Summary: Over the upcoming decade, AI-provisioned intelligence will become almost free, and instantly available. AI won’t just assist professionals — it will take over much of the work as a packed instant service provider. Welcome to the age of boundless skill scalability, where services transform into software and economies grow beyond human imagination.
Imagine a world where professional expertise is a few clicks away—faster, smarter, and unbelievably affordable. This isn’t science fiction; it’s the next revolution in SaaS. AI becomes your consultant, designer, and researcher, redefining what’s possible in every corner of work. This will create a new model for service design, where services that used to be provided by humans are packaged up and delivered by software. We can call this “Service as a Software”, reusing the acronym SaaS that has long been used for “Software as a Service.”
(Watch the jazz music video I made based on this article. YouTube, 2 min.)
Software as a Service (the original SaaS) is a cloud computing model in which software applications are hosted by providers and delivered to users over the Internet on a subscription basis, eliminating the need for local installation and maintenance.
The infamous “cloud” is a buzzword for “somebody else’s computer” located in an unknown place and accessed over the Internet. (Ideogram)
Remote computing services began in the 1960s when IBM offered mainframe computing power to large organizations through time-sharing systems. The modern SaaS concept emerged in the late 1990s with companies like Salesforce and Concur offering internet-based applications. The industry gained significant momentum in the early 2000s, with Gmail’s 2004 launch marking one of the first mass-market SaaS products for consumers.
SaaS is now the dominant software delivery model, with remarkable market growth and adoption: The global SaaS market is projected to reach $720 billion by 2028.
My prediction is that the new SaaS, service as a software, will be dramatically bigger. It can potentially disrupt all knowledge work currently done by human business professionals. There are approximately one billion knowledge workers in the world, and making service into software to replace or augment that many expensive employees will likely be worth around $50 trillion. How fast will this happen? The low-hanging fruit (such as UX consultants, as I discuss below) may be replaced in 4-5 years, whereas bigger systematic changes to the world economy might easily take two decades. But in 20 years, I expect service as a software to contribute $50T to the world economy, whereas software as a service “only” contributes $5T.
About 10x more gain from flipping the model!
Examples:
Patient being diagnosed by an AI model instead of a human doctor. (Research shows that AI already has higher diagnostic accuracy than human physicians.)
People having their will written by AI instead of a human lawyer. (This has long been possible with pre-AI legal software using templates, but AI will allow for much more flexibility and understanding of legal subtleties, as well as the ability to read and understand the client’s financial statements.)
Product companies conducting user interviews by AI instead of hiring a UX consultant. (This is already happening. For example, luggage company Away used AI service Outset to interview 78 customers about their reasons for choosing hard-sided versus soft-sided suitcases. It used the findings to design web pages for a new line of soft suitcases with a 25% higher conversion rate than their previous pages.) User interviews were the number-one user research methodology employed by UX researchers in a recent study, so conducting them at scale with AI is particularly useful.
When considering scenarios like these, don’t limit your imagination to the capabilities of current AI, which still performs professional services at a rudimentary level. Think of the Ph.D.-level AI we will have in 2027 — or the super-intelligent AI expected by 2030.
The service of conducting and analyzing a large series of user interviews can be packaged up and done by AI software. (Leonardo)
Corporate Examples
Even though the most powerful service-as-a-software products may require 2027-level AI capabilities, we already see initial successes with current-level AI. A few examples of companies that successfully implemented AI-based solutions to automate service and consulting work:
Hilton Hotels: AI-powered chatbots assist guests with booking, check-in, and personalized recommendations, improving guest satisfaction and operational efficiency.
Delta Airlines: AI analyzes customer data to predict potential service issues, improving passenger satisfaction and loyalty.
DHL: AI streamlines delivery routes, resulting in faster delivery times, reduced fuel consumption, and improved operational efficiency.
JPMorgan Chase: JPMorgan Chase implemented AI to enhance risk management and fraud detection processes using their Contract Intelligence (COiN) platform, which uses machine learning to review legal documents and extract essential data points.
And here are some current AI products that act as “consultants-in-a-box”:
Bloomfire: An AI-based knowledge management platform that captures and organizes collective knowledge, making it accessible across the organization.
HubSpot Customer Platform: Suite of marketing, sales, customer service, CMS, operations, and commerce software doing marketing automation, sales CRM, customer service, content management, and operations management.
Salesforce Einstein Analytics: An AI-driven predictive analytics tool that analyzes customer data to predict behavior and preferences to create targeted customer experiences.
UiPath: Robotic process automation (RPA) that automates repetitive tasks by mimicking human actions.
QuickBooks Online: AI-driven accounting software that automates expense tracking, invoicing, and financial reporting.
Your team of executive consultants has been shipped in and unboxed. The full team of world-class experts works for $20 per month. (Midjourney)
These two bullet lists are another example of service as a software: I asked Google’s “Deep Research” product to scour the literature for recent examples of professional service work being taken over by AI.
Traditionally, a company that wanted to understand a market shift like professional services automation would pay $10 million to a company like McKinsey to study and summarize the research. Now, $20 to Google does the trick.
In fairness to McKinsey, their report would have been much more elaborate than the 9-page document with 38 references I got from Google Deep Research. Crucially, it would have come with a slide deck with snazzy graphics that a consultant with a highly polished stage presence would present to your executive team. (I used to be a consultant — I know what clients pay for: compelling stage presence and snazzy visuals are often more valued than the depth of the insights, because convincing the executives is the most important part of most consulting projects.)
However, by 2030, the AI report will likely be more insightful than McKinsey’s and will still only cost $20. It may even be presented by much-evolved AI avatars with a more charismatic stage presence than any human, except possibly a few Oscar-winning actors who’re likely too stupid to present consulting findings.
Intelligence is Free
Conceptualizing the future of service-as-a-software requires us to flip our understanding of much of economic value creation. High intelligence used to be rare and expensive, meaning that it would only be applied to the most important and valuable problems.
For the sake of argument, let’s define “high intelligence” as IQ 130, or the smartest 2% of the population in rich countries. Major intellectual breakthroughs require more brainpower, but most advanced knowledge work in business can be done at that level, assuming decent education and experience.
Whoa, that’s an important assumption I just added: education and experience are incredibly challenging and time-intensive for humans to acquire. It takes a full decade beyond high school before someone becomes a skilled business professional. And if that person wants to change profession, poof, there goes at least 5 more years for reskilling and gaining appropriate new experience.
Thus, not only are we limited to 2% of the population for the more challenging knowledge work, but we’re further limited because most of these people specialize in something different than the problem at hand. All these restrictions on the supply of intelligent and skilled knowledge work translate into high prices in the form of salaries for regular staff or hourly rates for consultants.
Economics 101 says that the higher the price, the less will be bought. Demand drops until it hits the clearing price on the supply-demand curve. The underlying need is still there, but it doesn’t translate into actual demand in the economy to use high-powered staff to solve minor problems. You don’t pay a consultant $100K to solve a problem worth only $1K.
But if the price of expertise dropped by a factor of 1,000x? Now, you might pay $100 to gain $1K, so my hypothetical project will get done. That change will play out over the next decade as AI makes intelligence almost free.
It may seem odd for me to claim that “intelligence will be free” in a month that saw the introduction of OpenAI’s o1-Pro service costing $200 per month. An unprecedently high price for AI that’s better at reasoning than the standard models costing $20/month. Sounds a lot as if higher intelligence costs more.
However, first of all, $200/month is a trifle compared to many of the complicated problems that AI can now tackle. The o1-Pro service can often substitute for the work of professionals, costing at least $20,000 per month. So, even the higher price for fancy AI corresponds to a drop of 100x in the cost of solving many complex problems. (Again, lower price = more of that work will now be done.)
But more importantly, the cost of AI is not flat. For example, the cost of running GPT-4 level AI was $36 per million tokens when ChatGPT 4 launched in March 2023. As of December 2024, the price for GPT 4o-mini is $0.15 per million tokens. (The current 4o-mini model is about as capable as the original GPT-4 model.) This means that in 21 months, the price has dropped by a factor of 240x, corresponding to an annualized rate of change of 2,300%.
Not fully free, but almost.
It’s safe to expect that whatever improved AI we’ll get in 2025 will be expensive at first, because it will require 100x the compute of 2023-level AI. Similarly, the 2027-level AI that’s expected to equal a team of 200 Ph.D.s in a box will also cost more in the beginning — though less than even the measly salaries most Ph.D.s command. In any case, this doesn’t matter to our long-term thinking, because less than two years later, those prices will likely drop by at least 100x. (I.e., something that costs $5 in 2025 will cost 5 cents in 2027 and 0.05 cents in 2029. Close enough to free.)
Case Study: Consulting Advice on E-Commerce Gift-Giving UX Design
At the recent ADPList AI Summit, Mark Angelow from UX consultancy agency O3 mentioned that he often uses AI to create a short overview for clients of best practices for specific design issues, such as how to design a certain navigation element. Perplexity is optimal for this use case.
I would add to Angelow’s recommendation that you can specify the target audience when asking for a best-practice overview. For example, “write an overview of best practices for UX design of gift-giving features on an ecommerce site, targeting the overview at readers who are experienced UX designers.” Or, use the same prompt, but “… targeting the overview at readers who are not UX specialists but experts in online retailing. Emphasize the business impact of following these best practices.”
(I picked gift-giving in e-commerce because Baymard Institute recently published a nice article about this design problem. However, while this article is good, it can never be customized to the exact circumstances of your specific design problem and target audience, the way an AI-written overview can be.)
Here’s Perplexity’s overview of the best practices for designing a gift-giving user experience:
------------ Begin Perplexity Answer ------------
The optimization of gift-giving features in e-commerce requires careful attention to both functional and emotional aspects of the user experience. The implementation should acknowledge the dual-purpose nature of gift transactions, serving both the gift-giver and recipient effectively.
Key Design Principles
Early Visibility: Present gifting options during the product discovery phase, ideally on the product page itself, to prevent user abandonment due to uncertainty about gifting capabilities.
Streamlined Designation: Enable gift designation both in the cart and during checkout to reduce anxiety and prevent premature abandonment. The ability to mark items as gifts should be consistently available throughout the purchase journey.
Personalization Options: Custom gift messages are essential, with 53% of gift-giving users opting to include personalized messages during checkout. The message composition interface should include pre-written templates while allowing for customization.
Implementation Guidelines
Gift Card Experience: Create a dedicated path-to-purchase for gift cards with clear navigation and comprehensive options:
Separate landing page showcasing all gifting options
Multiple delivery methods (digital and physical)
Flexible value selection combining preset amounts with custom input
Minimalist recipient information collection
Friction Management: While the core experience should be efficient, introduce intentional friction points where they enhance thoughtfulness and decision quality. The interface must particularly accommodate time-sensitive purchases with intuitive navigation and clear delivery options.
Technical Considerations
Dynamically adjust shipping field labels for gift-marked orders
Provide detailed explanations of available gifting features
Enable multi-item gift designation in cart
Ensure mobile-friendly redemption processes for digital gifts
The success of gift-giving features hinges on balancing transactional efficiency with emotional resonance, recognizing that each gift purchase represents not just a transaction but a relationship-building opportunity.
------------ End Perplexity Answer ------------
I’m particularly impressed that Perplexity linked to that Baymard Institute article I linked above, using it as the (credited) source of the statistics that 53% of gift-giving users include a personal gift message during checkout.
Why not simply read that Baymard article in the first place? After all, Baymard Institute is the recognized world leader in e-commerce usability. For one, their article runs 2,891 words, which may be too much for your client. In contrast, Perplexity’s overview is less than 10% of this length, clocking in at 270 words.
Perplexity wrote this overview at a 16th-grade reading level, corresponding to a bachelor’s degree, but then I did ask it to target a readership of UX professionals (somewhat like you, dear reader).
In contrast, Baymard wrote at a 13th-grade reading level, corresponding to targeting college freshman students. As actual consultants who live from selling insights, Baymard prioritizes usability, also of its advice.
Speculation: UX Services as AI Software
Product design is a small profession, with about 3 million UX professionals worldwide. In contrast, the world has an estimated 13 million physicians, 29 million software developers, and 94 million teachers, so disrupting those professions will benefit the world economy — and people’s health — by much more than any improvements we can make to the product design process. However, UX is my domain of expertise, so I’ll use it as an example to speculate on how the service-as-a-software transformation may play out over the next 10 years.
We already saw how current AI can provide decent customized design advice, written to target your exact audience (the design team, stakeholders, or anybody else) with just-in-time information about your current design question. Similarly, leveraging the metaphors of working with AI as a teacher or a coach will allow us to use AI to upskill staff with those same empowering qualities of being just-in-time and customized to your exact circumstances. As these abilities improve, they will disrupt the traditional lucrative services of selling UX training courses on a day of the instructor’s choosing and which teach generic content. AI will cause a similar disruption of the traditional promotional service of publishing generic articles on the day of the author’s choosing (admittedly, what I’m doing right now! I’m writing myself out of a job, so it’s lucky that I’m publishing these articles for free anyway).
Just-in-time personalized advice and training are more valuable and cheaper than similar content offered as a service. And personal coaching on your exact work problem as you’re working on it is even more valuable, and will still be cheap when provided by an AI coach. Higher value and lower cost equals vastly more consumed. Just this one change from services to software will hugely strengthen the world’s product design teams and allow them to accelerate skill growth for new staff. This is necessary to support the increase from 3 M UX pros today to 10 M in 2034 I predict. We can’t grow these many new skilled professionals, this quickly, the old-school way.
As an aside, why do I predict more than 3x growth in the product design field in just a decade? Because our work products are set to become dramatically cheaper, with at least a 10x drop in price per unit of design or unit of research insight, as UX teams accelerate at a much faster pace than the productivity growth possible for individual staff. UX work is fast becoming a commodity, which is great for the world, because we need immensely more design and usability, which will only happen if high-quality product design is cheap.
Currently, all the steps in the UX design process are overly expensive, with work processes reminiscent of a Saville Row bespoke tailoring shop — handcrafted, one stitch at a time. Moving many of these services — fully or partly — to being done by AI will accelerate the design process and lower the cost of each step.
Cheaper = more bought. This bears repeating: that’s why the AI takeover won’t cause unemployment among product design professionals. The world’s need for better design is incredibly far from being met, and this need will grow many times with the explosion in software development as our engineering colleagues accelerate their building and shipping pace even faster than the design profession can keep up. Software is eating the world, but it hasn’t even finished having breakfast yet.
(When I mention the enormous growth in software development over the next decade, I sometimes get the pushback that “we don’t need 50 more competing applications that do the same as Photoshop.” However, this is a naïve perspective on what developers will actually be producing. Most new software will be in-house and/or narrowly vertical, not consumer apps. Even so, many new categories of consumer apps will emerge as AI enables new things to be done cheaply. For example, almost any application for grownups might get parallel versions for children with simpler and more fun features. And even established application categories will benefit from competition, even if not necessarily 50 more variants. For example, I recently switched from Adobe Premiere to CapCut for my video editing, enticed by its better AI capabilities. But most important, the vast majority of new development will be spent on quality improvements, not on parallel products that do the same. We can design many new interaction styles with better usability, but they must be implemented. The more advanced an interaction, the more lines of code it typically requires. Similarly, advances in Generative UI will lead to interactions that morph to each user’s specific needs but also need to be implemented.)
Much user testing and other user research will probably be conducted by AI. (Note that AI can only perform the role of test facilitator. The test users still need to be actual human customers.) We’re already close to the time when this will happen, though AI intelligence must advance at least one more generation. But by 2027, if not 2025, AI-facilitated research will be possible.
Furthermore, much of the UI design will be produced by AI through Generative UI products. Coupling AI-facilitated user research with AI-generated user interfaces allows us to tighten the loop for iterative design to the dream of redesigning the product after each individual test user. (This ultimate version of discount usability is sometimes called “RITE” for Rapid Iterative Testing and Evaluation.)
Every time our AI has completed a test session, it’ll analyze the data in a few seconds and generate recommended redesign options that it hopes will improve the usability of the product. A few more seconds for the design AI to chew on these recommendations, and it’ll generate an updated user interface. As with all iterative design, it’s only our hope that the “improved” design actually works. Since hope is not a strategy, we’ll need to test again. But we may only need a minute between each test session. Iterative design at light speed!
When AI conducts user testing, analyzes the findings, and produces findings that are handed over to another AI to immediately generate a redesign, a revised product design can be available for another test session in less than a minute after finishing the previous session. Finally, iterative UX design can move at rocket speed. (Midjourney)
In fact, the model I just presented, where finishing a user session, analyzing it, and redesigning are separate discrete steps that happen sequentially, may be outdated with integrated AI that does both usability testing and redesign. Why not redesign the UI on the fly, during the test session with the participant who causes us to hypothesize that a different design would work better? Let’s make that UI change right now, so that the user doesn’t have to struggle for the rest of the session with something we know we want to change.
The faster we can run through the design-test-redesign-retest iterative design loop, the higher the user experience quality of the final product design. (Leonardo)
Learning Service Skills
How will advanced AI acquire the skills to perform professional services? Maybe all that’s needed is a next-generation foundation model. After all, these all-in-one AI models have been trained on all human knowledge, including that of any specific domain. For example, right now, general GPT o1 is better at diagnosing patients than human physicians. o1 has read all the medical textbooks and research papers that have ever been published, but it hasn’t been optimized specifically for medicine. Simply knowing so much theory is enough to outperform human doctors.
It's also possible that AI will need specialized training to perform better in targeted domains. I suspect that this will be true in product design, since most of a senior human professional’s skills come from accrued experience from practical design projects. Very little comes from training courses, books, and research papers.
If we need specialized AI and not just an advanced general foundation model, maybe training the specialized AI on in-depth domain knowledge will suffice. Or maybe reinforcement learning from large numbers of instances will be required. (Ideogram)
Is it enough to train AI on UX knowledge? Or do we need reinforcement learning where AI practices on thousands of cases and gets told whether it made the best design decisions or research analysis? I also suspect that we’ll need this additional step to get a high-performing AI to facilitate user research, derive good findings and redesign recommendations, and then actually design a better UI.
A company like UserTesting.com might invest a few million dollars in having experienced UX professionals from poor countries review thousand of hours of usability test footage and annotate them with the events that should trigger particular usability insights and redesign recommendations. We have the data and we know how to analyze it.
Reinforcement learning for design is harder, since there’s no full agreement as to what constitutes good design. Even so, we can probably still have several experienced product designers annotate each of thousands of design drafts with their good and bad design elements.
Even if all this training ends up costing ten or twenty million dollars, it would still be a good investment if it ends up accelerating the performance of the product design profession by 10x through better specialized AI.
In conclusion, service become software; intelligence becomes ubiquitous. The boundaries of human work dissolve as AI transforms labor into a scalable, automated force. Welcome to a world where expertise flows like water, reshaping the economics of knowledge forever.
(Watch the jazz music video I made based on this article. YouTube, 2 min.)