AI Systems of Action

Introduction: The Shift from Tool to Labor
In the era of SaaS (Software as a Service) 1.0, software primarily served as a tool, a system of record. It was a digital layer designed to help us track tasks, visualize data, and collaborate more efficiently. While valuable, it largely depended on human input and decision-making to drive workflows.
SaaS (Service as a Software) 2.0 marks a fundamental and transformative shift. With the advent of AI, software is no longer just a passive application; it's rapidly becoming an active laborer, a system of action. AI's ability to autonomously perform tasks, like filing reports, generating content, writing code, and much more, unlocks immense value in workflows that were previously too manual, fragmented, or low-margin to justify significant software investment.
This shift is opening up entirely new multi-billion dollar software categories in a wide variety of different verticals ranging from logistics, auto, manufacturing, etc. Because AI can genuinely absorb labor, rather than merely facilitate it, the businesses being built today will be significantly larger, spanning a wider variety of verticals and possessing far deeper revenue potential than those of the previous generation.
Note: These are just my current thoughts, a working hypothesis shaped by what I’ve been seeing, reading, and thinking about. I’m not claiming certainty, but I’ve come to believe that having a point of view, even if imperfect, is better than sitting back and being a skeptic. I believe in this case, but not all, waiting puts you on the defensive. It can make you slower to recognize the moment when something truly new shows up.
I believe when you have a thesis, a well thought out one, based on proximity to truth, and deep reasoning, not just projection, you are more prepared. You can spot patterns faster and you can lean in with conviction when something aligns. It doesn’t mean you’ll always be right, but it puts you in a position to act, not just react.
And in early-stage investing, that speed and readiness matters.
SaaS: The Interface Layer for Business
The first wave of Software-as-a-Service, emerged in the late 1990s and early 2000s, evolving from the preceding Application Service Provider (ASP) model to introduce multi-tenant, subscription-based web applications delivered online.
Companies like Salesforce, which pioneered the SaaS approach for CRM, and Concur, which pivoted from CD-ROM software to a cloud subscription model, were at the forefront of this movement. Thus leading to the size of the global SaaS market, as more people could start to create companies, because of these cloud providers, making it cheaper.
The core characteristics of SaaS 1.0 is centric user interfaces, comprising forms, tables, and reports, and dashboards that provided visibility and facilitated collaboration. This model replaced perpetual software licenses with a monthly or annual subscription fee. Primarily designed for office workers, managers, and other roles within the knowledge economy, these platforms allowed teams in finance, HR, and sales to share data and engage in basic collaborative workflows without the need for local installations or manual syncing.
While SaaS 1.0 was successful in simplifying processes and offering an affordable, scalable alternative to on-premise infrastructure, it possessed notable limitations. Decision-making remained sluggish because these systems could surface data but could not initiate actions autonomously, meaning all significant business logic and choices still depended on human oversight.
In simple terms: software helped human workers be more efficient.
- Salesforce didn’t sell your deals; it helped you track them.
- OpenTable didn’t seat customers; it gave hosts a better dashboard.
- Zendesk didn’t solve problems; it just routed tickets.
All of these are "tools-for-humans" models, the value was in enabling people to perform work more effectively. But humans were still at the center of the process. This worked for decades, but the costs, training, turnover, inconsistency, never left.
Consequently, tasks such as follow-ups, approvals, and cross-system coordination required manual effort, which limited both speed and scale. Furthermore, many of these early platforms were built on monolithic architectures that tightly coupled the user interface, logic, and data, thereby restricting flexibility and impeding integration.
Within the broader history of labor, SaaS 1.0 mirrored a trend of technology as a tool for augmentation rather than automation. It effectively digitized manual work and made information accessible, but it functioned as an "interface layer" that structured data for human consumption, leaving all critical thinking and complex tasks to be performed by people.
The Inflection: Service as a Software: AI That Works
Marc Andreessen’s 2011 essay, “Why Software Is Eating the World,” marked a turning point in how we understood technology’s role in the global economy, software wasn’t just a tool, it was becoming the foundation of entire industries.
Fast-forward to today, and AI agents are taking that thesis one step further: if software ate the world, AI agents are now digesting it.
Where software once empowered humans to work faster and more efficiently, AI agents are beginning to do the work themselves, planning, deciding, and executing tasks across systems without human input. The same forces Andreessen described, global internet access, scalable cloud infrastructure, and cheap computing, have set the stage for this next wave. But this time, it’s not just software transforming industries, its autonomous agents transforming labor and redefining productivity.
Modern AI (especially large language & vision models) does two things simultaneously to allow for this:
- Collapses unit costs of cognition – pattern-matching, summarising, classifying approach $0 marginal cost.
- Latent-demand unlock - the agent works 24/7
Take our portfolio company, Hostie, as an example. Hostie is a voice AI agent purpose-built for restaurants. It answers the phone, takes reservations, handles large parties and corporate buyouts, responds to customer requests, all with the polish and personality of a human host, 24/7.
The result?
- Fewer missed calls
- No more missed revenue
- No staffing shortages at the front desk
- Happier customers
- Better margins
This differs fundamentally from traditional restaurant SaaS, most software systems were tools: point-of-sale software, shift schedulers, reservation platforms. These tools supported human workflows, but they didn’t replace the work itself. A dashboard could show call volumes, but it couldn’t answer the phone. A reservation system could process online bookings, but someone still had to handle overflow or last-minute changes by phone.
Humans remained in the loop, monitoring, managing, and reacting. Hostie removes that requirement. It doesn’t support the workflow; it is the worker. The result is a restaurant operation that runs with fewer missed calls, fewer staffing constraints, and a more consistent guest experience, driven not by more labor adding cost, but by more leverage.
Take our other portfolio company, Maneva, which builds video-to-action AI agents for factory floors. Think of them as skilled digital operators who see, decide, and act in real time. Some of their customers use Maneva daily to inspect 100% of their product, without adding headcount (again collapsing the unit cost of cognition), without complicated integrations, and without high-maintenance hardware.
The result?
- Higher quality control (4 bars/second visual inspection vs. 1 in 100 manually)
- Fewer system breakdowns
- Freed-up technician time
- Lower camera maintenance costs
- Seamless real-time actions (counting, verifying, flagging defects)
Here’s the shift, instead of hiring teams of engineers to build brittle, one-off automation scripts, operators can now deploy general-purpose agents. A $20 webcam and a tripod is all it takes, and within days, the system calibrates itself, begins learning the task, and starts executing, like a line worker that improves with every shift.
Just like restaurants, for years, manufacturers have relied on tools, quality dashboards, machine monitoring systems, and control room alerts. These systems helped operators track what was happening, but they didn’t take action. A camera could record a defect, but not respond to it. A dashboard could show slowdowns, but someone still had to intervene.
Maneva’s agents become experts in a specific task over time, like a line worker who gets better every shift. And unlike brittle one-size-fits-all computer vision models, these agents adapt, improve, and generalize across use cases. The result is a factory that runs with fewer errors, fewer idle minutes, and fewer human bottlenecks.
More Examples of Unlocked Opportunity
Hostie and Maneva represent just two examples within our portfolio, but there’s a world of opportunity unlocked by AI agents’ ability to navigate complexity & unstructured data that lends itself well to applications like:
- Real Estate Appraisal: Automax AI
- Healthcare: Voice Agent like Abridge
- Customer Service: Sierra AI
- Legal: Harvey
- General purpose voice: Vapi, Retell
Markets Are Bigger Than You Think
When AI agents don’t just assist but act, the boundaries of what software can do expand dramatically. The next generation of software, powered by advanced AI agents, is moving beyond mere digitization to active, autonomous execution. This leap unlocks a vastly larger prize: the $11 trillion U.S. labor market. Suddenly, the Total Addressable Market (TAM) for software companies isn't just a company's software budget, but its payroll.

The classic equation for market size is simple:
TAM = (# of Paying Customers) x (Average Spend per Customer)
Historically, software's reach was limited by complexity and cost, capping both sides of this equation. Only certain companies could afford sophisticated software or labor to help them. AI agents change this calculus, by drastically lowering the "cost of cognition", the expense and effort required to perform “complex” tasks, making advanced capabilities accessible to both large enterprises and underserved segments like SMBs, and expanding the total addressable market for software:
- The Long-Tail: Small and medium-sized businesses, and even prosumers, can now afford AI-driven services that previously required expensive labor or specialized hires.
- Take Superpanel, one of our portfolio companies, they help contingency fee plaintiff law firms evaluate and sign 3–5x more legal cases without hiring additional staff. Traditionally, case intake required large teams to screen hundreds of low-quality submissions for the rare high-value case. Superpanel replaces much of that cognitive labor with an AI agent that autonomously collects and analyzes case data, surfacing the most promising leads instantly. By unlocking this capacity for smaller or resource-constrained firms, Superpanel doesn’t just improve productivity, it expands the market by making sophisticated intake possible for firms that previously couldn’t afford it.
- Enterprises: While large companies have always had access to sophisticated software, AI agents fundamentally change how they can deploy cognitive capabilities across their organizations. Rather than being constrained by headcount, budget cycles, or the scarcity of specialized talent, enterprises can now scale cognitive work instantly and precisely where needed, from customer service operations managing thousands of simultaneous interactions to internal functions like legal contract review and compliance monitoring. Most importantly, AI agents enable enterprises to experiment with cognitive capabilities that were previously prohibitively expensive. This dramatically expands the enterprise software market by making previously "nice-to-have" capabilities into feasible core functions. Tasks that companies knew would add value but couldn't justify the human capital investment, like analyzing every customer interaction for sentiment, reviewing every vendor contract for optimization opportunities, or providing personalized recommendations to every prospect, suddenly become economically viable, unlocking demand for cognitive capabilities that enterprises always wanted but could never afford to implement comprehensively.
We’re starting to see the ripple effect, enterprise clients are pushing back on $30M consulting projects, because AI is doing the same work faster, cheaper, and without the army of consultants.
When clients can plug in various AI to get to answers faster, thus leading to less labor from consultants they can now request for smaller scopes.
And if you’re building agentic software in enterprise, you’re not just competing with legacy SaaS budgets, you’re coming into a $11T labor market.
Emergence of new business models
In traditional SaaS, companies charged per seat or flat monthly fees. But when software starts doing real work, answering calls, inspecting products, completing tasks, that model breaks.
We’re seeing a shift to usage-based pricing as the new default, and it makes sense. When software performs a task, it consumes compute, bandwidth, and model cycle, much like paying a worker for time spent. So pricing based on minutes, messages, or tasks executed is both logical and fair. But what’s emerging now is even more aligned: usage-based pricing, enhanced by outcome-based bonuses.
- A core rate reflects the work done (e.g. time spent or volume processed)
- An additional incentive kicks in when the system delivers value (e.g. revenue-generating or cost-saving actions)
Usage-based core (to reflect AI effort and infrastructure) + Outcome-based bonus (to reflect business value delivered)
This hybrid model ensures:
- Customers pay only when the system performs.
- Vendors are rewarded for outcomes, not just activity.
- Trust builds quickly because the ROI is transparent and tied to real-world results.
This is the new pricing stack for AI-native services: metered execution with performance-linked incentives. And it’s going to dominate.
General vs. Domain Specific AI:
As AI continues to reshape software, two distinct models have emerged: general-purpose AI and domain-focused AI. Each reflects a different strategy for how intelligence is deployed, scaled, and monetized. On one side are tools built for breadth, fast-moving, horizontal products that serve anyone and everyone with universal utility. On the other are systems built for high specific domains, slow-burning, vertically integrated solutions that solve complex problems in niche, high-stakes environments.
General-Purpose AI: Fast, Wide, and Competitive
General-purpose AI tools are built for horizontal appeal. They target universal tasks that cut across nearly every profession. Think of Otter for meeting notes, Gamma for presentations, or Perplexity for research. Their power lies in their accessibility and immediate utility for a massive user base.

- Advantages: The primary advantage is speed, these tools benefit from frictionless, bottoms-up growth loops. An individual developer starts using a coding assistant like Cursor, loves it, and their whole team adopts it. This viral adoption leads to rapid user acquisition and incredibly fast feedback cycles, allowing for quick product iteration.
- Trade-offs: This speed comes at a price, the low barrier to entry and vast potential market make this a fiercely competitive space. The risk of commoditization is high, as countless competitors can emerge to offer a slightly better or cheaper version of the same function. It's a race for scale, where owning a user is much harder than owning an entire workflow.
Domain Focused AI: Slow, Deep, and Durable
Domain focus AI takes the opposite approach, instead of serving everyone, it aims to become the indispensable system for a specific, high-value industry or function. These tools tackle painful, complex, and often heavily regulated problems in fields like payroll, compliance, law, medicine, or manufacturing.
- Advantages: The core strength is its defensibility, by deeply integrating into a company's most critical workflows and data, it creates incredibly high switching costs. These tools aren't just assistants; they are agentic systems capable of judgment and execution in high-stakes environments, unlocking markets that traditional software could never touch. This creates a powerful moat, where defensibility is built through deep workflow and data integration.
- Trade-offs: The path to this deep integration is long and arduous. Go-to-market cycles are slow, requiring significant effort to navigate entrenched industries and build trust. In settings where a mistake can have serious consequences, like medical diagnostics or legal discovery, a startup must earn the trust of its users. Consequently, early traction for these Domain focus AI companies may look small and unimpressive compared to its general counterparts. However, once embedded, that growth compounds deeply, leading to a durable, long-term business
Intermezzo, another portfolio company, is a textbook example of this, it embeds itself deep in one of the most painful, complex, and regulated functions in the enterprise: global payroll. Unlike legacy systems that rely on brittle logic, manual compliance updates, and country-specific engines, Intermezzo is rebuilding payroll from first principles as a programmable, AI-native backend.
It's not a UI layer or a bolt-on tool, it’s the trusted infrastructure that HCM platforms and enterprises can plug into. Its defensibility comes not from fully brand, but from deeply embedded into HR platforms payroll functions where accuracy and compliance are non-negotiable.
The Stack Isn’t Ready (Yet)
As adoption of AI agents accelerates, so do the friction points. Beneath the surface of these promising tools lie foundational challenges, hallucinations, flawed reasoning, traceability gaps, and unresolved security concerns that builders must confront head-on. These aren’t reasons to pause; they’re the realities of building with a stack that’s still maturing.
The Reasoning Gap: Hallucinations and Flawed Logic
At the forefront of current AI limitations are the dual issues of hallucination and poor reasoning. AI agents, particularly large language models (LLMs), often generate false or misleading information with confident delivery, despite lacking true reasoning or contextual understanding.
This is especially problematic in high-stakes, enterprise settings where accuracy and traceability are essential. While solutions like Retrieval-Augmented Generation (RAG) are helping, they’re not sufficient alone.
An approach occurring is product-level traceability to uncover hallucinations and have the user part of the process, where you’re building systems that expose the agent’s reasoning step-by-step.
For example, our portfolio company, an AI SOX audit agent, has embedded full transparency into its platform by showing how its AI tests each control: users can view the exact evidence used, the logic applied, and even override decisions. This audit trail ensures accountability and builds trust in environments like compliance and financial reporting, where a single hallucinated result could undermine the entire system. Instead of hiding the AI’s process, their company makes it inspectable and correctable.
Lingering Security Concerns
Furthermore, the domains of security, compliance, and auditability for AI systems are still playing catch-up. As AI becomes more integrated into critical applications, ensuring the security of the models and the data they process is paramount. New vulnerabilities and attack vectors specific to AI are emerging, and robust frameworks for compliance with regulations and the ability to audit AI decision-making processes are still being developed.
Building the Future Mid-Flight
Despite these significant challenges, the demand for AI agent solutions is undeniable. While the underlying AI stack is not yet fully mature, the market is not waiting. Companies are actively developing and deploying AI-driven SaaS products, navigating the existing limitations while simultaneously pushing the boundaries of what is possible, because the pain is real:
- A freight manager juggling emails, spreadsheets, and tribal knowledge to route a single truck
- A law firm drowning in missed leads because no one’s manning the phone
- A CFO hiring two extra analysts just to consolidate accounting data across 12 systems
These aren’t edge cases, they are everyday problems. And the promise of “software that works like a person”, that actually does the job, not just tracks it, is too powerful to ignore.
Conclusion
The story of software is shifting from system of record to system of action, from passively tracking work to actively performing it.
Yes, the stack is still forming. Yes, the models will get better. But the gap between what’s possible and what’s deployable is exactly where great companies get built. Founders willing to live in the friction, understand the constraints, and still ship, those are the ones who’ll define this next wave.
If you’re building the agentic platforms of tomorrow, we’d love to hear from you. Email: chinwuba@behindgeniusventures.com