On August 25, 2026, McKinsey published its State of AI 2026 report. Nine in ten organizations now use AI in at least one business function. Forty-four percent scale it across the enterprise, up from thirty-eight the year before. And then the sentence that should have landed like a bell. The share of respondents who attribute measurable EBIT impact to AI use held at thirty-seven percent, unchanged from the 2025 survey.
Two weeks later, on September 8, Bain & Company released its sixth annual Global Technology Report. The lead number was $4.7 trillion. That is the size of the profit pool Bain expects AI to reshuffle across ninety-two industries between now and 2035, more than triple the profit shift the Internet caused in twice the time. Bain then wrote the sentence executives should paste above their desks. The Internet was a distribution technology. AI is a production technology.
Read those two findings in sequence. Adoption doubled. Scaling accelerated. Profit sat still. And the technology under all of it turns out to reshape what a company makes, not how it reaches the customer. If your AI transformation roadmap for executives assumes that adoption produces transformation, the reports just told you it does not, and named the reason.
The wrong line went up.
Two reports, one signal
Neither report is anti-AI. McKinsey's headline stat is that eight in ten enterprises report at least one function scaling AI, and it treats that number as evidence of maturation. Bain forecasts an economic shift larger than the arrival of e-commerce. Both are bullish. Which is why the flat EBIT line is the diagnostic.
McKinsey's own framing does the work. "The organizations that translate individual productivity gains into lasting enterprise-level financial performance are likely to be those that transform their businesses, not just adopt AI tools." That sentence is a concession. The prior year of the same survey read as though adoption WAS the transformation. This year McKinsey admits it was not.
Bain says the same thing in different vocabulary. Dunigan O'Keeffe, the lead author of the September 8 report, told reporters that CEOs need "a clear prediction of where their industry is headed with board-level commitment to act." He also said speed matters more than most CEOs realize, because moving early gives you more data, which improves results, which lets you move earlier still. Both consultancies converged on the same finding by different roads.
Then there was the MIT Project NANDA report from the prior August, titled The GenAI Divide, which found that ninety-five percent of enterprise generative AI pilots produced no measurable return on investment despite roughly $30 to $40 billion of enterprise spend. That report named the failure mode plainly. Pilots bought from specialist vendors and stitched into real workflows succeeded around sixty-seven percent of the time. Internal builds and generic tool deployments succeeded one-third as often.
Three studies. One signal. Adoption is not the return. Integration is closer, but only when it hits the actual production layer.
The roadmap most executives are running
Look at almost any AI transformation roadmap for executives and it has the same three phases. Phase one is enablement. Buy Copilot licenses, spin up an internal ChatGPT, run a training program. Phase two is pilots. Pick five to ten use cases, run six weeks each, measure lift. Phase three is scaling. Take the winners, replicate them across business units, and count them as transformation.
This is a good SaaS rollout plan. It is a bad AI plan.
The plan works when the technology being adopted is a distribution technology. If you are buying Salesforce, phases one and two and three make sense. Adopt the tool, pilot a team, scale the deployments that improve close rates. The work being done underneath is the same. You are distributing better information to the same salespeople doing the same job.
The plan collapses when the technology under it changes what the job is. That is what Bain means by production technology. When the underwriter no longer reads applications and instead reviews model output, the job stopped being what it was in phase one. When the analyst no longer runs queries and instead audits an agent's reasoning traces, the job stopped. When the customer support tier one no longer resolves tickets and instead handles the ten percent the agents escalate, the job stopped.
Phases one and two never anticipated that. They picked the vendor and the pilot on the assumption that the work would stay in place while a helper sat on top of it. The helper landed. The work stayed. The EBIT line stayed with it.
Where the thirty-seven percent lives
The ceiling is not a mystery. McKinsey's own data suggests that where the ceiling breaks, the same conditions repeat. Redesigned end to end. Owned by a business leader with P&L authority, not by a shared services group. Tied to a specific production output the company sells or delivers. Not measured by tool adoption or license counts, but by unit economics of the thing that comes out the other side.
Only a small fraction of scaled deployments meets that description. The MIT NANDA study estimates it at five percent. McKinsey's ceiling has moved a point or two inside individual functions but not in the aggregate. That is what a design choice looks like at scale. Every enterprise has run roughly the same playbook and produced roughly the same outcome. If the playbook were the constraint, the outcome would vary.
The playbook is the constraint.
Why board-level commitment is not the fix
O'Keeffe's line about "board-level commitment to act" is correct and slightly misleading. Board-level commitment to the current roadmap produces more of the current roadmap, faster. Ninety percent of the boards that have committed to AI have committed to the enablement-pilot-scale sequence. Every quarter the reporting gets thicker. The EBIT line does not move.
What board commitment actually needs to be is commitment to the redesign. Which is a harder sell to a board, because redesign is not measurable in the first two quarters. It is measurable in the second year, when the finished process runs at half the cost and twice the throughput of the pre-AI baseline, and the P&L catches up. Boards that fund only what shows up in the next earnings call cannot fund a production redesign.
The commitment has to sit with the redesign. It has to accept that the first eighteen months will not read like a normal SaaS rollout. It has to defend that inside the finance function against every quarterly counter-argument.
An inverted AI transformation roadmap for executives
Flip the sequence. Start where a production plan starts, which is at the output. The order that appears to work, based on MIT NANDA's five percent and the shape of every successful case study I have seen up close, is roughly this.
Start at the unit of value
Pick one thing your company sells, or one output it delivers to a paying customer, and describe the current production of that thing end to end. A loan decision. An audit opinion. A shipped SKU. A dispatched technician. A published research report. A signed contract. Describe the unit itself, at the level of the paying customer's transaction.
Every step of the current production has a cycle time, a labor cost, a defect rate, and a set of people who touch it. Write it down. If you cannot write it down, you do not know your business well enough to redesign it with AI, and no pilot will teach you.
Redesign the process before you pick the vendor
With the current production written down, ask what the same output looks like if a model takes the intermediate steps and a human takes only the judgment calls at the boundaries. Cycle time drops. Labor unit cost drops. Defect rate probably rises before it falls. The organization has to accept the temporary regression.
This is the step almost no roadmap includes. Most executives skip from "we should use AI" to "which vendor." The vendor is chosen against a wish list of features, not against a redesigned production process, so the vendor's tool sits next to the old process instead of replacing pieces of it. The old process is faster and cheaper to keep running than to rip out. The tool becomes ornamental.
Choose the vendor against the redesign
Now the vendor selection has real criteria. Does the tool implement the specific steps you plan to replace. Does it integrate with the specific systems that hold the specific data. Does it produce output in a form the human at the judgment boundary can actually audit. Anthropic, OpenAI, Google DeepMind, and every capable open-weight lab from Meta AI to Mistral will meet some criteria and miss others. Pick against the redesign.
Vendors picked this way tend to look small on paper and carry weight inside the company. Vendors picked from a wish list tend to look large and drift.
Rebuild the role
The person who used to do the intermediate steps now audits the model's output for the judgment steps. That is a different job. It needs different training, a different compensation structure, and a different span of control. If you do not rebuild the role, the person continues to do the intermediate steps the old way while the model runs in parallel producing outputs nobody uses.
Rebuilding the role is where most transformations die. HR does not have a template. The manager does not want to redraw span. The finance team cannot classify the new headcount. So the process reverts. The model becomes a dashboard.
Measure the output
The metric that determines whether the redesign worked is unit economics of the output. Cost per loan decision. Cost per audit opinion. Cost per shipped SKU. Not licenses issued, not queries answered, not prompts written.
McKinsey's flat thirty-seven percent EBIT number exists because the industry is measuring adoption, which is easy to move, and reporting it as transformation, which is a different variable. Executives who present adoption as a proxy for return will keep landing at thirty-seven percent for the next three surveys.
What this means for a $4.7 trillion profit pool
The Bain framing forces the point. If the coming decade reshuffles $4.7 trillion of global profit and does it 1.7 times more widely than the Internet, most of that profit shift will land inside industries where a few firms redesign production before their competitors do. In the industries that do not, it lands with new entrants that had no legacy production to defend.
The stakes make the flat EBIT line more dangerous than it looks. A firm sitting at thirty-seven percent is not "not moving." It is losing ground against the small number of firms that have already inverted the roadmap and are compounding data, results, and speed on top of a redesigned process. Two years of that compounding produces a gap you cannot close by buying more licenses.
The McKinsey survey field-dated May through June of 2026. Bain's forecast covers a decade. Those are not the same time horizon, and the reason both consultancies converged in the same month is that the horizon is now visible from where every executive is standing.
Why architecting this needs a partner
The steps above are simple to write and hard to run. Vendor sales cycles pull the roadmap back toward tool selection. HR pulls it back toward training programs. Finance pulls it back toward measurable-this-quarter pilots. The center of gravity of every organization pulls toward the old sequence, because the old sequence is what its incentives and reporting structures were built for.
Redesigning a production process against a $4.7 trillion transition is not a tool purchase. It is architecture work. You need a partner who has seen the sequence break in real companies, knows which handoffs kill the redesign, and can defend the eighteen-month P&L dip in front of the board. A partner who inverts the roadmap on purpose.
Agor AI Advisory does that work. We start at the unit of value, redesign the process before we pick the vendor, rebuild the role before we roll out the tool, and hold the measurement to the output. The three reports from the last thirty days say the same thing about what does not work. Something has to say the opposite for the industries that refuse to be on the losing side of the profit shift.
Sources
- The state of AI in 2026, McKinsey Global Survey, August 25, 2026
- McKinsey says enterprise AI is finally 'on the road to ROI', The Register, August 25, 2026
- AI puts $4.7 trillion of profits at stake, Bain & Company, September 8, 2026
- AI transformation set to shift corporate profits by $4.7 trillion, CIO Dive, September 2026
- MIT: 95% of enterprise AI pilots fail to deliver measurable ROI, Healthcare IT News
- MIT report: 95% of generative AI pilots at companies are failing, Fortune, August 18, 2025