.png)
We spend a lot of time with B2B operators trying to scale AI across their organizations. And almost every one of those conversations lands in the same place: "Individual AI tools are delivering real value... but we're not seeing that value compound into enterprise transformation."
The gap keeps showing up in the data, too: 60% of enterprise AI pilots never make it to production.1 In this post, we’ll explore why AI adoption is stalling, and how software can meet people where and how they work.
A note upfront: We’re investors in Liminal, a company building solutions to this problem. But first, let’s look at the broader pattern we’re seeing across enterprise AI.
Scaling AI Beyond the Individual
The individual-level story of AI productivity is real. Employees feel more empowered and productive. Tasks get done faster. According to SaaS benchmarks from 800+ leaders across the industry, over two thirds of software startup leaders believe AI is measurably improving productivity, and over half of companies confirmed that AI spurred headcount reduction — in theory, you can do more work with fewer humans.
The belief in the technology is clearly there. But the next challenge is moving from individual productivity gains to enterprise-wide transformation.
That's where agentic AI comes in: autonomous systems that can orchestrate multi-step workflows, make decisions across organizational boundaries, and operate without constant oversight. The potential is substantial, with McKinsey projecting $2.6-$4.4 trillion in productivity gains from agentic AI.2
As more capital and headcount flows into these initiatives, leaders are starting to ask uncomfortable questions about whether all this activity is actually compounding into anything that moves the business forward.
This is the adoption paradox: individual AI tools work well, agentic AI promises substantial value, yet enterprise transformation feels elusive.
It's not a capability problem. Individual tools deliver exactly what they're designed to deliver. The issue is that scaling individual productivity to agentic enterprise transformation requires a different approach — one that current models weren't built to solve.
Why Agentic Deployments Stall
From an operator's perspective, agentic AI adoption rarely fails because people resist change. It fails because most agentic deployment approaches force a choice between scale and relevance.
We typically see agentic deployments fall into one of two paths:
The first is prediction-driven deployment. Leadership tries to decide upfront what should be automated, rolls out agents or workflows at scale, and hopes they map to how people actually work. Sometimes they do. More often, they don’t… especially once work gets cross-functional, messy, or context-dependent (which is most real work).
The second path flips the burden. Instead of predicting workflows centrally, organizations ask employees or IT teams to assemble automation themselves. In this version, savvy engineers, marketers, sales professionals and more use tools to automate how they work. Two marketers may have the same end goal, but their AI automations could look completely different from one another.
This solves for customization, but only if you have unlimited time, deep technical expertise, and a tolerance for constant maintenance. Most teams don’t.
In practice, both approaches hit the same wall: you either get scale without relevance, or relevance without scale. Adoption stalls somewhere in the middle, and AI becomes “helpful” rather than transformational.
Resolving the Paradox
Through our work with B2B operators and research into enterprise AI deployment patterns, we've learned this agentic AI adoption problem won't be solved by better prompts, more agents, or another layer of tooling. It requires a different architectural approach: one that inverts the current model entirely.
One emerging category that’s addressing this paradox: behavioral agent automation platforms (BAAPs).
BAAPs flip the deployment model. Instead of asking organizations to define workflows upfront, they observe how work actually happens — what people do repeatedly, where they struggle, what's high-value or high-friction — then assemble agentic capabilities based on that behavioral evidence.
Automation doesn't start with a design session or a rollout plan. It starts with observation.
Mechanically, that means capturing interaction signals through lightweight interactions and across the tools a team already uses, such as CRM, email, calendar, browsers, shared docs, and internal dashboards. The platform builds a behavioral model at the metadata level: which actions recur, where handoffs break, which decisions get made by whom, on what cadence, and where work crosses functional boundaries. Automation candidates surface from those patterns, grounded in real workflow data rather than a hypothesized process diagram built in a conference room.
Let’s look at a few examples of how a platform that can observe, learn, and automate at the individual level might be used across an organization.
Example 1: Sales Automation + Intelligence
One of your sales reps needs to prepare for discovery calls, and has a platform that’s observing their behavior across channels. Because the platform is connected to their calendar, CRM, Slack, email, and external sources like the web, it can see how that prep actually happens. Before every high-value call, the rep pulls account history, searches for internal context, researches attendees, and drafts an agenda — four or five times a week. It’s high-value work, but it’s repetitive.
The platform recognizes the pattern, identifies it as an automation opportunity, and surfaces it to the user. With a single approval, it assembles an agentic capability that automatically generates a complete 'deal intelligence dossier' two hours before each meeting, pulling together all the right context without the rep needing to ask.
Example 2: Finance Month-End Reporting Automation
You have a financial analyst that needs to run their month-end report. In this instance, you also have the same platform that’s able to observe behavior across all channels. The platform notices that your analyst runs month-end reporting the same way for the second month in a row: pulling data from three systems, normalizing formats, running a variance analysis on the most recent income statement, and building a deck.
Instead of requiring IT to predict this workflow and build it manually, the platform observed behavior, called on its episodic memory to recognize this as a pattern over multiple cycles, and then surfaces it as an automation candidate based on how this analyst works. A workflow is created, freeing up time for the analyst to refine and check worth rather than reformatting and manually pulling in necessary data.
From an operator's standpoint, that matters more than it might sound:
- BAAPs don’t assume everyone with the same title works the same way
- They don’t require massive change-management efforts to force standardization
- They don’t ask already-overloaded teams to become automation architects on the side
Instead, individualized automation emerges naturally from lived behavior. And because that observation happens across the enterprise, it can scale.
This is the real unlock. Productivity gains stop living at the edges and start compounding through the system without blowing up governance, security, or oversight. Individual workflows stay flexible while enterprise control stays intact.
That’s how the individualization vs. scale tradeoff finally breaks.
Building the Next Evolution
We think this architectural shift is foundational to where enterprise AI is headed.
The next phase isn’t about bolting more tools onto existing systems. And it’s not about shifting even more responsibility onto end users. It’s about platforms that can discover what should be automated, and then automate it automatically.
That’s the pattern we’re watching closely, and it’s why BAAPs are becoming central to how we think about enterprise AI adoption at scale.
Which brings us back to the company we flagged at the top: Liminal. We've backed them through multiple rounds because they're pioneering the BAAP approach we believe enterprise agentic AI needs. They've published a detailed foundational paper laying out how behavior-driven automation works in practice and why observation-first orchestration represents a meaningful architectural shift. It’s worth reading, especially if you're trying to understand why AI adoption feels easier to start than to scale.
The paradox isn't that AI doesn't work. Individual tools deliver real value. Agentic AI holds transformational potential. But bridging that gap requires systems designed to adapt, not just tools designed to assist.
1 IBM Institute for Business Value. 5 mindshifts to supercharge business growth. IBM Corporation, May 2025.
2 McKinsey & Company. The economic potential of generativeAI. McKinsey & Co, June 2023.
