The Bottleneck Is Human Attention: Answer Yield Management

TL;DR: The only real bottleneck in AI right now is human. Not "AI replaces everyone" — something colder: humans cannot inspect the volume AI produces. Companies shove more work onto seniors with models. Seniors drown in review. Juniors lose the dojo. The way out is not more heroic eyeballs. It is manufacturing discipline on intelligence output — specs, automated tests, sampling audits, rollback. I call it Answer Yield Management.

I am James, CEO of Mercury Technology Solutions. Hong Kong — August 2026

Yesterday I sat in an AI advisory session with senior leadership at a Global IT management.

Two questions kept colliding in the same room:

1. Will software engineers lose their jobs?
2. Who should actually use AI if we want ROI this quarter?

The market loves the first question because it is dramatic.

The second one is the one that breaks organizations.

The wrong conclusion people reach

People flatten this into:

"AI will replace humans."

Too vague. Too brutal. Mostly useless as an operating plan.

The actual failure mode is more specific:

AI output scaled. Human verification did not.

You can generate a week's work before lunch. You cannot honestly review a week's work before lunch.

That gap is the bottleneck. Not GPU shortage. Not "the model isn't smart enough." Human attention, judgment bandwidth, and sign-off capacity.

Who is actually exposed

My take in that room was blunt.

If a software engineer has crossed into architect-grade capability — system boundaries, failure modes, tradeoffs, what "good" looks like under load — their seat is not disappearing. Their job is mutating upward.

The people in real trouble are the ones stuck in the middle of the old ladder:

• just graduating
• not enough scar tissue
• waiting for seniors to apprentice them through real production pain

Because the corporate dojo is being dismantled in real time.

The old model:

Senior teaches junior → junior absorbs judgment → junior becomes senior.

The new local optimum for the P&L:

Senior + AI → 5× throughput this sprint → visible benefit this quarter.

Why spend scarce senior hours mentoring when those same hours plus agents ship five times more product?

I am not celebrating that. I am describing the incentive.

Able people, overworked people

Push that logic one step and you get the trap every serious company is in:

• Senior workload explodes
• Mentorship becomes "we'll do it later"
• New-hire pipelines thin out
• The people who can still judge quality become the single point of failure

Able people used to multi-task. Now able people overwork — courtesy of AI.

And this is not a software-only story. Same pattern hits design, legal ops, research, marketing, finance analysis — any role where generation got cheap faster than verification got automated.

This is the K inside the firm

I have been writing and speaking about the K-shaped economy under intelligence inflation for a while — including in New Opportunities Under Intelligence Inflation.

People hear "K-shaped" and think countries and asset classes.

Look inside one org chart.

Resources concentrate on the upper arm:

• higher pay for proven seniors, or
• the budget that used to hire juniors gets converted into compute and tools for seniors

Double transfer. Upper arm compounds. Lower arm starves.

World-scale K-shape. Firm-scale K-shape. Same geometry.

Then competition between firms forces the upper arm to strap on more AI — and a new ceiling appears.

Even the best senior + the best agent stack still collides with the fact McKinsey-style cognitive load charts have been screaming:

Human brains fatigue. AI does not sleep. You will never out-read a machine that never stops writing.

So what now?

Steal the factory floor

Manufacturing solved "too many units for human eyes" decades ago.

Chips on a line do not get inspected one-by-one by a sage with perfect vision. The system does this:

1. Define the spec — what counts as good
2. Automate the tests — machines check machines
3. Sample audit — humans watch the process, not every unit
4. Rollback on failure — bad lots do not ship; root cause feeds the line

The human job upgraded from "stare at every die" to "design the inspection system."

AI output is the same physics with prettier tokens.

If your operating model is "senior engineer line-reviews every AI commit," you have reinvented manual visual inspection at semiconductor scale. That is not craft. That is how you burn out the only people who can still tell good from garbage.

Answer Yield Management

I call the fix Answer Yield Management.

Software verification has to become layered infrastructure, not a hero activity.

Risk tier

Default path

Human role

Low

AI checks AI against hard specs

Spot-check the checker

Medium

Automated suite + statistical sampling

Review samples + drift

High

Full gates + named human sign-off

Accountable judgment

Unstructured "just vibe with the PR" is how you die slowly. Zero inspection is how you die fast. Graded yield control is the only living path.

Stanford's AI Index 2026 put a number under the discomfort: AI agents on OSWorld-style real computer-use tests jumped from roughly 12% to ~66% success. Celebrate the jump if you want. Also notice the residue — in structured evaluations you are still looking at failure on the order of one in three.

That is not "almost AGI, ship unattended." That is a factory with a known defect rate.

You do not respond to a 33% fail mode with vibes. You respond with process.

What this looks like when it works

I run a version of this on our own stack.

Most of mtsoln.com is written with AI in the loop. I still have a functioning nervous system at the end of the day because the structure is graded:

• architect sets the spec
• automation is the bouncer
• humans only escalate on serious red lights

Important reminder, said without romance:

AI will almost always leave holes — even on simple tasks. Weird gaps. Confident nonsense. Locally pretty, globally wrong.

For seniors, that means the career center of gravity moves again:

From producer → to problem-setter.

The more AI output floods the channel, the more valuable the person who can define:

• what done means
• what error budget is acceptable
• what auto-rejects below threshold
• what requires a human name on the line

Musk's line about taste in the AI era is directionally right — and in engineering, taste is not aesthetic mood board energy.

Taste is architectural judgment. What belongs in the system. What must never ship. What failure looks like before it has a ticket number.

Models are still weak at that. That is the moat.

The junior problem, without the lullaby

I will not sell you the soft story.

The dojo was demolished. Companies will keep choosing senior×AI throughput over patient apprenticeship whenever the quarter is on fire. That incentive is real.

But this is also the first era where every junior has a 24/7, infinite-patience sparring partner sitting next to them.

Used to be: the company forced you to train. Now: you force yourself — or you fall off the lower arm of the K.

AI is an exploit for the self-directed. AI is a catastrophe for people who only move when pushed.

The K-shape does not stop at job title. It runs through mindset.

Proof, not theory: four–six weeks → eight hours

I will put a real delivery on the table.

A client site: https://www.eastwood.page

Old path for this class of site — 4 to 6 weeks end to end:

Phase

Time

Understand business + priorities with client

2 days

Content / narrative

2 days

MVP build → client review + sign-off

4–6 days

Align with developers

2 days

Complete development

7–10 days

QA

2–4 days

Cloud deploy

2 days

Revised path with Answer Yield Management + agentic production:

Narrative → build → roll out to cloud → client sign-off.

Eight hours. End to end.

Not because someone typed faster. Because the unit of work stopped being "hero crafts every pixel and every paragraph under manual inspection" and became:

spec integrity → generated production → automated gates → human judgment only where it still pays.

That is the same manufacturing upgrade wearing a website costume.

What to do this week

If you lead eng or digital delivery:

1. Stop measuring AI ROI as tokens generated. Measure escaped defects, cycle time, and senior hours spent on pure re-reading.
2. Force seniors to externalize taste into specs. Acceptance criteria, error budgets, auto-reject rules. If it only lives in their head, you do not have a system — you have a bottleneck with a pulse.
3. Install tiered review. Low / medium / high. No more universal line-by-line cosplay.
4. Buy compute for judgment leverage, not just generation volume. More output without more yield control is how you industrialize technical debt.
5. Rebuild a thin dojo on purpose. Bounded quests for juniors inside the automated factory — or accept you are eating your seed corn.

If you are individual IC:

• Below architect: your emergency project is judgment under load, not more prompt tricks.
• At architect: your job is becoming yield owner — the person who designs how wrong answers die before customers see them.

The only path I currently trust

Complete non-inspection is suicide. Complete human inspection is death by overwork. Graded, manufacturing-grade control of AI answers is the surviving middle.

The bottleneck is human attention. The upgrade is Answer Yield Management. The career move is from making answers to defining what a good answer is allowed to be.

Build the inspection system. Or drown in your own throughput.

Mercury Technology Solutions: Accelerate Digitality.