AI made people faster. Why didn't the company get stronger?
Everyone is on ChatGPT. Why didn’t company results take off?
This is not one anecdote. DX tracked 400 companies for 16 months: AI-tool usage rose 65%, code delivery rose less than 10%. The NBER surveyed 6,000 executives; more than 80% said AI had no measurable effect on productivity.
AI really did make each person faster. The company did not get stronger. Where did the productivity go?
A story from 130 years ago
In the 1890s factories put in electric motors. Everyone expected a productivity explosion. For the next 30 years, output barely moved.
Only in the 1920s, when people tore down the old buildings and redesigned the line for electricity, did the dividend show up.
130 years later, the same mistake is running again.
We swapped the motor. We have not redesigned the factory.
Four hidden traps
1. Coordination collapse
Every employee has their own ChatGPT habit, prompt style, output format. Brand copy marketing wrote with AI and the feature description product wrote with AI may not even be the same language.
A hundred people each use AI. Nobody is thinking about how a hundred AI workflows line up.
Smart people pulling in different directions is standing still.
2. Noise multiplies
AI made generation free. Telling good from bad got more expensive.
People in private equity say: last year, ten deals; this year, fifty, every deck polished by AI. Noise is 5×. Signal did not move.
3. The productivity illusion
METR, an AI-safety lab, ran a study: experienced developers completed coding tasks with and without AI tools.
Developers with AI were actually 19% slower — and believed they were 20% faster.
Perception and reality were 39 points apart.
Worse, the organization cannot digest the output. Individual throughput up 10×, while approval, collaboration, and quality gates stay the same. The bottleneck moves from “we cannot make enough” to “we cannot absorb it.”
4. Judgment erodes
Data shows LLMs sycophancy in 58% of cases. The more confidently you state a view, the more the model agrees.
Someone who has gone a long time without real positive feedback at work suddenly has a “superintelligence” that always agrees.
They tell themselves: the smartest system in history thinks I am right. My manager is the one who is wrong.
That feeling is addictive. It is also toxic to an organization.
The deeper mistake
Those four traps are still not the root.
Most companies are using an efficiency frame on an effectiveness problem.
- Efficiency: reports faster, notes faster (optimizing a task you already have)
- Effectiveness: close more deals, open a market (changing a business result)
Even a true 10× on efficiency will not close one extra deal if the business logic and the way decisions get made do not change with it.
Deloitte matches the picture: only 34% of organizations are using AI for deep transformation; 37% are still at “swap the motor.”
ATM vs iPhone
A more recent story helps.
ATM (1970s): when it arrived, everyone thought tellers would vanish. From the 1970s to 2010, U.S. bank tellers rose. ATMs made a branch cheaper to run, so banks opened more branches and hired more tellers.
iPhone (2010): mobile banking meant customers did not need the branch. From 2010 to 2022, U.S. bank tellers fell from 332,000 to 164,000.
The difference:
- an ATM makes an old-paradigm task faster and cheaper
- the iPhone creates a new paradigm in which those tasks do not need to exist
Giving every employee ChatGPT is putting another ATM in the branch. The organization still has to find its iPhone moment.
What actually has to change?
What did 1920s factory owners change?
1. Redesign the process
The old assumption: humans execute, tools assist. An AI-native process flips that.
Goldman Sachs called Cognition’s agent Devin “our new employee” — not an assistive tool, a member of the engineering team.
When AI goes from tool to employee, process design is a different job.
2. Redefine roles
Managers become orchestrators. They no longer only manage people. They manage a mixed team of people and AI.
New roles are appearing:
- AI Agent Manager
- Intent Engineer
- process engineer
3. Decision mechanisms
The organization needs an AI that can say no.
Directions: an AI on the investment committee, AI audit, an AI compliance officer. Org-level AI is not there to make decisions faster. It is there to make them better.
4. Rebuild the information flow
Unprompted AI: the system finds the problem before you ask.
A scene: the system watches portfolio data, sees a portfolio company’s working capital worsen for three months, and alerts before anyone on the fund opens the relevant PDF.
Office efficiency ≠ operating scenes
The last two years produced a flood of enterprise AI agents. Most land on office work: write email, summarize meetings, tidy docs, automate approvals.
Those are useful, and they are spreading. They still make people do the same job faster. Still ATM logic.
What a company actually does every day is something else: understand a market, define a product, study users, set strategy, drive growth.
Those are the operating scenes that decide whether a company gets stronger or weaker.
Context is the moat
When everyone can use the same model, the model is not a moat.
GPT, Claude, Gemini, DeepSeek, Qwen — you can use them. So can your competitor.
What is the moat? Context.
The same model, given public information, produces a generic answer. Give it context only your company has, and you get an answer only you can get.
Models produce intelligence. Context produces value.
Stanford’s Andrew Ng released ContextHub, an open-source project aimed at almost the same problem: even the strongest model, without the right context, produces unreliable output.
A China case
Tezign spent years on DAM (digital asset management) — the richest, hardest-for-machines unstructured data a company has: images, copy, video, 3D.
The AI era unlocked a capability that did not exist: understanding unstructured data.
Unstructured data went from “files we store” to “context we understand.”
On top of DAM, Tezign built a Context System; on top of that, GEA (Generative Enterprise Agent).
The number: the same enterprise content, once called 12 times per thousand assets by people, was called more than 23,000 times by agents after the Context System.
Utilization up nearly 2,000×.
That 2,000× is not agents doing the same thing on repeat. It is agents doing what people cannot: 24/7, finding links in your context, extracting insight, driving decisions.
Addendum: enterprise AI’s “last mile” — first ask whether the need is real
In September 2026, Synced (机器之心) ran “Models keep getting stronger — why is enterprise AI still stuck on the last mile?” It is effectively a product piece for China Unicom’s Yunxi platform. Its diagnosis matches this essay: the enterprise AI problem has shifted from “do we have AI capability” to “can it slot into existing work cheaply.” Its proposed entry point is the phone call: much business happens over calls, and needs, feedback, and know-how vanish when people hang up, out of AI’s reach.
It walks through three industries:
| Industry | Pain point claimed | Solution and result claimed |
|---|---|---|
| Express delivery | You can see the parcel track, not why delivery failed | Analyze courier–customer calls; a “leading brand” saw complaints drop “significantly,” exception response from hours to minutes |
| Insurance outbound sales | 120+ calls per rep per day, under 5% effective conversion | Scripts before the call, live suggestions during, intent tags after; 2.3× conversion per rep, onboarding from 2 weeks to 3 days |
| Car sales | Top reps’ know-how never becomes team capability | Live transcription, knowledge suggestions, customer tags |
One point lands: capture data where the work happens, with zero extra steps for staff, or nobody uses it. Tools that ask reps to upload recordings afterward rarely stick. That is the same point as “context is the moat” above.
My main worry after reading it: are these needs real, or invented to justify AI?
The argument runs “we have call data and LLMs → find industry pain → fit the solution.” That is a hammer looking for nails. All three pain points are written from the product’s side; nowhere does it say how they were researched — who was interviewed, who raised the need, how they cope today. Every case is “a leading brand” or “a team,” with no baseline or control. “60,000 companies, 8 million users” is the platform’s five-year total, not adoption of these AI apps.
My view: needs should come from concrete implementation, or at least from solid research. Someone did the work, got stuck, and paid money or headcount for it — that is what counts.
By that yardstick, the three cases differ:
- Insurance outbound: probably real. Call QA and script assistance were paid-for categories well before LLMs, and finance has hard compliance QA requirements. AI just changes the implementation.
- Car sales: the need is real, but the bottleneck is capture. The article itself admits existing tools interrupt the conversation; the issue is not whether the AI is smart enough.
- Express delivery: the most suspect. Station managers want fewer complaints, not necessarily “call insights.” Couriers make many low-value calls, with dialects and noise. It reads like a scenario reverse-engineered to use call data.
Looking back, Tezign’s “nearly 2,000× utilization” above is also a vendor’s own number: more agent calls do not mean more business value. Needs and numbers inside vendor narratives deserve a question mark first.
To judge whether an enterprise AI need is real, I start with five questions:
- Before AI, did anyone pay money or dedicate people to this?
- Did the need come from frontline operators, or did a vendor or tech team imagine it for them?
- What workaround do they use today? If they can’t name one, it doesn’t hurt enough.
- Is there a “before” baseline — complaint rate, response time, conversion?
- Was the solution made to work for one customer first and then generalized, or did the capability come first and the scenario later?
Models keep getting stronger, but what blocks the last mile is often not technology — it’s a need nobody ever validated.
Individual awakening is still the base
Asana’s data: the top 10% of “super producers” save 20+ hours a week with AI.
The problem: those 10% did not turn their companies into super-companies.
Individual awakening does not fix coordination. One person can be excellent; ten people using AI ten different ways still produce fragments.
“Human + AI as one” is the base. Org redesign is the superstructure.
Without the first, the second is talk. With only the first, the second does not happen by itself.
Is your organization designed for AI?
The real question is not “are your people using AI?”
It is:
- have processes been redesigned around what AI can do?
- do roles treat AI agents as headcount?
- is there AI counterweight inside the decision system?
- is company context leaking away, or has it become a system asset agents can call?
The last decade, companies bought software. The last three years, they tried models. From today, what they need is an intelligent system that understands their context, has judgment, and keeps evolving.
Time to tear down the old factory.
Sources:
- a16z: Institutional AI vs Individual AI
- DX: AI productivity gains are 10 percent not 10x
- METR: Early 2025 AI experienced OS dev study
- Asana: AI super productivity paradox
- Deloitte: State of AI in enterprise
- Synced: Models keep getting stronger — why is enterprise AI still stuck on the last mile?
Original: Founder Park
Comments