Of Termites & Tokens

Maybe the future of work is… termites? Let me explain.

Right now a lot of companies are trying to use AI to automate low-throughput workflows - essentially replacing human tasks with AI workflows (or “agents” if you want to get all fancy). The story goes that AI can make it faster to build a landing page, or faster to respond to each customer support ticket. This is an efficiency story “let’s replace our expensive humans with cheap tokens!”

But I think this is boring and mundane. The more interesting question is what happens when AI changes the throughput of the workflow itself? Don’t make one landing page. Make 1,000. Don’t summarize one customer interview. Summarize every customer interaction continuously. Don’t run one competitor analysis. Monitor the market in real time.

This is where “workflow automation” feels too small and unimaginative. The opportunity is not to build faster, smaller machines - but to become a colony: a mixed population of humans, agents, bots, scripts, dashboards, workflows, alerts, memories and permissions, all sensing and acting through local signals.

In the colony company - where are the pheromone trails?

The Margin Trap

Fundamentally a lot of the AI transformation discussion is still stuck inside the old unit economics of work. The process remains the same shape, but one step gets cheaper. A human used to draft the thing, now AI drafts the thing. A human used to analyze the thing, now AI analyzes the thing.

But this is unimaginative - and actually not what a lot of companies need or want. It’s been possible to make workflows more efficient long before AI: outsourcing and automation have been around forever. But instead of making workflows more efficient, most of the time companies chose to invest in growth.

Making an existing process more efficient is a margin play: same output, cheaper. But a lot (most?) of businesses I’ve been involved with would far rather keep margin consistent and drive top line growth.

I think workflow automation is a margin trap: AI makes cost reduction so legible that we miss the weirder growth move. We ask how to do the existing workflow with fewer people, fewer steps and less time. The better question might be: what happens when this workflow is no longer scarce?

A landing page used to be a project. A customer research synthesis used to be a deck. A competitive scan used to be a quarterly exercise. But if these things become cheap enough, they stop being discrete projects and start becoming continuous organizational senses. The company doesn’t “do research.” It has research metabolism. The company doesn’t “make campaigns.” It has a campaign ecology.

AI doesn’t just reduce the cost of work. It changes the carrying capacity of the organization.

Termites and Tokens

Colonies like ants and termites and bees operate with a radically different coordination cost structure. They don’t plan and operate like a typical organization but rather operate through distributed local action. The intelligence is in the relationship between actors and environment.

Maybe this is a better image for agentic work - not AI replacing existing workflows but rather agents swarming across the organization expanding the throughput of every single workflow - every task becomes 1000 variations, judged, evaluated and tested through a mixture of computers and humans. But happening at a frequency and volume that changes the organization’s shape.

The question becomes less “which jobs can AI replace?” and more “what kinds of collective behavior become possible when small acts of cognition become cheap?”

James March’s exploration/exploitation frame is useful here.

A central concern of studies of adaptive processes is the relation between the exploration of new possibilities and the exploitation of old certainties. Exploration includes things captured by terms such as search, variation, risk taking, experimentation, play, flexibility, discovery, innovation. Exploitation includes such things as refinement, choice, production, efficiency, selection, implementation, execution.

Adaptive systems that engage in exploration to the exclusion of exploitation are likely to find that they suffer the costs of experimentation without gaining many of its benefits. They exhibit too many undeveloped new ideas and too little distinctive competence. Conversely, systems that engage in exploitation to the exclusion of exploration are likely to find themselves trapped in suboptimal stable equilibria.

As a result, maintaining an appropriate balance between exploration and exploitation is a primary factor in system survival and prosperity. 

There are tradeoffs between efficiency and growth. Replacing humans with tokens is mostly exploitation: refinement, efficiency and execution. Augmenting humans with tokens is more exploratory: new work, new markets, new experiments, new forms of organizational attention. Efficiency has a cleaner story than growth. Cost savings are immediate and CFO-friendly. Growth from new workflows is fuzzier. It requires imagination.

I’d argue too much of the discussion so far is on efficiency and exploitation - and not enough on growth. What new opportunities are enabled by AI?

Pheromones of the Organization

In an era of agent swarms, the useful concept is stigmergy: coordination through traces left in the environment. Ants have pheromone trails. Organizations have CRM fields, Slack reactions, issue queues, tags, alerts, dashboards, approval states, confidence scores, meeting notes and customer transcripts.

In a low-throughput organization, these signals are mostly coordination residue. In a high-throughput agentic organization, these signals become the operating system. An agent notices a churn pattern because a CRM field changed. Another agent drafts a follow-up because a support ticket crossed some threshold. A dashboard triggers a budget review. A sales transcript becomes product research. A Slack thread becomes a workflow.

This is where “make 1,000 landing pages” gets interesting. The point is not 1,000 landing pages as slop. The point is whether the organization has the pheromone system to learn from 1,000 landing pages. Which variants attract attention? Which segments respond? Which claims create trust? Which messages produce confusion? Which pages should be killed? Which ones should reproduce?

Smart folks are already paying attention to “company world models”. This is going to be a very interesting line of discovery and change how organizations operate - but/and maybe we can skip the step where we wire up Slack/Jira/Drive to build a company world model and just skip straight to the event stream?

If AI can see the flow of customer transactions - maybe the entire company strategy can be re-derived?

Throughput, Not Just Output

The Theory of Constraints gives us another useful frame of reference. Local efficiency is not the same thing as system throughput. It’s not valuable to optimize every step. The point is to understand the constraint that governs the whole system’s ability to create value. A useful overview of Theory of Constraints frames throughput as the rate at which the system generates money through sales.

So the question is not: can AI increase output? Of course it can. The question is: output of what, into which constraint, with what feedback loop?

Most AI workflow automation is cost-world thinking. It asks whether we can reduce operating expense in this step. The better question is throughput-world thinking: can we increase the rate at which the system creates value?

Company As Colony and New Management Problems

Gareth Morgan’s Images of Organization gives us the classic metaphors: organizations as machines, organisms, brains, cultures, political systems, psychic prisons, flux and instruments of domination. Maybe agentic organizations unlock a new metaphor: organization as colony.

This is where management starts to look less like process optimization and more like ecology. The job is not just to assign work, but to design the local rules, signals, constraints and feedback loops that let useful work emerge without letting the colony eat itself.

Organizations as colonies raises a bunch of new questions to consider:

• What is the org chart when everyone can do anyone’s job?
• When each indiviudal can produce more, how do you keep coordination costs in check?
• How do you keep a real-time map of strategy when the landscape is shifting constantly?
• How do you design for systems of work, not individual outputs?
• What skillsets are valuable in a world of agent swarms and colonies?
• When agents have to decide how to act, which constraints and values do they refer to?

Erik Brynjolfsson’s Turing Trap is useful here: we have strong incentives to build AI that imitates and substitutes for human labor, even when the more valuable path may be augmenting what humans can do.

As noted above, both automation and augmentation can increase productivity and wealth. However, an unfettered market is likely to create socially excessive incentives for innovations that automate human labor and provide too weak incentives for technology that augments humans.

The first fundamental welfare theorem of economics states that under a particular set of conditions, market prices lead to a pareto optimal outcome: that is, one where no one can be made better off without making someone else worse off. But we should not take too much comfort in that. The theorem does not hold when there are innovations that change the production possibilities set or externalities that affect people who are not part of the market.

The margin story of AI is: replace humans with tokens. The growth story is: use tokens to make new forms of human agency economically viable.

But once that happens, management changes. In a machine org, management assigns work. In a colony org, management designs the conditions under which useful work emerges. What can act? What can decide? What can spend money? What can contact a customer? What can rewrite a document? What needs approval? What leaves a trace? What can reproduce?

That last question matters because the danger is not one agent doing one thing badly. The danger is a pattern of agency replicating across the organization. A bad prompt copied into ten workflows. A broken metric used by five teams. A temporary automation becoming infrastructure. A hallucinated summary becoming a board slide. A vibe-coded script quietly becoming a system of record.

Colonies are powerful because local behavior scales. Colonies are dangerous for the same reason.

Maybe the first wave of AI adoption is about margins. The second wave is about throughput. The third wave is about new organizational forms.

–

Editor’s note: I just started a new job at Alephic so expect more AI and future of work takes!