Why do applications for agents feel... underwhelming, and what can the adoption of electricity teach us about transformational change?
Getting full value out of LLMs and agents requires fundamentally redesigning how your organization handles information, decisions, and workflows, not just automating what you already do.
Part 9 of learning in public about multiagent systems, this time reflecting on why agents are underwhelming and what we can learn from history. This article is informed by 40+ interviews with experts, extensive reading, and my own opinions. (Disclaimer: My promise is that these write-ups are written by me, a real human, rather than just an LLM!)
You can find slides here and my LinkedIn here.
Executive Summary
Drawing on the analogy of electricity’s forty‑year journey from invention to system‑level impact, I frame agentic AI adoption in three phases from drop-in replacements for existing processes (point solutions), siloed augmentations of existing processes (application solutions), and the transformational frontier of system solutions.
From this history I’ve distilled four lessons that guide how I’m thinking about future winning companies: (1) obsolete constraints disappear, (2) new bottlenecks emerge, (3) winners redesign every assumption, and (4) mental models lag behind technology.
Despite all the excitement around AI agents and automation, most executives I speak with share a nagging feeling that their AI investments aren't delivering the transformational impact they expected. And as a consumer, I’m often underwhelmed. Sure, the chatbots answer customer queries faster, and the document processing saves some time, but where's the revolutionary change that was promised?
Fundamentally, I’m just doing the same processes (booking travel, gathering information from various sources, organizing my inbox, preparing for meetings) faster, but we haven’t upended anything yet.
Inspired by one of my favorite books that I highly recommend, Power and Prediction, I will join the chorus:
We are engaging with the most incremental opportunities of agents but have yet to understand their transformative insight.
This exact pattern played out over forty years when electricity was "the future," and understanding that historical precedent reveals exactly where we are with AI today and what's required to unlock its true potential. The story of electricity's adoption offers a precise roadmap for why current AI tools feel incrementally useful rather than transformational, and more importantly, how organizations can leapfrog to the system-level changes that create lasting competitive advantage. In short:
Getting full value out of LLMs and agents requires fundamentally redesigning how your organization handles information, decisions, and workflows, beyond just automating what you already do.
What I mean when I say “multiagent systems”
Before diving into history, let me clarify what I mean by multiagent systems Multiagent systems are agents (usually driven by LLMs) that coordinate directly with each other to accomplish complex tasks, share information, divide labor, and adapt strategies in real-time without constant human orchestration.
Think of the difference between having five smart assistants who each help with one task versus having five specialists who talk to each other, hand off work seamlessly, and figure out the best approach together.
What makes this transformative is that unlike traditional automation that follows preset rules, these agents negotiate with each other, divide complex problems into subtasks, share context and learnings, and adjust their approach based on outcomes. A customer service multiagent system might have one agent analyzing sentiment, another researching account history, a third generating solutions, and a fourth checking compliance. All work in parallel and sharing insights in milliseconds.
Companies like Klarna are already using multiagent systems to handle customer service inquiries that previously required multiple departments. Their agents go beyond answering questions to coordinate to resolve billing disputes, process refunds, update accounts, and schedule follow-ups, reducing resolution time from days to minutes.
The key insight is multiagent systems enable decentralized decision-making at machine speed, which fundamentally changes what kinds of organizational structures become possible.
The introduction of the lightbulb was less than illuminating
When Edison demonstrated the light bulb in 1879, it was promised that electricity would change everything. The technology was obviously superior to gas lamps and steam power in terms of lumens, cleanliness, and reliability. Yet it took until the 1920s for half of American households to have electricity.
That was forty years between proof of concept and widespread adoption.The technology was ready, but the systems, infrastructure, and mental models for using electricity hadn't been invented yet.
Business historian Alfred Chandler documented how early adopters of electricity fell into predictable patterns. Most companies simply replaced their steam engines with electric generators, keeping everything else exactly the same. Same factory layouts, same workflows, same organizational structures. They got marginally better results through lower energy costs and less noise but saw nothing revolutionary.
Three specific mistakes shaped this forty-year delay:
They kept factory layouts designed for steam's limitations. Multi-story buildings with machinery clustered around central power shafts made sense for steam. With electricity, this layout was unnecessary, but they were fundamentally limited by a constraint that no longer applied.
They ran electric motors like steam engines. When steam was up, everything ran. Factory managers used electricity the same way (all on or all off). This completely missed electricity's ability to power individual machines on demand, which was one of its primary benefits.
They optimized for the wrong metrics. Companies measured success by fuel cost savings rather than production flexibility, missing electricity's true value proposition.
The breakthrough came when manufacturers realized electricity went far beyond replacing steam to eliminating power as a constraint entirely.
This shift in mental models, more than any technological advancement, unlocked electricity's transformative potential.
This is precisely where most organizations are with AI today. They're in what researchers Ajay Agrawal, Joshua Gans, and Avi Goldfarb call "the between times," which is that crucial period after witnessing a technology's power but before achieving its system-level impact.
When we think of transformative technologies as augmenting existing means of work, we miss the point. They enable entirely new ways of organizing work, making decisions, and creating value, but only if you're willing to rebuild systems around their unique strengths rather than forcing them into existing frameworks.
The three phases every transformative technology follows
The electricity adoption pattern reveals three distinct phases that every revolutionary technology goes through, and AI is no exception.
Phase 1: Point solutions (where most AI sits today)
Early factories that adopted electricity simply replaced their central steam engine with an electric motor. Very little changed at the meta-level, with factories using the same massive central power source, belt-driven machinery, and factory layout designed around steam's limitations. The benefits were real but modest and were reflected in lower fuel costs, reduced maintenance, and cleaner air in the factories.
Today's AI equivalent is ChatGPT for writing emails, automated data entry, sentiment analysis tools, basic chatbots, etc. These point solutions slot into existing workflows without changing anything fundamental about how work gets done. They're useful, but they don't transform your business model or competitive positioning.
Point solutions feel underwhelming because they are.
They're designed to work within current systems with minimal disruption, not to be revolutionary. That incrementalism is inherently their virtue, in that you can make minor changes and confirm their value.
This is the local vs global maximum problem. From your perspective, you’re taking all the steps towards the best opportunities you have. But the real challenge might be that as you take tiny steps, you’re missing out on the global maximum elsewhere.
Phase 2: Application solutions (the emerging opportunity)
The next wave of electricity adopters got more sophisticated. Instead of one central motor, they installed individual electric motors on groups of machines. This enabled modular operation, such that if one machine stopped, others could keep running. Better, but still constrained by the same factory layouts and workflows designed for steam power.
We have some concrete examples of these application solutions. Companies like Morgan Stanley have deployed AI agents that work together to analyze market data, generate research reports, and provide trading recommendations. While impressive, these systems still operate within traditional departmental boundaries. The agents are more sophisticated than chatbots, but they're not yet reorganizing how decisions flow through the organization.
These application solutions create measurable improvements in efficiency and accuracy. They're where most "AI-mature" organizations are heading in 2025, but they're still working within fundamentally unchanged organizational structures and procedural silos.
Phase 3: System solutions (the transformational frontier)
It literally took 40 years for electricity to become revolutionary. Forward-thinking manufacturers completely redesigned their factories around electricity's unique advantages. Instead of clustering heavy machinery around a central power source (then the steam shaft, as was required), they could spread operations across lightweight, single-story buildings. Instead of running all machines continuously, they could power equipment on-demand.
This enabled Henry Ford's assembly line, which was impossible with steam power because it required flexible positioning of workers and equipment in linear workflows rather than circular arrangements around central engines. It allowed factories to relocate from expensive city centers to suburban areas with cheaper land and larger indoor spaces.
It fundamentally changed the economics of manufacturing from optimizing around power constraints to optimizing around information flows and production logic.
Here's what system-level transformation might look like for AI. Imagine a financial services firm where multiagent systems fundamentally restructure how investment decisions happen. Instead of analysts → portfolio managers → risk committee → execution, you have specialized agents continuously evaluating opportunities, modeling risks, and executing trades within parameters, with humans setting strategy and handling exceptions. The organizational hierarchy flattens because information no longer needs to be compressed as it moves up levels.
Another example: A logistics company that moves from hub-and-spoke routing (designed when communication was expensive) to dynamic, real-time routing where thousands of agents negotiate optimal paths based on current conditions, package priorities, and capacity. The entire organizational structure shifts from regional hierarchies to fluid networks of coordination.
System-level transformation is where we're heading.
Organizations that rebuild their decision-making, information flows, and operational structures around AI's unique capability will decouple prediction and information synthesis from human judgment.
Instead of using AI to make existing processes faster, these organizations will need to redesign entire workflows around AI's ability to synthesize huge amounts of data, simulate outcomes, model scenarios, act immediately, and coordinate complex multi-step reasoning.
The Specific Lessons History Teaches
The electricity transformation offers four specific lessons that apply directly to AI adoption:
Lesson 1: The Constraint That Disappears
With steam power, physical proximity to the power source constrained everything. With electricity, power could go anywhere. The constraint simply vanished.
With human-mediated processes, information processing speed and capacity constrain everything. With multiagent systems, these constraints vanish such that huge amounts of data can be synthesized in fractions of time, and decisions become the new bottleneck. Organizations that recognize this first will rebuild around what becomes possible when information processing is no longer a bottleneck.
What this means practically:
Your monthly planning cycles exist because humans need time to gather and process information.
Your approval chains exist because information is fragmented across people.
Your departmental structure exists because humans can only hold so much context.
None of these are required when agents can process millions of data points continuously.
Lesson 2: The New Constraint That Emerges
When power transmission stopped being a constraint, new constraints emerged: production flow efficiency, worker expertise, and supply chain coordination became the new bottlenecks.
For agents, the new constraints are already visible:
Data quality becomes critical when bad data can trigger thousands of automated decisions.
Objective alignment becomes crucial when agents optimize ruthlessly for whatever metrics you give them.
Governance frameworks designed for human decision-makers break down when agents make millions of micro-decisions.
When information processing stops being a constraint, data quality and access become the new bottlenecks. The winners will be organizations that build infrastructure for these new constraints before they become crisis points.
Lesson 3: Winners Redesign Everything
The factories that thrived were the ones that recognized electricity enabled entirely new organizational possibilities and rebuilt their operations accordingly.
"Redesigning everything" goes beyond better chatbots or faster processing. It's asking:
If information were free and instant, why would we have departments?
If agents can coordinate perfectly, why do we need middle management?
If analysis happens continuously, why do we have quarterly reviews?
The organizations that win will be the ones that question every structural assumption. CrewAI, one of the leading agent orchestration platforms, is a fascinating example of designing a company from scratch with agents.
Lesson 4: Mental Models Change Slower Than Technology
The 40-year gap existed because mental models about manufacturing were stuck in the steam era. Factory owners couldn't imagine layouts that weren't constrained by central power distribution.
Today's mental models about organizational design are stuck in the era of information scarcity. We still think in terms of human-speed decision cycles, departmental information silos, and sequential approval processes. We organize work around human limitations even when agents don't share those limitations. Breaking these mental models is harder than implementing technology. We can't yet imagine organizations that aren't constrained by human information processing limitations.
What System Transformation Actually Looks Like
Let me make this concrete with insurance, an industry ripe for system-level transformation.
Traditional insurance operates on "economics of risk transfer," whereby customers pay premiums, insurers collect money upfront, pay claims later, and profit from actuarial predictions about aggregate risk. The core business model is betting on statistical models of future losses in an actuarial tango.
Here's how the three phases play out:
Point Solutions (happening now): AI chatbots that handle claims processing faster, computer vision that assesses damage from photos, and predictive models that improve underwriting accuracy. To be clear, these are useful improvements to existing processes, with measurable ROI, but they're incremental.
Application Solution (emerging): End-to-end automated underwriting that pulls data from multiple sources, assesses risk, prices policies, and issues coverage without human intervention… but still within the traditional insurance model.
System Solution (the transformation): Complete business model transformation around risk prevention. AI enables a fundamental shift to "economics of risk prevention," using continuous monitoring, predictive modeling, and proactive intervention to prevent losses rather than just paying for them after they occur.
The key insight is that we’re talking about AI enabling a fundamentally different business logic.
Practical Implementation
The electricity analogy gives us a roadmap. Here's how to use it:
1. Find Your Steam Engines & Legacy Constraints
Pick one core process and ask:
What information moves between people and why?
Which steps exist only to share context?
What would break if everyone had perfect information?
As a thought experiment: is there a sequential review that serves to only add information? Or do additional stages improve the outcome, such as by catching errors?
To be clear, as someone working in healthcare, some constraints still have value!
2. Look for a Bounded Application
Choose a bounded process where you can test multiagent coordination and where there are minimal consequences from error (aka where you can accept some risk):
Clear inputs/outputs
Good data availability
Forgiving failure modes
Hold your agents’ hands to start, with agents making recommendations and humans making decisions. Compare outcomes. Your humans might be less consistent than you think.
3. Identify the New Constraints That Emerge
Assume a perfect world in which your data flows freely. What are the new things that break, such as:
Data quality: Bad data + automation = scaled disasters
Objective alignment: Agents optimize literally. Be precise about goals
Governance: Who's accountable when agents collaborate on decisions?
4. Challenge Mental Models
A big challenge here is unlearning assumptions. In every planning session, ask:
What would a company with perfect information do here?
Which of our processes exist because humans have limits (e.g. reasonable expectations around working hours, work-life balance, etc)?
If we were starting today, would we organize this way?
So what’s next?
The electricity precedent offers both encouragement and urgency. The encouraging news is that you're not behind if your AI initiatives feel incremental. You're exactly where most organizations were in electricity's early decades, where they were getting real but limited value from point solutions.
But we’ve got more urgency here, which is probably why you’re reading this article (and thanks for reaching this far!). The organizations that question fundamental assumptions now, that pilot multiagent coordination today, that build new governance structures this year will operate at a different clock speed than their competitors.
The path forward requires shifting from asking "How can AI improve our current processes?" to:
"How would we organize if information was perfect, decisions instant, and coordination free?"
Our between times moment is now. Will you be the factory that installed better motors, or the one that reimagined what a factory could be?