The Second-Order Effects of AI-Augmented Organisations

Jun 8, 2026

Once AI colleagues are real, span of control collapses, hierarchies flatten, team-size doctrines die, and the new failure mode is attentional drift, not cognitive overload.

In a previous post I argued that the right way to model an AI-augmented organisation is as an organisation: an org-chart but with human and AI workers, each with a role and a responsibility, with a human accountable at the top. That framing is a useful design tool, but what might happen if you adopt this?

Traditional organisations typically have a hierarchical shape and are ultimately resource constrained by time or money. Deeper organisational structures burn time through communication and coordination. And all organisations are constrained by money.

With AI colleagues, both constraints relax. And you can structure your org however you like, optimising for your specific context. Flatter orgs become more viable, because one human can directly manage many AI colleagues. Deeper orgs are also viable, because adding layers only costs GPU time.

The rest of this post is about what follows from taking that seriously.

From cognitive to attentional burden

In human businesses, a business theory called span of control suggests that a manager can supervise roughly three to eight subordinates, depending on how much cross-communication needs supervising. Information technology stretched that number, because digital communication is easier to monitor. The underlying constraint is cognitive: the manager must understand the domain well enough to listen, plan, decide, and act, repeatedly.

In AI-augmented businesses, with powerful cognitive LLMs, this is no longer a limit. Language models are capable of understanding high-level, fuzzy domain concepts. Given the right guidance, context, and connectivity, they can excel in many information-centric tasks. The human designer still needs domain expertise to specify a good task and provide good guidance, but the day-to-day burden is attentional, not cognitive. How much work your organisation gets done depends on the attention span of the humans at the approval gates.

Software engineering with coding assistants is probably the clearest example so far. Software engineering with AI assistance consists of two repeating approval gates: specification and review. Engineers must specify the work well enough to provide guidance about the goals of the work and codebase context. When the agent finishes, the engineer reviews the solution and either provides more guidance or accepts the code. There is still a cognitive burden - "is this code good?" - but it is lower than before, because AI is writing the code. The attentional burden - "hey, human, look what I've done!" - is higher, because you may have several assistants working in parallel.

Automating approval gates with AI managers

In an AI-augmented organisation the optimal span of control number changes dramatically. Work is no longer mostly cognitive, and is pushed towards the attentional limit at the approval gates designed into the org. Organisational velocity becomes proportional to the number of human touchpoints. An optimal AI organisation minimises the number of human approval gates.

Which begs the question: can the approval gates themselves be automated? Often, yes. Traditional ML techniques are useful here. You can quantify the error rate of any decision gate, including the rate of mistaken human approvals. You can also train a model to predict approval decisions, and in its error rate can be lower than the human equivalent.

All processes have an error budget. If an AI sends a letter to a sales prospect who turns out not to need the service, that is an error, but a low cost one. In healthcare, the equivalent error has very different stakes. The architect of the organisation must design the trade-off explicitly. For gates with a high error budget, engage an AI as the gate. In an AI org this is actively encouraged, because it raises velocity and reduces the burden on humans.

The end of organisational hierarchies

Traditional organisational structures are hierarchical because information must be summarised, reported up the chain, and acted upon. With a span of control of around ten, a 1,000-person organisation needs at least three layers (10 x 10 x 10). Add a C-suite and a board and you are at five or more. These hierarchies are typical even today, but they exist because of the antiquated desire for human information management.

At each layer, detail is necessarily lost, to keep the volume of information manageable for a human. But if AI-augmented organisations can reach a much higher span of control, the need for layered information management drops, to the point where hierarchies are no longer mandatory.

Instead, organisational structures will be designed to define accountability. Hierarchies may still emerge, but as a business-specific shape driven by accountability, not by historical expectation.

Beyond copy-paste teams

Structures based on team size (pods, squads, tribes, two-pizza teams) all derive from span-of-control heuristics that no longer apply. Teams, if a team is required at all, will be drawn around domain-specific capabilities or services. I believe domain-driven design will become increasingly important as AI-augmented teams deliver business value within bounded contexts. And the root of each bounded context will be at least one human who is accountable for it.

This does not mean every capability has to be independent. There is a strong argument for duplicating positions with the same role across teams. The resource bottleneck is minimal, unification helps with org-architecture management, and sharing org-wide learnings improves global effectiveness. A project-manager-like role in one team will look very similar to the same role in another, perhaps with team-specific tweaks like a different Slack channel or context source.

Eventually I expect a templating layer to emerge for org sub-structures: a common set of roles for software delivery, a default pattern for project management, the (lack of?) need for a C-suite, and so on.

Roles and departments will reshape

Plenty of other shifts follow from this, depending on which capabilities are commoditised first. For example:

• IT and HR converge, because provisioning a worker and provisioning a workspace become the same act.
• Engineering and operations blur further, because the agent doing the work may also be the operational artefact. A customer-facing chatbot can feed its own learnings back into making a better chatbot.
• Internal tooling and platform teams merge into something closer to agent-ops teams.
• Procurement starts buying tokens and model capacity instead of SaaS seats.
• Legal has to handle new AI-specific scenarios around liability, IP, and disclosure.
• Finance has to invent a cost centre that is neither payroll nor capex.
• The CIO/CSO remit expands to cover agent behaviour, not just human and system security.

And new roles emerge to manage AI organisations:

• Agent manager: supervises and steers a set of AI subordinates.
• Org architect: develops and maintains org-wide architectural patterns.
• Role designer: designs and develops new agent roles.
• Agent SRE: keeps the AI workforce healthy.
• Context developer: ensures the AI workforce has the right information at the right time.
• AI resources manager: evaluates whether agents are doing their job (i.e. evals).
• A variety of domain-specific human-in-the-loop roles for process gates that cannot or should not be automated.

Ambiguity causes drift

The biggest threat to all of this is ambiguity, and the drift it causes. Ambiguity at any level, from the smallest task up to the largest strategic decision, causes the business to drift. The direction of drift is what matters. An intentional, good drift is framed as improvement and learning. But even good drift can take a business towards goals its stakeholders do not agree with.

AI workers, left to their own devices, use a combination of their latent knowledge, their role, and their persisted experience to guide their actions. A poor role definition, poor context, or a poorly-matched model all introduce ambiguity, which can derail the workflow the worker was designed to enact.

What makes this worse is the absence of resource constraint. The AI worker can work, and work, and work some more. This is very visible today in "vibe-coded" software, where an initial seedling of an idea sprouts legs, arms, and other appendages best left unnamed.

The same happens in businesses. You start with one customer, product, or service hypothesis, a simple idea, and attempt to test it. Over time, new questions are asked and new ideas formed. AI orgs make it trivially easy to test these hypotheses, so the same thing happens, faster. The business you end up with looks nothing like the original idea. You are left with a colossal Frankenstein attempting to test all the things.

This is not a new problem, and the solution is not new either. Design the org to be decoupled and cohesive. Remove the parts that are no longer relevant. Kill experiments once they are complete. Make humans accountable.

Towards accountable orgs

On that note, I'm coming to the conclusion that one of the best operating models is "if you build it, you're responsible for it." This is basically the DevOps model, but previously the decision to build something came from from a senior business or product leader. They were ultimately responsible for putting resources (or pressure!) in the right locations. But since that role has moved somewhat towards the engineer, thanks to AI coding assistance, I think this becomes even more important.

The alternative is to fall for the call of the coding Siren; to code and code until an itch is scratched, only to leave it to die on the rocks.

Encouraging a strong business responsibility to the thing developers are accountable for can only be a good thing.

This has the added benefit of forcing engineers to sunset products they don't want to continue with. Or at least hand it off to someone that is excited by it.

Structure your org around bounded business contexts. Businesses will undoubtedly go through enormous change. Make humans accountable for what they produce.

Next: begin rolling out Helix-Org, the place to define AI-augmented organisations, to alpha users.