What should investors think about an exploding industry sector that makes useful, but highly perishable, products that are hugely expensive to build and costly to operate? The conventional wisdom is that we are in the early stages of an historic IP land grab where the first handful of companies to stake claims will build durable AI franchises running their frontier foundation models, trained on pricey Nvidia chips, profitably inferencing away in gleaming new gigawatt scale data centers. But right now, the only companies making money are selling picks and shovels for this AI gold rush — the chip vendors, power suppliers and data center operators. The more glamorous model developers are all growing revenue, but losing money, and collectively burning mountains of cash.
Regarding the future of AI, let’s stipulate that you’ve heard a lot of opinions in the past few years about how fast things will change, how big the market will be and how society will be transformed by this new foundation for modern computing. The most extreme version of this excitement predicts the near-term achievement of artificial general intelligence (“AGI”), where machines become smarter than people at all material tasks. In this scenario, superintelligent AI leads to the restructuring of the world economy and mass unemployment as the fundamental role of labor is redefined. Although a tantalizing prospect, we expect a different scenario to unfold.
We aren’t doubters about the core appeal of AI, only skeptical about how fast we can travel on the rocky road of implementation and new technology deployment. We know from personal experience the many challenges. This makes us realists about timing, costs and technology integration paths. Very specifically, we are dubious about the current capital spending plans for the most high-profile participants and incredulous about valuations for most of the venture-backed players operating at super scale. In our view, these may be among the riskiest investment wagers ever made due to the size of the financings and associated valuations when measured against the modest use cases to date, relative immaturity of the companies, and their nascent business models.
You’ll recollect that, in our white paper published last year, we asked, “What’s all the fuss about AI?” Despite the attention, publicity and plaudits, we concluded:
- The AI phenomenon may actually be UNDER-HYPED. AI-based technology will make a bigger impact than you think BUT, and this is crucial, will take much longer than you imagine to do so — adoption won’t all happen in the next three years; it might take a full generation.
- The last 30 years have been about getting machine learning and AI to work; the next 30 years will be about implementation, which will require nothing less than the wholesale upgrade and renovation of our entire computing infrastructure. That’s a huge project and can’t be accomplished overnight.
Why will it take so long? We know the whole AI juggernaut feels like a technology supernova, but we would underscore that this development and adoption cycle is a very big bucket of change, maybe the largest ever imposed on a mature, now highly integrated world economy. In short, we believe:
- New technology models can evolve quickly, but the associated infrastructure will take time to fund, construct and optimize. We had cars before roads and a national highway grid. We had trains before tracks and a transcontinental railroad system. AI will inevitably be more of the same because we are lacking most of the necessary supporting infrastructure around chips, power, and data center capacity – all currently in short supply.
- The real constraint is inside companies, many of which have fragmented technology stacks, legacy systems and a myriad of change management challenges. There is a general lack of preparedness to trial, integrate and scale the new technology. Domain-specific tuning for new models also slows the adoption pace. Not surprisingly, restaurants need different features than jet engine suppliers. Lawyers want something different than machine tool operators.
- Perhaps most poignantly, while technology can evolve rapidly, human behavior does not. Aside from the necessary infrastructure build-out and corporate practices, cultural habits evolve more slowly than most people think. People are stubborn and habits create inertia.
For context and as a reference point, it’s useful to think about the effect of social friction on the otherwise remarkable rise of e-commerce. Amazon was founded 31 years ago in 1994, and since that time has led a relentless expansion of online retailing across the economy and around the world. But now over three decades later, with thousands of competitors, massive infrastructure investment in trucks and warehouses, and despite its much-deserved success, total e-commerce from all suppliers across every merchandise sector is only just over 15% of retail sales in the US (Census, 2025) and similar in Europe (Forrester, 2025). That’s all to say, change happens slower than you think, even for very good ideas and highly successful business models.
THREE KEY DEVELOPMENTS
Said another way, we believe the road to full AI implementation is long with many way stations, where new model innovations, operating efficiencies and user insights about the highest value applications will pace and modulate the implementation progress. Producers AND customers both need to be “venturesome.” That’s been the case for all major technological transformations, including canals, railroads, electricity, semiconductors, computing and most recently, the internet.
But change is happening, and it’s momentous. Here are the three key AI developments that matter most:
- Agents;
- Rise of digital labor; and
- The coming automation boom.
These three concepts define what’s most important about the next chapter of the AI Revolution, which in turn, is central to BayPine’s overall digital transformation thesis. Note that none of these developments require AGI, which we believe cannot be reached on our current development path simply with more scaling (i.e., larger models, more tokens, faster chips, bigger data centers). Real AGI will require new approaches for capturing and modeling intelligence, expanded datasets for training, and breakthroughs around deep reasoning that transcend the current probabilistic foundations of the leading LLMs. We’re not saying it’s impossible, just not imminent.
Rather, we see the related complex of agents, automation and digital labor as part of a predictable evolution, foreshadowed by technical progress, growing familiarity, lower costs and more compelling business use cases. To reprise, here’s what we predicted a year ago about AI progress. These are the immediate antecedents:
WHAT WE PREDICTED
- An investment mania: That’s definitely underway. We’ve already seen this year the largest and most expensive venture financings ever done at Anthropic, XAI and most prominently at OpenAI, a company that has set a record for the most money ever raised at the highest valuation by a private company. AI-centric public companies like Figma are trading at over 18x revenue and Palantir is selling at 107x price/sales. And of course, AI chipmaker Nvidia is now the most valuable company in the world. The mania is in full bloom, and many are speculating that we are in an AI “bubble” (Forbes, Nvidia, 2025).
- Rise of specialty models: New versions are proliferating from all the leading vendors, including open source, mini-models, OEM editions, and reasoning versions. Many different product choices are emerging (Markus, 2025).
- Dramatic price reductions: This is happening now, led by the Chinese, who are charging 5-20% of what US competitors are asking. For example, the DeepSeek base model is priced at 14 US cents (DeepSeek R1 Review, 2025) and Alibaba at 41 cents (Solutions, 2025) per million tokens for OEMS vs. ChatGPT 5 at $1.2 per million tokens and $15 per million tokens (OpenAI, 2025) for OpenAI’s reasoning model. It’s dollars going to pennies.
- Galloping improvements in functionality: New offerings like ChatGPT 5 have recently improved math processing, coding skills and reasoning abilities along with a slew of other features. But even as the various frontier models become more capable, there is less differentiation than you might think. The usability issue now is all about performance, product packaging and market segmentation (Forbes, OpenAI Releases GPT-5—Here’s What’s New with the AI Model Behind ChatGPT, 2025).
- Finally, we predicted the first wave of commercial customer utility would revolve around personal productivity tools and embedded features, primarily powered by chatbots. These are appearing everywhere – now in Apple Siri, Amazon Alexa and Google Assist, and in enterprise software apps like Salesforce, Oracle, SAP and thousands more (Business Case Studies, 2024).
All that’s happening now, unfolding as we had anticipated.
NOT SO SECRET AGENTS
So, what’s happening next? Maybe not what you’d expect IF you are simply imagining more and better chatbots. We are now on the cusp of a major change in how we experience AI. The conventional view in many business circles is that AI is a power tool for information workers. And to be fair, a significant portion of the first wave of applications have appeared as personal productivity tools – exemplified by the near ubiquity of chatbots like Claude, ChatGPT and Gemini, plus all the very capable Chinese competitors such as DeepSeek V2, the Qwen Series from Alibaba, ERNIE from Baidu and many more. Unfortunately, this easy familiarity has obscured the current rise of agentic capabilities and coming revolution in task, process and full-blown job automation.
What’s the difference between chatbots and agents? You might think of it broadly as the difference between chatbots “saying” and agents “doing.” The generative AI chatbots are increasingly functioning as a lowest common denominator computing interface for white collar workers, where the AI is primarily used to create new content (e.g., text, speech, images, video) based on user prompts. That’s extremely useful for generating ideas, doing research, helping draft contracts or preparing corporate presentations. There are thousands of other uses, but the base utility revolves around “saying,” where the technology functions as a trusty assistant and, maybe in the future, as your primary interface to mail, web browsing, spreadsheets and composition tools.
Agents, in contrast, focus on autonomous decision-making and independent task execution – “doing.” Agents aren’t a different thing than AI, just a more evolved form built on the same intelligent LLM foundations as the chatbots, but with more tools and capabilities added, allowing machines to take action – first in the digital domain and increasingly in the real world.
Agents are already making complex stock market trades – seeing more data, recognizing patterns and reacting faster than a human can. In our investing world, think about the extraordinary growth of quant hedge funds where investment insights and data analytics are now reflected in rapid response algorithms. There are payables agents working in controllers’ offices reviewing expense claims, checking receipts, and making payments. AI customer service will soon be able to handle 80% of what humans previously managed at call centers, including complex technical queries in multiple languages, diagnosing problems, and documenting the issues and resolving them (Gartner, 2025).
These new autonomous agents are capable of making independent decisions and executing them without human oversight. The best agents require no or very limited intervention. Pause to consider that. No other technological invention in our history, however important, has ever achieved that capability before. Not printing presses, not light bulbs, not vaccines, not atomic energy, not semiconductors.
The best agents not only act autonomously, but they can also tackle complex problems, plan, test, react, adapt to their environment and continue learning. They have personalities and can consult, remember, and communicate clearly. And through sensors and robotic articulation these agents are growing increasingly comfortable operating beyond the digital realm in our physical world — not a software-defined playpen or artificial reality but live in the chaotic stew of real life. These are called world models, trained not just on textual information, but increasingly on data from robots, LIDAR sensors and videos that inform the models about everything from gravity to weather and physics to football.
Maybe the easiest place to see that future is with autonomous vehicles, driverless cars making their debut with Waymo, the autonomous taxi company started by Google ten years ago. These sensor-equipped taxis are now operating in six US cities and delivering close to 400,000 rides a week – with no drivers. The company has already logged more than 10 million rides covering over 100 million miles (Reuters, 2025) with 80% fewer injury-causing crashes (Waymo, 2025) and 90% lower insurance claims than human drivers in their same geographies (Reinsurance News, 2024). Technology doesn’t just work; it’s far safer than the human alternative. And when surveyed, 70% of passengers who’ve tried both prefer the driverless option to Uber or Lyft (PRNewswire, 2025).
DIGITAL LABOR
We are convinced that the rise of agents and their close cousins, robots, will have a profound effect on the shape of the modern economy and structure of the global workforce. In fact, we would encourage you to think about agents and AI-powered robots as digital people, not software programs or mere machines.
So, if you think about AI agents as digital colleagues, more like a new life form than a gadget, here’s the key question. Will AI augment human capabilities or replace labor? AI is already doing both, but the relative weight of the two trajectories has wildly different implications. The augmentation story, making doctors better at diagnosing medical problems or helping lawyers draft better briefs, is an uplifting notion. And comforting in the expectation that these new capabilities might be like cell phones, MRI machines or ATMs, making jobs easier to perform, employees more efficient and society more productive. It’s all those things and a great toolset. But it’s not just that. Because agency is destiny.
Said starkly, we are ALSO probably trending toward the replacement paradigm because the economic rationalization for employing agents instead of humans is so easy to understand and fund. If automating a giant distribution center can reduce employee count and associated labor costs by 60%, that pays for a lot of AI-enabled robots (Contimod, 2025). Not surprisingly, Amazon already has more robots than workers in their warehouses and distribution centers, over one million (TechCrunch, 2025). These robots are getting more capable and less expensive every year. We spoke recently with the CEO of a Fortune 500 business who told us that, by adopting AI agents, his customer service costs have gone from growing 10% a year to declining 20% annually, a 30 percentage-point margin swing. As for capability, the bots can handle 95% of inbound calls with better satisfaction rates than human operators (Gartner, 2025). They also work 24 hours a day, take no vacations, don’t complain, get hernias or require complicated benefit packages. Think of what that “always-on” calculus means for managers choosing between man and machine.
THE AUTOMATION BOOM
But it won’t just be routine warehouse jobs, customer service representatives, security guards, truck and taxi drivers, waiters, and clerks. Agent-based automation will also be coming for high-end professional services – meaning lawyers, financial analysts, project managers, human resources executives, and graphic designers. Or said another way, anybody who works on a laptop.
- Software programmers may be the best current example of the replacement paradigm already underway. Employment in the profession has flat-lined and is dropping in many sub-specialties, even as demand for coding continues to increase in almost every industry. Microsoft laid off 9,000 people while announcing record revenues and profits (Times, 2025). Microsoft CEO, Satya Nadella, said 20-30% of their code is now being written by AI (LlamaCon, 2025). Marc Benioff, CEO of Salesforce, announced the company won’t be hiring any new software engineers in 2025 because they are achieving 30% productivity gains with various AI coding technologies (Techradar, 2025). And Mark Zuckerberg is aiming to have 50% of his code written by AI in 2026 (Digit, 2025).
- Finance will be another early casualty where essentially all the major decisions are data points about the handling of money – who owns it, where to keep it, how to lend or borrow it, how much interest is owed, when to pay it back, and where to send it. It’s neat, tidy and highly numeric. Now, it can also be infinitely complex. It’s perfect raw material for the AI combine to harvest. As an early marker, Ramp, a five-year-old start-up that builds agents to automate financial operations like expense management, just raised $500 million at a $22.5 billion pre-money valuation (Newswire, 2025).
STRUCTURAL CHANGES IN THE LABOR FORCE
The nearest recent global employment disruption of this scale may have been the China shock, when the West accelerated outsourcing initiatives and ultimately admitted China to the WTO in 2001. Free trade with this Asian juggernaut and access to that country’s huge low-cost labor pool, combined with NAFTA, gutted US and European manufacturing. Over the past three decades (1993-2023), manufacturing jobs have declined from 15% to 8% of all US employment (FRED, 2025). Think of the political strife, rust belt social pathologies and population dislocations caused by this precipitous decline.
Let’s contemplate the services sector of our economy, which is much bigger than manufacturing. It’s 80% of total employment when you include healthcare, so almost 10x the employment numbers for manufacturing (Statista, 2025). Now think about outsourcing accountants and financial clerks to agentic AI companies like Ramp, software coding to companies like Cursor, and drivers to companies like Waymo. Let’s dimensionalize that. There are approximately three million accountants (Accountants-BLS, 2025) and financial clerks (Clerks-BLS, 2025), 4.4 million software engineers (Scalers, 2025) and four million taxi drivers, chauffeurs (Drivers-BLS, 2025) and truck drivers (Trucking.org, 2025) in the US — a total of 11.4 million service jobs at risk in just these three sub-categories. That’s almost as many jobs as the 12.7 million manufacturing jobs left in the United States. You get the picture about the potential economic, social and political dimensions of this drama, which is already in motion and will play out inevitably over the next decade.
We frequently label techno-skeptics as modern Luddites. They were the people who protested the introduction of automated looms in English textile mills two hundred years ago. During this time of great change, it’s useful to remember the Luddites weren’t wrong about the impact of the frightening new technology on labor. Jobs for weavers dropped by 80% from 1800 to 1830, and wages fell 40% in just six years between 1804 and 1810. Unemployment in factory towns led to destitution, hunger and widespread protests. Unfortunately, in this age before welfare and unemployment insurance, it took almost four decades to tame the social distress and re-absorb the displaced workers. It’s an expensive lesson we can’t afford to relearn because we are about to unleash a generation of ferocious automation in our own economy (World History, 2023).
THE BIG OPPORTUNITY
Should we despair, maybe even panic? Will there be mass unemployment? Of course not. And for three main reasons:
- First, while remarkably good and getting better rapidly, AI models are still challenged to capture the full range of human capabilities. The models presently lack full intuition, future vision, and the ability to manage “net new,” what we might think of as entirely novel situations and breakthrough thinking. We humans still have plenty to offer our customers, co-workers and employers in the creative realm.
- Second, as amazing as they are, AI models struggle to manage context across time, domains and mediums. People, by contrast, are very good at integrating sources of data, judging their relevance and assessing quality of inputs. We have the ability to assemble information from memory, analytical processes, lived experience, conversations, sensory inputs, social interactions and emotional cues. Computers, so far, are largely constrained to their narrow, short-term and mostly textual knowledge bases assembled from training data. That structural gap is the real constraint around machines learning and eventually emulating more of what humans can do.
- Third, in every major technology revolution before this one, our economy has adjusted and re-employed the people who’ve lost jobs to machines, automation and evolving business models. In the last century, our own economy saw employment in the agricultural sector drop from 40% of working Americans in 1900 (GilderLehrman) to under 2% today (Economy, 2023). But despite these dislocations, recessions and the Great Depression, we’ve always returned to full employment.
The very good news is that AI will allow us to create entirely new jobs, redefine what companies do, and build whole new kinds of businesses. For example, Meta recently agreed to pay $15 billion for half of Scale.AI, a company that labels, creates and manages synthetic data sets for training AI models, a brand-new concept (Verge, 2025). AI training and consulting firms are also booming, and there is a dire shortage of application engineering talent for integrating AI technologies into the workplace.
The $30 trillion US economy needs to adjust by educating our population for the new reality and relentlessly re-tooling our businesses to thrive in an AI-animated world. And that’s the core of the opportunity for BayPine – helping our best companies prosper in a digitally denominated future.
You make digital progress by doing it – not studying the possibilities, hiring consultants or forming committees. There is no substitute for setting priorities, picking a target, rolling up your sleeves and executing. BayPine’s portfolio companies have adopted a comprehensive digital transformation agenda driving measurable improvements across each of their businesses.
Here is a small sample of how we are using AI to deliver improved operating results and bottom-line value across our portfolio:
- Penn Foster, a digital learning platform, has deployed AI voice agents for learner support – providing tailored guidance, curricular support, encouragement, and more.
- Mavis, an automotive services tire business, is deploying a new store operating system that uses AI to monitor supply and demand, then dynamically adjust prices, to drive staffing levels and to manage capacity utilization at the individual shop level.
- CenExel, a network of integrated clinical research sites that recruits and monitors patients in clinical trials conducted by pharmaceutical and biotechnology companies, has deployed AI to streamline compliance and quality assurance in its core trial processes.
- QualDerm Partners, a multi-state dermatology management services organization, is using AI to help providers make more accurate diagnoses and to analyze demand patterns in new markets to determine future clinic locations.
- Harbor Global, an IT services business advising law firms and corporate legal departments, is building and shipping AI capabilities to its customers, including tools that advise during negotiations, automate legal research, and build competitive intelligence.
- POLYWOOD, a high-density polyethylene outdoor furniture brand, is using GenAI tools to create personalized text and images for more effective digital marketing.
The success formula for employing AI takes senior sponsorship, usually from the CEO, central control by a technology-fluent domain expert as an internal advocate, and active project coordination with business owners led by a chief transformation officer. BayPine supports companies in these efforts by building digital fluency across senior leadership teams, helping them navigate change management challenges and providing an experienced hand on the tiller for complex technology implementation projects. Finally, we preach monomaniacal focus on a few key initiatives delivered with an overwhelming sense of urgency. You simply have to be in a hurry.
Real digital transformation using AI requires planning, concerted focus, active change management and substantial resource investment. AI-led change is not automatic, easy or cheap. Don’t confuse the real personal productivity improvements of ChatGPT subscriptions, which are terrific, with the opportunity to re-imagine and comprehensively renovate the business models of high-potential companies.
Citations
Accountants-BLS. (2025, August). Accountants and Auditors. Retrieved from Accountants-BLS: https://www.bls.gov/ooh/busine...
BusinessCaseStudies. (2024, December). AI-Powered Personal Assistants: A Comparative Analysis. Retrieved from Business Case Studies: https://businesscasestudies.co...
Census. (2025, August). Quarterly Retail E-Commerce Sales. Retrieved from Census: https://www.census.gov/retail/...
ClearTax. (2025, October). World GDP Ranking 2025 List. Retrieved from ClearTax: https://cleartax.in/s/world-gd...
Clerks-BLS, F. (2025, August). Financial Clerks. Retrieved from Financial Clerks-BLS: https://www.bls.gov/ooh/office...
Contimod. (2025, March). 47+ Warehouse Statistics: Industry, Automation, & More For 2025. Retrieved from Contimod: https://www.contimod.com/wareh...
DeepSeek R1 Review. (2025, June). Retrieved from Apidog: https://apidog.com/blog/deepse...
Digit. (2025, May). Mark Zuckerberg says AI will write most of Meta’s AI code by 2026. Retrieved from Digit: https://www.digit.in/news/gene...
Drivers-BLS, T. (2025, August). Taxi Drivers, Shuttle Drivers, and Chauffeurs. Retrieved from Taxi Drivers-BLS: https://www.bls.gov/ooh/transp...:
Economy, T. G. (2023). USA: Employment in agriculture. Retrieved from The Global Economy: https://www.theglobaleconomy.c...
Forbes. (2025, October). Nvidia. Retrieved from Forbes: https://www.forbes.com/sites/t...
Forbes. (2025, August). OpenAI Releases GPT-5—Here’s What’s New With The AI Model Behind ChatGPT. Retrieved from Forbes: https://www.forbes.com/sites/a...
Forrester. (2025, February). European E-Commerce Sales To Surpass €565 Billion By 2029. Retrieved from Forrester: https://www.forrester.com/pres...
Gartner. (2025, March). Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues Without Human Intervention by 2029. Retrieved from Gartner: https://www.gartner.com/en/new...
Geekforgeeks. (2025, August). ChatGPT 5 Explained Everything: What's New and How to Use. Retrieved from Geekforgeeks: https://www.geeksforgeeks.org/...
GilderLehrman. Statistics: Trends in American Farming. Retrieved from GilderLehrman: https://www.gilderlehrman.org/...
India, M. (2025, April). Mark Zuckerberg Wants AI To Do Half Of Meta's Coding By 2026. Retrieved from Mashable India: https://in.mashable.com/tech/9...
Markus, A. (2025, October). Why the Rise of Specialized Open-Source AI Models Deployed at the Edge Will Redefine AI Infrastructure in 2025. Retrieved from Artur Markus: https://www.arturmarkus.com/wh...
Newswire, P. (2025, July). Ramp Raises $500 Million at $22.5 Billion Valuation to Accelerate AI and Build the Future of Finance. Retrieved from PR Newswire: https://www.prnewswire.com/new...
Openai. (2025). Openai Pricing. Retrieved from Openai: https://openai.com/api/pricing...
PRNewswire. (2025, June). Obi Releases First In-Depth Analysis of Waymo's Entry Into the Rideshare Market. Retrieved from PRNewswire: https://www.prnewswire.com/new...
Reinsurance News. (2024, December). Waymo shows 90% fewer claims than advanced human-driven vehicles: Swiss Re. Retrieved from Reinsurance News: https://www.reinsurancene.ws/w...
Reuters. (2025, July). Alphabet's Waymo picks up speed as Tesla robotaxi service expands. Retrieved from Reuters: https://www.reuters.com/busine...
Scalers, T. (2025, March). How many software engineers are in the US? Retrieved from The Scalers: https://thescalers.com/develop...
Solutions, C. D. (2025, January). LLM Price War: Alibaba’s Qwen-VL-Max at 16% of GPT-4o’s Price, But Overseas Sales Challenges Loom. Retrieved from Ctol Digital Solutions: https://www.ctol.digital/news/...
Statista. (2025, July). Distribution of the workforce across economic sectors in the United States from 2013 to 2023. Retrieved from Statista: https://www.statista.com/stati....
TechCrunch. (2025, July). Amazon deploys its 1 millionth robot, releases generative AI model. Retrieved from TechCrunch: https://techcrunch.com/2025/07...
Techradar. (2025, February). Salesforce CEO says no plans to hire more engineers, as AI is doing a great job. Retrieved from Techradar: https://www.techradar.com/pro/...
Times, T. E. (2025, August). Microsoft lays off thousands, makes $27 billion profit in Q2 — CEO Satya Nadella doubles down on AI mania. Retrieved from The Economic Times: Microsoft job cuts due to AI: Microsoft lays off thousands, makes $27 billion profit in Q2 — CEO Satya Nadella doubles down on AI mania - The Economic Times
Trucking.org. (2025). Trucks moved roughly 72.7% of the nation's freight by weight in 2024. That is just one of many statistics calculated and tracked by American Trucking Associations' economics department. Retrieved from Trucking.org: https://www.trucking.org/econo...
Verge, T. (2025, June). Meta is reportedly making a $15 billion bet on AGI. Retrieved from The Verge: https://www.theverge.com/news/...
Waymo. (2025). Safety. Retrieved from Waymo: https://waymo.com/safety/#:~:t....
WorldBank. (2025, January). Employment in industry (% of total employment) (modeled ILO estimate) - United States. Retrieved from WorldBank: https://data.worldbank.org/ind...
WorldHistory. (2023, March). Luddite. Retrieved from WorldHistory: https://www.worldhistory.org/L...