Enterprise AI reasoning substrate: PwC

Matt Sekol

Lead Driver, US Microsoft Alliance, PwC US

Email

Key takeaways:

Enterprise software still just records business activity. AI is exposing that recording and reasoning are different jobs, and bolting generative AI onto existing apps only makes recording faster, not smarter.

True cross-functional reasoning can't live inside any single application — it requires a shared "substrate": a unified layer of data, models, and governance that sits beneath and beyond individual apps.

Microsoft's version of this substrate combines OneLake/Fabric for unified data, Microsoft Foundry for frontier model access, and Entra ID/Purview/Agent 365 for governance.

As the substrate matures, SaaS applications are sorted into three postures: those that respect the substrate, those that claim to be one, and "walled gardens" that don't.

Four functions — risk, customer experience, cybersecurity, and sustainability — reveal whether an organization's substrate is ready.

Executive Summary

Enterprise software has evolved enormously over the past 25 years, from data centers to virtualization to SaaS. Its fundamental job has not. It still exists to record business activity: transactions, customers, inventory, employees, and the operational footprint of the business itself.

What has changed is the enterprise around those applications. Collaboration has created new processes and interlocks. Data sets have been progressively abstracted from their applications to match how companies actually work. You can find this in data lakes and warehouses cobbled together across the enterprise.

AI is now exposing the limits of this architecture. Enterprise software vendors are layering generative AI into their products, but that reasoning stays locked inside the application as an extension of the recording exercise, not a break from it. Ask the CRM about a customer's buying pattern and you save an hour, but the reasoning never gains organizational context and stays in the silo.

True enterprise reasoning is the next leap. It brings datasets together the way people and processes already work: to decide, forecast, recommend, and act. It cannot live inside any single application. It should live beyond them, in a shared reasoning substrate: the unified data, model, and governance layer that AI reasons on top of.

That is why the enterprise decision is moving past which app to buy, bringing scrutiny to lingering software contracts. The question leadership teams are now asking, or should be, is which substrate will carry the company forward because the substrate, not the application, is where the next decade of enterprise capability can compound.

Apps alone can’t reason like the organization does

Every enterprise today runs on purpose-built applications. CRM owns the customer view, ERP runs transactions and operations, HR systems own the employee record, and industry-specific solutions run core operational functions.

Yet corporate reasoning is inherently cross-functional. This became unmistakable as sustainability officers tried to measure carbon emissions and found themselves pulling information from disparate, disconnected systems to assemble a single organizational view. The same pattern had been quietly defeating customer-360, risk aggregation, and supply chain visibility efforts for far longer.

Many of today's enterprises function cross-functionally, but their technology stacks do not. Enterprise software produces localized, siloed intelligence. Everyone from front-line workers to executives should have context-rich, systemic intelligence.

So far, the response has been new AI-enabled software features. These are the final throes of the siloed application stack, not the next phase of it. The bolt-on is useful for incremental experience improvements; it cannot deliver the cross-functional reasoning the enterprise needs. For that, companies still fall back on the same cobbled-together lakes and warehouses they've long leaned on, which is an architecture not designed for this work.

Sustainability was a first indicator as a newer workload attempting to pull company data together. As AI advances, every cross-functional question across customer, risk, supply chain, workforce, exposes the same architectural gap. The application silos, the security posture, and the fragmented data estate that used to belong to IT are now business problems with deadlines attached.

The New Substrate: Lakehouse, a Frontier LLM, and Governance

The application layer is not where AI-driven organizational insights emerge. Nor is it where the future of employee and agent collaboration takes shape. Those will surface through a new layer of connected enterprise data, and the tooling required to build it inadvertently subverts to the traditional enterprise stack, abstracting the application layer in favor of intelligence.

This new reasoning layer requires three things that AI-enabled applications typically can’t deliver on their own:

A lakehouse that holds one governed copy of the enterprise's structured and unstructured data, paired with a semantic layer that understands organizational context across the surfaces where work happens. In Microsoft's stack, that is OneLake within Fabric, with Microsoft IQ: Fabric IQ for business semantics, Work IQ for M365 context, and Foundry IQ as the unified retrieval layer for agents providing the meaning the substrate runs on.

Models capable of cross-domain inference, accessible through an enterprise control plane rather than embedded inside individual SaaS apps. In Microsoft's stack, that is Microsoft Foundry, with frontier models from OpenAI and Anthropic alongside a broad and growing model catalog, each of them accessible under one governance posture.

Identity, lineage, auditability, observability, and policy applied to data, models, and agents alike. Governance can no longer live inside individual applications; once reasoning crosses application silos, governance has to cross them too. In Microsoft's stack, that is Entra ID, Purview, Agent 365, and the Foundry control plane working together, now extending to agent-specific controls as agents become the unit of reasoning.

Every serious enterprise platform is now converging on a version of this pattern: AWS with Bedrock and Anthropic, Google with Vertex and Gemini, and Salesforce reframing its entire platform as an agent-accessible substrate with Headless 360.

Microsoft's substrate is distinctive in spanning data, productivity, and agent runtime under one identity and governance plane, which is the combination that cross-functional reasoning often demands. The decision in front of leadership teams is which substrate to commit to, because the substrate, not the app, is where the next decade of enterprise capability will compound.

What Happens to the Application Layer

With a new substrate in place, enterprise software stops being the center of gravity. What remains useful about an application is largely its data model, and for vendors with proprietary models, the very thing that kept renewals coming becomes the thing that ends them. This is already visible in how vendors are repositioning. New capabilities around enterprise applications, like APIs, MCP tools, or CLI commands are growing, so agents can operate the platform without a browser, a recognition that the application's value now lies in what agents can reach, not in what humans can click.

SaaS is sorting into three categories:

Apps that expose their workflow, data models, and UX over company-owned data in the lakehouse. These compound value because they respect the substrate.

Vendors large enough to credibly position their own platform as a substrate. These bets often succeed for customers already heavily invested in that vendor's domain, but force the enterprise into a substrate decision shaped by one functional area rather than by how the company works across functions.

Vendors that insist on owning their data and bolting on their own AI as a value-add. Walled-garden vendors often find their solutions progressively disintermediated by agents that read and write directly to the substrate. Meanwhile, their data models, the one remaining asset of value, are simply replicated into the company's semantic layer and rendered redundant.

As contracts come up for renewal, IT, business, and procurement teams will likely be asking where each vendor falls.

The disintermediation runs deeper than the application layer. For the past thirty years, enterprise software represented the durable enterprise asset, replacing physical assets as value moved to intangibles. Going forward, the governed data layer, the semantic enterprise content, and the model's capabilities form a new digital portfolio.

In the near term, companies can buy the app, rent models, and own data. Over the long term, apps can churn, models can churn faster (frontier model generations are now compressing to weeks rather than quarters), and the substrate can accumulate more context. The asset that compounds is the one the enterprise owns.

This is where the payoff of enterprise agents lives, and why the substrate matters. A Copilot or Foundry-governed agent that can reason across CRM, ERP, HR, and other business data is only possible if those datasets live somewhere the agent can reach. Apps that wall off the data wall off the agent. When this happens, the organization falls back to human processes to inefficiently bypass them. That is what is at stake.

Three shifts in company posture follow, and they represent real changes in buying patterns:

Stop evaluating AI features app-by-app. Evaluate whether the app feeds and respects the substrate.
Treat Microsoft Fabric (or whichever lakehouse the enterprise commits to) as a board-level asset, not an IT line item. This is the final elevation of IT from cost center to value driver, for now.
Renegotiate or exit contracts with vendors whose AI strategy depends on keeping enterprise data inside their walls and scrutinize substrate-ambitious vendors for whether their substrate claim fits how your enterprise reasons across functions.

The procurement question is moving from "what does this app do?" to "can this app support our substrate?"

What the Canaries Reveal

Organizations are cross-functional by nature, but some functions encounter this shift earlier and more acutely than others. These are the functions where the substrate is tested first, because they simply cannot succeed inside any single application. If they struggle, it is an early indication that the substrate isn't ready, no matter what AI projects are sprawling across the company.

The canaries

Can the CRO see counterparty, operational, reputational, and climate risk in one frame? If not, risk reasoning is still isolated in spreadsheets.

Can an enterprise agent see a customer across CRM, support, billing, and product telemetry in a single reasoning pass? If not, the agent is likely just a chatbot about opportunities and pipelines.

Can the CISO correlate identity, data lineage, and AI agent activity in one governance plane while applying controls to agents themselves, not just to the data they reach? If not, AI governance is theoretical, and the broader security posture probably is too.

Can the CSO produce a supplier-emissions view without a multi-week scramble across internal systems and supplier surveys? If not, the substrate is fragmented.

Leadership implication

These four canaries reveal more than a substrate gap. Each reveal whether the executive team is operating with a shared architectural vision. If risk, customer experience, security, and sustainability are pursuing their own siloed data strategies, the substrate typically never forms, and the agents built on top of it can inherit the same fragmentation the organization is trying to escape.

The architecture diagram is a leadership imperative, not an IT artifact.

From Recording to Reasoning

The recording era is not ending because applications failed or were too cumbersome to manage. It is ending because the modern work of the enterprise has outgrown them. What sits ahead is a deeper shift than digital transformation delivered: true company-level transformation toward democratized reasoning.

This is no longer a forward-looking argument. Microsoft has now publicly committed to it. The hyperscalers and the largest application vendors are repositioning around it. The substrate components, like unified data, frontier reasoning, and governance that crosses agents and applications, are shipping into general availability this quarter, not next year. The quiet inversion described at the start of this briefing is no longer quiet.

Enterprises will adopt AI. Not all of the enterprises will reason. The difference lies in architecture, not the aspiration. It may start at the LLM, but the value accrues in the substrate.

Models can keep improving. Applications can keep churning. The decision that compounds across both is the platform decision, including how the substrate is governed, operationalized, and activated. That decision is now in front of leadership teams, and many of them have more of it already in place than they realize.

The quiet inversion is no longer quiet, and the signal can only grow stronger.