· Chris Daily

A BELITHE PERSPECTIVE: Building AI Strategy

A note on what this is and what it isn't. In June 2024, CompTIA published a thoughtful whitepaper titled Building AI Strategy. It did a good job of laying out the state of enterprise AI adoption as it stood in the…

A BELITHE PERSPECTIVE: Building AI Strategy

By Chris Daily

Table of Contents

Jump straight to References ↓

About this article

A note on what this is and what it isn't. In June 2024, CompTIA published a thoughtful whitepaper titled Building AI Strategy. It did a good job of laying out the state of enterprise AI adoption as it stood in the first half of that year, and it remains a useful reference for leaders starting to think about how AI fits into their organization.

Two years is a long time in a field moving this quickly. What follows is not a revision of that paper. It's a new essay written in the same spirit — pragmatic, evidence-led, oriented toward people who have to actually make decisions — but from a different vantage point. It draws on the same underlying questions CompTIA raised, and it takes CompTIA's 2024 survey data as one source among many. It also draws on the full wave of AI research that has landed since: Stanford's 2025 and 2026 AI Index reports, McKinsey's State of AI 2025 survey, Lightcast's analysis of more than 1.3 billion job postings, the Federal Reserve's productivity research, the European Union's AI Act and its rollout, and peer-reviewed studies from the National Bureau of Economic Research and others.

The angle is human-centered. That is not a marketing word here. It is a commitment to the idea that AI is best understood as an amplifier of what people can do — not a substitute for what they do — and that the most important decisions about this technology are decisions about people. Every claim in this document that rests on research carries a footnote; a full references list appears at the end.

This is a perspective post. The views are mine. The evidence is everyone's.

Starting with people

Most writing about AI starts with the technology. The chips. The models. The billions in capital. The race between companies, the race between countries. That framing is true enough as far as it goes. It just puts the wrong thing at the center.

The right starting point is a woman at a call center in the Philippines whose productivity jumped by more than a third when her employer gave her an AI assistant that had quietly absorbed her senior colleagues' best practices [1]. Or a small-business owner in Indianapolis looking at a generative AI tool and wondering whether it will save her four hours a week or put her bookkeeper out of work. Or a teacher in rural Liberia who has never had reliable internet and who will either be reached by this wave of technology or be left behind by it. Those are the people AI is already reshaping, and a thoughtful strategy has to keep them in view.

We are past the point where anyone serious can debate whether AI matters. Eighty-eight percent of organizations now use AI in at least one business function [2]. But adoption is not the same as value. Only about six percent of companies have turned that adoption into meaningful enterprise-level financial returns [3]. The gap between those two numbers is the real story, and it is a story about people and organizations more than it is a story about technology.

This paper works through what we know, what we don't, and what a leader — of a company, a classroom, a nonprofit, a community program — can actually do about it. I have tried to be honest about uncertainty where uncertainty is warranted and clear where the evidence is clear. I have tried not to oversell. The technology is remarkable. The hype around it occasionally is not. And underneath both sits a set of genuinely hard choices that deserve more care than the news cycle has been giving them.

What we're actually talking about

"AI" has become one of those words that means everything and therefore nothing. A finance team says they are "using AI" and could mean anything from a forecasting model that has been running since 2018 to a browser tab that their analyst opened this morning. A marketing agency says they are "AI-powered" and could mean they use a writing assistant, or they have retrained an entire model on their client data. Before a strategy conversation can go anywhere useful, we have to tighten up what we're naming.

A working definition: AI is software that produces outputs based on probability rather than fixed rules. That line separates it from everything that came before. A traditional program does what it was programmed to do. An AI system does what it has learned is likely to be the right answer. That distinction has huge practical consequences — it is why AI systems can surprise you in both directions, sometimes with a brilliant answer and sometimes with a confidently wrong one.

Within that definition, two families dominate the current wave. Predictive AI pores over historical data to forecast what is likely to happen next — who is likely to default on a loan, which customers are likely to leave, where fraud is likely to occur. Generative AI, the newer and more dramatic branch, produces new content: text, images, code, audio, video. Both are useful. They solve different kinds of problems. Confusing them is one of the most common ways a strategy conversation goes off the rails.

A much older field than the headlines suggest

The field did not spring into existence in November 2022 when ChatGPT launched. Alan Turing posed the foundational question — can a machine think? — in 1950 [4]. Arthur Samuel taught a computer to play checkers in 1959 and, along the way, gave us the phrase "machine learning" [5]. Joseph Weizenbaum's ELIZA in 1966 demonstrated that even simple pattern matching could feel uncannily like conversation [6]. DeepMind's AlphaGo beat a world-champion Go player in 2016 — a moment that convinced a lot of people who had been sitting out the AI conversation that something fundamental had changed [7]. What is different about this wave is not the underlying idea. It is the compute, the data, and the accessibility. A high school teacher now has access to tools that ten years ago would have required a research lab.

Where the field is heading next

A newer pattern worth naming is agentic AI — systems that plan and carry out multi-step tasks on their own, rather than answering one prompt at a time. Mentions of agentic AI in U.S. job postings grew more than 280 percent in a single year [8]. The direction of travel is from AI that responds to AI that acts. It is also where the governance questions get sharpest, which we'll return to.

Why the definition matters in practice

Two consequences follow, both important. First, because outputs are probabilistic, human review is not a nicety. It is load-bearing. A system that is right 95 percent of the time still hands you a wrong answer roughly one time in twenty, and those wrong answers often arrive wearing the same confident tone as the right ones. Second, AI is rarely a product in itself. It is a layer inside other products. When a leader says "we are going to adopt AI," they almost always mean "we are going to adopt software with AI features." The interesting questions are about those features, not about AI in the abstract.

What the productivity research really shows

The honest version of the productivity story is more interesting than either the utopian version or the catastrophist one.

The agents who got better

Start with what is solid. Researchers at Stanford and MIT spent a year watching more than 5,000 customer-support agents at a Fortune 500 software company — most of them working from contact centers in the Philippines. Half were given access to an AI assistant; half were not. The agents with AI resolved 14 percent more cases per hour on average. For the novice and low-skilled agents — the ones who had been on the job a few months — the gain was 34 percent [9]. Senior agents saw almost no benefit.

Read that carefully, because it tells a human story that is easy to miss. The AI did not replace the senior agents. It compressed the experience curve for the newer ones. A two-month rookie started performing like a six-month veteran. The tacit knowledge the senior agents had built up in their heads — how to handle an angry customer, which escalation paths worked, what to ask a confused caller — was being distilled into the AI's suggestions and passed along to the people who most needed it. The machine was not doing the veterans' job. It was transmitting their expertise.

The broader pattern

Other studies have found similar patterns in similar domains. Stanford's 2026 AI Index report summarizes gains of 14 to 15 percent in customer support, 26 percent in software development, and up to 50 percent in marketing output [10]. The gains are real but they cluster around tasks where AI can function as a kind of patient coach. In domains that require deep reasoning or long-horizon judgment, the evidence for gains is thinner. Some early research is starting to hint that heavy reliance on AI may actually slow skill development over time — a concern worth taking seriously as we think about training the next generation of workers.

The part that gets missed

Then there is the home, which is the part of the story hardest to measure and possibly the most important. A recent study by Michael Blank of Stanford, Gregor Schubert of UCLA, and Miao Ben Zhang of USC tracked browsing behavior across more than 200,000 U.S. households before and after ChatGPT launched. The productivity gains on digital tasks at home — job searches, travel planning, comparison shopping, personal administration — ran between 76 and 176 percent [11]. That is more than double, and sometimes nearly triple, the previous speed.

It is a staggering number and it points to something that enterprise-focused analyses often miss. AI may be arriving first and fastest in the parts of life our economy has never been good at measuring — helping a parent draft an IEP letter for a child with learning differences, helping a small-business owner write a grant application, helping someone who cannot afford a lawyer understand a contract they have been asked to sign. The Federal Reserve estimates that generative AI contributed roughly a 1.1 percent boost to U.S. labor productivity by late 2024 [12]. The household number is much larger than the workplace number. That tells us the bottleneck on workplace gains is not the tools. It is the organizations.

The bottleneck worth dwelling on

That bottleneck deserves more attention than it usually gets. Nearly 90 percent of organizations are using AI somewhere [13]. Only about 6 percent are capturing meaningful enterprise-level returns [14]. The difference between the two groups is not that the winners bought better tools. The winners redesigned the work. They changed how jobs were defined, how decisions flowed, how performance was measured. The losers bolted AI onto processes that were never designed to accommodate it, and then wondered why the ROI never arrived.

There is a leadership lesson in that, and it is uncomfortable: the hard part is not the technology. It is the organization. Most leaders would rather buy a tool than redesign a workflow. AI rewards the ones willing to do both.

The harder problems

Ask most technology leaders what is going to slow them down on AI, and they will start with the technical list: compute costs, data readiness, integration complexity. These are real. They are also not the hardest problems.

CompTIA's 2024 survey asked businesses of every size what their biggest AI adoption challenge was. In every size category — small, medium, and large — the top answer was the same, at 46 percent: figuring out the right balance between humans and AI [15]. Not infrastructure. Not budget. Not vendor selection. The human question.

That answer should tell us something. Leaders sense, correctly, that the decisions with the biggest long-term consequences are decisions about people. Which tasks should AI take on? Which should stay human? Who decides? Who is accountable when the AI is wrong? How do we tell a worker their job is changing — and mean it, instead of sliding toward "we're giving you an AI so we can cut headcount later"? These are not problems a technology budget solves.

Underneath the human question sits a cluster of technical problems that are genuinely hard. In roughly the order I think they deserve attention:

Data you can actually trust

Every AI system is downstream of its data. If your customer records are fragmented across eleven systems and three spreadsheets, your AI is going to produce fragmented answers. The earlier hype cycles around "big data" and analytics left most organizations with data infrastructure that was more aspirational than functional. AI makes that debt visible in ways that are often embarrassing. It is not unusual for an AI pilot to quietly turn into a data cleanup project, which is exactly the project most organizations have been putting off for a decade.

Security and privacy, with attackers now using AI too

The defenders do not have a monopoly on this technology. AI-assisted attacks rose 72 percent between 2024 and 2025, and the average cost of a breach involving AI has crossed $5.7 million [16]. Gartner projects that by 2027, 40 percent of AI-related data breaches will come from employees using unsanctioned consumer AI tools with corporate data — what the industry has started calling "shadow AI" [17]. Stanford's AI Incident Database recorded a 56 percent jump in AI-related incidents in a single year [18]. Most organizations are aware of these risks and still moving faster than their mitigations.

Governance that is now the law

For much of 2023 and 2024, "AI governance" was a boardroom conversation about principles. It is not anymore. The European Union's AI Act came into force on August 1, 2024, with obligations for general-purpose AI providers applying from August 2, 2025 [19]. Penalties for prohibited AI practices reach €35 million or 7 percent of worldwide annual turnover — higher than GDPR [20]. In the United States, Colorado's AI Act (SB24–205) takes effect on February 1, 2026, and California's AI training-data transparency law (AB 2013) is already in force [21]. Organizations operating across borders are now operating under multiple regulatory regimes at once, and the landscape is moving fast enough that what was optional last quarter may be mandatory next quarter.

The evaluation problem

How do you know if an AI output is right? In domains with clear ground truth — did the model identify the fraudulent transaction correctly? — this is tractable. In domains where "right" is a judgment call — was this the right tone for that customer? did this summary capture what mattered? — the evaluation problem becomes a workflow problem. Somebody has to read the output and decide. Somebody has to have the skill and the time and the authority to push back. A lot of AI rollouts have foundered on that last word.

What unites these obstacles is that none of them are solved by buying a better model. They are solved by building better organizations — ones with cleaner data, tighter security, real governance, and workflows that keep humans in the loop where it matters.

The choice most leaders actually face

There is a decades-old debate in enterprise technology: buy or build? For AI, the interesting answer is that most organizations should do neither, at least not in the way the debate usually frames it.

Very few companies should be training their own foundation models. The economics are prohibitive — Alphabet alone reported more than $150 billion in annual capital expenditure in 2025, much of it AI infrastructure [22]. That is not a game a mid-sized business plays. A slightly larger number of organizations will want to buy standalone AI products — a transcription tool, a research assistant, a code generator — and integrate them into their workflows. For specific, well-defined problems, this is a reasonable path.

The pattern most organizations will actually experience

But the pattern that will define most organizations' AI experience is neither of those. It is what is already inside the tools they use every day. When Microsoft added Copilot to Excel and Word, that was AI arriving. When Salesforce added Einstein, that was AI arriving. When Adobe added generative fill to Photoshop, that was AI arriving. Enterprise surveys bear this out: around two-thirds of organizations expect to get their AI primarily through features embedded in tools they already use, rather than through separate AI products [23].

This is good news in a way. It means most leaders do not need a separate "AI budget" for most of what they want to do. They need to know which of their existing vendors are shipping AI features, whether those features are good enough to rely on, whether the data flowing into them is well-governed, and whether their teams are trained to use the features responsibly.

Where the real work lives

What this also means is that the hard work is not procurement. It is integration into the work itself. McKinsey's analysis of why a small number of companies are capturing outsized value from AI comes down to one finding that cuts through the rest: the high performers are nearly three times more likely than their peers to have redesigned workflows around AI, rather than inserting AI into workflows designed for a pre-AI world [24]. Seventy-three percent of product-development teams are not using AI agents at all [25]. The ones that are, and that have rebuilt how they work to take advantage of them, are pulling ahead.

In practice, that looks less like a technology initiative and more like an organizational-design exercise. Which decisions will the AI make? Which will it recommend? Which will it escalate? Who reviews what, and on what cadence? What happens when the AI is wrong — and how do we learn from that fast enough to prevent the next one? The answers vary by organization, by function, by risk tolerance. They cannot be outsourced to a vendor. They have to be worked out by the people doing the work, with leaders who are paying attention.

The single most useful thing a leader can do in the next twelve months is probably not to buy a new AI tool. It is to pick one workflow that matters, bring together the people who run it with the people using the AI features already available, and redesign that workflow from the ground up. Measure the before and the after. Share what you learn. Do it again.

Skills, dignity, and the work ahead

A few years ago, "AI skills" was a niche category on resumes — something data scientists had. Not anymore. Lightcast, which analyzes around 1.3 billion job postings, found AI-related postings rose 109 percent from 2024 to 2025, on top of a 73 percent rise the year before [26]. AI-related skills now appear in 2.5 percent of all U.S. job postings — a 297 percent increase over the past decade [27].

This is not a technology-department story anymore

The headline number that should focus the mind is this: 51 percent of job postings requiring AI skills are now in occupations outside of IT and computer science [28]. Marketing managers. HR professionals. Salespeople. Teachers. Financial analysts. Chemists. The demand for AI fluency has spread into every corner of the professional economy, and it is accelerating fastest in Human Resources (up 66 percent year over year) and in education and training (up roughly 200 percent for generative AI skills specifically) [29].

Employers are backing up their job descriptions with money. Postings that mention at least one AI skill advertise salaries 28 percent higher on average than comparable postings — about $18,000 more a year. Postings that mention two or more AI skills carry a 43 percent premium [30].

Here is what that means in human terms. A worker who invests the time to become reasonably fluent in AI — not "AI engineer" fluent, but "can use these tools well, knows their limits, can evaluate their outputs, understands the risks" — is adding a measurable amount to their own economic value. A worker who does not is quietly losing ground. That is not an abstract labor-market statistic. It is a personal career equation. And it applies as much to the accountant in Indianapolis as to the software engineer in San Francisco.

The equity question

This is also, in my view, where the human-centered case for AI education lives. The gap between the workers and communities who will learn to use these tools fluently and the ones who will not is shaping up to be one of the defining economic divides of the coming decade. If that gap tracks existing fault lines — wealth, geography, the accident of which school you went to, whether you have reliable internet at home — the next wave of technology will entrench the inequalities the last wave was supposed to close. If it does not, that will be because people worked to make it not.

Democratizing AI literacy is the kind of goal that sounds abstract until you are sitting across from a small-business owner who has never used a prompt and does not know where to start, or a community-college instructor who wants to teach her students these tools but has no curriculum and no time to build one. The tools themselves are cheap now. Training is the bottleneck. Whoever solves the training at scale — and solves it for the people the last wave of technology left behind — will shape what this wave becomes.

What the skill demand actually looks like

For organizations trying to build internal capability, the skill demands fall into four rough buckets: data and prompts (about 41 percent of AI-related postings), systems operations and architecture (26 percent), cybersecurity (15 percent), and coding (11 percent), with a long tail of governance, ethics, and domain-specific skills making up the rest [31]. The big surprise in that list is how heavily weighted it is toward working with AI — crafting inputs, evaluating outputs, managing data — rather than building AI. The era when "AI skills" meant "someone who writes neural networks" is ending. The era when it means "someone who can work well with a system that learned to predict" is here.

The part that won't go away

One more thing is worth saying on skills, and it is a little personal. There is a line attributed to Pasteur — fortune favors the prepared mind [32]. It is usually quoted in the context of scientific discovery. It applies here too. No one gets to opt out of this transition. But the preparation is available to anyone willing to do the work, and the preparation is not just technical. It is also about judgment — knowing when to trust the tool, when to push back on it, when to turn it off entirely. Judgment is a human skill. It always will be.

A human-centered close

I'll end with three things I believe, stated plainly.

One: this is real

AI is a major technological transition, comparable in scale to the shifts brought by electricity, the internet, and mobile computing. The evidence on adoption, investment, and productivity is too consistent to read as hype. Private AI investment grew more than 127 percent in 2025. Billion-dollar funding rounds for AI companies nearly doubled [33]. Whatever your prior view of the technology, this is not a fad that goes away if you wait it out.

Two: the outcomes are not predetermined

Every technology wave has been shaped — for better and for worse — by the choices that leaders, governments, workers, and communities made about how to use it. The internet could have been a commons. Mostly it isn't. Social media could have strengthened civic life. Mostly it hasn't. We are still early enough in AI's arc that the choices in front of us matter. Who gets trained. Who gets protected. Which problems we point the tools at. Whether we treat people as partners in the transition or as overhead to be reduced.

Three: human-centered is not a slogan

The frame that matters most is human-centered. Not "AI-first." Not "AI-transformation." Human-centered. That means designing systems that expand what people can do rather than replacing what they do. It means keeping humans in the loop on decisions that carry consequences. It means investing in training and access so that more people — not fewer — get to participate in what this technology becomes. It means being honest about risks and slow to trust claims of effortless productivity. It means remembering that every statistic in this paper represents actual people: the customer-support agent who got better at her job, the small-business owner who finally got a grant written, the student in a rural district who suddenly has a tutor available at midnight.

What to do today

If you are a leader reading this, here is the short version of what the evidence suggests you do:

  • Pick one workflow that matters, and redesign it with AI rather than adding AI to it. Measure the before and the after.
  • Treat data governance, cybersecurity, and AI governance as one integrated program. Regulators now do, and attackers always did.
  • Invest in AI literacy across your whole organization, not just in the technology function. The wage premium is larger in non-technical roles for a reason.
  • Keep a human accountable for every consequential decision. If you cannot name that person, the decision is not ready to be automated.
  • Tell the truth to your people about what you are trying to do. The fastest way to lose a workforce's trust on AI is to be vague about whether it is meant to help them or replace them.

If you are a worker reading this, here is the short version for you: spend some time every week using these tools — not to be impressed by them, but to figure out where they help you and where they do not. Notice the errors. Notice the blind spots. Notice what they free you up to do that you could not before. The people who will do well in the next ten years are the ones who developed informed opinions about these tools while they still had time to do so.

No debate will stop this wave from rolling in. But the shape of what it leaves behind — who it lifts, who it leaves — is still being written. The work of writing it well is the work of putting people first.

References

All sources cited in the footnotes above. URLs verified as of April 2026.

  1. Brynjolfsson, E., Li, D., & Raymond, L. R. (2023; revised 2025). Generative AI at work. NBER Working Paper №31161. National Bureau of Economic Research. https://www.nber.org/papers/w31161
  2. McKinsey & Company. (2025, November). The state of AI: How organizations are rewiring to capture value. Global Survey. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
  3. McKinsey & Company. (2025). The state of AI in 2025. See Note 2.
  4. Turing, A. M. (1950). Computing machinery and intelligence. Mind, LIX(236), 433–460.
  5. Samuel, A. L. (1959). Some studies in machine learning using the game of checkers. IBM Journal of Research and Development, 3(3), 210–229.
  6. Weizenbaum, J. (1966). ELIZA — A computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36–45.
  7. Silver, D., Huang, A., Maddison, C. J., et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529, 484–489.
  8. Stanford HAI & Lightcast. (2026). Annual AI Index 2026 — Labor market chapter. Stanford University. https://lightcast.io/resources/research/stanford-ai-index-2026
  9. Brynjolfsson et al. (2023; revised 2025). See Note 1.
  10. Stanford HAI. (2026). The 2026 AI Index Report — Economy chapter. Stanford University. https://hai.stanford.edu/ai-index/2026-ai-index-report/economy
  11. Blank, M., Schubert, G., & Zhang, M. B. (2026, February). The household impact of generative AI: Evidence from internet browsing behavior. SSRN Working Paper №6311439. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6311439
  12. Bick, A., Blandin, A., & Deming, D. (2024–2025). The rapid adoption of generative AI. Working papers, Federal Reserve Banks of St. Louis and San Francisco.
  13. McKinsey & Company. (2025). See Note 2.
  14. McKinsey & Company. (2025). See Note 2.
  15. CompTIA. (2024). 2024 AI progress report. Downers Grove, IL: CompTIA. Companion publication: CompTIA. (2024). Building AI strategy (whitepaper).
  16. Total Assure. (2026, January). AI cybersecurity statistics in 2025 (synthesizing IBM Cost of a Data Breach Report 2025 and CrowdStrike 2025 Global Threat Report). https://www.totalassure.com/blog/ai-cybersecurity-stats-2025
  17. Gartner, Inc. (2025, February). Predicts 2025: Privacy in the age of AI and the dawn of quantum. https://www.gartner.com/en/newsroom/press-releases/2025-02-17-gartner-predicts-forty-percent-of-ai-data-breaches-will-arise-from-cross-border-genai-misuse-by-2027
  18. Stanford HAI. (2025, April). The 2025 AI Index Report — Responsible AI chapter. Stanford University. https://hai.stanford.edu/ai-index/2025-ai-index-report
  19. European Parliament and Council. (2024, June 13). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (the AI Act). Official Journal of the European Union, L series (2024, July 12). https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  20. European Parliament and Council. (2024). AI Act, Article 99 (Penalties). See Note 19. See also European Commission. (2026, January). AI Act: Governance and enforcement. https://digital-strategy.ec.europa.eu/en/policies/ai-act-governance-and-enforcement
  21. Colorado General Assembly. (2024). Senate Bill 24–205, Consumer Protections for Artificial Intelligence Act. https://leg.colorado.gov/bills/sb24-205. California State Legislature. (2024). Assembly Bill 2013, Generative Artificial Intelligence: Training Data Transparency. https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202320240AB2013
  22. Stanford HAI. (2026). The 2026 AI Index Report — Economy chapter. See Note 10.
  23. CompTIA. (2024). 2024 AI progress report. See Note 15.
  24. McKinsey & Company. (2025). See Note 2.
  25. McKinsey & Company. (2025). See Note 2.
  26. Lightcast. (2025, July). Beyond the buzz: Developing the AI skills employers actually need. https://lightcast.io/resources/blog/beyond-the-buzz-press-release-2025-07-23
  27. Stanford HAI & Lightcast. (2026). Annual AI Index 2026. See Note 8.
  28. Lightcast. (2025). Beyond the buzz. See Note 26.
  29. Lightcast. (2025). Beyond the buzz. See Note 26.
  30. Lightcast. (2025). Beyond the buzz. See Note 26.
  31. Stanford HAI & Lightcast. (2026). Annual AI Index 2026. See Note 8; Lightcast. (2025). Beyond the buzz. See Note 26.
  32. Pasteur, L. (1854, December 7). Inaugural lecture, University of Lille. Original: "Dans les champs de l'observation, le hasard ne favorise que les esprits préparés."
  33. Stanford HAI. (2026). The 2026 AI Index Report — Economy chapter. See Note 10.

[1] E. Brynjolfsson, D. Li, and L. R. Raymond, "Generative AI at Work," NBER Working Paper №31161 (Cambridge, MA: National Bureau of Economic Research, April 2023; revised 2025). The study observed 5,179 customer support agents at a Fortune 500 software firm whose contact centers were located primarily in the Philippines; agents with AI assistance resolved 14% more cases per hour on average, with a 34% gain for novice and low-skill workers.

[2] McKinsey & Company, "The State of AI in 2025," Global Survey (November 2025). Survey of 1,993 organizations; 88% report using AI in at least one business function, up from 78% the prior year.

[3] McKinsey & Company, "The State of AI in 2025" (see note 2). Roughly 6% of firms are classified as "high performers" capturing disproportionate enterprise value; only 5.5% report material EBIT impact attributable to AI.

[4] A. M. Turing, "Computing Machinery and Intelligence," Mind, Vol. LIX, №236 (October 1950), pp. 433–460.

[5] A. L. Samuel, "Some Studies in Machine Learning Using the Game of Checkers," IBM Journal of Research and Development, Vol. 3, №3 (July 1959), pp. 210–229. Samuel is widely credited with coining the term "machine learning" in this paper.

[6] J. Weizenbaum, "ELIZA — A Computer Program For the Study of Natural Language Communication Between Man And Machine," Communications of the ACM, Vol. 9, №1 (January 1966), pp. 36–45.

[7] D. Silver, A. Huang, C. J. Maddison, et al., "Mastering the Game of Go with Deep Neural Networks and Tree Search," Nature 529, pp. 484–489 (January 2016).

[8] Stanford Institute for Human-Centered AI and Lightcast, "Annual AI Index 2026 — Labor Market Chapter" (Stanford University, 2026). Mentions of the "Agentic AI" skill cluster in U.S. job postings rose from 0.06% of postings in 2024 to 0.23% in 2025 — a more than 280% year-over-year increase, representing roughly 90,000 U.S. postings.

[9] Brynjolfsson, Li, and Raymond, "Generative AI at Work." See note 1.

[10] Stanford HAI, "The 2026 AI Index Report — Economy Chapter" (Stanford University, 2026). Peer-reviewed field studies report generative AI productivity gains of 14–15% in customer support, 26% in software development, and up to 50% in marketing output. Gains are smaller for tasks requiring deeper reasoning, and recent evidence suggests heavy AI reliance may carry long-term learning penalties.

[11] M. Blank, G. Schubert, and M. B. Zhang, "The Household Impact of Generative AI: Evidence from Internet Browsing Behavior," SSRN Working Paper №6311439 (February 2026). Analysis of browsing data from more than 200,000 U.S. households between 2021 and 2024 implies productivity gains of 76%–176% on digital tasks such as job search, travel planning, and personal administration. Also available as a Stanford Institute for Economic Policy Research (SIEPR) working paper.

[12] A. Bick, A. Blandin, and D. Deming, "The Rapid Adoption of Generative AI," working papers of the Federal Reserve Banks of St. Louis and San Francisco, 2024–2025. The authors estimate generative AI contributed roughly a 1.1% increase in U.S. labor productivity by the second half of 2024 relative to 2022.

[13] McKinsey & Company, "The State of AI in 2025." See note 2.

[14] McKinsey & Company, "The State of AI in 2025." See note 3.

[15] CompTIA, "2024 AI Progress Report" (Downers Grove, IL: CompTIA, 2024). Survey of 514 North American tech and business professionals. "Determining best AI/human interaction" was the top-ranked adoption challenge across all company-size segments (46% of large, medium, and small firms alike).

[16] Total Assure, "AI Cybersecurity Statistics in 2025" (January 2026), drawing on the IBM Cost of a Data Breach Report 2025 and the CrowdStrike 2025 Global Threat Report. AI-assisted attacks rose 72% year-over-year in 2025; the average cost of an AI-powered breach reached $5.72 million.

[17] Gartner, Inc., "Predicts 2025: Privacy in the Age of AI and the Dawn of Quantum" (February 2025). Gartner projects that 40% of AI-related data breaches will arise from cross-border generative AI misuse — typically unsanctioned ("shadow AI") use of consumer tools with corporate data — by 2027.

[18] Stanford HAI, "The 2025 AI Index Report — Responsible AI Chapter" (Stanford University, April 2025). The AI Incident Database recorded 233 AI-related incidents in 2024, a 56.4% year-over-year increase, spanning privacy violations, algorithmic bias, misinformation, and system failures.

[19] Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (the "AI Act"). Official Journal of the European Union, L series (12 July 2024). Entered into force 1 August 2024; general-purpose AI (GPAI) provider obligations began applying on 2 August 2025.

[20] EU AI Act, Article 99(3). Fines for prohibited AI practices may reach €35 million, or 7% of total worldwide annual turnover for the preceding financial year, whichever is higher. High-risk system violations may reach €15 million or 3% of turnover under Article 99(4).

[21] Colorado Senate Bill 24–205, the "Consumer Protections for Artificial Intelligence Act" (signed 17 May 2024; effective 1 February 2026); California Assembly Bill 2013, the "Generative Artificial Intelligence: Training Data Transparency Act" (signed 28 September 2024; effective 1 January 2026). A growing number of additional U.S. states have enacted or introduced AI-related legislation.

[22] Stanford HAI, "The 2026 AI Index Report — Economy Chapter." Major cloud providers accelerated AI-related capital expenditures in 2025; Alphabet (Google's parent company) reported annual capital expenditures exceeding $150 billion.

[23] CompTIA, "2024 AI Progress Report" (see note 15). Among surveyed firms (n=492), approximately two-thirds expect to acquire AI primarily through features embedded in business tools they already use, rather than by purchasing standalone AI products or developing AI internally.

[24] McKinsey & Company, "The State of AI in 2025." High-performing firms are nearly 3× more likely than peers to have fundamentally redesigned workflows as part of their AI efforts. See note 2.

[25] McKinsey & Company, "The State of AI in 2025." Approximately 23% of respondents report scaling agentic AI somewhere in the enterprise; 73% of product-development respondents report no AI-agent use at all. See note 2.

[26] Lightcast, "Beyond the Buzz: Developing the AI Skills Employers Actually Need" (Moscow, ID: Lightcast, July 2025). Analysis of more than 1.3 billion job postings; AI-related postings rose 109% from 2024 to 2025, on top of a 73% rise the prior year.

[27] Stanford HAI and Lightcast, "Annual AI Index 2026" (see note 8). AI-related skills are mentioned in 2.5% of all U.S. job postings — up 55% year-over-year and 297% over the past decade.

[28] Lightcast, "Beyond the Buzz" (see note 26). As of 2024, 51% of U.S. job postings requiring AI skills are in occupations outside IT and computer science. Generative AI postings for non-tech roles have grown more than 800% since 2022.

[29] Lightcast, "Beyond the Buzz" (see note 26). Human Resources recorded the fastest annual growth in AI-skill demand at 66%; education and training showed the highest growth in generative AI skill demand at roughly 200%.

[30] Lightcast, "Beyond the Buzz" (see note 26). Postings mentioning at least one AI skill advertised salaries 28% higher on average — approximately $18,000 per year — than comparable postings without AI skills; postings mentioning two or more AI skills carried a 43% premium.

[31] Stanford HAI/Lightcast analysis of U.S. job posting data, as reflected in the Annual AI Index 2026 and Lightcast's "Beyond the Buzz" (2025). Approximate distribution of AI-related skill demand: data and prompts (41%), systems operations and architecture (26%), cybersecurity (15%), coding (11%), other (7%, including governance, ethics, and domain-specific skills).

[32] Commonly attributed to Louis Pasteur, inaugural lecture at the University of Lille, 7 December 1854: "Dans les champs de l'observation, le hasard ne favorise que les esprits préparés." ("In the field of observation, chance favors the prepared mind.")

[33] Stanford HAI, "The 2026 AI Index Report — Economy Chapter." Private AI investment grew 127.5% year-over-year in 2025, with generative AI capturing nearly half of all private AI funding; newly funded AI companies rose 71%, and billion-dollar AI funding events nearly doubled. See note 10 for full reference.

Chris Daily is the Managing Director of InnoPower LLC and a Certified AI Master Trainer. Explore the full course library at belithe.com.