Apply to Register Here

Founded in 2015, Activant is a global investment firm with $1.5 B AUM that partners with high-growth founders and teams that are transforming commerce.

Activant

We’re always looking to learn about new perspectives, new research, and new ideas.

Get in touch

Disclaimers & Terms of Use

Privacy Policy

New York

334 Brannan St, 
San Francisco, 
CA 94107 USA

San Francisco

Berlin

17 Jamieson St
Gardens
Cape Town, 8001
South Africa

Cape Town

17 Old Kings Highway South, 
Darien, 
CT 06820 USA

Darien

Harish is providing the Amazon Prime fulfillment experience to everyone else through a combination of AI and an asset-light network of warehouses across the United States.

Alex is automating workflows for the world’s largest enterprises by using process mining to build the connective tissue for businesses.

Mayan is delivering Autonomous IT to the largest companies in the world through systems of intelligence.

Carsten built one of the largest commerce platforms in the world, which powers enterprise merchants and retailers across the globe and now serves as the SAP commerce engine.

Soups is stopping financial crime in its tracks by building the first end-to-end platform for fraud, risk, and compliance.

Adam is enabling restaurant owners to take back their customer relationships by digitizing their entire online and offline operations in a single platform.

Activant is a research-led growth firm that invests in relentless founders

Read past issues

Monthly Newsletter

Browse Research Library

Research

Home

Since 2015, we have invested in category defining companies during their most critical phases of growth.

Our portfolio

Companies

Activant is a research-led global investment firm that partners with high-growth companies.

We partner with founders who have won their initial battles and are ready for the next challenge.

About

A team built to support and advise yours.

Latest Research

Founders who aspire to build something bigger than themselves chose to partner with us at their most critical point.

Team

Andrew grew up in Northern Ireland on either side of the Good Friday Agreement in a family that lived in the thick of it. He first visited the U.S. for a summer camp in Vermont, and later moved full-time to attend Harvard University, where he studied behavioral economics.

Andrew co-founded an ed-tech platform in 2014 and ran the company for two and a half years before joining Activant in its first fund. As an early team member, Andrew has been part of scaling Activant alongside its portfolio companies. He runs the firm's New York office and is a connector of ideas and people.

"For as long as I can remember, I’ve always wanted to build companies. Partially for the curiosity, partially for the competition, but mostly to push the boundaries of what's possible. I count my lucky stars every day to wake up and work alongside some of the best founders in the world. There is no greater privilege in my book."

Edtech startup focused on development of personalized critical thinking curriculums for students.

This isn't a prediction piece, and it's not a definitive playbook. It's a framework—one we've been developing, testing, and refining over the past 18 months in close conversation with founders, operators, and teams across our portfolio. The pace of change in AI is relentless, but some patterns are proving consistent. We've written this to share the lens that's helping us cut through the noise and spot what matters: where software is heading, what's now possible, and what it takes to build enduring companies in this emerging era. We hope it's useful—not as a map, but as a compass.

We're at the beginning of a once-in-a-generation shift in enterprise software. For decades, enterprise value accrued to Systems of Record—rigid databases wrapped in UIs, designed to store structured data and enforce workflows. These systems, born in the 1970s and scaled through the SaaS era, now underpin trillions in global economic infrastructure. But they weren't designed for the world we now live in, one defined by overwhelming data volume, constraints on human bandwidth, and rising expectations for software that doesn't just inform, but acts.

The future belongs to a new software paradigm: Systems of Intelligence. Systems of Intelligence (SOIs) represent a complete rethinking of how software creates value: they ingest unstructured data, learn from context, integrate across silos, and take autonomous action to drive real business outcomes. This transition will unlock new markets, collapse bloated service layers, and compress time-to-impact in ways legacy systems simply can't match.

This shift is not incremental. It is a replatforming. One that replaces dashboards with decisions, workflows with outcomes, and manual middleware with intelligent automation. And it is happening faster than most incumbents, or even most investors, realize.

The implications extend far beyond software itself. We're witnessing the early stages of what may be the largest economic expansion in enterprise technology since the introduction of the personal computer. As SOIs automate not just software workflows but entire service categories, they have the potential to expand software's share of the economy from its current 1-2% of GDP to 10% or more—while simultaneously growing the overall economic 

In what follows, we unpack this transformation in three parts: why legacy systems are breaking, what the SOI paradigm looks like, and how builders can design for this new reality—and win.

Systems of Record: What They Got Right, and Wrong

Systems of Record (SORs) earned their place in the enterprise stack by solving a first-order problem: digitize the paper trail. Starting in the 1970s and 1980s, software giants like Oracle and SAP introduced revolutionary systems that finally replaced endless filing cabinets and paper-based processes with digital alternatives. ERP, CRM, SIEM, ITSM, HCM, and FMS platforms emerged to capture structured information, enforce compliance, and scale standardized workflows across organizations of unprecedented size and complexity.

And it worked—for a time. SORs provided structure, compliance, and scale in a world where digital alternatives simply didn't exist. They became the backbone of modern business, handling everything from payroll to supply chain management to customer relationships. Today, SORs account for over $400 billion in annual software 

But those very strengths have become their limits. The foundational assumptions that made SORs successful in the 1980s and 1990s—that data would be structured, that workflows would be predictable, that humans would be the primary interface between systems—no longer hold true.

SORs Are Built for Manual Inputs, Not Context

They depend on manual data entry and rigid schemas designed around the limitations of human input and 1980s database technology. Every field, dropdown, and form was defined not by what would produce the richest context, but by what a human could realistically be asked to enter by hand. The richness of reality—calls, conversations, documents, decisions, and the thousand small interactions that actually drive business outcomes—remains invisible to these systems.

Even worse, the narrow band of structured data that SORs are built to capture is patchy at best. Sales teams under-update CRM systems; workflows depend on fields that are blank, outdated, or inconsistently used. Studies consistently show that critical fields in major SORs like Salesforce are incomplete 60-80% of the time. Enterprises today are flying blind, relying on brittle databases that reflect only a fraction of what is actually happening.

As a result, SORs are brittle. To make these systems work, entire layers of human middleware have grown up around them, complete with roles, certifications, and business models built on navigating their complexity. That ecosystem has become a powerful constituency with a self-preservation instinct. The harder the system is to use, the more people are paid to make it usable.

It's a fragile equilibrium propped up by inertia, not innovation. The entire consulting-industrial complex, from Accenture to Wipro, owes its scale to the complexity of these systems. When software requires armies of consultants to implement and maintain, something has gone 

Long contracts, massive switching costs, and years of custom logic keep SORs in place. This stickiness comes from lock-in, not loyalty. Ask any operator, "Do you like your CRM?" and most will pause. The software is tolerated, not loved. Companies like Veeva, the leading life sciences CRM, famously tout 95% gross dollar retention, but this is a testament to how difficult it is for customers to leave once they're embedded, rather than customer 

The switching costs are deliberately prohibitive. Migrating from one SOR to another typically requires 12-24 months, costs millions of dollars, and carries significant risk of data loss or business 

  Vendors have optimized for retention through friction, not through value creation.

As a result, with customers effectively locked in, legacy providers have raised prices, rather than delivering productivity improvements. SaaS pricing has decoupled from broader economic indicators, rising at 8-12% annually, roughly four times the rate of 

 When incumbents can raise prices faster than they ship features, the market is ripe for disruption.

Enterprise software users are running out of patience. Three structural shifts are exposing the brittleness of legacy systems and accelerating the timeline for replacement.

The ChatGPT moment wasn't just a technological breakthrough—it was a reset in user expectations. The consumer bar is now the enterprise bar. The disconnect is stark: employees use sophisticated AI tools in their personal lives, then return to work to fill out forms in systems that feel decades out of date.

The productivity costs are staggering. Employees lose 80 hours per year to underperforming IT and 180 hours per year to slow, legacy 

 Across 100 million US knowledge workers, that's approximately $1 trillion in lost productivity 

 AI-powered optimizations could drive savings of up to $240 billion from software and hardware asset management alone, according to Activant benchmarks.

No wonder boards everywhere are demanding AI strategies—and consultants are cashing in to help people implement them. But now they need results.

The economic model of SORs is breaking down. Why pay $2,000 per terabyte per day for Splunk when AWS offers comparable storage for cents? Why maintain armies of system integrators just to keep a Salesforce instance alive when data quality is poor and user adoption is low? Why license 1,000 seats when only 100 users are active—and only five are getting meaningful value?

Enterprise software spend has reached $8.7K per employee in 2024-25, up 20-30% in just one 

 For large enterprises, that's hundreds of millions in annual costs. CFOs and boards are asking harder questions about ROI, especially when compared to the immediate productivity gains visible from AI-native tools.

Legacy systems once benefited from having all the data. But AI-native tools are catching up fast—both by living alongside existing SORs and by generating and inferring context in real time from email, Slack, video calls, PDFs, and logs.

The dirty secret about enterprise data is becoming impossible to ignore: what once looked like a moat is increasingly a liability. The numbers tell the story: nearly 50% of data workflows are still manual, 87% of data science projects never reach production, and only half of available data is being used to extract 

 Meanwhile, 33% of business users lack confidence in the quality of data provided to them—hardly the foundation needed for AI 

As data generation accelerates, any system depending on manual input is fighting a losing battle. Data creation is expected to grow 80% over the next three years, to become nearly 100x the amount generated in 

 The winners will be systems that can ingest this flood of data in real time, while legacy moats drown in their own complexity. More on this below.

SOIs go beyond just replacing software, and are actually encroaching on the much larger services layer that sits on top of it.

 Services represent 10%+. The $541 billion IT outsourcing market, the armies of system administrators, the endless integration projects—all exist because legacy SORs are too brittle to operate without human 

SOIs flip this model. These systems resolve issues at the source, eliminating the need for manual tickets and workarounds. They surface the right connections across tools without relying on consultants to wire them together. And they begin delivering measurable value immediately—long before a traditional rollout would even be complete.

The economic case is clear: replace expensive service providers while improving efficiency. And unlike traditional rip-and-replace projects, SOIs can start by living alongside legacy systems, capturing services spend while gradually taking over workflows. This isn't just about better software. It's about AI expanding software's footprint from a niche slice of the economy into critical, GDP-scale infrastructure.

The SOI stack represents a fundamental reimagining of how enterprise software should work, built around three foundational layers that work together to create truly intelligent systems.

Data is simultaneously the #1 moat for SORs and the #1 blocker for AI in the enterprise. This paradox defines the current moment: legacy systems hoard data behind rigid schemas and proprietary formats, while AI systems require rich, accessible, real-time context to function effectively.

In a recent study, 96% of analytics leaders reported being held back by data management challenges, including storage, access, quality, state, and flow of the 

 33% of business users are not confident in the quality of data provided to them, and 44% don't have visibility into their data pipelines. But this challenge is also the 

 The best SOIs excel in the data layer through two key capabilities that traditional SORs cannot match.

A. Platforming: Stitching Disparate Systems

The modern enterprise uses thousands of tools—up 2.5x in seven years. Large organizations manage over 2,000 different software systems, each with its own data model and integration requirements. The average security team resolving alerts requires gathering context across 50 different security 

SOIs excel at reducing these silos. Consider shipping a laptop to a new employee—a workflow that spans HR systems, IT asset management, procurement, facilities, identity management, and financial systems. Traditional SORs force manual handoffs at each step. SOIs unify this into a single, intelligent process.

, an Activant portfolio company, demonstrates this in financial services, unifying the previously siloed fields of fraud, risk, and compliance. Their platform combines payment flows, device intelligence, behavioral signals, and third-party risk indicators into a dynamic context graph that enables real-time, autonomous decision-making. Sardine creates a unified view of what previously would have been managed by separate teams, managing different tools, to allow for faster, more accurate decisions.

Traditional SORs were designed around structured data entry—forms and fields that humans could populate. But the most valuable information exists in unstructured formats: emails, documents, video calls, chat messages, and informal interactions that actually drive business outcomes.

The scale is massive. Internal company data is growing 63% monthly from 400+ sources, yet only half is being used to extract 

 90% of internal data is unstructured, limiting knowledge sharing across 

 SOIs unlock this previously invisible data to build a far richer picture of operations than traditional SORs ever could. And this is even before we layer on the mountains of new unstructured data (and context) being created in AI-native tools from note-takers like Granola, to outbound calling platforms like 

Voice is emerging as both a critical data source and interface. In many industries, especially SMBs, the phone remains the dominant channel for customer interactions and internal handoffs. Recent advances in transcription and speech synthesis have made 24/7 voice agents economically viable—transforming inaccessible streams into rich automation opportunities.

We're seeing this across verticals: Stuut uses voice AI to automate accounts receivables, a highly manual process today. Studies show voice AI prevents the loss of the 83% of customers who will move to other restaurants after hitting voicemail more than 

 Healthcare systems deploy voice agents for appointment scheduling. Legal firms generate case notes from client consultations. Tools like Vapi and Bland have made this infrastructure accessible even for nontechnical teams.

SOIs are extending beyond the "office world" into physical environments with video and IoT data. Over 1.5 billion video surveillance cameras generate 5,500 petabytes daily—a largely untapped data 

 Vision-language models now enable SOIs to process video as structured, queryable input for real-time insights into safety, equipment failures, and operational inefficiencies.

Companies like Teton embed computer vision in healthcare patient rooms, detecting falls and monitoring vital signs automatically. In manufacturing, cameras confirm physical arrivals while sensors verify package contents, triggering actions only when real-world state is verified. This isn't just digitizing workflows, it's making environments themselves observable and responsive. And while this space remains early, some of the best founders and engineers are developing critical infrastructure to bridge the gap between software and physical assets. A great example is Realworld, which is building a perception and reasoning engine for physical systems. Think DataDog for the real world.

2. Context & Control: The Intelligence Layer

Intelligence requires context, and unleashing intelligence requires controls. While legacy systems were designed around departmental silos, SOIs bridge organizational boundaries and reason across previously disconnected domains.

This capability is critical as organizational knowledge walks out the door daily. 14.8 million workers are expected to retire between 2024 and 2029, cutting 10% of the workforce in 18 major 

 Even worse, 57% of Baby Boomers say they have shared less than half the knowledge needed to do their jobs, and 21% admit they have shared none of 

Traditional knowledge transfer methods—documentation, training programs, mentorship—simply aren't keeping pace with knowledge loss. SOIs fight back by capturing, connecting, and preserving institutional knowledge automatically.

A. Knowledge Graph: The Organizational Brain

A strong DataOps foundation can be leveraged to create something far more powerful than a traditional database: a dynamic knowledge graph that serves as the organizational brain: a living model of how the organization actually works, learns, and evolves.

Modern SOIs build rich, contextual maps that connect people, processes, data, and outcomes in ways that reflect true organizational complexity. 

, an Activant portfolio company, exemplifies this in IT infrastructure management. When a server experiences issues, the system understands which applications depend on it, which business processes might be affected, which teams should be notified, and what remediation steps are most likely to succeed based on historical context.

This knowledge graph becomes more valuable over time. Every interaction and outcome feeds back into the system's understanding, creating a compounding advantage that traditional SORs cannot match. The more an SOI is used, the more powerful—and intelligent—it becomes.

Most importantly, SOIs can reason across departmental boundaries. A customer complaint might reveal a product issue, supply chain problem, or training gap. A security incident might indicate fraud patterns or compliance risks. SOIs identify cross-divisional patterns that no single system could detect and take proactive action before problems escalate.

B. Controls: Security and Governance at Scale

Security is the #2 blocker to AI adoption, demonstrated by 84% of companies that see cybersecurity as the primary roadblock to implementing AI 

 This concern is well-founded: 64% of companies report being "under pressure" to adopt Generative AI, creating a tension between competitive necessity and security 

The stakes are higher than ever. New data is growing rapidly—estimates suggest that the number of sensitive data files organizations must manage will increase approximately 5x by 

 Even well-architected security policies can become inundated with this growing stock of data to manage over time, yet 66% of staff believe that their data has already outpaced their ability to secure 

Organizations understandably fear the leak of sensitive data into AI models, which are environments into which they have minimal visibility and which they can't control. That's why they need robust access controls and security built into their platforms from the ground up.

Traditional permissioning systems were designed for humans accessing discrete applications. SOIs operate differently—they access data across multiple systems, reason about complex relationships, and take autonomous actions on behalf of users. This requires reimagining identity and access management. Consider that 84% of organizations have experienced identity-related security breaches, with an average cost of $4.9 million, while 99% of cloud identities are "excessively privileged" with only 2% of granted permissions actually 

SOIs require sophisticated controls that grant contextual permissions based on current needs, audit all actions with full traceability, implement dynamic access controls that adjust based on risk levels, provide kill switches for anomalies, and ensure compliance with data protection legislation in 137 countries where GDPR penalties can reach up to 4% of global 

 Financial services, healthcare, and government sectors require complete audit trails and demonstrated compliance with evolving regulations.

This auditability is essential not just for compliance, but for building trust. When AI systems make decisions affecting real business outcomes, organizations need to understand how and why those decisions were made.

3. Autonomous Systems: The Execution Layer

The ultimate goal of SOIs is not just to understand, but to act. This is where they differ the most fundamentally from traditional SORs, which are limited to generating reports, dashboards, and alerts. SOIs close the loop by taking autonomous action to drive real business outcomes.

This capability is desperately needed. Organizations are drowning in alerts and overwhelmed by operational decisions. Take security for example. A typical Security Operations Center faces 4,500 alerts daily, with Fortune 100 companies seeing 

 Most are repetitive, low-fidelity "noise" requiring manual clearing. Between 50% and 99% of these alerts are false positives, yet analysts believe up to 100% of routine alerts could be 

The human cost is unsustainable. The security industry employs 5.5 million people but declares a workforce gap of 4 million. Every security employee is doing the work of 1.7 

 Analysts get desensitized, demotivated, and miss true threats—up to 67% of alerts are completely 

SOIs solve this by automating routine responses and escalating only when human judgment is required. They go beyond detection and actually resolve issues—the ultimate outcome that organizations care about.

A. Systems Designed for Autonomous Outcomes

Building truly autonomous systems requires rethinking how systems should be designed when they don't depend on human intervention for routine decisions. Activant portfolio company 

 is an excellent example of how the foundational layers of the stack enable this in cybersecurity, automating threat detection, investigation, and response across some of the world's largest enterprises. The platform stitches together inputs from disparate systems, including SIEM logs, endpoint detection tools, and threat intelligence feeds, injecting intelligence only at points where it can operate reliably and avoiding compounding effects where minor hallucinations snowball into major issues. The system understands when it has sufficient context to act autonomously and when it needs human judgment. And then it takes proactive action—blocking malicious IPs, isolating compromised devices, and escalating critical incidents to human analysts. Critically, Tines closes the loop. It executes remediation autonomously when appropriate, documents actions and reasoning, and learns from outcomes to improve future responses, creating a virtuous cycle that reduces burden on human analysts.

Building systems that operate autonomously requires balancing AI flexibility with enterprise predictability. The best SOIs achieve this by being deterministic where it matters and generative where it adds value. They use AI to interpret complex inputs and reason about responses, but execute through reliable, well-tested workflows that produce consistent outcomes.

SOIs represent a fundamental shift from deterministic to non-deterministic software, from static dashboards to real-time intelligence, from rigid UIs to headless platforms capable of multi-modal interaction. The experience moves from human-in-the-loop to human-on-the-loop, from human middleware to autonomous outcomes, from non-intelligent software to AI agent colleagues with memory of our styles and 

SOIs co-exist alongside existing SORs and generalist AI platforms, connecting multiple SORs across domains like ServiceNow, Workday, SAP, and Splunk into unified knowledge graphs. They run autonomously in the background or surface in front-ends, scaling headlessly across multiple form factors. 

 exemplifies this: 92% of their platform operates autonomously. The remaining 8% surfaces only what truly requires human expertise and judgment, keeping humans on-the-loop rather than in-the-loop.

The illustration above is purely to highlight that SOIs will exist at various depths of the stack, and as headless platforms, will surface on the front-end to the user in a variety of ways.

Model Context Protocol (MCP) connectivity has accelerated the gravitational pull away from traditional SOR UIs by opening systems to model providers. You can now connect HubSpot to OpenAI and query knowledge directly—why ever log into your CRM if you can query in Claude? This creates an opening for best-in-class SOIs to own customer engagement. While OpenAI or Anthropic could compete, the best SOIs are positioned to drive reliable outcomes through specialized, secure design for enterprise context. SOIs will be critical infrastructure whether owning full-stack UI/UX, accessed via Claude, or embedded in vertical SORs.

It is worth flagging that it won't always be smooth sailing as some SORs will resist data access—as we're seeing with Salesforce and 

 It mirrors fintech: banks fought Plaid despite customers owning their data, just as businesses own their data regardless of which SOR houses it. The stickiness of bank accounts (customers rarely switched despite low net promoter scores) parallels SOR lock-in today. Competition ultimately improves products for everyone—many SORs will emerge stronger.

While the conceptual advantages of SOIs are compelling, actually executing on this vision is extraordinarily difficult. The gap between concept and successful implementation is littered with well-funded startups that understood the opportunity but couldn't execute effectively.

First-Contact Magic: Overcoming Buyer Inertia

Software inertia is one of the most powerful forces in enterprise technology. Even when the technical case for replacement is overwhelming, organizational inertia can delay adoption for years. Switching systems means retraining teams, rewriting workflows, migrating data, and taking personal risk. For organizations to endure that disruption, the new system must deliver a step-change improvement that feels instantly superior—something that feels like magic.

The most successful SOI companies ensure that this magic is experienced by the customer during their first experience using the platform—the "first-contact." This typically manifests as pre-populated intelligence (Activant portfolio company 

 combines TMS with AI workflows that automatically pre-fill data from emails and calls), instant problem resolution (Lightyear reduces months-long telecom projects to days through AI coordination), or invisible automation (Sardine makes thousands of fraud decisions per second without human intervention) that are highly superficial in the UI.

The information contained herein is provided for informational purposes only and should not be construed as investment advice. The opinions, views, forecasts, performance, estimates, etc. expressed herein are subject to change without notice. Certain statements contained herein reflect the subjective views and opinions of Activant. Past performance is not indicative of future results. No representation is made that any investment will or is likely to achieve its objectives. All investments involve risk and may result in loss. This newsletter does not constitute an offer to sell or a solicitation of an offer to buy any security. Activant does not provide tax or legal advice and you are encouraged to seek the advice of a tax or legal professional regarding your individual circumstances.

This content may not under any circumstances be relied upon when making a decision to invest in any fund or investment, including those managed by Activant. Certain information contained in here has been obtained from third-party sources, including from portfolio companies of funds managed by Activant. While taken from sources believed to be reliable, Activant has not independently verified such information and makes no representations about the current or enduring accuracy of the information or its appropriateness for a given situation.

Activant does not solicit or make its services available to the public. The content provided herein may include information regarding past and/or present portfolio companies or investments managed by Activant, its affiliates and/or personnel. References to specific companies are for illustrative purposes only and do not necessarily reflect Activant investments. It should not be assumed that investments made in the future will have similar characteristics. Please see “full list of investments” at https://activantcapital.com/companies/ for a full list of investments. Any portfolio companies discussed herein should not be assumed to have been profitable. Certain information herein constitutes “forward-looking statements.” All forward-looking statements represent only the intent and belief of Activant as of the date such statements were made. None of Activant or any of its affiliates (i) assumes any responsibility for the accuracy and completeness of any forward-looking statements or (ii) undertakes any obligation to disseminate any updates or revisions to any forward-looking statement contained herein to reflect any change in their expectation with regard thereto or any change in events, conditions or circumstances on which any such statement is based. Due to various risks and uncertainties, actual events or results may differ materially from those reflected or contemplated in such forward-looking statements.

The Replatforming of Enterprise Software

Systems of Intelligence

We are hardly the first to claim a new technology is poised to upend work as we know it. With every wave of innovation comes a chorus of such pronouncements, many of which turn out to be hyperbole. Technology may make our jobs easier, but most of the time, the core of what we do remains fundamentally the same. But spend five minutes exploring ChatGPT, or any comparable Large Language Model (LLM), and it’s clear that this wave is different.

The reason is simple. Most technologies are simply tools that help us accomplish tasks and be more productive. AI, on the other hand, can think – or at least provide the appearance of thinking. It’s like having another knowledge worker on the payroll (for a fraction of the cost).

As we explore the world of AI Agents and the potential to shape how we work, we will share three perspectives that we hold at Activant and illustrate what we think is possible through a hypothetical example.

Activant Perspective 1: LLM-enabled software will autonomously accomplish highly complex tasks with increasingly less guidance.

This is the beauty (and threat) of LLMs: they can do work we thought would long be the purview of humans. It’s unsurprising, therefore, that many people worry their job will soon be automated away by AI. In the short term, this will lead to dislocation. But technological revolutions are never zero-sum; in fact, they have always been accretive. As workers adapt, they become more productive, and the companies they work for can do more with less. Some stop there, but most businesses opt to invest the gains back into their infrastructure and people. In the end, revenues go up, companies do better, and the economy grows.

That said, we are still in the very early days of this cycle – and the early days of the technology itself.

Activant Perspective 2: LLMs alone are not enough to achieve Perspective 1.

An LLM can’t complete most tasks on its own. It is merely a dense set of attention networks: essentially, statistics on a massive scale. Yes, they can engage in human-like “thought,” but to generate real value from any of these models, a human would have to integrate them into actual workflows and make use of their decision-making competence.

For example, imagine that you hired workers to lay bricks. You could lead them to a stack of bricks and explain where and how you want them laid. They could perfectly understand your instructions, but understanding alone isn’t enough because the workers would also need to be able to lay bricks.

Bricklayers need to have more than just cognition. They need to have arms, legs, and physical strength. Furthermore, they need a memory to recall how you wanted each of the bricks to be arranged, and a way of correcting any errors so they don’t cause the eventual structure to fail.

If LLMs are to extend beyond mere Q&A they need (1) a suite of tools to impact the real world, (2) memory repositories to remember what actions they’ve taken, and (3) auto-critique algorithms to error correct along the way.

When all these components come together, an “Agent” is born.

 An AI Agent is orchestration software that combines an LLM with memory, tools, and auto-critique mechanisms.

 An AI Agent is a highly advanced bot that can complete work that has been assigned to it in natural language.

Though still in their infancy, Agents are poised to fundamentally reshape the way we work. They will become natural extensions of our organizations and complete day-to-day tasks like newly hired employees. They will help us get work done, and even collaborate with others on our behalf. Eventually, they will be able to work together, breaking down tasks into atoms of work able to be handled by a full “office” of Agents, some of which orchestrate, while others specialize.

That is why this wave is different. Working together, Agents will be able to handle increasingly complex tasks, replacing knowledge workers in a variety of fields. As they do, what it means to work – and to lead – will change dramatically, as both will increasingly depend on a mix of human and human-like capital.

Activant Perspective 3: Agent teaming is the future of autonomously completed, complex work – 

Agent ecosystems have four key components:

providing “conscious thought” within AI agents

external APIs facilitating task completion and agent interaction with the real world

 data storage mechanisms that retain and recall information

the ability to correct mistakes when completing tasks

 – agents can effectively “team” and autonomously complete complex work. The exhibit below describes the individual components of agent ecosystems, and we will attempt to bring things to life using a hypothetical example. Let’s consider the job of an Activant Investment Analyst!

Re-imagining the role of an Activant Investment Analyst

Among many other things, we often rely on our analysts to create market maps, which requires them to:

Find companies that exist in that industry,

Compare these companies quantitatively and qualitatively, and

Place them on a two-dimensional grid according to this comparison.

Until recently, this was almost inconceivable as a task for machines but let’s explore how we might put our “Agent Analyst” to use.

1. Understanding the industry (and task): Large Language Models

For our Agent Analyst to understand the task at hand, it needs to comprehend instructions in natural language. While LLMs provide that ability, requests literally come in many different shapes and sizes. To compare foundation models and model providers, one would likely consider the number of parameters in a model, whether it is open-source or closed-source, and the provider modality. One might also consider whether it includes “Fine-Tuning-as-a-Service” (FTaaS), the process of training a pre-trained model on subject-matter-specific data to produce a model that performs better in specific use cases.

When comparing foundation models, the number of parameters in a model is often a good approximation of output quality. As a rule of thumb, the more parameters a foundation model has, the better its performance across different benchmarks such as reasoning, web browsing, or even game 

 Although fine-tuning a model can boost its performance on a particular task, larger models with more parameters tend to have better generalization skills, making them more useful.

As seen in the exhibit on the following page, model sizes are increasing exponentially while model performance is also increasing, but not at the same rate.

At the heart of any LLM lies its architecture: a blueprint dictating how various attention heads and layers are organized. There are also "weights,” which are numeric values assigned to parameters in the neural network, refined and adjusted through training to enhance performance. Training involves feeding data into the model, allowing it to learn pattern matchings, and then adjusting its weights based on prediction accuracy measured via various benchmarks. It's through this process that the model obtains its ability to understand and generate context dependent and human-like text.

, openly share their architecture, weights, and sometimes even their training methodologies. This approach not only reveals the intricacies of their design but also democratizes access.

Conversely, some entities opt for a more guarded approach. Closed-source models, such as GPT-4 by OpenAI, protect the details of their architecture, weights, or training processes, presenting them as a kind of “black box.” Instead of allowing direct model access, they employ APIs as intermediaries, thereby controlling and often monetizing the interaction between the end user and the model.

Since this article’s initial publication, the AI landscape has evolved significantly, and we have updated our Activant Market Map of foundational models. Some changes are significant. For example, OpenAI was previously placed on the open-source side, but the company has since focused on closed-source offerings and has deprecated many of its former open-source models.

Provider modality relates to the data type a provider offers models for. This includes various transformations such as text-to-video, text-to-text, text-to-image, and image-to-text, among others. Companies that operate in just one domain are labelled as “uni-modal”; by contrast, “multi-modal” companies provide models spanning multiple domains. As AI Agents continue to evolve, we expect a growing emphasis on modality in foundation models (more than just LLMs).

Today, OpenAI, Google, and Anthropic dominate the industry, but the next few years will see considerable competition as investment in R&D increases in a proverbial arms race. Given this uncertainty – and the promise of better models always on the horizon – it’s hard to imagine agents relying on a single provider. From an investment perspective, backing one LLM provider is essentially a vote of confidence in its R&D team, but past performance is not necessarily indicative of future results.

Having understood the industry and the task, the Agent Analyst now needs the ability to search for companies. In other words, the Agent Analyst needs a tool.

Simply put, tools are snippets of code that LLMs can use to create, modify, and utilize external resources to execute tasks they couldn’t otherwise complete on their own.

Most tools are problem specific. In this case, all the Agent Analyst needs is a connection to Google or SerpAPI, a prominent wrapper for Google’s Search API. Unsurprisingly, other tasks will require different tools. Game this out, and you quickly arrive at the need for toolkits and tool aggregators. As tasks become more complex, agents will rely on not one but a variety of tools to accomplish a single objective. Through tool aggregators, Agents will be able to access the resources they need programmatically and as those needs arise.

Single tools refer to companies or services that specialize in just one function, such as searching the internet, booking a flight, or creating audio with a realistic voice. While these offerings do have certain strengths, integrating them into a toolbox, and thus into agents, is often cumbersome. Agent libraries provide the orchestration framework that combines all the components in the Agent ecosystem, and in some prominent examples toolboxes need to be set up before an Agent itself is even initialized. This means that if there's a need to integrate additional tools later, the agent would have to be initialized all over again. Given these considerations, embedding single tools into an Agent's workflow can be quite labor-intensive and time-consuming.

By contrast, tool aggregators offer Toolboxes-as-a-Service (TaaS). Theoretically, aggregators can compile an infinite number of tools and distribute them via a single, unified API. In practice, they typically provide tools that have been identified by an Agent as necessary to complete a given task. If, however, an Agent requires additional tools, it can access them without restarting.

A subset of tool aggregators, LCNC companies could provision tools and/or provide tool-building functionality for 

 As long as they offer an API, agents can leverage these solutions as they do tool aggregators, integrating (and even custom-building) tools into their workflows on the fly.

Many LCNC platforms have existed prior to the advent of LLMs and already have large libraries of powerful, ready-built tools to draw upon. Several companies, including 

, have begun offering Agents-as-a-Service (AaaS), using these libraries to jump ahead of competitors, many of whom are starting from scratch. Others may choose to offer APIs to facilitate quick adoption within the agent ecosystem and more reliably maintain their market share.

Because of their ease of integration and ease of use, we believe tool aggregators and API-based LCNC platforms are the best positioned to take advantage of this new agent ecosystem. As platforms like Llama Index and Relevance AI grow, Agents will be able to dynamically access a vast library of toolsets without needing to update their structure or code, increasing both their functionality and value.

As our Agent Analyst has progressed, it has gained the capability to discern tasks and search for companies. However, identifying a company means nothing if an Agent can’t remember it for later use. Agents require a dedicated memory component that can store company names and their associated descriptions.

Over the years, the software industry has seen a multitude of data-structure innovations, from traditional SQL-based relational tables to modern streaming services like Kafka. Yet, to truly grasp the intricacies of Agent development and the depth of virtual cognition, it's helpful to draw parallels with human memory.

At a high level, human memory is split into three core categories: sensory memory, short-term (or working) memory, and long-term memory.

In humans, sensory memory can be thought of as the direct inputs from our various senses of taste, smell, sight, touch, and hearing. These are perceived subconsciously, processed, and then filtered for relevance.

sensory memory captures the responses generated from the tools AI Agents utilize.

Office

Hometown

Residence

Andrew
Steele

Partner

Office

Hometown

Residence

Research

Systems of Intelligence

Towards the AI Agent Ecosystem (Updated)

Usage-Based Billing

Towards the AI Agent Ecosystem

The Next Chapter of Embedded Finance

Check this Out - Part 2

Check this Out - Part 1

The Dominoes are Starting to Fall

Relevant Experiences

Office

Hometown

Residence

Andrew Steele

Partner

Office

Hometown

Residence

Research

Systems of Intelligence

Towards the AI Agent Ecosystem (Updated)

Usage-Based Billing

Towards the AI Agent Ecosystem

The Next Chapter of Embedded Finance

Check this Out - Part 2

Check this Out - Part 1

The Dominoes are Starting to Fall

Relevant Experiences

Andrew
Steele