History taught us that Superintelligence arrives one domain at a time. It starts with the boring and bounded domains first. Later no one can pinpoint exactly when it happened.
Here by “superintelligence,” I mean what Stockfish is to chess. Being decisively better than the best human who has ever lived within one bounded domain. We have seen superintelligence before and several times already. This article is about what someone would have to own to take the next domain.
It was 1997 in a conference room in New York. One of the greatest chess players of all time, Garry Kasparov, had lost to a machine. That loss was easy to dismiss as a trick of brute force.
20 years later, AlphaZero taught itself the game just from the rules. It destroyed the strongest hand-written engine on Earth after playing itself millions of times. Today the best chess player is not a person. It lives on your iPhone as an ordinary consumer software.
Code had its Tuesday in San Francisco around 2023. Not long ago engineers were publicly arguing about whether AI could write software. Suddenly a generation of them couldn’t remember how they’d worked without it.
You can watch it happen inside @cursor_ai today. The machine writes based on human intent. Then the human approves.
The intelligence is enormous and the ceremony is almost nothing.
Two unusual properties
Chess and code shared two unusual properties. A clean + abundant reward signal and someone standing in the loop to collect it.
A reward signal gives a machine-readable answer to a simple question: Did that work?
In chess the result is available millions of times over as a form of win, loss or draw. Similarly we have unit testing for coding. We can build system that will try something, continuously check if it worked or not and keep adjusting at scale. At that moment, it can learn from far more attempts than any human possibly can.
This is what “The Bitter Lesson” by Rich Sutton talks about. A general method that can scale with computation will beat systems built from hand-coded human expertise. The two that scale arbitrarily are search and learning. Unfortunately learning only pays when the right data exists. For reinforcement learning the right data is a record of what the system tried and whether it worked.
Sutton wouldn’t extend the lesson the way I’m going to do it here. He doesn’t think LLMs are following it at all. They learn from human text rather than from their own experience. What he wants are agents that never stop learning from the latter. I’m not settling for that. I’m making a narrow point about where reward signals come from.
DeepSeek-R1 got better at math thanks to reinforcement learning on verifiable rewards. The model can think for itself to get its answers checked against ground truth. We can make progress by finding another domain with checkable answers.
So for the next valuable domain we need to have (a) enough reward signals and (b) an embedded system to collect it.
She was the best rep on the team
She carried around fifty accounts every quarter. Her memory exceeded far beyond the fields in the CRM. She knew which champion was secretly looking for new jobs and the exact phrasing that she could use to unblock procurement freezes. It took her five years to learn all of these, and it lived entirely in her head.
Then one day she decided to take a job somewhere else. Every pattern and hard-won instinct walked out the door with her on her last day of working.
The company kept the system of record and lost the knowledge.
Most importantly the company lost the human judgment of their best team member. If we type down the sequence of events from this, it would look something like:
- Everything valuable she knew was a pattern learned from years of trial / errors and real world outcomes.
- It was never written down or shared so it evaporated the moment she left.
- Multiply her by every best employee of yours who has quit across every company for 20 years.
- The result is an enormous pool of reward data that companies failed to capture.
Reward signal for revenue is known to be harder to isolate.
Reward signals for revenue knowledge is harder to isolate
Closed-won itself is a terrible training signal for any system of records. That label arrives at the end of a six-month cycle. Sixty touches from five people may have contributed. There is no clean way to isolate the one that mattered. Pricing / ICP, and market conditions can change while the deal remains open. You also see only the action someone took but not the email they could have sent instead.
Chess and code are great to work with because a system can run a billion attempts in a box for free. AlphaZero learned through self-play without studying a human game. A sales cycle cannot be self-played. A synthetic buyer’s “yes” carries no revenue. A simulated pipeline cannot produce a real reward.
What revenue / a lot of knowledge work does have is a second source of feedback (The most important one.)
Human Approval
When an agent proposes an email / next step, a person approves or edits it. That decision is a “preference label.”
It was human preference that helped turn rambling GPT-3 into something we wanted to talk to for hours. RLHF collects preferences between outputs and trains on them. Every approval in a revenue workflow can create one of those labels on actual knowledge work with context and evidence.
Approval carries its own bias too. A busy rep may rubber-stamp a draft, while a picky rep may reject one over a comma. The approval answers, “Would I sign my name to this?” but it does not tell you whether the buyer will respond. A system trained only for approval will produce flattering, safe emails that reps accept and buyers ignore. The field calls this reward hacking. It’s a common failure mode.
Approval and real world outcomes correct different errors. Approval is immediate but it is a proxy. Outcome arrives later and reflects what happened during the deal. Approval teaches the system what a great rep would send. Reply and stage advancement show which approved actions worked. So training against both signals gives the system a better chance to improve.
An edit here contains more information than an approval. A yes says only that the draft was good enough. When a rep changes five words, then cuts the second paragraph to replace it with something human, and adds the detail about the CFO’s board meeting. The diff shows what was wrong and what the rep preferred for that account.
Cursor captures this richer signal. Alongside GitHub’s stored code, Cursor observed the code being written (accepts, rejects, and what the human typed instead.) A useful learning loop for AI should record a professional correcting a proposal in real time.
GitHub stored the code. Cursor stood in the room while it got written. Salesforce stores the deal, and nobody is standing in the room.
You Can’t Scrape Human Judgments
Large models were built by consuming trillions of words from the internet. That means finished answers or finished code. Models increasingly know the same things as they are trained on the same public record. Pretraining has consumed much of what humanity published.
The internet rarely recorded the decisions behind its outputs. A great knowledge worker might look at a stalling deal and decide (based on something she cannot name) to skip the polite follow-up and call the CFO directly. The activities may get tracked in the system of record while the judgment that selected it happens in her head and disappears.
The next frontier is judgment captured at the point of decision with the outcome attached.
We have every research paper Einstein published and no recording of how his mind chose those moves over the thousand he discarded. Humanity kept the output and lost the selection process. For most of history, judgment moved through apprenticeship. You stood beside the master, watched thousands of decisions, and slowly absorbed what they could not explain. That process did not scale because it lived in behavior rather than text.
An outcome loop turns that apprenticeship into a repeatable system. Every time an expert knowledge worker approves or rewrites an agent’s proposed move, the system captures a judgment on an ideal action and connects it to the result. Pretraining learns from outputs, while post-training can incorporate judgments and outcomes. Those judgments were absent from the public internet. So the only way a company can capture them has to be inside the work itself, while it’s happening real-time.
“But doesn’t Salesforce already have the data?”
There are two ways to know a football game. You can receive the final score, 24 to 17, or watch every play. You can analyze the coverage that broke, and the moment the game completely turned.
Salesforce gives you the score (Or any system of records). Its record is a static and human-typed log of what happened. This is often entered late by someone with little incentive to document it well. Salesforce did not take the action, so naturally it cannot identify what caused the outcome. Salesforce knows a deal closed but cannot tell
- which of forty touches mattered,
- in what order, or why.
The incumbent threat comes from companies close to the loop. Gong has the meeting recordings. Google and Microsoft own the mailbox and calendar where much of the signal originates. Salesforce and HubSpot are shipping agents through their established distribution.
The race depends on where the human approves the work every day. A system that observes the conversation collects transcripts. We need a system that proposes the next move and gets judged and collects preferences. Those preferences become training data because the system participated in the work instead of storing a record underneath it.
The product reps can collect a richer signal than the database that stores the resulting email.
The model was never the moat
Foundation models are commoditizing, and recent releases make the trend visible. Moonshot released the full weights for Kimi K3, a 2.8-trillion-parameter model with a million-token context window. It placed first in the Frontend Code Arena while costing less than closed frontier models. Today it can run on your own hardware as a download.
Thinking Machines also released Inkling just a couple days ago, and said its value lies in being a good base to specialize. The company is betting that models adapted to an organization will beat one-size-fits-all systems on that organization’s work.
Capable intelligence is becoming cheap to download. A company’s edge will come from the proprietary signal used to specialize it.
In a world of a million models
When capable models are abundant and free to download. It’s only natural that companies will ship fleets. A shared open base plus a lightweight adapter per tenant, maybe per segment, workflow, or knowledge worker. Ten thousand customers could mean ten thousand models, each shaped by the approvals and corrections of the people it serves. The product becomes a factory that keeps producing specialized models.
The checkpoint loses value quickly. A company may use K3 today, Qwen tomorrow, and something better in the autumn. Every base-model release erodes the value of a fine-tune tied too closely to one checkpoint.
Three assets survive rebasing: logged trajectories, a reward model trained on domain preferences, and an evaluation suite that can identify the best candidate for a specific job. Those assets transfer to future base models.
Capturing the reward signal requires someone to be present where the judgment happens.
Permission is another constraint. The system needs access to the mailbox, calendar, and pipeline data, with standing to propose real moves involving real money. A startup earns that permission slowly.
Evaluation also becomes scarce when a company must distinguish among a million available checkpoints inside one domain. Owning that evaluation gives the company power over the fine-tuning market that depends on it.
Most AI discourse focuses on the reward signal. Permission and evaluation within the workflow create more durable positions for enterprises.
What compounds, and what doesn’t
A flywheel claim is easy to make and easy to puncture.
“You can’t pool it.”Customer contracts can forbid training on one company’s content for another company’s benefit, and they should. Content stays inside the customer boundary. What can cross that boundary is policy. Abstract state-to-action patterns, action schemas, and reward-model weights learned from preference structure rather than prose.
Your emails stay yours. A general lesson about how to respond to a nineteen-day silence after a pricing conversation belongs to the domain.
“Environments-as-a-service will commoditize this.” There’s now an industry selling exactly that: simulated enterprise sandboxes, synthetic trajectories, reward signals on demand, sold to labs and Fortune 100s alike. But a simulated buyer can’t approve anything. No money attaches to a synthetic stall, no career rides on a fabricated forecast, and no rep stakes a reputation on the decision. Simulations cannot manufacture approvals on real accounts with real revenue attached.
“Workflow lock-in is the real moat.” Workflow lock-in matters. The surface where people approve work becomes sticky for ordinary operational reasons before training data adds value. A company must win that position through product quality and distribution. The data begins compounding after the workflow earns adoption.
How to Build Superintelligence
The order matters in revenue and in any other domain.
Start by instrumenting the workflow. Log the full tuple: account state, candidate actions, chosen action, the human verdict, the reason for rejection, the edit diff, the two-week outcome, and the eventual result. Teams that skip this step can spend a year building models on incomplete records, with no way to recover the missing history.
Once logging is in place, build the evaluation before the model. Create a frozen set of real situations with expert-preferred actions and score it consistently. Otherwise, reinforcement learning can exploit a hole in the reward function and hide the failure for a quarter.
Supervised fine-tuning on accepted and corrected actions comes next. It is cheap and buys much of the early perceived quality, including voice, format, and house style. An open base plus a fine-tuning platform works well here, and specialization pays early.
Those collected preferences can then train a reward model. For the same account state, approved actions beat rejected ones, and shipped edits beat original drafts. Once the model can score candidates offline, it can rank them at inference without changing the policy. Generate five drafts and ship the best one.
Reinforcement learning comes after the reward model, using fast, verifiable outcomes such as replies, reopened conversations, booked meetings, and stage advancement. Closed-won remains a calibration signal rather than the daily reward.
Better open models will arrive, so the system needs a regular rebasing schedule.
Onboarding 1,000+ companies
There is always a department where team members have more accounts that they can realistically cover. They have time to work, maybe a third of them. You had the deal that went dark last Tuesday needed a nudge, but you were in back-to-backs all day, and by the time you looked up it was Thursday. The lead you meant to enrich is still cold. You update the CRM on Sunday night and only because you are forced to do so. It’s a physics problem. There’s one of you and there’s too much.
Every one of those accounts gets an agent on it, watching every day. The follow-up that’s four days late is drafted and waiting for your yes before you’ve finished your coffee. The deals rescue themselves. You never login to Salesforce but it becomes more accurate than it ever was. The Sunday-night data entry is gone.
Buyers have gotten very good at spotting machine-written outreach, mailbox providers have gotten very good at filing it, and consent rules mean volume without judgment is a liability rather than a strategy. The era where more sends meant more pipeline is over. It ended because everyone got the same cheap tools at once.
Superintelligence is the thing you’d sign your name to. It has read every email in the account, every call, every deal your company has won or lost. It writes the follow-up that you’d have written yourself spending hours. That’s the bar we need. It should write the email well enough that you read it and think “yeah, that’s right,” and press yes.
Everything is built around that “YES”.
The agent does the work and you keep the judgment.
It never sends without you. It never closes the deal, you do. It never reads the room on the call, you do. It deletes the manual part of the job you hate so you can spend your hours on the part only a human can do.
Every time you press yes, or fix a line before you do, it learns your instincts. After a few dozen corrections it writes like you.
You stop being one rep with sixty accounts and become one rep with sixty agents, and you’re the one orchestrating them.
The best engineers became the orchestrators of agents and they didn’t lose their jobs to Cursor. Now they ship ten times what they used to. The reps who will win the next five years are the same story.
Stop opening Salesforce
The mentioned paradox here is that, you will stop opening Salesforce and Salesforce becomes more accurate than it has ever been. Both happen because humans were never the right tool for keeping the record. Take them off that job and the record fixes itself.
That changes three numbers a revenue leader already manages.
Ramp and turnover. The rep who quit took five years of institutional knowledge with her, and her replacement spent much of a year relearning what the company once knew. Companies pay for that loss and call it ramp.
Capturing judgments turns institutional knowledge into infrastructure. The system retains what previous reps taught it, and new hires inherit that memory. Ramp can focus more on learning the product and less on reconstructing the accounts. Turnover stops resetting all of that knowledge to zero.
Coverage. Your best rep works twenty accounts well while the other forty decay. Across a team, much of the pipeline you already paid to generate receives the neglected version of your company.
An agent on every account can work the existing pipeline every day. Better coverage plugs leaks in the pipeline you have already paid for, which may be cheaper than generating new opportunities.
Headcount. A mid-market rep costs $150,000 to $250,000 fully loaded and spends a fifth to a third of the week on data entry, follow-up hygiene, enrichment, and chasing. For a twenty-rep team, that can put roughly a million dollars of annual payroll into work a custom agent can handle.
Adding a rep adds more administrative work, more ramp, and more knowledge that may leave later. Agents can grow output without adding the same payroll burden, which lets buyers compare their cost with labor rather than a conventional software budget.
The team still keeps the control. Agents stage the work and hold judgment calls for human approval. A serious deployment needs enough context to reason over the whole relationship and the discipline to ask permission. Across the team, that can finally give the leader a CRM number worth trusting.
Throughout history we’ve lost the judgment of our best people the moment they leave, Octolane is where that stops.
We’re building superintelligence at Octolane. An agent on every account that finds the next move, does it with your approval, learns from your corrections, and gets measured on whether the deal moved or not. If you run revenue on Salesforce or HubSpot, you can connect read-only in ten minutes, with no migration requirement, and see what it finds in your pipeline.
Further reading
The technical grounding, if you want to check the claims yourself.