How AI stacks are rewriting the foundations of business

The shift from on-premises computing to software as a service modified the technology model and compelled information technology organizatios to modernize the way it builds, buys and operates software.

It also reshaped how software vendors price, deliver, and add value. But for many buyers, SaaS didn’t fundamentally change how the corporate made money or how work got done day after day. The business became more agile and IT became less of a friction point, however the operating model of the enterprise largely stayed intact. SaaS transformed software corporations and IT departments first – and mostly stopped there.

How this AI wave is different

Artificial intelligence is different. In our view, this wave reaches past the IT function and into the core mechanics of the enterprise – how decisions get made, how work gets executed, how risk gets governed and the way capital gets allocated. The rationale is most corporations still run as federations of departments with their very own application stacks, their very own data models and business logic trapped inside systems that were never designed to coordinate as one. Those systems might be “deterministic” inside a silo, but across silos the enterprise runs on reconciliation – spreadsheets, meetings, approvals and tribal knowledge. That’s not an IT problem. That’s an organizational tax.

The promise of AI is to eliminate much of that tax. Not by sprinkling copilots onto yesterday’s apps, but by bringing probabilistic intelligence into the enterprise in a way that’s governed by deterministic constraints. Frontier models are the core engine of this shift – they’re getting more capable and more functional, and they’re going to remain a linchpin of the longer term software stack.

But model power alone doesn’t solve the enterprise problem. The winners will likely be the organizations that construct an entire system around those models – one which connects to existing deterministic applications, creates a shared truth layer, controls agents as they take motion, and uses human feedback to repeatedly improve.

For this reason we consider the upside is so large. Enterprises that get this right won’t just run more cheaply – they may run in a different way. They are going to scale with less proportional labor growth, compress cycle times from insight to motion, and begin to behave more like platform corporations, with compounding advantage that’s difficult for competitors to repeat.

On this Breaking Evaluation, we use George Gilbert’s model to explain how that transformation unfolds – and why it requires an entire stack revolution to supply:

• A strategy to preserve and extend deterministic applications while probabilistic systems reason across them;
• A System of intelligence layer that harmonizes enterprise truth in real time so agents can act with confidence;
• An agent control and engagement loop that keeps humans within the approvals, exceptions and learning cycle.

Enterprise software today and the deterministic myth

The slide below gets to the core tension we keep coming back to in Breaking Evaluation” Enterprises want deterministic outcomes, but they try to bolt probabilistic systems onto an environment that isn’t actually deterministic in practice. Generative AI is probabilistic by nature. If agents are going to take motion confidently, the encircling system has to supply guardrails, shared semantics and predictable recovery. The issue is that the “deterministic software stack” most enterprises point to is actually a jungle of disconnected application islands as shown here.

On the left above are the appliance domains (enterprise resource planning, customer relationship management, finance, supply chain management, human resources, security, analytics, manufacturing and more). In the center is the actual integration layer enterprises depend on today – human experts with their very own tools (for instance spreadsheets) interpreting exceptions, meanings, approvals and reconciliations using tribal knowledge. On the appropriate are the business friction that falls out from that approach:

• Delayed truth: The “truth” shows up late since it’s reconciled after the actual fact.
• Conflicting semantics: The identical term can mean various things across systems and departments.
• High coordination cost and a number of human hours wasted: Meetings, spreadsheets, rework and escalation paths change into a part of and essentially define the operational model.
• Manual recovery: When something breaks, humans reconstruct state and intent.

There’s a degree here that we predict is deceptively necessary. Although the applications themselves are deterministic, the quantity of human semantic glue required to make end-to-end processes work means the final result becomes effectively probabilistic. Same inputs don’t reliably produce the identical outputs because interpretation, exception handling and cross-department coordination vary by person and circumstance. We compare this to a craft-work economy – before the assembly line, every part was barely different, so every finished product was barely different. That’s what enterprise operations seem like today in lots of corporations – deterministic machines, craft processes.

The implication is economic. When outcomes depend upon human glue, costs rise with revenue because scale normally means more coordination, more exceptions, more specialists and more management overhead. The goal, in our view, is to make knowledge work more repeatable: Add leverage to the economics while also increasing quality. There’s also a very important nuance we’d like to emphasize in that the destination isn’t “industrial sameness.” If this is completed appropriately, corporations can get repeatability and customization – differentiated outcomes delivered with the economics of repeatability.

Let’s ground this in easy terms: Craft economics shows up as labor. At volume, marginal economics look more like a services business than a software business. That’s the reason we keep using the phrase service as software. The thesis is that as these islands get harmonized and integrated, more corporations can operate with software-like marginal economics – and that changes operating models and business models, not only IT architectures. In the very best case, corporations start behaving like platforms – and that dynamic shows up in every industry.

Before debating which model, which agent framework or which toolchain “wins,” enterprises must internalize what this slide is actually saying. The present stack is held together on the seams by people. The brand new platform has to automate the seam work – semantics, reconciliation, coordination and recovery – so agents can act inside a system that’s finally deterministic where it matters, while still profiting from probabilistic intelligence where it adds leverage.

The service as software (SaSo) tech stack

Let’s dig into the technology model and the necessary changes we see coming. That is where the “service as software” idea becomes real. The shift from on-prem to SaaS forced vendors and IT teams to re-architect and retool – however the impact mostly stopped there. What this technology shown within the slide below implies is that agents will change the client’s operating model, not only the seller’s delivery model. Any company that desires agents to do greater than automate small tasks has to rebuild how the business coordinates work across silos – because end-to-end outcomes don’t live inside a single application boundary.

At a high level, the slide above lays out a full stack for an agentic enterprise – and it’s deliberately not only a brand new AI layer bolted onto an existing x86-powered, general-purpose stack. It’s a model where deterministic and probabilistic systems coexist, where the agent loop is grounded in shared semantics, policy and real-time state, with human feedback. Let’s step through the blocks:

• Historical system of truth: The information platform world (snapshots, governed metadata, “what happened” analytics).
• Real-time system of truth: Systems of record (operational truth, but still siloed by app boundaries).
• System of intelligence (SoI): A dynamic model of the business that harmonizes entities, meaning, rules, and state across the enterprise.
• System of agency: Agents that perceive, reason, determine, act and learn – orchestrated against business outcomes, not only app workflows.
• System of engagement: The human and application touchpoints — people, dashboards, notifications, approvals – and critically the feedback loop that teaches the system over time from the reasoning traces of humans.

The underside line is service as software isn’t a an large language model bolted onto a CRUD database. It’s a replatforming of how enterprises represent truth, coordinate motion and compound learning – and it requires a brand new middle layer that almost all enterprises do not need today.

System of intelligence (SoI): A brand new, high-value layer within the AI stack

We consider the green layer is the cash layer – and it’s also the missing layer. Everyone talks about agents. Everyone talks about orchestration. But agents that aren’t grounded in a shared enterprise model mostly produce local automation – not enterprise transformation.

For 60 years, we built silos – analytic silos and operational silos. If an agent goes to transform the business, it has to perceive and reason across those silos. Otherwise you’re just accelerating work contained in the same fragmented structure. That’s why the SoI sits between “systems of truth” and “system of agency.” It’s the piece that turns a pile of applications right into a coherent enterprise.

This can also be where the slide’s system of intelligence details change into critical. The SoI is what upgrades the enterprise from retrospective dashboards to forward-looking guidance, addressing not only what happened but:

• Why did something occur?
• What’s prone to occur next?
• What should we do now?

Our view is that you just don’t get credible answers to those questions at enterprise scale without harmonizing semantics and rules across applications. Business logic is trapped today inside ERP, CRM, human capital management, security platforms, data stacks and departmental workflows. “Single source of truth” has been promised for many years by the tech industry; but the truth is humans still reconcile the seams – with tribal knowledge, meetings, spreadsheets, exceptions and escalations.

One nuance that’s easy to miss is the SoI isn’t only “data as an asset.” The industry learned that lesson many years ago. The following step is rules as an asset – extracting and normalizing the business rules embedded in operational systems so the enterprise can execute and reason consistently across domains. That’s a significant a part of what “digital twin of the enterprise” actually means in practice – a real-time representation of state, entities and processes, not only a prettier analytics layer.

The important thing takeaway we wish to emphasize is that the SoI is the control point that lets agents act confidently across the enterprise. Without it, you’ll see a whole lot of agent pilots and point automations – and little or no durable, cross-functional transformation.

System of engagement and agency

This stack slide also makes a degree that always gets lost in agent platform marketing. Specifically, learning isn’t optional. The system has to enhance because it runs. That’s why we keep coming back to the “system of engagement” box within the diagram. We would like to make clear this isn’t social media engagement. It’s a feedback system that’s proprietary to a firm.

The sensible dynamic is that agents will operate until they hit uncertainty – then humans step in. The bottom line is what happens next. We envision a loop where exceptions change into training fuel – the system learns from human approvals, corrections, reasoning traces and escalations. For instance, early Tesla Autopilot wasn’t good since it was perfect on day one. Reasonably it got good because thousands and thousands of edge cases were captured, learned from and fed back into the system. Enterprises need the equivalent learning loop for business operations.

This can also be where the “system of agency” box earns its cred. It’s not only task automation. It’s adaptive agents orchestrated against outcomes, with the power to perceive state, reason, determine, act after which learn. If the SoI is the shared semantic truth and business-state layer, the system of agency is where work gets executed and coordinated – and where the business advantages start showing up in throughput, cycle time and organizational leverage.

The important thing point we wish to emphasize is that the agent era is not going to be won just by higher prompt engineering. It is going to be won by enterprises that construct a closed loop model with shared state (SoI), governed motion (agency) and continuous learning through engagement.

The green layer land grab

A key strategic point that we consider is directionally right is that each major platform vendor is now jockeying for the green layer – since it becomes the highest-value piece of real estate within the stack. The controversy is which corporations have it and where it lives. Specifically:

• Does it emerge inside application silos (CRM, ERP, IT service management, HCM – each constructing their very own “mini-SoI”)?
• Or does it change into an enterprise-wide layer that sits above and harmonizes across those systems?

We contend that the one company that has shown real traction constructing something like that is Palantir Technologies Inc.– however the model is pricey and never yet broadly repeatable. We liken what Palantir are doing to SAP SE’s early days where a heavy, customer-by-customer buildout was obligatory. Palantir must depend on forward deployed engineers at roughly $950,000 per 12 months per engineer. This isn’t a mass-market operating model. That’s why we regularly bring Celonis SE into the discussion since it’s attempting to show process-level harmonization into something more repeatable and accessible across the silos.

This can also be where pricing power enters the story. If you happen to own the layer where agents actually execute end-to-end work – and you’ll be able to see that work – you’ll be able to price on outcomes. If you happen to’re below that layer (feeding data, feeding context, supplying tools), the pricing looks so much more like consumption and utility pricing. Salesforce Inc.’s bid to push toward “Headless 360” so any agent can access customer context without the user interface is an interesting strategic move and signals a broader shift away from seat-based pricing toward usage and outcomes as agents change into the users.

We cite a useful discussion from VeeamON because it pertains to the brand new AI stack. Chief Executive Anand Eswaran put forth the argument that the missing layer in AI is “data and AI trust.” The more nuanced calibration in our view is that trust, compliance, privacy and security are essential – but they’re supporting infrastructure across the SoI, not the SoI itself. The SoI is the missing AI layer that can drive productivity. These other elements are a part of the “supporting forged” of the SoI that make agents secure to deploy at scale. The green layer remains to be the system that represents enterprise state and semantics – and without it, “trust” doesn’t have anything coherent to control.

The underside line is this can be a land grab for the longer term platform layer. The winners will likely be the businesses that could make the green layer real, enterprise-wide and repeatable – not only bespoke – while pairing it with the harnesses that make agent-driven operations governable and secure.

The north star: A full-stack digital twin

The slide below is where the conversation focuses on the specified end state. It’s not only a cool architecture diagram – it’s a north star for a way the AI stack has to evolve if enterprises actually want agents making decisions, taking motion and improving over time without blowing up the business. The slide calls it The Full-Stack Digital Twin – Architecting the Expertise Refinery and that phrase has intending to us. The purpose isn’t that each company is about to construct a twin. The purpose is that the enterprise goes to must manufacture intelligence – after which refine it into repeatable outcomes the best way a factory refines raw inputs into products.

At a high level, the slide is split into two halves:

• The deterministic digital twin at the underside – the “scaffolding” – where the enterprise defines what’s true and what’s allowed;
• The cognitive digital twin at the highest – the “crystallization” – where the system captures how experts make judgment calls when the foundations aren’t enough, after which learns from it.

And the important thing idea is straightforward but difficult to execute. Specifically, you don’t get the cognitive twin without the deterministic twin. The context graph talk that went viral earlier this 12 months – “tacit knowledge,” “tribal knowledge,” “why behind decisions” – won’t stand by itself. In our view, it only becomes economically viable when it’s integrated with the deterministic foundation that defines the state of the business.

Deterministic digital twin first – then the ‘why’

The five layers on the left side of the slide are essentially a maturity ladder. The underside two layers are the deterministic foundation – and that’s where George thought he was “done” earlier within the 12 months until the industry began fixating on the context graph problem: What happens when rules break down, conflict or don’t exist?

Here’s the structure the slide above lays out – and it’s price being explicit because each layer builds on the one below it:

• Layer 1 – Mapping layer brings canonical identity across systems. That is the “Rosetta Stone” step – the identical business object recognized across dozens of disconnected apps.
• Layer 2 – Rules layer brings prescriptive business and regulatory rules. The mandatory constraints – sequencing, approvals, compliance.
• Layer 3 – Institutional memory is the evidential authority – capturing the record of how experts reasoned about past decisions, not only what they decided.
• Layer 4 – Decision guidance is the advisory authority – synthesizing that memory into recommendations and confidence, in context, in real time.
• Layer 5 – Learning and feedback is the adaptive authority – scoring reasoning quality, detecting drift, and feeding learning loops so the system improves repeatedly.

The necessary nuance is that enterprises can’t jump straight to Layers 3 to five by “sprinkling AI on top.” The cognitive piece only works when it’s anchored to the deterministic state of the business – otherwise you’re just collecting narratives and not using a machine-verifiable foundation.

The associated fee problem and why deterministic scaffolding changes it

George’s research is grounded in a principle that we predict many enterprises are still underestimating. Specifically, frontier labs are spending enormous sums capturing reasoning traces – essentially paying for expert teaching at scale. Enterprises aren’t doing that today except in narrow pockets because the fee and burden of teaching is just too high.

The claim here isn’t that enterprises must act like frontier labs. The claim is more practical in that when you arrange expert teaching contained in the deterministic twin, business process by business process, you cut the surface area of ambiguity. Experts don’t have to clarify every part – they explain the exceptions, where the foundations stopped being sufficient. That’s where the economics change, and why this “expertise refinery” concept isn’t only a research idea – it’s a path to turning judgment into an asset.

What’s actually within the ‘scaffolding’

That is where we wish to push hard, because “scaffolding” is becoming a hand-wavy term out there. Vendors need to position themselves because the missing layer. Veeam Software Group GmbH, for instance, is leaning into “trust” – compliance, governance, security, recovery – and we pressed on whether that trust layer is a component of the scaffolding or something else contained in the SoI. We also identified that Dell Technologies Inc.’s Dataloop acquisition has a knowledge graph component, and will contribute to its cyber resilience practice. It’s hard to consider other data protection players resembling Cohesity Inc., Commvault Systems Inc. and Rubrik Inc. aren’t considering along similar lines.

Regardless, the migration path starts with what customers are asking for immediately: Dimensional semantics. Metrics and dimensions. Standard definitions of things resembling bookings and RPO. That’s real work, but it surely’s not the tip destination.

The destination is what we call out as stateful rules – the moment when rules aren’t just definitions living in a catalog, separate from the information, but are combined with the live state of the business. That’s when the digital twin becomes real, because only then can the enterprise ask and answer questions like:

Why did this occur, what’s prone to occur next and what should we do next?

If rules and state are separate, you don’t have a system that may reliably answer those questions. You could have documentation.

This has potential implications as well for recovery. Today, recovery occurs at the information level (for instance, recuperate a file or dataset). But there’s no notion of process logic in that recovery. In the longer term, business resilience would require granular recovery of the state of the business – including not only what an agent did when, but additionally why an agent took an motion and the corresponding logic behind it.

Governance shifts from ‘who can see what’ to ‘what are you able to do’

We also dig into governance because that is where the agentic era challenges old assumptions.

Governance today is basically resource-based – who can access this dataset, this column, this row. That model doesn’t disappear. But it surely isn’t sufficient once agents are doing work, since the whole point of an agent is you don’t know exactly what it would do ahead of time.

So governance has to evolve into something that’s intent-based – policy encoded into the dual that constrains what actions an agent is allowed to take inside an activity space. For instance, which tools it might probably invoke, what actions it might probably perform, what boundaries it might probably’t cross. In our view, that is the start of the control plane that makes agents secure at enterprise scale.

This “North Star” slide is the blueprint for where the AI stack is heading – and why the “missing layer” debate is getting noisy. The deterministic twin is what makes agents secure. The cognitive twin is what makes them useful at scale. And the critical components of trust, governance, resilience, recovery – becomes the ingredients that allow enterprises move from experiments to real operating model change. We consider the vendors that understand their role in that system – and don’t confuse scaffolding with the system of intelligence itself – have a likelihood to expand their total available market dramatically as enterprises start treating “state of the business” as something that have to be managed, governed and eventually recovered with the identical seriousness as data.

The economics of a brand new operating model

The slide below brings us back to business impact. The left side says capital expands – AI factory capex stacks on top of (and eventually eclipses) the legacy x86 refresh. The suitable side says coordination labor falls – reconcile, interpret, approve, recuperate, integrate, monitor – because an AI semantic layer starts automating the “human glue” that holds fragmented enterprise software together. In our view, that’s the actual platform shift – not “more servers,” but less human coordination wrapped around brittle systems.

Tokens change into a P&L line item

Jeff Clarke put the premise on the table at his Dell Technologies World 2026 keynote this week: Tokens change into a line item on the profit-and-loss and the operating model changes. We agree. Enterprises will get the productivity boost one in every of two ways: They either construct AI factories and manufacture tokens internally, or they tap intelligence through application programming interfaces and neoclouds. Either way, the purpose is identical: Scale revenue without scaling labor.

That’s why we keep coming back to the stack discussion from the prior slides. The AI factory produces intelligence, however the operating model impact is seen when that intelligence starts coordinating what humans do today on the seams – reconciling data, adjudicating exceptions, managing recovery, pushing approvals and translating insight into motion in real time.

Client Zero proof: Dell compresses weeks into minutes

A reputable here wasn’t a vendor promise – it was Doug Schmitt, Dell’s chief information officer and president of services describing how Dell itself is changing the best way work gets done. The instance is familiar:

A cross-functional meeting (logistics, parts, finance, sales, go-to-market) gets stuck on a basic query because the information doesn’t reconcile. People argue about whose numbers are right. Someone gets assigned to “go find the reality.” Every week goes by. They arrive back with a brand new cut. It’s still not the appropriate cut. Now it’s email tennis. Two weeks later, perhaps you finally converge – and by then the business has moved.

Schmitt’s point was that Dell built a knowledge mesh (its term for connecting their data together) and now shows up with a “single version of the reality.” Within the meeting, they don’t wait every week – they prompt and reprompt in real time until they get the appropriate answer (for instance, “not city-level, county-level”). The loop compresses what used to take days or perhaps weeks right into a single session – 15 to twenty minutes, perhaps a half-hour. That’s an example of an operating model changing within the room.

From org charts to shared truth and final result alignment

The “why” behind that Dell anecdote is the difference between:

• A bunch of highly productive individuals, and;
• A company with a shared model of the state of the business that orients everyone’s activity toward collective outcomes.

In our view, that is where the system of intelligence earns its value. It’s not a dashboard. It’s not “a greater warehouse.” It’s a platform that matches the inputs (information+agents+human expertise when needed) to the final result the business is trying to attain. Once we say “platform,” we’re being literal – core functions are additive and compatible with the model.

We also challenge a narrative that’s getting oversold immediately – the “single-person billion-dollar company” trope. We’ve heard versions of this movie before (the “solo merchant” era of early web hype) and what we got wasn’t a universe of solo merchants – what we got was Amazon. The rationale is personal productivity isn’t the identical thing as coordinating planning, control and resource allocation for collective outcomes. Firms exist because coordination is difficult – and useful.

So the directional bet isn’t “corporations get smaller.” The directional bet is that winning firms get greater and more platform-like because they’ll coordinate agents and folks through a system of engagement on top of shared truth – without counting on the old org chart as the first coordination mechanism. In that world, the org chart gets subordinated to a hierarchy of business metrics:

• activity-level metrics;
• process-level metrics;
• North Star metrics.

And when you try this, you get the platform economics we laid out:

• High fixed costs to rise up the system of intelligence and encode expertise;
• Low and declining marginal costs because human expertise gets pulled in mainly for edge cases;
• Compounding advantage as every transaction and each expert teaching session adds to the system’s value.

Bottom line

Our view is the “AI operating model” is a migration from human-driven coordination across fragmented systems to AI-mediated coordination built on shared truth. That’s why Clarke’s prediction of tokens on the P&L resonate. It forces leadership teams to treat intelligence like a production input, not an experiment – and it forces the enterprise to confront the actual bottleneck which is coordination cost.

Next week we’ll go deeper and address the query: How far do data platforms evolve into the system of intelligence – and what happens when “observability” stops meaning “Datadog for apps” and starts meaning “Datadog for agents,” with the information volumes, evals and learning cycles that include it.

Image: theCUBE Research

Disclaimer: All statements made regarding corporations or securities are strictly beliefs, points of view and opinions held by SiliconANGLE Media, Enterprise Technology Research, other guests on theCUBE and guest writers. Such statements aren’t recommendations by these individuals to purchase, sell or hold any security. The content presented doesn’t constitute investment advice and mustn’t be used as the idea for any investment decision. You and only you might be liable for your investment decisions.

Disclosure: Most of the corporations cited in Breaking Evaluation are sponsors of theCUBE and/or clients of theCUBE Research. None of those firms or other corporations have any editorial control over or advanced viewing of what’s published in Breaking Evaluation.