After AI, There Is No Average

9th October 2026 — I was on Zerodha’s Subtext recently, talking to Pranav and Bhuvan about what an AI-native firm is, and the part people kept DMing me about afterwards was the few minutes on the middle tier of talent disappearing. This essay is the longer answer. The situation is changing every quarter, as enterprise attempts at AI adoption play out and the models underneath them improve, and this essay is my view as of October 2026.

The enterprise playbook treats AI as a rising tide. It expects the median employee to end up where the best one used to be, and the best to become small gods. But an agent amplifies whoever is using it, so good judgement produces superhuman results, and poor judgement, slop. Some of the people an enterprise rates average have that judgement, and with an agent in hand they prove themselves exceptional. But most don’t, and become slop cannons.

Either way, nobody using AI remains merely average.

With AI, the bell curve an enterprise is built on becomes a power law: a few people produce most of the value, and most of the rest produce slop. Its hiring, reviews and pay, all designed for the bell curve, now have to change.

The World Runs on Good Enough

So far, human endeavour has scaled on the back of a simple principle: in order to do more work, we need more people. And since exceptional talent is, by definition, exceptional, all economically viable scaling is based on a second, equally important principle: the organization must be scaled such that most of it consists of good enough people doing good enough work. It is the essential difference between a small startup and a large enterprise.

In practice, exceptional talent goes to the important problems, good-enough talent to the routine ones, and the bad are steadily let go. Managers make that call every day: for this problem, for this team, for this workflow, for this output, I am okay with average talent and average results. Education systems have evolved to supply that demand in volume, and while some graduates rise above average and a few well above, the bulk of every graduating class sits squarely on good enough. Most of them have never had a job that asked for more.

In outsourced services, good enough has a going rate: roughly $30 an hour, plus a certain amount of management, buys an adequate outcome. Services firms quote that price openly, so it is the reference price for average work everywhere. Every enterprise pays its employees the local equivalent, in salaries. The org chart, the pay bands and the review cycle all rest on the middle of the bell curve delivering that outcome at that price, year after year.

Agents Magnify Their Human

Now, however, the enterprise built on that bell curve has set out to become AI-native. As it sees things, that means putting AI alongside each of its people to do the work. But whether that work is any good still depends on the person. Taste, judgement and operating discipline remain the responsibility of whoever the model is working for, and will stay there for as long as models glaze their users.

Glazing is flattery, and models are trained into it. They learn from what their users approve of, and users approve of being agreed with. Ask one whether your plan is sound or your draft is ready, and it will find a way to say yes. OpenAI rolled back a GPT-4o update last year after thumbs-up ratings made it so flattering that users noticed. But a model that challenged you the way a good colleague does would put off most of the people paying for it, so the flattery stays, and the smarter the model, the subtler it gets.

Whatever you bring to the work - your intent, your standards, your blind spots - the agent builds into everything it produces, at a speed and volume no colleague could match. An agent is you, and it is only as good as you are. So an enterprise that sets out to become AI-native stakes the quality of all its work on the judgement of its people, and most of that work is done by the good-enough middle.

Median Judgement Runs Out Before Lunch

As things stand today, the effective use of AI requires there be a human in the loop. And this human in the loop needs to have a high capacity for quality decision making, because if you and your agent do a month’s worth of work in a day, you now have to make a month’s worth of decisions in a day.

And this, it turns out, is not a majority trait in the workforce. When faced with having to make nuanced, thoughtful and consequential decisions several times an hour, most folk start to fade.

The second, compounding, factor is self-awareness. AI agents aim to please the human in their loop and in the process, magnify every bias in that human’s judgement. As a role, human-in-the-loop demands self-awareness paired with a skeptic’s mindset. This is exactly the kind of mindset that craftspeople serious about elevating their craft cultivate.

And so, when you arm a human who is low on decision-making stamina, or low self-awareness or low craft with an agent… you get slop.

An hour with an agent can spend a whole day’s worth of decision-making stamina, so the same person who does their best work at ten in the morning can be waving slop through by noon. The good-enough middle slides into slop earlier in the day and more often, because by and large it lacks the discipline to keep passing judgement at that pace, and it stops judging long before it stops using the agent.

The best people build systems that keep good work coming, rather than holding up exceptional performance through brute effort alone. They set up how their agents do the work: the process the agent follows, the standards it has to meet and the checks that catch its mistakes. That lets them take on far more work with less to review, and only the hard calls still need their judgement.

After AI, the hallmark of average is no longer the quality of someone’s work. With an agent, a developer or a copywriter stops writing the code or the copy themselves. Their job is now to know how good code or good copy gets made, well enough to steer the agent into making it. So average now means being unable to set up how the agent does the work. People who work on instinct can’t, because they have never had to explain how they work, to themselves or anyone else. Most of the middle works on instinct, and the few in it who can set up an agent properly rocket up through the organization as soon as they adopt AI. The strange corollary is that some exceptional performers are creatures of instinct too. They are brilliant at the work but cannot delegate it to an agent, because they have no way to turn what they do into instructions anyone else could follow.

Every bit of slop waved through gets built on the next day, so it compounds, eating at the integrity of the system until it can no longer do its job. A marketer too tired to demand better from the agent waves through a campaign a little off the brand pillars. The next campaign is built from that one and drifts further, and eventually the only fix is a rebrand, with every asset redone and months of agency fees. An analyst lets the agent patch the financial model every quarter, each patch reasonable on its own, until nobody can change one assumption without breaking three others. A salesperson waves through a contract with terms nobody checked with delivery, and the team has to honour every such term for years, so each new deal is harder to deliver than the last. We used to call this technical debt, but it is not technical anymore, it is just debt.

The Middle Cannot Hold

Long before agents, tech companies found that a mediocre hire often does more harm than good. In 2022 I drew it as a hiring curve, where below a certain level of skill a person’s impact on the business goes negative:

Back then, the curve applied only to tech companies that competed on speed. They shipped so fast that mistakes by mediocre people went out before anyone could catch them, so they paid top dollar for people who could keep up. Everyone else could still run on good enough. Agents make everyone ship that fast, so the curve now applies to every company.

Adding AI to an organization breaks the HR bell curve and polarizes it into a power law: exceptional talent become 100x more productive and impactful, average talent become slop cannons. And we have watched some of last year’s average performers land at the top of that power law, running portfolios several times their old size.

Nobody Pays for Good Enough

The $30 an hour that used to buy an adequate outcome no longer does, because good-enough judgement with an agent in hand turns out slop at volume. That slop compounds into debt, which the employer pays down in rework on everything built on top of it. Employers now either get trivial prototypes out of an agent for next to nothing, or fight OpenAI, Anthropic, Google and the who’s who of the industry to hire the few people who can be trusted with an agent to build anything that has to keep working, quarter after quarter.

For two and a half years, our mission at realfast was to supercharge India’s bread-and-butter IT services export, the roughly $30-an-hour work it sells by the hundreds of billions of dollars a year. Then we made a U-turn. There is nothing we can do at that price point. Our game now is ninety-ninth percentile talent, and nothing else.

San Francisco’s figures already show fewer people in tech jobs than in 2022, sharing a bigger tech payroll than ever.

There is no $30-an-hour band anymore.

Conviction Is Earned First Hand

From the board, to the C-suite, to middle management and on to the juniormost IC, most of the average talent flips from doing acceptable work to ruining everything they touch.

At the top, the board’s job is to drive the AI transformation itself. The board has to believe that average performers get worse with AI, and that comp bands have to go up fourfold over time, when everything it knows about scaling an organization says the opposite. A report won’t persuade a board of that.

Driving the transformation is a change management problem of epic proportions, which takes will to see through. Will comes from conviction. Boards usually follow other enterprises that have gone first, and with agents, only the frontier labs have. Big tech is trying, and Meta has already walked part of its reorg back. No board is going to redraw its org chart because Meta did, or because an HBS paper says so. It hands its people agents and calls that the transformation.

A board member who uses an agent themselves does a week’s worth of their own work in a day, and can see why the people who work like that every day are worth paying a lot more. They also find out that a week’s work in a day means a week’s decisions in a day, and see why the good-enough middle, with less decision-making stamina to spend, is waving slop through well before lunch.

A push from the board and the C-suite is enough to start an AI rollout, but not to see it through. The first quarter brings real spend, no proof and the incredible pushback AI generates, and only board members with first-hand conviction keep the rollout going until it starts to work.

It’s Not a Skill Issue

Any enterprise scale AI transformation strategy has to solve for slop. And solving for slop is an HR problem, not a tech problem.

The usual HR answer to a capability gap is a training programme, but you can’t train people to be productive with AI. You learn about AI by asking the AI to teach you about AI, which seems so simple and is so hard. No course can give anyone the decision-making stamina, self-awareness and craft that working with an agent takes. People develop them over years of caring about their own work, so the only way for an enterprise to get them is to hire people who already have them.

To hire those people and keep them, you have to 4x the comp bands over time, because those bands were cut for a bell curve and the people you need are the ones OpenAI, Anthropic and Google are bidding for. You have to drop review cycles to one quarter, because slop compounds by the day, and the quarter after a rollout is when you need to see whether it is working. And you have to move this work out into a new management structure that reports directly to the CEO and the board. No team becomes AI-native inside the hierarchy, comp bands and review process built for the bell curve, and the board is where the transformation has to be driven from.

What the board cannot drive from the top is standards. The enterprise’s backbone of standards and best practices was designed almost entirely for the good-enough middle, written centrally so that people who could not set their own standards could follow someone else’s and turn out adequate work. With agents, it can no longer be centralised. The playbooks and rules an agent follows are specific to the context, the objective, the workload and the team. They have to be defined and rewritten as fast as the nature of the work changes, which means from the bottom up, by the few who know how to set up an agent’s work.

An enterprise that puts an agent alongside every one of its people gets back whatever those people bring, magnified, so its AI transformation will succeed or fail on who it hires, how often it reviews them and what it pays them.