-
IN PHOTO: Briones Industries President Joey Briones
Culture & Code
For the past few years, the AI
conversation has been driven by a seemingly straightforward assumption:
more
AI should mean more productivity.
If one AI agent helps, then five should
help more. If five specialized agents can perform different tasks, perhaps
twenty can transform an entire function. Add an AI researcher, analyst, writer,
scheduler, recruiter, finance agent, project manager, and an AI Chief of Staff
to coordinate them all, and the idea of a personal — or even enterprise-wide — digital
workforce suddenly starts to look plausible.
But there is another side to this story.
Imagine managing ten human employees.
You do not simply receive ten times the output. You also inherit ten sets of
priorities, communication needs, handoffs, exceptions, performance issues,
coordination demands, and accountability questions. The additional capability
is real, but so is the work required to organize it.
Why should an organization made of AI
agents be completely different?
As AI systems become more specialized,
persistent, autonomous, and interconnected, organizations may discover that the
cost of AI is not limited to compute, tokens, software subscriptions, or
infrastructure. Another cost begins to appear — less visible, but potentially
just as important:
the cost of coordinating all that intelligence.
I call it the
Orchestration Tax:
The Orchestration Tax is the additional coordination, context management, verification, exception handling, and governance effort required to turn multiple AI capabilities into reliable organizational outcomes.
AI can dramatically reduce the cost of
execution while simultaneously increasing the cost of coordination. And once
that happens, the real productivity equation begins to change.
Intelligence
Is Becoming Cheaper. Coordination Is Not
Generative AI is making certain kinds of
intelligence increasingly abundant. Research that once required hours can be
produced in minutes. First drafts appear almost instantly. Analytical work,
coding, monitoring, forecasting, summarization, and content creation can
increasingly be performed — or significantly accelerated — by machines.
That is an extraordinary development,
but abundance creates a new bottleneck. Someone — or something — still needs to
decide
who does what, what happens first, which agent receives which
context, where one agent's responsibility ends and another begins, which output
should be trusted, when work should escalate, and who ultimately owns the
result.
The more intelligence we deploy, the
more coordination decisions we potentially create around it. This produces a
paradox:
The cheaper intelligence becomes, the more valuable coordination may become.
The next phase of AI transformation may
therefore be less about acquiring additional intelligence and more about
learning how to
organize intelligence without drowning in the complexity
created by organizing it.
The
Real Productivity Equation
The conventional AI productivity story
is simple:
More AI → More Output → More Productivity
But that equation ignores the work
required to make multiple AI systems operate reliably together. A more
realistic formulation may be:
Gross AI Capability – Orchestration Tax = Net AI Leverage
The distinction matters because a
twenty-agent system may be technologically impressive while still being
organizationally inefficient. If it requires constant human monitoring,
extensive context preparation, repeated corrections, complicated handoffs, and
multiple layers of verification, the productivity gain may be far smaller than
the apparent volume of AI activity suggests.
The meaningful question is therefore not
“How much AI capability have we deployed?” It is
“How much useful
value are we getting after accounting for the effort required to coordinate,
verify, govern, and maintain it?”
That is what I would call
Net AI
Leverage — and it may ultimately be a far better measure of whether AI
transformation is actually working.
Where
the Orchestration Tax Comes From
The Orchestration Tax does not come from
one source. It accumulates across several different kinds of organizational
load.
Coordination Load arises because specialized agents must
divide work, sequence activities, transfer outputs, avoid duplication, and
resolve conflicts. A research agent may hand work to an analyst, who sends it
to a writer, who passes it to a verifier, who returns it to an orchestrator
before the final result reaches a human. Each handoff may improve
specialization, but each also creates another opportunity for context to be
lost, misunderstood, delayed, or duplicated.
Context Load emerges because every agent needs the
right information — but not necessarily all available information. Too little
context produces poor output; too much creates noise, cost, confusion, and
sometimes conflicting instructions. As agent systems grow, organizations must
decide what context should be shared, what should remain role-specific, what
should persist, and what should never be accessible at all.
Verification Load arises because AI systems still make
mistakes. One agent can check another, but eventually someone must determine
whether the final output is actually trustworthy. Adding verifier agents does
not eliminate verification; it merely changes where it occurs. This is why an
old principle remains relevant even in highly automated systems:
you can
outsource the thinking, but you cannot outsource the understanding. If
nobody governing the system understands the work well enough to recognize when
it is wrong, verification becomes ceremonial.
Exception Load becomes visible when reality stops
behaving normally. Unusual customer requests, missing data, conflicting
instructions, system failures, regulatory issues, or decisions outside an
agent's authority all require escalation or recovery. Automation works beautifully
in the normal flow; organizational complexity often reveals itself at the
edges.
Finally,
Governance Load grows as
agents gain access and authority. Which agent can send an email? Approve a
transaction? Change a customer record? Delegate to another agent? Access
sensitive information? Which actions require human approval, and who remains
accountable when something goes wrong?
Taken together, these coordination,
context, verification, exception, and governance demands form the hidden cost
of turning AI capability into reliable work.
The
Coordination Explosion
There is another reason this matters:
the complexity of a multi-agent system does not necessarily grow neatly with
the number of agents.
One agent is relatively easy to
understand. Five specialized agents may already require explicit handoffs, role
boundaries, and escalation rules. Fifty interacting agents can create an
entirely different management problem because every additional agent potentially
introduces new dependencies, relationships, permissions, messages, and failure
paths.
I call this the
Coordination
Explosion:
The point at which increasing the number and
interdependence of AI agents creates coordination complexity faster than it
creates useful capability.
At that point, the digital workforce
begins to
reproduce one of the oldest problems in human organizations:
bureaucracy.
Welcome
to the Agentic Bureaucracy
Human organizations accumulate
bureaucracy in familiar ways — unnecessary approvals, duplicated work,
excessive reporting layers, too many meetings, unclear ownership, and processes
designed largely to manage other processes.
Agentic organizations can develop
equivalent pathologies.
An agent calls another agent. That agent
asks a third. A verifier checks the output. An orchestrator consolidates it.
Another agent reviews the consolidated version. A human eventually approves the
result. What began as a straightforward task now passes through six layers
before anything useful happens.
This is
Agentic Bureaucracy:
When the system designed to automate work becomes so complex that managing the automation becomes work in itself.
There is an obvious irony here. We may
succeed in automating bureaucracy only to discover that
AI can create
bureaucracy of its own.
The technology changes. The
organizational problem does not.
The
Mega-Agent vs. Agent Swarm Paradox
This creates an important design
tension.
At one extreme is the
Mega-Agent —
one AI expected to know everything, use every tool, manage every workflow, and
solve every problem. As its scope expands, so do context overload, confusion,
and execution errors.
The obvious response is specialization:
separate agents for research, analysis, finance, scheduling, content,
operations, and other responsibilities.
But specialization creates its own
problem. The Mega-Agent suffers from
context overload; the Agent Swarm
can suffer from
coordination overload.
Neither extreme is ideal.
The real design challenge is to find the
point between them: enough specialization to improve performance, but not so
much that the cost of orchestrating the specialists overwhelms the value they
create.
This is where AI architecture begins to
look remarkably similar to
Organization Design.
The
Minimum Viable Agent Organization
In good Organization Design, we do not
create a new role, reporting layer, department, or committee simply because we
can. At least, we should not.
The same discipline should apply to AI.
Instead of asking
“How many agents
can we deploy?”, organizations should ask:
“What is the smallest number of agents required to deliver the outcome reliably?”
I call this the
Minimum Viable Agent
Organization:
The simplest Human × AI architecture capable of producing the required outcome at the required level of quality, speed, risk, and accountability.
The principle is deliberately
straightforward. If one agent can reliably perform the work, do not create
five. If three specialized agents materially outperform one, use three. If
another agent adds more coordination burden than capability, do not add it.
The objective is not to create the most
impressive swarm. It is to create the
simplest architecture that works.
Every
Agent Needs a Reason to Exist
That suggests a practical discipline for
agent design. Before creating another agent, leaders should be able to explain
its unique responsibility, why an existing agent cannot perform it, what inputs
it requires, what output it produces, who consumes that output, when it must
escalate, and what authority it actually holds.
If those answers are unclear, the
organization may not need another agent.
This sounds very similar to workforce
planning because, increasingly, it is.
Agent creation should follow the same
discipline as headcount creation: do not add a role simply because you can.
Agent sprawl may eventually become as
real an organizational problem as headcount sprawl.
From
Span of Intelligence to Span of Orchestration
This also extends another idea I have
explored in
Culture & Code:
Span of Intelligence.
Traditional management talks about
Span
of Control — how many people a leader can effectively supervise. In an
AI-enabled organization, that expands into
Span of Intelligence — how
much human and artificial intelligence a leader can effectively mobilize.
Multi-agent systems introduce yet
another dimension. Even if an executive theoretically has access to hundreds of
agents, how many workflows, decisions, exceptions, dependencies, and
interactions can that executive — or an AI Chief of Staff — meaningfully
understand and govern?
That creates what I would call
Span
of Orchestration:
The amount of distributed Human × AI activity that one leader or orchestrator can effectively coordinate without losing visibility, control, or accountability.
Access to more intelligence does not
automatically mean the ability to manage it. A leader may technically command
one hundred agents while meaningfully understanding only a fraction of what
those agents are doing.
And that creates another risk.
The
Oversight Illusion
In previous
Culture & Code
discussions, I have argued that as AI becomes more autonomous, humans
increasingly move
Above the Loop. The model is compelling: AI executes
while humans set direction, establish guardrails, exercise judgment, and remain
accountable.
But what happens when one human is
suddenly Above the Loop of hundreds of workflows and thousands of agent
interactions?
Technically, human oversight still
exists. Practically, it may not.
I call this the
Oversight Illusion:
The belief that human governance exists simply because a human formally sits above the system — even when the system has become too complex for meaningful human understanding or intervention.
A dashboard is not the same as
understanding. An approval button is not the same as judgment. And a human
being formally accountable for an outcome they cannot realistically comprehend
is not meaningful governance.
Human Above the Loop only works when
humans retain
sufficient visibility and understanding to know what the
system is doing, where the risks lie, which exceptions matter, and when
intervention is necessary. Otherwise, oversight becomes ceremonial.
The
Real Productivity Question
All of this changes what organizations
should measure.
The number of AI agents deployed is not
a meaningful measure of maturity. Neither is the number of workflows automated.
In fact, celebrating agent count may eventually become the AI equivalent of
celebrating headcount: impressive only if we ignore whether the organization
actually needs it.
A better measure may be
Orchestration
Efficiency:
The amount of useful organizational output generated
relative to the effort required to coordinate, verify, govern, and maintain the
AI system producing it.
If a twenty-agent swarm creates a report
that one well-designed agent could produce just as reliably, the larger system
is not necessarily more advanced. It may simply be more complicated.
The more useful question becomes:
Did AI make the work simpler — or did it merely move the
complexity somewhere else?
That question should sit at the center
of every agentic transformation.
Maximum
Leverage at Minimum Necessary Complexity
The best AI architecture will therefore
not maximize any single variable. It must balance at least four things
simultaneously:
·
Capability
— can the system
perform the work?
·
Simplicity
— how many moving parts
are actually required?
·
Reliability
— can the system
consistently deliver the desired result?
·
And
Governability — can humans still understand, monitor, and control what
is happening?
The objective is not maximum automation,
maximum autonomy, or maximum agent count.
It is:
Maximum Leverage at Minimum
Necessary Complexity
That may become one of the most
important design principles of the Agentic Enterprise.
The
Organization Design Lesson
There is a broader lesson here.
Traditional Organization Design has spent decades trying to eliminate
unnecessary layers, duplicated roles, unclear accountability, fragmented
processes, excessive spans, and bloated bureaucracy.
Agentic Organization Design will need to
solve many of the same problems.
The actors may increasingly include AI
agents instead of — or alongside — humans, but the organizational principles
remain surprisingly familiar:
clear roles, minimal layers, clean handoffs,
explicit decision rights, manageable spans, and accountability that does not
disappear between boxes.
The Agentic Enterprise may therefore be
less about having the most sophisticated AI architecture and more about having
the most disciplined one.
More
Intelligence Is Not the Same as More Leverage
We entered the AI era assuming machines
would eliminate work, and in many cases they will. But if organizations
continuously add agents without redesigning how those agents fit together, we
may simply replace execution work with
coordination work.
The winners of the Agentic Era may
therefore not be the organizations with the most agents. They may be the
organizations that need the
fewest agents to create the greatest amount of
reliable value.
That is the deeper lesson of the
Orchestration Tax. More intelligence does not automatically create more
leverage; it creates more possibility. What turns that possibility into
organizational advantage is the ability to structure it without drowning in complexity.
More intelligence creates possibility. Better orchestration turns possibility into leverage.