Last week, I visited a company that specializes in logistics systems.@OpenLedger
The office was pretty outdated.
The AC was ridiculously cold.
I was sitting there waiting for their tech lead to join the meeting, and the two ops guys next to me were quietly bickering.
It wasn't about the project.
They were arguing over a form.
It was only later that I got the full picture.
They've been tidying up their historical knowledge base internally recently.
Some people think:
"Don't keep all those manual correction records from before."
On the other hand, one person doesn't want to delete anything.
Because that's the experience their team has gradually built over the past few years.
The whole scene felt really off at the time.
On one hand, they're afraid of the hassle.
On one hand, it's hard to let go.
I casually asked a question while having dinner:
"Why are you suddenly so concerned about historical data now?"
The person in charge put down his chopsticks and said something I still remember.
He said:
"Data was just a record before, now data might become evidence."
I was thinking about this sentence all the way home the other day.
Because of the past many years.
Internally, enterprises rarely manage 'experience' seriously.
Especially those:
Customer service remarks.
Abnormal order processing.
Manual corrections.
Temporary patches.
Manually circumventing processes.
These things were previously quite casual.
Even when many old employees leave.
The company won't feel like they lost anything.
Because the default logic has always been:
Experience belongs to people.
Not an asset.
But after AI starts entering enterprises on a large scale.
This logic is slowly changing.
Especially recently many companies started to feed historical data back to Agents, and the issues suddenly got complex.
Because the model won't differentiate:
Which are standard processes.
Which were just temporary fixes by a certain employee back in the day.
The result is.
Many 'gray experiences' that were previously unnoticed are now starting to be amplified.
And the most troublesome part is:
Once these things are learned by the model.
Who is ultimately responsible?
A few years ago when I was working on a supply chain system, I saw a scene that was quite similar.
At that time, many companies had just started automating approvals.
At first, everyone thought:
Efficiency will be higher once the system is online.
But later the biggest problem wasn't the system.
But rather:
No one dares to sign anymore.
Because many things used to rely on implicit understanding.
Once the system records.
Many gray operations immediately become 'traceable.'
Later many companies began to see particularly absurd phenomena.
Everyone would rather the processes slow down.
They also don't want to leave too complete records.
I've been feeling this more and more recently.
The AI industry is starting to have that flavor now.
Especially this set of attribution and behavior recording logic that has been pushed forward recently, makes me easily associate it with internal audits in enterprises.
Many people now understand attribution.
Still stuck at:
"Who contributes data, who gets the profit."
But in the real corporate world.
Another thing will happen sooner.
That is:
Will the data left behind become a liability in the future?
This thing is actually quite sensitive.
Because in the past internet era, the data management logic of most enterprises was very rough.
To put it bluntly:
As long as it runs, it's fine.
Especially many legacy business systems.
It's filled with:
Temporary fields.
Manual remarks.
Interpersonal operations.
Circumventing rules.
Everyone used to not care.
Because no one thinks this data will be 'understood' again in the future.
But AI is different.
AI will restructure historical behaviors.
Even re-inferring.
This leads to a particularly real issue.
In the past many:
"Defaults exist but no one investigates" type of things.
In the future, they might all be re-examined.
I've recently started to hear some subtle changes.
Some companies are starting:
Limit export of historical knowledge base.
Cleaning up old customer service records.
Delete abnormal remarks.
Individual review of training data.
On the surface, it's called data governance.
But many times.
It's actually about 'doing subtraction' in advance.
Because everyone is starting to realize:
AI will no longer just be a production tool.
It might still be:
Long-term memory systems.
The most interesting thing here is.
Many enterprises used to have the biggest advantages.
It's precisely that:
"Flexible."
Processes can be flexible.
Clients can be treated specially.
The system can be manually fixed.
But once AI starts learning these things.
Organizations will discover for the first time:
It turns out in my own experience system, there are so many unclear areas.
I've recently felt it quite clearly.
Many companies have already started entering a new contradiction.
On one hand, hoping AI understands the business better.
On the other hand, they fear AI knowing too many business details.
Because the business is more real.
The more gray areas.
And once the gray area is recorded long-term.
Organizations will definitely start to get anxious.
So I'm increasingly skeptical now.
In the future, many enterprises might see a new position rarely seen before.
Not an AI trainer.
Also not a data analyst.
But rather:
"Data internal audit."
Specifically responsible for:
Which data can feed the model.
What experiences can't be kept?
Which historical records need to be deleted.
Which behaviors cannot be learned long-term.
Sounds exaggerated.
But many industries have already shown signs.
Especially in industries like customer service, logistics, supply chain, and finance where the processes are particularly complex.
Because the truly effective experiences in these industries.
Many times it wasn't completely compliant to begin with.
In the past, no one dug deep.
Because experience only exists in human brains.
But AI is slowly fixing these 'grays in the human brain.'
And once it's fixed.
Many organizational relationships will change.
The most valuable places for many old employees.
It's actually:
Knowing when to handle things 'outside the rules.'
But in the future, companies might become increasingly afraid of having these things replicated by models.
Because once scaled.
Risks will also be scaled.
This is also where I've been rethinking OpenLedger more recently.
On the surface, it appears to be solving:
Attribution.
Profits.
Data value.$OPEN
But on another level.
It's actually forcing enterprises to face a question seriously for the first time:
Which experiences,
What should be permanently preserved.
Many people now believe the next phase of the AI industry will involve model competition.
I actually feel more and more.
The truly explosive ones.
It could be:
Enterprises are beginning to redefine 'what data is worth keeping.'
Because of the past internet era.
The default logic has always been:
The more data, the better.
But in the AI era, we might see for the first time:
Some data,
Organizations will wish it had never existed.
This thing will become increasingly complex.
Especially once attribution paths, behavior records, and long-term training start to mature.
New boundaries will definitely emerge within enterprises.
What was most feared before was data loss.
In the future, it might start to become:
Data is retained too completely.