A submission by CIO Consultant Rocky Vienna
The AI value-creation problem has not stalled because the models are insufficient. It has stalled because organizations are asking AI to improve work they have not redesigned. Two recent pieces of evidence, a Stanford study conducted inside Google and a formal commercial announcement from Anthropic, point to the same conclusion.
What the Stanford Research Inside Google Found
In March 2026, Martin Gonzalez of Google DeepMind published findings from an 18-month Stanford study conducted inside Google on how employees adopt AI. The researchers found that most people were stuck in what they called ‘simple substitution’: swapping existing tasks for AI alternatives without changing how the work was structured. The effort to learn the tool often exceeded the payoff.
The employees who got real value did the opposite. They applied a product management mindset, identified the bottlenecks blocking their work, then selected the AI tool that cleared those specific blockers. They redesigned the workflow first and chose the tool second.
The finding is individual-scale, but the structural logic applies at the enterprise level. Companies treating AI as a productivity overlay get pilots. Companies redesigning workflows around what AI can do get returns.
What Anthropic and Blackstone Actually Announced
Anthropic’s May 2026 announcement with Blackstone, Hellman & Friedman, and Goldman Sachs is primarily a commercial event, a new delivery vehicle for AI demand that Anthropic’s CFO says is “significantly outpacing any single delivery model.” Blackstone’s President and COO Jon Gray names the specific bottleneck: the number of highly skilled implementation partners. That is a delivery capacity problem. The inference underneath is harder to ignore. Anthropic is not building a services firm because the models need more work. They are building one because the gap between model capability and enterprise deployment is not closing on its own.
Neither source was framed as a statement about enterprise AI strategy. Both point to the same structural gap. The constraint is not the intelligence. It is the organizational infrastructure required to put that intelligence to work inside load-bearing business processes.
The Actual Constraint
PwC’s 2026 Global CEO Survey found that only 12% of CEOs say AI has delivered both cost and revenue benefits, while 56% report no significant financial benefit to date. That is not because the tools are unavailable. It is because adoption and transformation are not the same thing.
Most enterprises already know how to optimize processes. They have Lean belts. They have Six Sigma certifications. They have process maps, transformation offices, and twenty years of consulting decks filed somewhere on SharePoint.
Methodology is not what is missing.
What is missing is consent.
The workflows closest to EBITDA are usually the hardest to re-architect because they are load-bearing. They move revenue. They manage risk. They satisfy customers. They keep the quarter intact. Changing those workflows creates local pain before it creates enterprise gain. It exposes fragile handoffs. It threatens status. It changes roles. It asks managers to trade a known operating rhythm for a future-state P&L that has not been proven yet.
So companies pilot. They form AI councils. They buy tools. They automate the edges. They call it transformation because the alternative would require admitting that the real value sits inside work no one has permission to disturb.
The EBITDA does not come from employees writing faster emails. It comes from compressing cycle times, reducing rework, increasing revenue per employee, improving throughput, and redesigning handoffs that used to require human coordination.
Five EBITDA Value Levers
A useful AI value-creation conversation should start with these five questions.
| VALUE LEVER | AI WORKFLOW REDESIGN QUESTION | EBITDA LINK |
| Labor leverage | Which steps can be automated, compressed, or moved to lower-cost human oversight? | Lower operating expense or higher revenue per employee. |
| Cycle time | Where does work wait for data, review, approvals, exceptions, or handoffs? | Faster throughput, faster cash conversion, and higher operating capacity. |
| Quality and rework | Where do errors create downstream review, remediation, customer friction, or compliance exposure? | Lower cost of poor quality and improved retention. |
| Revenue capture | Which bottlenecks prevent teams from pursuing, pricing, quoting, onboarding, or servicing demand? | Higher conversion, expansion, utilization, or share of wallet. |
| Management visibility | Which workflows lack measurable leading indicators, exception tracking, or owner accountability? | Better forecasting, faster intervention, and tighter resource allocation. |
The Forcing Function
This is why the Anthropic-Blackstone signal matters. It is not a statement about model capability. It is a structural signal about where the work of AI value creation actually lives.
PE ownership does not magically make AI easier. In some cases, it can make transformation too narrow if the mandate becomes cost takeout instead of operating-model redesign. But PE does change the permission structure.
When a sponsor sets a 24- to 36-month value creation clock and an operating partner has authority to mandate workflow redesign rather than merely suggest it, the political problem flips. The default stops being ‘protect the current operating model’ and becomes ‘produce the next valuation step or explain why.’
Operators inside PE-backed companies are no longer asking permission to rebuild the work. They are being asked why they have not.
That does not mean every AI initiative should start with disruption. It means the companies that see real returns will be the ones that connect AI to the workflows that actually move enterprise value. Andreessen Horowitz’s enterprise AI analysis makes a related point: adoption is clearest where the work is text-heavy, repetitive, measurable, and supported by human-in-the-loop review, specifically coding, support, and search. In other words, ROI appears first where the work can be redesigned, measured, and governed.
The Board Questions That Follow
Do not ask only: Which AI tools are we using? Ask: Which load-bearing workflows are we willing to redesign?
Do not ask only: How many pilots are live? Ask: Which use cases have an accountable owner, a measurable EBITDA lever, and permission to change the process?
If you operate inside a PE portfolio company, the question is not whether your processes are AI-ready. The question is whether your sponsor will give you cover to rebuild the workflows that actually produce EBITDA before the next company in your sector does it first.
The model conversation is not over. But for boards and operators, it is no longer sufficient.
The next AI advantage will not belong to the company with the longest list of pilots. It will belong to the company willing to redesign the work.
