1 was about operating models and talent, Part 2 is about money: the token economics of Agentic AI. As organizations move from human-centric opex (salaries, benefits, overhead) to agent-centric opex (tokens, APIs, compute, orchestration), they discover that the path is anything but linear.

On paper, the vision is enticing: a future mix where 20–30% of today’s headcount opex is replaced by spending on tokens, models, and data. In practice, enterprises face a messy middle: overlapping costs, surging token usage, unclear ROI per task, and governance models that lag the technology curve.

The goal of this second part is to reinterpret those challenges through your Agentic AI framework – not as obstacles, but as design levers that large enterprises can use to build structural advantage.

The cost-per-task paradox: Falling prices, stubborn bills

The research describes an uncomfortable reality: even as model prices fall sharply (roughly an order-of-magnitude per generation), the effective cost per task is often staying flat. Three forces drive this:

• Most organizations stay on the latest frontier model instead of downgrading once a new generation arrives.

• Tokens per query rise as agents tackle more complex, multi-step workflows involving tool calls, error correction, and large context windows.

• Usage expands rapidly once teams experience what agents can do, leading to more workflows being automated or augmented.

The net effect: token prices per unit decline, but total token consumption multiplies, leaving the overall bill “stubbornly high.” That is the cost-per-task paradox.

From an Agentic AI viewpoint, this is not surprising. As the capability frontier moves, enterprises naturally push more decision-making and autonomy into agents. The work itself becomes more complex, not less. Without intentional design, cost follows complexity.

Three levels of token economics: CEO, GM, and individual

The attached material frames token economics as a three-level problem: CEO, general manager (GM), and individual user.

• At the CEO level, the destination is clear: massively higher productivity and a shift in opex composition. Leaders are impatient with organizational friction, not with AI’s potential. In Agentic AI language, they are trying to move the entire enterprise from tool-usage to agent-orchestration.

• At the GM level, the problem is budgets and speed. Pilots show strong results, but scaling to thousands of users requires approvals, security reviews, and redesigns that don’t fit quarterly cycles. Token spend doesn’t neatly map to existing line items, creating friction in P&L management.

• At the individual level, token usage is extremely skewed. Early data indicates that the top 5% of users often consume more tokens than the remaining 95% combined. These are the “hero users” – superagents in human form – who often say, “I don’t need a team; they slow me down.”

In your Agentic AI framework, these three levels correspond to:

• Strategic agency (CEO level): where to bet, how far to shift decisions into agents, and how to sequence change.

• Operational agency (GM level): how to fund and govern an AI compute budget that cuts across traditional functions.

• Individual agency (user level): how to empower superusers without blowing up cost, and how to bring the rest of the organization across the experiential chasm.

Recognizing these three vantage points helps large enterprises avoid one-size-fits-all approaches to token economics.

What would actually swing the economics?

Several variables could radically alter the trajectory of token economics – positively or negatively. Through an Agentic AI lens, each is a design choice or external constraint that leaders must actively monitor and shape:

• Token cost trajectory: If frontier pricing falls by 10x per generation and usage does not scale proportionally, the economics tip quickly. So far, however, organizations keep shifting to newer frontier models while increasing task complexity.

• On-premise inference: Running open-weight models on enterprise-owned silicon could shift costs from opex to capex. This is underexplored but potentially a major lever, especially for high-volume, repeatable workloads.

• Model right-sizing: Matching model complexity and cost to task value is critical. A draft internal email does not need the most advanced reasoning model, whereas a customer-facing financial forecast might.

• SaaS and platform lock-in: Major enterprise platforms (CRM, ERP, ITSM, HR) control the data fabric. Their willingness to expose agent-friendly APIs and shift to consumption-based pricing will either accelerate or slow the opex shift.

• Organizational clock speed: Enterprises that decouple AI and org design decisions from annual planning will move faster than those anchored to rigid fiscal cycles.

• Regulation and trust: Even if agents can perform complex compliance or risk tasks, human sign-off is still required today. The speed of regulatory adaptation and trust-building sets an upper bound on agent autonomy.

• Quality at scale: Pilots are easy; running thousands of concurrent agent workflows with robust monitoring, error correction, and fallback mechanisms is hard.

Agentic AI enterprises treat these as a portfolio of levers, not as fixed background conditions. They experiment with on-premise inference, push vendors on APIs, build robust monitoring for quality at scale, and actively manage model selection rather than defaulting to a single model for everything.

Designing an AI compute budget: Treat tokens like cloud in 2015

One of the most actionable recommendations in the research is to create a dedicated AI compute budget, rather than funding tokens opportunistically from existing line items. This mirrors how leading enterprises treated cloud migration 10–15 years ago: as a transformation initiative with its own governance, metrics, and guardrails.

In an Agentic AI framework, this dedicated compute budget becomes the financial backbone for your agent fabric. It allows you to:

• Make deliberate trade-offs between frontier vs last-generation vs open-weight models.

• Fund high-impact, high-complexity use cases that require premium reasoning, while keeping routine workflows on cheaper models.

• Align token spend with strategic outcomes (e.g., time-to-market reduction, NPS improvement, risk reduction) rather than viewing it as generic “IT cost.”

Without this, AI investments become fragmented, and token spend is vulnerable to mid-quarter cuts that kill momentum. For large enterprises, this is a structural decision: treat AI compute as a shared, strategic utility, not as discretionary functional spend.

Portfoliating your models: Stop flying first class for every trip

The research suggests that organizations that “model-match” (choosing the right model for each task) can see 3–5x cost differences compared to those that use a single frontier model for everything. This is exactly where Agentic AI architectures can shine.

A practical portfolio approach looks like this:

• Frontier models for high-stakes, high-complexity tasks where quality and reasoning depth drive significant business value.

• Previous-generation or mid-tier closed models for moderate-risk, moderate-complexity tasks at scale.

• Open-weight models (possibly on-prem) for high-volume, low-risk tasks where latency and cost are more important than frontier capability.

From an Agentic AI perspective, this is essentially agent routing and arbitration. A top-level orchestrator agent decides which model to call based on the task type, required assurance level, and cost sensitivity. As the research notes, when a large telco reorchestrated its architecture so that a super-agent routed tasks to smaller, specialized models instead of pushing everything through frontier, it reported a 90% cost reduction and 3x throughput. That is token economics as system design.

Instrumenting cost per task: From blind spend to actionable telemetry

Perhaps the most important – and most underdeveloped – practice in enterprises today is instrumenting cost per task, per workflow, per outcome. Many organizations have no idea what it costs, in token and compute terms, to:

• Generate a proposal

• Resolve a Tier-1 support ticket

• Produce a first-draft contract

• Run a specific analytics scenario or simulation

Without this telemetry, enterprises are optimizing blind. In an Agentic AI framework, cost-per-task instrumentation becomes part of the agent observability stack. Alongside quality metrics, latency, and error rates, you track economic metrics at the same level of granularity. That enables:

• Rational decisions on where to apply frontier vs cheaper models.

• Real-time governance on superusers’ token consumption without blunt caps that kill high-value experiments.

• Dynamic routing and throttling based on budget constraints, not just technical constraints.

For large enterprises, building this metering early is painful but essential. Retrofitting cost observability after hundreds of agent workflows are live will be far more expensive.

Planning for the dual-cost transition: Paying for people and tokens

There will be a period – likely several years – where enterprises pay for both the legacy workforce and a rapidly growing token bill. The research is explicit: this is not a smooth glide path; it is a nonlinear transition full of cost overlaps and quarters where the math looks ugly.

In Agentic AI terms, this is the cost of parallel agency:

• Human agency remains fully in place (existing roles, org charts, and processes).

• Agentic agency is being layered on top (agents embedded in workflows, pods running faster, new AI-native initiatives).

Executives should treat this as a deliberate investment phase, not as a failed cost-saving exercise. That requires:

• Mapping where overlaps are most acute – which workflows will still have full human teams while agents come online.

• Setting explicit ROI expectations and timeframes for each major AI investment.

• Communicating to boards and investors that this is a structural transformation, not a short-term headcount arbitrage.

In practice, this looks like the “Next Monday” style prompts from the research: pull your top SaaS contracts and current token spend, instrument one end-to-end workflow, and use that as a reference point for broader planning.

Opportunities for large enterprises: From cost management to strategic leverage

If we step back, the emerging picture for large enterprises is not just about surviving token economics; it is about using them as a strategic lever in your Agentic AI journey.

Here are the core opportunities:

• Design a token-aware Agentic architecture: Agents route tasks to appropriate models based on value, risk, and cost. This turns token economics into a design parameter, not a post-hoc surprise.

• Use token telemetry to prioritize transformation: High-cost, high-frequency workflows become priority candidates for redesign, on-prem inference, or specialized models.

• Build a new governance layer around AI compute: A dedicated budget, clear guardrails, and cross-functional oversight avoid both overspend and under-investment.

• Leverage dual models to accelerate learning: While legacy and agentic models coexist, you can compare outcomes, costs, and cycle times, building a rigorous evidence base for scaling decisions.

• Signal to talent and vendors that you are an Agentic enterprise: By redefining roles, renegotiating SaaS contracts around API usage, and openly investing in agent-first capabilities, you position your organization as a preferred destination for top talent and cutting-edge partners.

For CXOs, the most important mindset shift is this: token economics are not a separate finance problem to be “handled” after AI deployment. They are integral to how you design Agentic AI into your enterprise – from architecture and operating model to talent and strategy.

Labels: Agentic Advantage, Toeknomics