The same case, proven outside the book
The Outside Evidence
A book arguing for its own method is a claim. The public record is a check. Six findings from independent researchers, industry benchmark reports, and reported incidents, none of them published by Certified Private AI, each with its source linked and its limit stated. Read them at the source. That is what they are there for.
Independent research, peer-reviewed
+14% issues resolved per hour
Assisted agents outperform, and your newest staff gain most
Brynjolfsson, Li, and Raymond studied 5,179 customer support agents at a Fortune 500 software company as an AI assistant rolled out in stages. Access to the tool raised productivity 14% on average, and 34% for novice and lower-skilled workers, while customer sentiment and employee retention improved and top performers barely moved. Published in the Quarterly Journal of Economics, 2025.
What it means for your business
This is the Rung 1 pattern measured in the wild: a person, an approved assistant, a review habit. The biggest gains went to the least experienced staff, the people a small firm can least afford to train slowly.
Limit: one company, one job type, and it measures assistance, not autonomy. Source: "Generative AI at Work," NBER Working Paper 31161.
Independent field experiment, Harvard and BCG
+40% higher rated quality, inside the frontier
The frontier is jagged, and that is the argument for rungs
In a field experiment with 758 Boston Consulting Group consultants across 18 realistic tasks, consultants using GPT-4 on tasks inside its capability finished 12.2% more tasks, 25.1% faster, at over 40% higher rated quality. On one task deliberately outside that capability, consultants using AI were 19 percentage points less likely to reach the correct answer than consultants without it.
What it means for your business
The same tool lifts one task and quietly misleads on a similar looking one. Choosing the rung per task, and keeping a human reviewer where the answer matters, is how a business keeps the gains and refuses the losses.
Limit: consulting tasks and one generation of models; the frontier moves as models change. Source: "Navigating the Jagged Technological Frontier," Harvard Business School.
Survey, 31,000 people in 31 countries
78% of AI users bring their own tools to work
Most knowledge workers already use AI. The open question is governance.
Microsoft and LinkedIn's 2024 Work Trend Index found that 75% of knowledge workers use AI at work, that 78% of those users bring their own AI tools rather than waiting for a company provided one, and that 60% of leaders say their company lacks a plan to implement AI.
What it means for your business
If your business runs on knowledge work, the odds are strong that AI use is already happening inside it, one personal account at a time. The decision still on your desk is whether that use runs on approved tools with written rules, or on personal accounts with none.
Limit: self-reported survey data from a vendor with AI products to sell; treat the direction as reliable and the decimals as approximate. Source: 2024 Work Trend Index, Microsoft and LinkedIn.
Industry benchmark, 600 breached organizations
$4.44M global average cost of a breach, 2025
Ungoverned AI now has a measured price
IBM's Cost of a Data Breach Report 2025 put the global average cost of a breach at $4.44 million, and $10.22 million in the United States. One in five breached organizations reported a security incident involving shadow AI, and those breaches cost about $200,000 more on average. Sixty-three percent of organizations had no AI governance policy in place.
What it means for your business
Shadow AI has moved out of the hypothetical column and into the benchmark data, priced. A written approved list, named owners, and records are the controls that keep your business out of that statistic.
Limit: averages across breached organizations, weighted toward large companies; your exposure is your own numbers, not the average. Source: Cost of a Data Breach Report 2025, IBM.
Reported incident, Bloomberg, April 2023
20 days from permission to leak, then a company-wide ban
Samsung: the leak was an employee trying to work faster
Weeks after Samsung Electronics allowed staff to use ChatGPT, engineers pasted proprietary semiconductor source code and confidential meeting notes into it. Samsung responded by banning generative AI tools on company devices and networks while it worked on controlled alternatives, after an internal survey in which 65% of respondents already believed such services posed a security risk.
What it means for your business
Nobody hacked Samsung. Helpful employees used a free tool with no approved path and no written rule. That is the exact failure an approved list and a controlled route exist to prevent, and it took under three weeks.
Limit: a single incident, with some details Samsung never disclosed; it proves the failure mode, not its frequency. Source: Bloomberg News, "Samsung Bans Staff's AI Use After Spotting ChatGPT Data Leak".
Global survey, 1,993 respondents in 105 countries
88% use AI somewhere; about one in three is scaling it
Nearly everyone has the tools. A third have begun scaling them.
McKinsey's State of AI 2025 reports that 88% of organizations now use AI regularly in at least one business function, up from 78% a year earlier, while roughly two-thirds have not begun scaling beyond pilots and only about 6% qualify as high performers reporting meaningful profit impact. Reporting on the survey points to the same separator: the high performers redesign workflows instead of layering AI on top of old ones.
What it means for your business
Access to AI is now something anyone can buy off the shelf. The scarcer asset is a documented, governed process wrapped around it, pointed at a measured outcome. That is precisely the thing Certified Private AI implements.
Limit: self-reported survey; "high performer" is McKinsey's definition, and correlation is not proof of cause. Source: "The State of AI in 2025," McKinsey, as reported by Information & Data Manager.
Watch the Researchers Say It Themselves
The app this page is modeled on leaned on a featured video every day. The habit worth copying is not its conclusion; it is sending people to watch the evidence themselves. Four talks on YouTube by or about the researchers behind the studies above, each labeled with what it can and cannot prove. They open on YouTube, in a new tab.
Video, Microsoft WorkLab podcast, 2023
Erik Brynjolfsson on How AI Will Transform Productivity
The lead author of the 14 percent study, in conversation: AI pays off when organizations restructure work around it, and augmenting people beats automating them away. The economist behind the first card above, explaining his own numbers and what they do not show.
Pairs with
The first evidence card, and Move 1 of the onboarding automation challenge below: measure the task before you delegate any of it.
Limit: a podcast conversation, not the paper itself. For the numbers, read the study. Watch on YouTube.
Video, study presentation
Ethan Mollick: Navigating the Jagged Technological Frontier
Wharton professor Ethan Mollick, a co-author of the Harvard and BCG experiment, walks through its design and result: large gains on tasks inside the frontier, a 19 percentage point penalty on the task outside it, and why that boundary is hard to see from inside the work.
Pairs with
The second evidence card, and Move 6 of the onboarding automation challenge: the test day that decides whether an agent earns wider scope.
Limit: one study, presented by its own author; models have moved since it ran. Watch on YouTube.
Video, Strange Loop podcast
How to Build an AI-First Organization
Mollick on the management side: most companies use AI to shave costs and think too small, while the organizations that gain redesign roles around it, keep humans in the loop where judgment lives, and treat adoption as a leadership job rather than an IT purchase.
Pairs with
Move 4 of the onboarding automation challenge: the written arrangement, owned by a named person, that turns scattered use into a governed system.
Limit: expert argument, not measurement. Weigh it as informed advice. Watch on YouTube.
Video, Insight Partners ScaleUp:AI, October 2025
The Jagged Frontier of Generative AI
Mollick with Insight Partners managing director Lonne Jaffe on what actually blocks adoption inside companies: not model capability, but leadership, organizational design, and incentives. Small accuracy gains can quietly multiply what an agent can safely carry, which is why rung decisions get revisited on a schedule instead of made once.
Pairs with
Move 7 of the onboarding automation challenge: the decision, made with numbers, to climb a rung or stay where the evidence says to stay.
Limit: a venture firm's event stage, and the host invests in AI companies. Discount accordingly. Watch on YouTube.