Every company I talk to is using AI to do the same work faster. Congratulations. So is everyone else.
Faster drafting, analysis and document work can create useful capacity. But competitors can buy access to the same models and tools. The speed itself doesn't stay distinctive for long.
When everyone automates the same tasks the same way, efficiency stops being an advantage. It becomes the new minimum. You haven't pulled ahead. You've just kept pace with a faster treadmill.
This is the efficiency trap. And most of the AI conversation is stuck inside it.
Faster Doesn't Mean Better
Research highlighted by UC Berkeley Haas found that AI can intensify work rather than automatically reduce it. When production gets faster, organisations may increase the volume and pace of work instead of creating room for better judgement.
The tools make more work possible. The operating model decides what fills that capacity.
I've watched this unfold in consulting, in recruitment ops and in engineering teams. The productivity metrics go up. The humans get stretched. And the quality of decisions quietly decays because nobody has time to actually think. Lies, damned lies and productivity dashboards.
Ungoverned AI adoption doesn't produce efficiency. It produces intensity with better optics.
Cutting Execution Before Fixing Orchestration
There's a mistake playing out at scale right now. Companies audit their cost base, identify where AI can replace output and cut the people who do the work: junior developers, content producers, call centre agents, document reviewers.
The problem? Execution wasn't the bottleneck. Orchestration was. That's the layer that coordinates humans and AI, sets standards, reviews quality and holds coherence across the operation. In most organisations, it's middle management, and it hasn't been retrained to run hybrid teams.
Cut execution before you fix orchestration and you lose the organisational memory, relationship capital and embedded judgment that held things together. You've removed the people who knew where the bodies were buried, and handed the shovel to a language model.
The same pattern is emerging in enterprise software. Companies are declaring their tools obsolete because AI agents can replicate features. The actual problem is the integration and workflow layer, not the tools themselves. Most won't die because agents replaced them. They'll die because they failed to evolve from interfaces into orchestration platforms.
The organisations getting this right are doing something unglamorous. They're training managers to review AI outputs, designing accountability for hybrid teams and building connective tissue between tools. Not "humans versus AI." Just competent plumbing.
Governance Turns Shared Technology Into a Distinctive Capability
I know. Governance sounds about as exciting as a compliance manual. Stay with me.
If your AI strategy is about capability, you're already commoditised. Every competitor has access to the same models, the same APIs, the same automation platforms. Capability is table stakes.
What separates organisations that sustain advantage from those that don't is how they structure the rules, oversight and accountability around AI. Done properly, this is the architecture that converts raw capability into outcomes you can trust, repeat and defend.
I think about it in four concentric layers.
The outer layer is strategic: risk appetite, accountability and oversight at board level. Who owns it when AI gets something wrong? What's the escalation path? If you can't answer these in under 30 seconds, you don't have governance. You have a policy document nobody reads.
Next is the operating model: translating strategy into daily behaviour. Usage policies, staff training, expectations that are specific enough to follow. This is where you manage shadow AI, the gap between what people are told to do and what they actually do at 10pm on a Tuesday.
Third is process: safeguards baked into delivery. Human review before AI outputs reach customers. Approval gates at decision points that carry real consequence. The discipline of treating AI as a productive but unreliable colleague who needs a second pair of eyes.
Fourth, at the core, is technical: output validation, content filters, prompt constraints, personal data detection and continuous monitoring. The safety net that catches what every other layer misses.
Skip any layer and the system leaks. Build only the technical controls without the operating model and you get well-filtered nonsense. Write policies without technical enforcement and you get good intentions with bad outputs. The whole thing has to hold together.
Sizing the Cost of Human Absence
The question I hear most: if AI can produce the output, why keep humans in the loop?
Here's how you answer it. Don't try to measure the value of human judgment directly. KISS: Keep it Stupid Simple. Measure what happens without it.
Start with error cost. Record the consequence when an unsupported claim, incorrect clause or miscategorised transaction passes through the workflow. The cost of removing a reviewer isn't only the salary saved. It includes the likelihood and consequence of errors the reviewer would otherwise catch.
Then look at trust. In professional services, when anyone can generate a plausible strategy deck, value shifts towards the person who stands behind the recommendation with context and accountability. Compare retention, corrections and escalation rates across workflows with different levels of human involvement. Treat the difference as evidence to investigate, not proof that one factor caused it.
Next, decision quality. AI makes production cheap, while judgement under ambiguity retains value. Track reversals, rework, exceptions and time to correct a decision. These measures show whether oversight is improving the workflow without pretending to price judgement directly.
Add these up. Error exposure, trust signals and decision rework give you a practical view of the cost of human absence. Compare that exposure with the reviewer time before changing the control.
From T-Shaped to Comb-Shaped
This brings me to a workforce shift that I think most people are underestimating.
The kind of governance I've described can't be done by specialists. It requires people who can connect strategy to operations, technology to culture, risk to commercial reality. You need someone who understands enough about engineering to evaluate an AI output, enough about finance to build the business case, enough about operations to anticipate what breaks on implementation and enough about people to manage the anxiety that comes with it.
For two decades, the career model was the T-shape: deep expertise in one domain, shallow awareness across others. That model is breaking. AI is collapsing the value of narrow depth-as-execution (knowing how to do the thing) while inflating the value of depth-as-judgment (knowing when, why and what it connects to).
The new model is the comb: multiple prongs of functional depth connected by a crossbar of integrative judgment. Not shallow generalism. Deliberate breadth with enough depth in each area to operate, not spectate.
Most hiring frameworks aren't built for this. Job descriptions still say "10 years in X." Promotion tracks still reward going deeper into one silo. The structures assume that bottlenecks are execution problems solved by specialist depth. AI has inverted that. The bottleneck is now integration, and the comb-shaped professional is the one who solves it.
This is where the freed capacity from AI should go. Not into doing more of the same, faster. Into reskilling people across functions and redeploying them into roles that require integrative thinking. Cross-functional rotations. AI governance task forces. Training that builds judgment across domains, not proficiency within one.
The organisations that do this will build teams that are structurally harder to compete against. The rest will have very fast, very efficient operations that nobody is steering.
What Leaders Should Do Next
Stop celebrating how much time AI saves. Start asking what that time is being reinvested into. If the answer is "more of the same, faster," you're in the trap.
Build governance as infrastructure, with the same rigour you'd apply to your technology stack. All four layers. Don't skip the middle two because they're harder to implement than a content filter.
Invest in your people's breadth. The comb-shaped professional, someone who can move between domains with enough depth to be dangerous in each, is the scarcest and most valuable hire in the market right now. Most HR functions aren't looking for them yet. That's your window.
At ELab AI, we help clients define the work, controls and measures around practical AI implementation. It requires more than models and automation. It requires people who can connect strategy, operations, technology and risk.
Use the business AI strategy guide to choose where to start, or read our AI training guide for the people and management capability required. Australian teams can map the same controls to the AI Safety Standard guide. Industrial operators can put a number on unmanaged tools with the hidden-cost framework. If you want to work through the operating model with us, tell us about your business.
Frequently Asked Questions
What Is the AI Efficiency Trap?
The efficiency trap appears when a business uses AI only to increase output and then fills the released capacity with more of the same work. Speed becomes the new baseline without improving judgement, customer value or the operating model.
Why Doesn't Faster Work Create Lasting Advantage?
Competitors can access the same models and tools. Advantage comes from the business context, knowledge and data, workflow, controls and human judgement built around them.
What Does AI Governance Include?
AI governance includes risk appetite and accountability, operating rules, human review, permissions, technical controls and ongoing monitoring. Each layer needs a named owner.
How Should a Business Use Time Saved by AI?
Decide before implementation. Released capacity may support more customer work, better review, learning, cross-functional problem-solving or avoided hiring. Measure whether that intended use happens.
What Skills Matter in an AI-Enabled Organisation?
People need enough depth to judge work in their own domain and enough breadth to connect it with operations, technology, risk and commercial context. AI can produce an output, but people remain accountable for deciding when it is useful.