What the Best Business AI Users Are Doing Different

Context

This episode uses KPMG's Q3 AI Pulse survey to answer a practical question: what are the companies getting real AI returns doing differently from the ones still experimenting. The host focuses less on which model is best and more on what sits above the models, including agent adoption, formal AI interfaces, sovereignty policies, routing, and cost management. That matters to product managers because enterprise AI advantage is shifting from tool access to operating design. The survey covered more than 2,100 senior leaders across 20 countries, and the comparison between experimenters and organizations with established AI ROI gives PMs a concrete maturity map.

The Big Idea

The strongest enterprise AI results are coming from companies that treat AI as an operating system problem, not a tool rollout. They build a management layer around AI, tie spend to value, and use that foundation to move from simple productivity gains toward revenue and resilience.

The important shift is that maturity no longer looks like having more licenses or more pilots. It looks like formal governance, controlled interfaces, multi-model strategy, agent adoption, and clear economic discipline.

Key Insights

Maturity now lives above models

• The clearest pattern in the KPMG data is that the best-performing organizations invest in the management layer around AI, not just the models themselves.
• Mature organizations are more likely to have formal, enterprise-wide approaches across experimentation, strategic planning, scaling, adoption, and ROI measurement.
• The host's argument is that this is the new template for business AI adoption. Competitive advantage is shifting from raw access to intelligence toward the systems that make intelligence usable inside a real company.

Agents reveal the adoption gap

• Among organizations with established AI ROI, 48% report significant employee adoption of AI agents.
• Among organizations still in experimentation, only 15% report significant employee adoption of AI agents.
• The number matters because agents are no longer a fringe experiment. In the host's framing, agents became real in 2026, and widespread agent usage is now a marker of organizational maturity.
• The broader lesson for PMs is that a company can look interested in AI but still be far from operational adoption if agents have not moved into everyday workflows.

Harness layers enable controlled scale

• 86% of organizations with established AI ROI have a formal AI harness layer. The host defines this as some structured interface through which employees interact with AI.
• Only 31% of experimenting organizations have that kind of formal layer.
• Mature organizations also push accountability upward. 53% place responsibility for AI-informed decisions at the C-suite level.
• The PM takeaway is that scale needs a designed interaction model. If employees access AI through scattered tools, unmanaged prompts, and one-off workflows, the company will struggle to govern quality, security, and repeatability.

Sovereignty and routing are differentiators

• Mature organizations are much more likely to think strategically about where data goes and which model should handle which job.
• 53% of organizations with established AI ROI have an enterprise-wide data sovereignty strategy. Only 8% of experimenting organizations do.
• 23% of organizations have already built model routing capability.
• The host connects this to recent platform moves. OpenAI's new enterprise marketplace lets companies spend committed budget on other open models while staying inside OpenAI's ecosystem. That is a sign that enterprise buyers increasingly want multi-model flexibility, even when vendors would prefer lock-in.
• The PM lesson is that model choice is becoming a product and platform decision, not just a procurement choice.

Goals are shifting past productivity

• Productivity is still important, but it is becoming a smaller share of the enterprise AI story.
• KPMG found that productivity as a key goal fell from 42% in Q1 to 37% in Q3.
• Other priorities are rising:
• Human-AI collaboration rose 4 points.
• Responsible AI, governance, trust, and security rose 4 points.
• Adaptability and resilience rose 3 points.
• KPMG also reports that a quarter of organizations are developing multi-agent systems and a fifth are orchestrating multiple agents.
• The strategic shift is that businesses are using AI less as a narrow efficiency tool and more as part of how the organization operates, responds, and grows.

Cost discipline gets more precise

• The host argues that the real shift is from cost control to AI economic management.
• Mature organizations are not just watching spend. They are trying to link spend to value:
• 48% of organizations with established AI ROI consistently assess value against cost.
• 77% of organizations with established AI ROI have AI cost monitoring dashboards, versus 43% of experimenting organizations.
• 73% of organizations with established AI ROI do cost review at the approval stage, versus 50% of experimenting organizations.
• 46% of organizations with established AI ROI use usage or token budgets, versus 31% of experimenting organizations.
• Importantly, tighter controls are not leading to less ambition. Average planned AI investment over the next 12 months rose from $186 million in Q1 to $210 million in Q3.
• For PMs, the point is that better measurement is enabling more confident investment, not just cheaper usage.

Mental Models & Frameworks

Experimenters versus ROI operators

This episode's most useful comparison is the split between organizations still experimenting and organizations that have already established AI ROI. Use that split as a maturity lens:

• Experimenters are still proving basic value. They may have pilots, scattered tools, or local wins, but they usually lack enterprise-wide systems.
• ROI operators have moved from curiosity to operating model. They are much more likely to have harness layers, sovereignty strategy, routing, cyber defense, dashboards, and approval processes.
• For a PM, this is a better way to assess AI maturity than asking whether the company has launched a chatbot or bought a model subscription.

The AI management layer

The host's core framework is that the most important work now sits above the model. That layer includes:

• A formal interface or harness for how employees use AI.
• Governance over data, access, and accountability.
• Model selection or routing logic.
• Security controls, especially when agents can act.
• Measurement systems that connect AI usage to business outcomes.

Use this model when your team is stuck asking which model to choose. The stronger question is whether the surrounding system is mature enough to make any model choice reliable.

Cost visibility to value linkage

A useful economic maturity ladder runs in three steps:

• Cost visibility: know usage, tokens, and tool spend.
• Cost control: add reviews, dashboards, and budgets.
• Value linkage: compare AI cost against time saved, risk reduced, or revenue created.

The episode argues that many companies are between the first and second stage. The most advanced organizations are starting to reach the third stage, which is why they can justify increasing AI investment instead of trimming it.

Decision Principles

Principle: Build governance before scale

• When: an organization wants to roll out agents or expand AI usage across teams.
• Why: the companies with established AI ROI usually have formal harness layers, sovereignty strategies, and clear accountability in place before AI becomes widespread. Scale without structure creates messy adoption and weak control.

Principle: Measure value, not spend

• When: reviewing AI programs, tools, or platform budgets.
• Why: token volume can rise while spend falls, as the Ramp data showed, so raw spend is a weak signal. What matters is whether AI cost is connected to business value such as productivity, resilience, or revenue.

Principle: Route models deliberately

• When: different workflows have different cost, quality, latency, or data sensitivity needs.
• Why: organizations are building routing and sovereignty strategies because one default model is rarely the best long-term answer. A deliberate routing policy reduces lock-in and improves economics.

Trade-offs & Nuance

Experiment freedom versus cost control

• Early in adoption, strict token or usage budgets can slow learning.
• The host explicitly warns that being too restrictive during the experimentation phase can be limiting.
• Later, once workflows are clearer and usage is scaling, budgets and reviews become more useful because they help connect spend to value instead of merely capping exploration.

Single vendor versus flexibility

• A single vendor is simpler to buy, govern, and support.
• A multi-model setup gives an organization more leverage on cost, performance, and data handling.
• The OpenAI marketplace example shows how vendors are responding to this tension. Enterprises want flexibility, while platforms still want to preserve ecosystem lock-in.

Efficiency gains versus revenue bets

• Efficiency is easier to prove because cost savings are usually visible first.
• Revenue-oriented AI strategies are becoming more common, especially in agent programs, but they are harder to attribute and usually require stronger product design.
• The most mature organizations are no longer treating efficiency and growth as a choice between two separate AI strategies. They are trying to pursue both at once.

Common Mistakes

Mistake: Treat rollout as adoption

Buying tools, approving pilots, or announcing an AI strategy is not the same as changing how work happens. The KPMG comparison shows that real ROI organizations are distinguished by formal systems and widespread usage, not by interest level.

Mistake: Add budgets too early

Putting tight usage or token budgets on teams before they have learned where AI creates value can shut down discovery. The better sequence is broad visibility first, then controls, then tighter value-linked management.

Mistake: Ignore security until agents

• AI agents expand what AI can do, but they also expand the risk surface.
• 58% of organizations with established AI ROI are putting AI-assisted cyber defense in place, compared with only 8% of experimenting organizations.
• The lesson is simple: once agents can take actions, security can no longer be treated as a later cleanup step.

Practical Application

Map your AI management layer

Create a one-page map of how people in your organization currently use AI.

• List every entry point, such as consumer chat tools, coding assistants, internal copilots, and agent workflows.
• Note who owns each entry point, what data flows through it, and what approval path exists.
• Mark missing pieces, especially a formal harness layer, routing rules, sovereignty policy, and accountability for AI-informed decisions.

This exercise turns vague AI adoption talk into a concrete operating design problem.

Segment use cases by value

Separate AI work into three buckets before you prioritize a roadmap:

• Productivity use cases, such as drafting, summarizing, or internal workflow automation.
• Risk and resilience use cases, such as governance, security, or reliability improvements.
• Revenue use cases, such as sales support, new product features, or agent-led growth motions.

Each bucket needs different success metrics. A single ROI story will hide important differences.

Measure value against AI cost

For each meaningful AI workflow, track both usage and business outcome.

• Track inputs such as model calls, tool spend, and support cost.
• Track outputs such as hours saved, cases resolved, conversion lift, or revenue influenced.
• Review the pair together monthly so AI economics do not get reduced to a token dashboard with no business context.

Pilot agents with security included

Choose one bounded workflow for an agent pilot, then build monitoring and security into the first version.

• Pick a workflow with clear inputs and limited action rights.
• Define what the agent can read, what it can do, and what human review still exists.
• Add security checks and incident monitoring before wider rollout.

That approach follows the pattern of mature organizations, which are pairing agent adoption with AI-assisted cyber defense.

Define a model routing policy

Write a simple policy for when different models should be used.

• Use one dimension for cost.
• Use one dimension for quality or reasoning depth.
• Use one dimension for data sensitivity or sovereignty needs.
• Use one dimension for latency.

Even a lightweight routing policy helps a PM avoid accidental lock-in and makes future platform choices easier.

Questions to Consider

• Which AI initiatives on your roadmap are still being judged by activity metrics, such as seats enabled or prompts run, instead of business value compared with AI cost?
• Where would a formal AI harness layer reduce the most risk or confusion in your organization: employee chat use, customer-facing copilots, or agent workflows that take actions?
• If your team expanded AI agent usage next quarter, what security controls would need to exist before an agent could safely read customer data or trigger real workflows?
• Which product or internal workflows genuinely need one default model, and which workflows would benefit from model routing based on price, performance, latency, or sovereignty requirements?
• Has your organization's AI strategy moved beyond productivity savings, and if not, what is one realistic revenue or resilience use case worth testing next?

Bottom Line

Enterprise AI leaders are pulling ahead by building the operating layer around AI, not by merely buying access to better models. The companies seeing ROI are formalizing how AI is used, governed, routed, secured, and measured, then using that foundation to expand from efficiency wins into broader business value.

Resources Mentioned