Genesis: Human Experience in the Age of Artificial Intelligence | Synthesis: The Superintelligence Protocol

Subscribe

Genesis: Human Experience in the Age of Artificial Intelligence | Synthesis: The Superintelligence Protocol

admin Avatar

By

Luke Soon

·

—

Agents Are the New Electricity; You Cannot Rent a Moat

Every competitor can rent the same agents from the same handful of labs. Agentic AI will not be your advantage. How you govern it might be.

I. The cruise ship, the kayak and the shared engine

MIT Sloan has published tidy advice for leaders, 5 ways to make agentic AI a competitive advantage. Large firms are cruise ships, it says; AI-native start-ups are kayaks. Launch a dinghy to explore the cove, then upgrade to a speedboat.

The image misses one detail. The cruise ship, the kayak and the speedboat now run on the same engine, rented from the same suppliers, at a price Epoch AI estimates is falling by between 9x and 900x a year. When everyone shares the engine, the engine stops being the reason anyone wins.

The five recommendations are sound: redesign the work, set realistic goals, treat governance as brakes that let you go faster, redraw the org chart around accountability, and accept that this is not business as usual. I would sign every one. But advice every firm can follow produces parity, not advantage. And the piece, for all its good sense, cites not a single number. So let me supply some.

II. The 6% problem

Start with the gap between adopting agents and profiting from them.

  • McKinsey’s August 2026 survey finds 40% of large firms now scaling agents, up from 27% a year earlier. Yet only 37% report any EBIT impact from AI, and the share of genuine high performers is stuck at about 6%.
  • BCG divides firms into a 5% “future-built” elite, 35% scalers and 60% laggards with minimal gains. The elite enjoy 1.7x the revenue growth, 3.6x the three-year shareholder return and 1.6x the EBIT margin.
  • 56% of CEOs in PwC’s 2026 Global CEO Survey see no significant financial benefit from AI yet; only 12% see both cost and revenue gains.
  • Forrester reports only 15% saw an EBITDA lift, and expects firms to defer a quarter of planned AI spend into 2027.
  • Gartner expects more than 40% of agentic projects to be cancelled by the end of 2027, and reckons only about 130 of thousands of “agentic” vendors are the real thing.

KPMG says 62% of large firms are now building or deploying agents. Capgemini finds 2% at scale. The adoption curve is steep; the advantage curve is flat. Kai-Fu Lee, never a pessimist, calls most corporate AI programmes theatre unless they move earnings. On this evidence he is generous.

StudyMeasureShare of firms
Adopting agents
PwC AI Agent Survey 2025adopting AI agents79%
KPMG AI Pulse, Q3 2026building or deploying agents62%
McKinsey State of AI 2026large firms scaling agents40%
BCG AI at Work 2026agents integrated into workflows30%
Capturing value
Deloitte State of AI 2026growing revenue through AI today20%
Forrester Predictions 2026saw an EBITDA lift from AI15%
PwC Global CEO Survey 2026both cost and revenue gains12%
McKinsey State of AI 2026AI high performers6%
Infosys and HFS 2026agentic AI pioneers6%
BCG Widening AI Value Gap 2025future-built firms5%
Capgemini Research Institute 2025agentic AI deployed at scale2%
Adoption is broad, advantage is narrow. Measures and samples differ by study; some cover AI broadly.

III. Electricity was never a moat

In 2003 Nicholas Carr told Harvard Business Review readers that IT doesn’t matter. He was wrong about its importance and right about its economics. Infrastructural technologies differentiate briefly, then become commodity inputs that everyone must have and no one can win on.

Jay Barney’s resource-based view says sustained advantage needs resources that are valuable, rare and hard to imitate. A frontier model, available by API to every rival, is valuable. It is not rare. It is copied in an afternoon. Next quarter it will be cheaper and better, for your competitor as much as for you.

And a rented advantage can be switched off. On 12 June 2026 a US export-control directive suspended Claude Fable 5 and Mythos 5 for foreign nationals, and Anthropic disabled both for every customer until 1 July. No data moved. The service simply stopped. Any firm whose edge ran on those models learned that it had been renting its moat from someone else’s government.

IV. Where the advantage actually lives

If the model is rented, the moat must be built from what the firm owns around it. The evidence points to four layers.

1. The loop, not the model. In 2024 Andrew Ng showed that GPT-3.5 scored 48.1% on a coding benchmark on its own, and GPT-4 67.0%. Wrapped in an agentic loop of reflection and tool use, the weaker GPT-3.5 reached 95.1%. The loop beat the model. McKinsey’s 2026 high performers are the same story at enterprise scale: nearly three-quarters have redesigned workflows, against about a quarter of everyone else. Bain finds leaders spend $4 on people and process for every $1 on technology. This is Prompts → Loops → Loop Governance in practice.

Case: Allianz Australia. Project Nemo uses seven agents for one narrow job, food-spoilage claims under AUD 500. Processing and settlement time fell by 80%, and the system was built in under 100 days. Allianz did not have a better model than its rivals. It had a better-bounded loop.

Case: Amazon. An agent upgraded tens of thousands of internal applications to Java 17, saving an estimated 4,500 developer-years and $260m a year. The task was dull, well defined and verifiable. That is precisely why it worked.

2. Judgment as the scarce input. The Harvard and BCG study of 758 consultants found AI raised quality by more than 40% on tasks inside its frontier, yet made people 19 points less likely to be right just outside it. Bain puts the arithmetic of agents bluntly: a model that is 97% accurate per step completes a 34-step task only 36% of the time. Someone has to know where the frontier runs and catch the rest. That someone is expensive and tired: BCG’s 2026 research on “AI brain fry” found affected workers make 39% more major errors. Meanwhile 29% of agentic leaders expect fewer entry-level roles, which is how the Frozen Workforce forms: we automate the apprenticeship that produced tomorrow’s supervisors. PwC’s 2026 Jobs Barometer shows the market already pricing this: a 62% wage premium for AI skills, and productivity growth of 163% in the top fifth of AI-exposed firms.

Case: Moody’s lending clients. A tier-1 US bank now generates 35 to 40% of each credit memo with agents, and banks in Asia and the Middle East report decision times down by up to 80% (Moody’s 4Q25 remarks). The agent drafts; the credit officer decides. The advantage is the officer’s time, redeployed to judgment.

3. Governance as speed. Here MIT Sloan is right, and I would push further. Governance is not the brake on advantage; it is the condition for it. Infosys finds 95% of firms have had an AI incident and 77% lost money, but responsible-AI leaders lose 39% less. PwC’s CEOs with strong foundations, including Responsible AI, are three times more likely to see meaningful returns. Yet only one firm in five has mature governance for autonomous agents (Deloitte), and KPMG finds confidence in governance rising to 73% while the share actually building controls into agents fell to 30%. Apply the enforceability test: a control that cannot be observed, measured and acted upon is decoration. A firm whose controls are real can let agents run where a rival’s must still ask permission.

Case: Ping An. AI service representatives handled more than 1.7bn interactions in 2025, 80% of the insurer’s service volume, and agents helped generate RMB133.2bn in sales (Ping An). Scale like that in a regulated industry depends on a governed channel, not a cleverer model.

4. Sovereignty of the switch. The durable asset is not the model but the ability to stop it, swap it and switch it back on. Beijing has understood this. China’s May 2026 rules on intelligent agents divide decisions into three tiers, those reserved to the user, those needing authorisation and those the agent may take alone, and give users the final say. Singapore’s IMDA framework ties every agent to a supervising human. MAS now expects kill switches and logs of agent reasoning. As I set out in my kill-switch essay, of 40 legal instruments surveyed, only seven carry binding stop requirements. The firms that can stop, swap and restart their agents on demand will keep running when their competitors’ suppliers go dark.

V. Where I part company with MIT Sloan

The lecturers say the biggest risk is failing to embed AI at all. That is half right. In 2026 the bigger risk is undifferentiated action: spending heavily to reach the same place as every rival, on the same models.

Case: Klarna and Commonwealth Bank. Klarna said in 2024 its assistant did the work of 700 agents; by May 2025 its chief executive conceded quality had fallen and began rehiring humans. CBA reversed 45 redundancies tied to a voice bot in August 2025 after admitting the jobs were not, in fact, redundant. Both cut first and redesigned later.

They are right that humans stay “on the hook” for tax, law and failure. But an accountability line on an org chart is only as strong as the person at the end of it. A named owner supervising forty agents with no time, no training and no rehearsed stop is not accountability. It is a scapegoat in waiting.

And the dinghy needs a warning label. Pilots are cheap, so firms launch hundreds; McKinsey found about 90% of vertical use cases stuck in pilot. A fleet of dinghies is not a navy.

VI. Five questions for the board

  1. The parity test. If every competitor had the same model tomorrow, which part of this would still be ours?
  2. The swap test. Could we move this workload to another model in 30 days, and have we tried?
  3. The stop test. Who can halt this agent mid-action, and when did we last rehearse it under load?
  4. The judgment test. How many agents does each supervisor oversee, and what is that person’s cognitive budget?
  5. The loop test. Which workflow did we redesign, rather than merely automate?

A programme that cannot answer these is buying capability, not advantage.

VII. My usual panel weighs in

  • Andrew Ng expects agentic AI to keep rising in value while calling the technology highly limited. His weaker-model-in-a-better-loop result is the empirical heart of this essay.
  • Erik Brynjolfsson found a 13% relative employment decline for 22 to 25 year olds in AI-exposed jobs, and says the most successful companies are the ones that measure well. Advantage goes to reinvention, not cost-cutting.
  • Kai-Fu Lee says the basic unit of the company will shift from a human to an agent, yet calls most corporate AI theatre unless it moves earnings: the optimist who still demands a P&L.
  • Fei-Fei Li reminds us that higher productivity does not automatically produce shared prosperity, and that AI has far more room to augment people than to replace them.
  • Zhang Ya-Qin calls 2026 the first year of agentic AI, and insists every agent needs a traceable responsible party. Accountability as a design rule, from Beijing.
  • Dario Amodei warns that half of entry-level white-collar jobs could go within five years, while his own company’s engineering guidance urges the simplest workflow that works.
  • Daron Acemoglu calls agents sold as one-to-many worker replacements a losing proposition, and caps AI’s ten-year productivity gain at about 0.66%.
  • Yoshua Bengio and Geoffrey Hinton warn that unrestrained agents already show self-preservation and may pursue hostile subgoals. Bengio’s answer, a non-agentic overseer, is an external check rather than a self-report.
  • Jensen Huang says IT will become the HR department of AI agents. That is Loop Governance as an operating function.

The panel agrees that the human job is moving from doing to directing, and that entry-level work goes first. It splits on speed, and on whether agents bring advantage or parity. The data above settle the second question for now: parity for most, advantage for the few who own the loop.

VIII. The Long-AND

This is not an argument for caution. It is Long-AND, not Short-OR: adopt aggressively and build what cannot be rented. The Fork ahead, Star Trek or Mad Max, will not be decided by who has the best model, because everyone will. It will be decided by who can govern what everyone can buy.

The winners will not be the firms with the most agents. They will be the firms that can say, with evidence, what their agents are doing, why, and how to stop them.

Short-term turbulence, yes. The long-term abundance belongs to those who own the loop.

Sources

Several large figures cited here cover AI broadly rather than agents alone, and some case studies are self-reported by companies or their vendors.

Leave a Reply

Short-term Turbulence for Long-term Abundance.

lumelalabs.ai