Beyond the Roadmap is Cheesecake Labs’ interview series with technology leaders about what they’ve learned that never made it into the slide deck. Our first guest is Jeremy Korst, founder and CEO of Mindspan Labs and founding co-author of the annual Wharton-GBK study of enterprise AI adoption.
Jeremy has seen technology adoption from both sides: as the executive who launched Windows 10 at Microsoft, and now as a researcher studying how enterprises put AI to work. His view is direct. AI rarely stalls on the technology. It stalls on ownership, on workflows built around old constraints, and on leaders who never decide what to stop doing.
About Jeremy Korst
Jeremy Korst is founder and CEO of Mindspan Labs, which works with organizations to turn AI vision into velocity. He is a Senior Fellow & Lecturer at The Wharton School and an Adjunct Assistant Professor at Columbia Business School. He is the founding co-author of the annual Wharton-GBK study of enterprise AI adoption, now in its fourth year. Before Mindspan he was an operating executive in technology and telecom, including launching Windows 10 at Microsoft and founding T-Mobile’s mobile broadband business. He is a regular contributor to Harvard Business Review and has been featured in WSJ, Bloomberg, Forbes, Fortune and Business Insider.
You’ve led a huge tech adoption from the inside (the Windows 10 launch) and now study adoption at Wharton. What is genuinely different about AI, and what is the same old pattern?
Most of it is the same old pattern. Adoption succeeds or fails on people and organization, and on whether the technology is pointed at the business’s actual priorities. The technology has to enable those things. It can’t substitute for them. That was true with Windows 10, where we could put the bits on a billion devices and still not change how a single person worked.
The people closest to the work see it too. In the fourth annual study of enterprise AI adoption from The Wharton School and GBK Collective, 65% of managers and above agree, at least somewhat, that people and processes hold their department back more than the technology does.
Two things are genuinely different. The first is the sheer pace. The capability you scoped against in January isn’t the capability you have in June, and too many of the planning cycles I see still assume it is. The second is identity. With AI, people aren’t only asking how their job will change. They’re asking what their job is, and that touches their sense of who they are at work. Leaders who treat that as a training problem miss it.
Enterprise AI use is now mainstream. Where are companies still not seeing returns, and what separates the ones that are?
The companies getting returns start from the fundamentals I described in the first answer: their people and organization, alignment with the strategy, and governance that turns ambition into shipped work. Then they do the hard part, which is redesigning the workflows themselves and aiming at growth and innovation, not just productivity. Productivity is necessary. It’s where most of the companies I work with start, which is exactly why it isn’t where advantage comes from.
The pattern I see most often in the companies that aren’t getting returns: they set out to save time, succeeded, and nothing arrived in the P&L. The hours went straight back into a workflow built around the old constraints.
The study shows how few have made the full shift. Only 7% of organizations are successful across all three of its success markers. Departments that have transformed their workflows report strongly positive returns 67% of the time, against 19% for the rest. And among managers and above, departments tracking at least one growth measure report strongly positive returns 35% of the time, against 13% for those measuring efficiency alone. Those are associations, not proof of sequence. But a growth target forces you to say what the AI is for.
Related: Why 95% of AI Pilots Never Reach Production (and What the Other 5% Do Differently)
You’ve written about a gap between how the C-suite and the teams doing the work see AI. Where does that misalignment most often break an initiative?
It breaks in the middle of the organization. In a Harvard Business Review article earlier this year, my co-authors Stefano Puntoni and Prasanna Tambe and I called it the messy middle: the layer between the executives funding AI and the people doing the work. In our previous year’s data, executives were far more likely than middle managers to report significantly positive returns, 45% against 27%.
The gaps haven’t closed, and there is more than one. In this year’s study, 87% of VPs and above say their organization has a clear long-term AI vision, against 64% of entry and mid-level staff. In our own analysis of the respondent data, managers and directors are the least likely of any level to say the C-suite owns AI agents: 27%, against 59% of VPs and above and 40% of entry and mid-level staff. The people expected to translate strategy into work are the least sure who is accountable for it.
That’s where initiatives stall. Executives approve use cases that look good from the top, and managers are left to make them work without clear ownership. Fix the ownership before you fund the next pilot.
Related: Trust at Scale: Why Governance and Adoption Are the Hardest Parts of AI
If the dashboard measures AI, you’ll get more AI. If it measures outcomes, you’ll get outcomes. — Jeremy Korst
If a CTO had to fix one thing in how their leadership team talks about AI, what should it be?
Stop letting “AI” be the subject of the sentence.
Nearly every leadership conversation I sit in contains a line like “we need to use AI in customer service.” Nobody can act on that. It has no owner, no decision inside it, and no way to be wrong.
Make the business outcome the subject instead. “We want first-contact resolution up ten points, and here are three places AI might get us there.” Now you’ve named a number, a function, and an owner who isn’t the CTO. The CTO’s job becomes feasibility and sequencing, which is the job they are actually good at.
This matters more for technical leaders than for anyone else, because the moment AI is the subject of the sentence, it becomes your initiative. You inherit the goal, the budget and the blame for an outcome you don’t control. Handing the subject back to the business isn’t passing the buck. It’s the only way the work gets prioritized against everything else the company cares about.
AI isn’t a strategy. It’s how you accelerate yours.
You’ve described leadership alignment as an amplifier. What exactly does it amplify, good and bad?
It amplifies whatever your organization already does with ambiguity.
Alignment doesn’t make anyone smarter about AI. It removes the need to guess. When a team knows which outcome they’re serving, they make a steady stream of small calls correctly without escalating: what to automate, what to leave alone, what “good enough” means for a model sitting in front of a customer. That’s the part that compounds, and from the outside it looks like speed.
Misalignment amplifies the reverse just as efficiently. Every ambiguous call gets escalated or guessed. You get teams solving one problem several ways, a governance review that becomes the place work goes to die, and a pilot portfolio that only grows, because nobody has the authority to kill anything.
The uncomfortable part is that you can’t tell which one you have from the top. Both look like activity. The signal is further down: ask three managers in different functions what the AI work is for. If you get three different answers, you have your diagnosis.
Build, buy or partner: how should a tech leader decide for AI capabilities in 2027, and when does each one backfire?
For most organizations the honest answer is all three. In my experience, few have all the internal talent and capability they need, and the technology isn’t waiting for hiring plans to catch up. So buy what is becoming standard, partner to move now on what you can’t yet do yourself, and build where the thing is genuinely yours: your data, your workflow, your customer relationship.
The part I see skipped most often is the plan to grow those capabilities internally over time. AI will be core to how the business runs, and you can’t outsource what’s core forever. A good partner helps you build that muscle rather than replacing it.
Each option has its own failure. Partnering backfires when you outsource the thinking rather than the hands: if the partner walks and nobody internally can explain why the system decides what it decides, you bought a dependency, not a capability. Buying backfires more quietly. You end up with a copilot in every tool you own, none of them aware of each other, and no one who owns the workflow they all touch. Building backfires when it rests on this year’s model assumptions and you budget to maintain it instead of rewriting it.
What is one metric you would put on a board dashboard to prove AI is paying off, and one you would retire?
Put on: the business outcomes the AI work is targeting, as the numbers themselves. Revenue from new offerings. Customer retention. Cycle time on a core process. Cost to serve. First-contact resolution. Every AI initiative should name which of those it moves and by how much, and the board should see progress against those targets, not against AI activity. You get what you measure. If the dashboard measures AI, you’ll get more AI. If it measures outcomes, you’ll get outcomes.
Retire: adoption. Seats, licenses, weekly active users, percentage of employees “using AI.” It was a reasonable proxy in 2024, when the open question was whether anyone would touch it. It isn’t now, because a rising usage number tells you nothing about whether the work changed.
How is AI changing the shape of product and engineering teams, and which roles matter more now?
The roles are converging. Design, product management and engineering used to be separate disciplines with handoffs between them. With new capabilities, one person can now prototype, spec and build across all three. The people who succeed have real depth in at least one of those domains and working capability across the others.
What matters more is anyone who can define a problem precisely, and anyone who can judge output they didn’t produce. A product manager who can write a crisp evaluation criterion is suddenly worth a great deal. So are senior engineers, but for a different reason than before. Not throughput, taste. Reviewing is the scarce skill now.
The other thing that separates the strongest people is a deliberately customer-centric approach. When building gets cheaper, the constraint moves to knowing what’s worth building, and that comes from working outside in: starting from the customer’s problem, not the capability.
What is a belief about AI adoption you held two years ago and have since changed?
I thought AI would be a powerful tool for knowledge and productivity. Help people find things faster, write faster, analyze faster.
It is that. What surprised me is how capable it is at transforming workflows and driving genuinely new work, not just speeding up the old work.
We’ve learned that firsthand at Mindspan. We run a team of AI agents alongside our people, each with a defined role across research, product marketing, social content, engineering and operations. It’s a force multiplier: a small firm can do work that used to take a much larger team.
The lesson that surprised me most is that the hard part isn’t the technology. It’s management. The agents need clear roles, clean handoffs, review of their work and someone accountable for what they produce. Those are the same questions every leader faces with a human team, and they’re the questions every organization deploying AI will have to answer.
What’s the one question you wish more tech leaders asked you?
“What should we stop doing?”
I get asked what to start constantly: which model, which platform, which use case, which pilot. Almost nobody asks what to kill. The stopping is where the capacity comes from. Most companies I work with don’t have an AI resourcing problem; they have a long list of initiatives and no mechanism for ending one.
The related question, and the harder one: “what are we not willing to change?” If the answer is the org chart, the incentive structure and the way decisions get made, you’ve just described your ceiling, and no amount of technology spend moves it. I’d rather a leadership team tell me that honestly in the first meeting than discover it in month nine.
Lightning round
- One book or resource: Co-Existence: The Next Phase of AI, by Ethan Mollick
- One tool he uses daily: the Mindspan agent team
- One prediction for 2027: Organizations will spend 2027 learning that autonomous agents aren’t plug and play. The hard questions will be about governing and managing them, and about how to build high-performing teams of people and agents working together.
Where does your company stand?
Jeremy’s point holds up in our work too: the hard part of AI is rarely the technology. It’s the data, workflows and ownership around it. Our free AI Readiness Assessment scores your company across seven dimensions, from data and workflow design to governance, and shows which of four stages you are in, from Foundational to Transforming.
For more of Jeremy’s research, read the 2026 Wharton-GBK AI Adoption Report and his Harvard Business Review article, “Managers and Executives Disagree on AI—and It’s Costing Companies”.
Beyond the Roadmap lands in inboxes every other Tuesday. Know a technology leader we should feature? Let us know.