Is SaaS Really Dying? The Essence of AI Agents Targeting the $30 Trillion Knowledge Economy
Jonathan Siddharth, CEO of Turing, framed his discussion on a podcast around the premise that "digital knowledge work performed in front of a PC is worth approximately $30 trillion."
His point of contention is clear: the bottleneck hindering AI progress has shifted away from computing resources or the volume of data "eating the internet," and toward human intelligence (expertise) itself.
Below, I will organize why SaaS is being shaken and where companies should compete, using his statements as a foundation and including concrete examples.
1. Deconstructing the "$30 Trillion Knowledge Economy" in Four Dimensions
Siddharth views knowledge work as a "four-dimensional matrix."
Industry(finance, healthcare, manufacturing, etc.)
Function(development, sales, accounting, HR, etc.)
Position/Role(CFO, FP&A manager, accounting manager...)
Workflow(board meeting material preparation, monthly closing, auditing, quoting...)
The point is that he broke it down to the level of "company = set of functions," "function = set of people (roles)," and "person = set of workflows," defining the "minimum unit (cell) to be automated." This allows AI to be discussed not in abstract terms of job categories, but at the granularity of "reading this document on this screen, using this tool, making this judgment, and creating this output."
2. The Four Pillars: Multimodal x Reasoning x Tools x Code
His hypothesis is simple:
"If you master the four areas of multimodal, reasoning, tool usage, and coding, you can do almost any job on a PC.".
2-1. "Buying shoes" is actually a "microcosm of business workflows"
The example used in the program is symbolic. Even just "making an AI buy shoes" requires:
Memory (foot size, preferences, budget, shipping address)
Screen understanding (product images, reviews, seller reliability)
Planning (questioning -> searching -> comparing -> purchasing)
Tool operation (e-commerce, extensions, payment, shipping settings).
In other words, the "simpler the task," the more it requires all four pillars. Conversely, an AI that can break through this can also be horizontally deployed to corporate business workflows.
2-2. Why code is necessary
To automate decision-making, such as "weighting the recency of reviews" or "observing the trend of ratings," small analytical code snippets are effective. To borrow Siddharth's phrasing, for an AI to "take action," it needs not only reasoning but also the "ability to calculate and verify."
3. The Bottleneck for AGI Has Become 'Human Intelligence'
Previously, the battle was about 'compute resources and data.' However, now that the 'easy data' derived from the internet has been exhausted, models lack the data to learn the chains of reasoning necessary for practical work. Therefore, Turing says they are involving experts like investment bankers, doctors, and lawyers to create data by recreating environments close to reality—in other words, expanding a 'data factory.'
In fact, the company is known for providing data to frontier model development firms, and its valuation has been reported in funding rounds.
Also, the 'GDPval' introduced by OpenAI is an evaluation framework that attempts to measure the economic value of AI based on 'artifacts created in practical work.' This can be seen as an attempt to redefine what it means to 'be able to do work' by returning to the outside of benchmarks (the workplace).
4. 'SaaS is dying' — Two meanings
Siddharth's 'spicy take' divides the reasons why SaaS is wavering into two categories.
4-1. Custom software development costs are falling
The difficulty of developing AI applications is decreasing, making it easier for in-house engineers to create tools for their own companies. As a result, the 'reason to buy general-purpose SaaS' weakens.
What is important here is the fact that corporate operations are 'slightly different.' LLMs are strong at these 'subtle differences.' As a result, the strength of SaaS—'it's hard to build, so buy it'—is easily undermined.
4-2. UI-centric SaaS will lose in the agent era due to design
SaaS until now has been based on the premise that 'people follow menus and operate via GUI.' However, as tool-calling agents become widespread, the main battlefield will shift from UI to the API/data layer.
The future he envisions is one where even recruiting (ATS) runs by 'just talking to an agent,' while only the 'state of the candidate DB' is maintained in the background. Here, the value of the 'operation screen' that SaaS has provided decreases relatively.
5. Implications for companies: Defense is the 'data layer,' offense is the 'learning rate'
For SaaS companies, practical measures can be summarized in the following two points.
Control the System of Record (source of truth data):
Of 'UI + DB,' the part that retains value is the DB side. Companies with strong data integrity, permissions, auditing, and history will be able to remain at the core even in the agent era.Increase the Learning Rate:
Organizations that can quickly discover patterns where models 'break in the field' (e.g., PDF tables cannot be read across pages, output collapses when turned into a presentation, etc.) and run the loop of evaluation -> data conversion -> improvement are strong. This is consistent with his assessment that 'the reason P/L doesn't change with AI adoption is not the model's fault, but because last-mile design is difficult.'
Conclusion
'SaaS is dying' does not mean that software will become unnecessary. On the contrary, work will be broken down into 'workflow cells,' and as AI enters those spaces, software will proliferate.
However, the center will shift from 'screens' to 'data and agents.' The next competition over the $30 trillion knowledge work market starts here.