The first phase of enterprise AI was about speed. The next will be decided by whether AI can operate inside the real constraints of business without creating more rework than value.
Emily Faracca
Multimedia Content Writer
Workday
Audio also available on Apple Podcasts and Spotify.
For a while, enterprise AI has been sold like a clean trade.
Let the machine take the rote and repeatable work.
Give people their time back.
Watch productivity climb.
But that framing leaves out a critical part of the story: in business, speed is only valuable if the work actually holds.
Across the market, the conversation is beginning to shift away from AI in a sidebar. The real test of enterprise AI was never going to be how quickly it could generate a first draft, but whether that output could survive contact with the actual conditions of enterprise work: permissions, policies, workflows, handoffs, edge cases, compliance requirements, and all the institutional logic that determines whether something is merely plausible or actually usable.
In a conversation on the Future of Work podcast between Jon Lexa, vice president of go-to-market for Sana, and Callie Kemmer, senior director of AI outbound and growth at Workday, that missing distinction came into focus.
Their discussion was not about whether AI is powerful. That part is settled. It was about why power alone has not translated cleanly into enterprise value, and what kind of architecture, adoption model, and leadership mindset the next phase of AI will actually require.
That is where the market is starting to mature. The most consequential shift in enterprise AI is not from no AI to more AI. It is from AI that generates to AI that executes. And those are not the same thing.
Report
Key to this conversation is to distinguish between what language models are built to do and what enterprise systems require.
The first wave of enterprise AI proved that models can generate at extraordinary speed. But that early success also created a distortion. It encouraged leaders to evaluate AI by how quickly it could draft, summarize, or recommend, rather than by whether those outputs could move through the business without creating a second layer of cleanup.
That is where much of the promised value has quietly eroded. Workday research found that nearly 40% of AI’s promised productivity is lost to rework, and that organizations realize more durable value when they measure outcomes achieved, not just hours saved.
The issue is not that AI cannot accelerate activity. It is that activity without enough contextual grounding often creates hidden verification work for the people downstream.
Lexa named the underlying reason: “Things have to be correct. So something that’s stochastic or probabilistic in nature, it doesn’t work that well in an environment where everything needs to be 100% right 100% of the time.”
That is the enterprise AI challenge in one sentence. Language models are probabilistic. Business processes are not. This is why so many teams feel both impressed by AI and burdened by it.
“Something that’s stochastic or probabilistic in nature, it doesn’t work that well in an environment where everything needs to be 100% right 100% of the time.” - Jon Lexa, VP of GTM, Sana
The deeper challenge is not intelligence alone. It is context in the enterprise sense of the word.
Consumer AI often treats context as prompt history or nearby data. Enterprise context is something else. It includes permissions, systems of record, process rules, role awareness, data lineage, and the institutional memory that determines whether an action is appropriate, complete, or risky.
Kemmer framed the adoption challenge accordingly, stating the importance of moving away from treating AI “like a destination” and toward something “woven into how we work” and “second nature.”
That is more than a design preference. It is an operating model question. If AI remains a destination, the employee still leaves the workflow, retrieves an answer, and then assumes responsibility for translating that output back into the business.
The model may be fast, but the human still carries the burden of interpretation, validation, and action.
Mature enterprise solutions reduce that translation layer and place intelligence closer to the systems and processes where decisions actually become work. That is the difference between AI that sounds capable and AI that can be operationally useful.
A great deal of AI commentary still assumes that incumbency is a weakness. The common story is that AI-native challengers will outrun established enterprise platforms because they move faster and carry less structural baggage.
But if the next phase of enterprise AI depends on governed action rather than clever generation, incumbent SaaS providers have a real structural advantage. They already sit near the system-of-record layer. They already contain years of embedded business process, established permissions, security models, and customer trust. In an agentic environment, those assets matter most.
Lexa argued that even in a future where agents are used every day, companies will still need guardrails, deterministic processes, and a way to capture the data those agents create.
“In order to capture that,” he said, organizations need “a system of record that is secure, that is governed appropriately, that has the right permission architecture for agents and for humans.”
Kemmer pointed to the same reality from another angle, suggesting that an agentic layer can expose bottlenecks and broken processes, but also warned that the foundational “table stakes” can be overlooked precisely when they become more important.
One of the weakest responses to AI efficiency is the assumption that productivity gains should naturally culminate in headcount reduction.
Lexa framed the issue more constructively. Looking at the kinds of gains AI can create, he said the more important question is what organizations do with the capacity that comes back.
“What do we do with excess capacity when we’re very productive?” he asked. “We can just do so many more things, which then requires us to have better judgment around how we spend our time.” He added that those gains should create more room “to exercise good judgment and develop taste and to reason better rather than getting stuck in mundane tasks.”
AI should not primarily be treated as permission to reduce labor before the work itself has been redesigned. It should be treated as an opportunity to redeploy human effort toward harder decisions, better prioritization, stronger collaboration, and more ambitious outcomes.
Data from Workday research confirms this direction. It concludes that when companies focus solely on speed, employees absorb the cost through rework and fatigue, whereas organizations that reinvest AI gains into people see stronger and more sustainable value.
“What do we do with excess capacity when we’re very productive? We can just do so many more things, which then requires us to have better judgment around how we spend our time.”— Jon Lexa, VP of GTM, Sana
The next phase of enterprise AI will not be won by whoever ships the most dazzling interface or generates the most content. It will be shaped by who can narrow the gap between output and action, who can operate inside trusted systems, and who can use efficiency gains to strengthen the enterprise rather than hollow it out.
That means grounding AI in systems the enterprise already trusts. It means reducing the distance between output and execution, and understanding that efficiency gains are only valuable if they lessen friction rather than relocate it. And it means recognizing that the more capable AI becomes, the more important it is to invest in the people who must guide, interpret, and govern what it does.
That is the standard leaders should be using now. Not “How much faster did AI make us?” but “How much of that work could the business actually trust, use, and act on?”
The AI conversation has moved past novelty and toward harder questions: where trust lives, why systems of record matter, what incumbency actually buys you, and how leaders should think about capacity once AI starts returning it.
Because if AI only gives organizations faster drafts and a reason to cut staff, then the grunt work has not gone away. It has simply been redistributed, often onto the very people expected to make the new model work.
Report
AI has been a gamechanger across nearly every industry and marketplace. But realizing the true value of AI requires buy-in between different stakeholders. In this AI Masterclass, learn how you can make the most of AI for finance, HR, and more.