90% of your work could be handled by AI. So why is it stuck at 30%?
TL;DR: Anthropic research shows AI could handle 90% of office tasks — but most businesses are only using it for 30%. The gap isn't about tools or effort. It's about whether your business has the architecture for AI to actually operate inside it. Here's what closing it requires.
I talk to founders and COOs every day. And there's a version of the same conversation I keep having, with leaders who are smart, ambitious, genuinely committed to making AI work, and stuck.
Their teams are using AI. Just not in the way that moves the needle.
Isolated tasks, vertical dabbles, a chat assistant someone remembers to open. The deeper opportunity, AI that's wired into the actual intelligence of the business, remains stubbornly out of reach. Nobody can quite explain why.
That gap is the stone in my shoe. It's why we built Decidr. And this week, Anthropic put a number on it.
Their new research measures the difference between what AI is theoretically capable of doing and what businesses are actually using it for.
In office and administration, AI could handle 90% of tasks today. Observed usage: around 30%. You have a machine that can do 100 things. You're using it for 30.
We didn't need the research to know this. We see it daily. But it confirms something important: this is not a niche problem. It's the central problem of this era of AI adoption.
This is not an AI problem
The gap doesn't exist because AI lacks capability. It exists because most businesses haven't built the internal infrastructure to connect AI capability to their actual work.
Look at where observed AI coverage is highest: computer programmers at 75% task coverage, customer service representatives, at 70%, not far behind. These aren't the categories where teams tried hardest.
They're the categories where the workflow architecture made it easy to slot AI in. Clean input-output loops. Discrete tasks. Existing infrastructure in the form of APIs and ticketing systems.
Now look at where the gap is widest: strategy, planning, business development, decision-making. The categories where founders and executives spend most of their organisational capital.
These don't have clean architecture. The logic that governs how work gets done, who owns what, when to escalate, which relationships matter, lives in people's heads. It was never documented.
And until it is, AI sits in the corner like an extraordinarily capable employee who has never been properly onboarded.
Buying more tools doesn't fix this. Neither does more training. The bottleneck is orchestration.
What closing the gap actually looks like
The businesses pulling ahead aren't spending more on AI. They're doing the structural work everyone else is avoiding.
They're treating knowledge capture as infrastructure. Around 80% of what makes a business run is "tacit", held in people rather than systems.
Undocumented decisions, unwritten escalation paths, institutional knowledge that walks out the door when someone resigns. Before agentic AI can act on your business, that knowledge needs to exist somewhere it can be accessed. Getting it out of heads and into structure is unglamorous. It's also the work everything else depends on.
They're mapping how work actually moves, not how they think it moves. AI doesn't improve vague processes. It amplifies them. The teams seeing real returns have invested in understanding the actual path a piece of work takes through their organisation before asking AI to operate inside it.
When CareerOne did this for recruitment, the result was a 65% increase in candidate match rates and 8X more approved applications for interview. The AI didn't produce that outcome. The architecture did.
They're building for inspectability.
The real transformation isn't AI answering questions faster. It's AI running workflows autonomously, making decisions, taking action, flagging exceptions, with every step visible and auditable. That's the distinction between a faster Google and an operating layer. It's what DecidrOS is built to be.
Are you in the gap?
Here's the diagnostic I use on every call. Three questions:
- Can AI access your business knowledge, not generic public data, but your data, your customers, your workflows, your performance metrics?
- Are your AI tools executing automated workflows, or waiting to be asked?
- When AI takes an action in your business, is there a record your team can inspect, learn from and improve?
If the answer to any of those is no, you're in the gap. And the longer you stay there, the harder it gets to close — because the businesses that close it first don't just improve their own performance, they raise the floor for everyone else in their market.
The gap is still the stone in my shoe. The difference now is we know exactly what it is — and how to remove it.
An earlier version of this post appeared on Duncan Brett's Substack.
See how Decidr approaches the architecture problem.
