We've spent a century training humans to work like machines. You're being sold a lie about AI replacing your workforce.
AI doesn't replace people who think. It replaces people forced to work like machines.
For a century we've organised businesses around scientific management - Taylorism - breaking complex work into simple, repeatable tasks. We asked people to follow scripts, suppress judgment, and run with the predictable efficiency of a machine. Now we've finally built the perfect machine for that job.
AI will automate execution, but humans must orchestrate outcomes. A colleague, Danilo Kreimer, asked online whether you'd still hire a human assistant in an age of advanced automation. My gut said yes - because a good assistant is a high-level orchestrator, doing work AI cannot.
01Getting our terms straight
The AI world doesn't agree on terms.
Focused digital specialists - summarise text, scrape a site, autocomplete code. Effective within boundaries, brittle beyond them.
Coordinate multiple models and workflows (LangGraph, Cursor). Excel at structured, repeatable processes; fail when context shifts.
The strategist who holds the vision - chooses the music, adapts to the room, drives a purpose.
This doesn't have to be a threat: give the robotic work to the machines and free people for strategic work.
02The reality of AI orchestration
An AI orchestrator is a coordination system for multiple models - it breaks down requests, routes them to specialised models, and synthesises results. Technically impressive. But where the vendor pitch ends, reality begins:
- The "70% problem": the gap between the demo and the deliverable. "It gets you 70% of the way there, but that last 30% keeps taking one step forward and two steps back."
- The failure loop: ask an AI to fix one bug and it creates two more - and over-reliance stops teams ever developing the skills to solve problems themselves.
- The hard numbers: MetaGPT hits 85.9% on HumanEval - meaning 14% of generated code fails basic tests; whole apps generate in under 7 minutes for under $1, but need real human work to ship.
- The security crisis: Copilot generates vulnerable code 40% of the time; AI-assisted repos are 40% more likely to leak secrets, and 70% of leaked credentials are still active two years later.
03Five capabilities AI can't touch (yet)
1Intent translation
AI follows instructions literally; a human reads what you actually meant.
Transcribes every card, finds each person on LinkedIn, and fires the same generic "great to meet you, let's connect" at all of them.
Reads the intent - nurture promising connections. Sorts into priority piles, sends personal notes to hot prospects, books intro calls with partners, and diarises a check-in on the top three.
2Context navigation
AI acts on explicit data; a human understands the unwritten rules of the organisation.
Books one three-hour meeting with Marketing, Sales and Product and sends a generic agenda.
Knows Sales feels ignored and Product thinks these meetings waste time. Runs short bilateral sessions first to pre-vet ideas and win buy-in, then a one-hour alignment where everyone already agrees.
3Relationship orchestration
AI processes transactions; a human builds relationships.
Drafts a correct, professional, completely forgettable "congratulations on your new contract" email.
Knows the client's CEO loves a particular local bakery. Sends a custom gift basket with a handwritten note: "Heard the news - you'll need the extra energy. Well done!"
4Creative problem-solving
When a tool breaks, AI halts - it's brittle. A human improvises another path.
Reports "Failure: cannot access data source." and stops.
It's for a board meeting, so they pivot: a cached copy of last quarter, scraped competitor press releases, and numbers straight from the sales team - combined into a directionally-correct report, on time.
5Strategic judgment
AI optimises the immediate metric; a human protects long-term value.
Ranks by upvotes and revenue impact, topping the queue with popular but minor feature requests.
Spots a no-revenue task: "refactor the fragile billing module." The only engineer who understands it leaves in six weeks - so they halt features and spend the sprint on the refactor, killing a massive future risk.
04The pattern of failure and success
MD Anderson's IBM Watson partnership burned $62 million over four years for a system that never treated a single patient. It started as a $2.4M, six-month project.
McKinsey's 2024 research shows up to 2x faster task completion and ~50% time reduction in documentation - but only when human oversight stays central.
05Expertise is changing
Traditional mastery meant the steadiest hands or the cleanest code. The most valuable experts now are orchestrators of AI swarms. The best radiologist is no longer the fastest reader of scans - it's the one who orchestrates multiple AI analyses while applying deep human judgment. The risk: a generation who can prompt but not understand.
06Questions for leaders
Sources: MetaGPT (ICLR 2024) · Pearce et al. (IEEE S&P 2022) · GitGuardian 2024 · McKinsey 2024 · Perry et al. (Stanford 2022) · IEEE Spectrum on IBM Watson
Book a Focus Call