The Fourth Tool | Eliran Keren

BCG surveyed 1,488 US workers earlier this year.

The finding that made headlines: heavy AI use is causing something researchers are calling “AI brain fry.” Mental fog. Slower decisions. 34% of affected workers actively planning to quit.

The finding that didn’t make headlines: the number four.

One AI tool to two — productivity goes up. Two to three — still improving. Three to four — productivity falls.

That’s not a rounding error. It’s a threshold. And nobody’s explaining what’s actually happening at it.

The thing everyone’s calling brain fry

The BCG study is careful about what it’s measuring. It’s not just AI use. It’s AI oversight.

Workers required to read through and interpret AI-generated content — rather than let agents complete tasks independently — reported 14% more mental effort, 12% greater fatigue, and 19% greater information overload. Not from doing more work. From reviewing what the AI did.

“People were using the tool and getting a lot more done, but also feeling like they were reaching the limits of their brain power,” said Julie Bedard, the BCG managing director who led the research.

The standard response to this finding: we need better AI. More autonomous. Less oversight-heavy.

That response is a category error.

What’s actually happening at tool four

Think about what adding a fourth AI tool means in practice.

You already have a writing assistant. An email summarizer. A meeting tool that captures action items.

Now you add a fourth. A research tool. An automated scheduler. A code helper. Pick one.

Each of those tools produces outputs. Drafts, summaries, suggestions, completed tasks. Each output is something you didn’t personally generate — which means you have to verify it, decide whether to act on it, and track whether that action completed.

With three tools, a skilled knowledge worker can manage that manually. It’s tight, but it’s tractable. The cognitive load stays inside working memory.

At four, it stops being tractable. Not because the brain hits a ceiling. Because the management overhead crosses a line. You are now spending more time administering the AI’s outputs than the AI saved you on producing them.

The question isn’t: “Why can’t people handle four AI tools?”

The question is: “What is a person actually managing when they use four AI tools without infrastructure to support that management?”

The loops nobody’s counting

Every AI action creates an open loop.

AI drafted the email — is it sent? AI captured the meeting — are those action items tracked anywhere? AI completed the research — did someone verify the citations before the doc went out?

When you do the work yourself, closing the loop is automatic. You wrote the email. You know whether it’s sent. The loop closes in the same motion.

When AI does it, the loop closes… when? You have to go back. Check. Verify. Follow up. That’s a deliberate cognitive event, not a byproduct of doing the work.

Three tools means managing three categories of open loops simultaneously. Four means four. Five means five. There’s no structural support for any of it — no layer that tracks what the AI produced, what’s been verified, what’s still pending a human decision.

This is not brain fry. This is an architectural gap expressed as exhaustion.

The coordination ceiling

The InfoWorld piece published this month named something important from the multi-agent systems angle: it’s not that the agents are failing. It’s that the coordination layer between them doesn’t exist. Systems that scale agent count without building coordination infrastructure hit cascading failures, not because of individual agent quality, but because of the absence of shared context and handoff infrastructure.

The same dynamic is playing out at the human level.

Knowledge workers are adding AI tools without building coordination infrastructure between them. No shared context layer. No system that knows what each tool produced and what still needs resolution. No infrastructure that surfaces the right oversight at the right moment instead of letting it accumulate.

The BCG finding isn’t that human brains max out at three tools. It’s that three tools is roughly where the unstructured coordination overhead exceeds the cognitive bandwidth most people have available for managing it manually.

At four, you’re past capacity. Not cognitive capacity. Coordination capacity.

Why the fix isn’t fewer tools

The reflex — use fewer AI tools — surrenders real value to reclaim breathing room.

The execution savings are real. AI compresses defined task time. Giving that back to reduce cognitive load is a bad trade.

The other reflex — get better AI, more autonomous, less oversight-heavy — just moves the problem. Fully autonomous AI still produces outputs. Those outputs still carry accountability. You still have to know what the system did on your behalf, even if you didn’t watch it do it.

The gap that’s creating the ceiling isn’t the quality of individual AI tools. It’s the absence of any layer that manages what those tools collectively produce.

You need a system that knows: which AI actions completed today, which outputs require a human decision, which loops are open, which commitments are pending verification. Not another AI tool. A coordination layer above the tools. The infrastructure that turns a stack of capable tools into a managed system.

Gartner predicted over 40% of agentic AI projects will be canceled by 2027 — not because the models aren’t capable, but because enterprises failed to build the governance and coordination layer around them. The productivity ceiling BCG just documented is the same failure at the personal scale.

The actual diagnosis

Brain fry is a real phenomenon. The symptoms are real. The quitting numbers are real.

But the name is wrong, and the name matters — because a bad diagnosis leads to the wrong treatment.

“Brain fry” implies the ceiling is cognitive. The fix is therefore cognitive: better habits, fewer tools, more breaks.

The ceiling isn’t cognitive. It’s structural. The tools have no coordination infrastructure above them. Every new tool you add without building that infrastructure expands the coordination debt. The fourth tool doesn’t break your brain. It breaks an unstructured manual system that was already at its limit.

The fix isn’t rest. It’s infrastructure.

Build the coordination layer. Not because AI is overwhelming — because AI without management architecture was always going to compound before it simplified.

Eliran Keren — Founder of Deeplica, building the coordination layer for the person at the center.

Sources: When Using AI Leads to “Brain Fry” — HBR, March 2026 · ‘AI brain fry’ is real — Fortune, March 2026 · AI agents aren’t failing. The coordination layer is failing — InfoWorld, April 2026 · Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027