Perhaps the best software is the one the client never sees.

I spent years trying to organize origination. I am starting to suspect the advantage is not in selling the tool, but in using it to deliver the service better than everyone else.

On Tuesday, Sequoia announced an investment in a company called Crosby. The description looks contrary to Silicon Valley habits: it is not merely a legal software company. It is a law firm that put engineers and artificial intelligence inside its own operation. The client does not buy a license to negotiate contracts better. He hands over the contract and receives the work done.

The difference looks small.

Economically, it can be enormous.

For decades, software was built to increase the productivity of some profession. The lawyer bought software and kept charging for the service. The accountant bought software and kept doing the accounting. The bank bought systems and kept selling financial products. The technology vendor kept the subscription; the professional kept the savings on the work.

Now another possibility appears. If technology can execute a large part of the process, perhaps the company that built the software should sell the result itself.

The question interests me because I spent the last few years building, almost without noticing, the same tension inside credit origination.

The more operations I do, the more evident it becomes that there are two jobs mixed into the same profession.

The first is rare.

Knowing the businessman. Building trust. Understanding the business. Sensing a capital need before it turns into urgency. Knowing which questions to ask. Deciding when a request deserves to go to market. Structuring. Negotiating. Choosing among bad alternatives and knowing how to say no when all of them are bad.

That work requires judgment.

The second job appears right afterward.

Asking for a document. Organizing the document. Finding out which version is correct. Building a data room. Recording debt. Researching lenders. Updating a contact. Forwarding material. Following up. Centralizing questions. Comparing term sheets. Controlling deadlines. Checking commissions. Updating the pipeline.

That work requires discipline.

We mixed the two because for a long time there was no alternative.

The professional who excelled at relationships also carried his own little administrative department in his head. It worked while the number of operations was small. Then his commercial success began to punish him.

More clients produced more back office.

More back office produced less time with clients.

Origination itself strangled the capacity to originate.

It is a bad architecture.

My initial reaction was the typical reaction of someone who likes technology: this has to become software.

I still think it does.

The question is for whom.

Perhaps for a while I made the classic mistake of confusing what can be a product with what should be a product.

A machine can be valuable precisely because it is not for sale.

Think of a good proprietary trading desk. Its risk system is part of the advantage. It does not necessarily make sense to turn it into SaaS for competitors.

A manager can develop excellent underwriting infrastructure. It may make far more money using it to buy assets than charging other managers a license.

An industrial company has proprietary processes to manufacture better. It does not have to turn the factory into a manufacturing course.

Why should an origination house act differently?

If I can build an engine that lets a small team organize more operations, see capital better, reduce rework, preserve history, and learn from every deal, perhaps the economically smarter application is to use it.

Not to sell it.

That conclusion became more uncomfortable as I started doing the math.

Software is seductive because recurring revenue looks clean. One client pays a subscription. Another comes in. The chart rises. Investors like it. Analysts can model it. The problem is that we need many clients paying relatively little to reach the economic value contained in a few large operations.

A R$ 1,000 monthly subscription produces R$ 12 thousand a year.

A single R$ 10 million operation at 1% of economics produces R$ 100 thousand of gross revenue.

I am not suggesting every operation pays 1%, nor that every operation closes. That would be childish. There are splits, costs, taxes, negotiations, different products and structures. Some operations die after weeks of work and pay precisely zero.

Even so, the order of magnitude deserves attention.

If the technology exists to make origination more productive, perhaps selling the technology means selling by subscription what I could use to participate in the financial result it produces.

The pickaxe may be excellent.

It is advisable to check the mine first.

There is also a strategic problem.

The better the software, the more capacity I hand to the user. If the user is a competing originator, I may be selling him exactly the infrastructure that could differentiate my house.

Perhaps that makes sense in a gigantic market where the capture by software exceeds the capture by the service.

I am not sure that is our case.

The private credit market is one where fees are charged on volumes far larger than the budget for the tools used to produce those volumes.

It is the old difference between selling equipment and participating in production.

The work is worth more than the tool.

That makes me look at Sequoia and Crosby in a less technological and more financial way.

What interests me is not that they used artificial intelligence.

It is that they chose the richer economic layer.

The client needs a contract negotiated. Crosby sells a negotiated contract.

It does not hand over a brilliant editor and wish you luck.

That formulation looks obvious after someone executes it.

Perhaps credit origination needs something similar.

The businessman does not want credit software.

He does not want a CRM.

He does not want a list of lenders.

Most of the time he does not even want credit in the abstract sense.

He wants to buy the competing company without decapitalizing the group.

He wants to finance an expansion.

To lengthen a debt that became short.

To buy a machine.

To release collateral.

To take advantage of a real estate opportunity.

To get through the off-season.

To reorganize the balance sheet.

Credit is merely one of the instruments used to get there.

If I sell software to the originator, I stay one layer away from that decision.

If I build the capital house itself, I sit next to the businessman.

That proximity is worth more.

It produces more information.

It produces more trust.

It produces more future operations.

And, perhaps above all, it produces a memory no external software company can obtain with the same depth.

When an operation passes through our house, we learn what a system sold to third parties rarely sees in full.

Why the client sought capital.

How much he asked for initially.

How much he actually needed.

What collateral he offered.

Which collateral we preserved.

Which institutions refused.

Why.

Which ones made a proposal.

What the capital cost.

How the client negotiated.

How long it took.

Which documentation created friction.

How the operation ended.

What happened afterward.

That is much more than a CRM.

It is a proprietary economic dataset produced by real work.

On the other side of the table, we also learn.

Which lender actually executes.

Which institution talks about a given sector but does not do it.

Which fund has appetite for a specific risk.

Whose declared minimum ticket differs from the minimum ticket in practice.

Which team answers quickly.

Which structure tends to reach committee.

Which collateral changes the pricing.

Who walks away when the process gets a little difficult.

That knowledge compounds.

Every operation leaves the next one less ignorant.

That may be one of the most important advantages of doing the service instead of merely selling the tool.

Software can record.

The work produces the data.

There is a difference.

And the capital market is becoming even more interesting for whoever can build that memory. In February, JPMorgan announced up to US$ 50 billion of commitment from its own balance sheet to direct lending, on top of partner capital. In the investor presentation, it describes its advantage in fairly clear terms: deep relationships with borrowers on one side and investors on the other, with the bank in the middle offering origination and structure.

Again the same division.

Capital looks for origination.

Origination looks for capital.

The institution that controls the relationship between the two controls an economically valuable part of the chain.

That point started to weigh on my decision.

If the greatest asset of our future house is a proprietary network of companies on one side and capital on the other, opening the engine as a platform to any user may destroy part of the scarcity we are building.

Not because information should be artificially hidden.

But because a capital relationship is more than a name in a directory.

There is trust.

History.

Confidentiality.

Behavior.

An institution takes my call because it knows I will not send just anything.

A businessman hands over documents because he knows the information will not circulate indiscriminately.

That balance loses value in an open marketplace.

Platform logic generally wants to maximize participants.

The logic of a good boutique may want exactly the opposite.

Few good clients.

Few good operations.

Suitable counterparties.

A rigorous process.

That may be the rare market in which well-used exclusivity is an operational characteristic and not a marketing device.

I do not mean the client has to beg to be served.

I mean the house has to be able to refuse an operation.

That capacity is essential.

An institution that has to close everything cannot advise anyone.

If the revenue depends only on placing debt, the incentive inevitably starts deforming the analysis. Every problem becomes financeable because every negative answer destroys a fee.

That is dangerous.

I want the engine to help us increase productivity without forcing us to increase volume indiscriminately.

Better to select ten excellent operations than to administer a hundred requests without conviction.

Technology should make the filter stronger.

Not make the garbage faster.

That principle looks especially important now that I have started looking at international capital more seriously.

In 2024 I concluded that capital went global while origination stayed local. A year later, the operational conclusion is less romantic: internationalizing an operation adds enough work to destroy the economics of a badly organized boutique.

Knowing a fund is not enough.

We have to understand ticket, currency, sector, jurisdiction, documentation, compliance, structure, and who actually decides. The Brazilian company has to arrive organized enough for someone who does not know the local context.

All of that could require an entire team.

Or a better engine.

This is where proprietary software starts changing the nature of the service.

Imagine a small team able to take on a company, organize financial data, classify documents, build a data room, maintain a history of lender appetite, compare structures, and follow the workflow with far less manual work.

The client does not have to pay less for that.

He is paying for the result.

The difference appears in the house's margin and speed.

That is the point software investors may have understood before traditional providers: automation does not necessarily have to reduce price. It can increase the productivity of whoever captures the existing price.

A conventional boutique grows by adding analysts.

An AI-native boutique can try to grow by adding automation before adding people.

I do not eliminate the senior professional.

I give him operational leverage.

That strikes me as far more interesting than creating a financial chatbot and calling it artificial intelligence.

The most important parts of origination still require judgment.

It is quite unlikely that I would want to delegate entirely to a machine the decision about which lender should receive a delicate operation from a family with whom I built a relationship over years.

But there is no romantic reason for a senior professional to lose fifteen minutes looking for the right version of a balance sheet.

The right frontier is between repeatable intelligence and responsible judgment.

Automate what is repeatable.

Use human beings where the consequence demands interpretation.

The better the engine, the more time should be left for relationships, strategy, and negotiation.

That is the metric.

Not the number of automated tasks.

Hours returned to valuable work.

Perhaps, in the future, much of the intelligence that is human today can also be incorporated into the system. That will depend on data, quality, and responsibility. I do not have to solve that question now.

There is enough banal work to remove first.

And every time we remove banal work, something interesting happens: the service starts acquiring software economics without ceasing to be a service.

That combination can be extremely powerful.

Revenue based on the value of the operation.

Marginal cost reduced by technology.

Proprietary data produced by execution.

A direct relationship with the client.

A capital network that becomes smarter with every deal.

That looks far harder to copy than a SaaS with similar features.

The code can be copied.

The history cannot.

The relationship cannot.

The reasons a hundred operations were refused do not appear in a public API.

Neither does a company's behavior during a negotiation.

That information is born of the work.

That is precisely why I am starting to think the engine should stay closed.

Not forever as a matter of principle.

Only while it holds more value as an operational advantage than as a product sold.

Optionality matters.

We can open parts of it in the future.

We can offer technology to specific partners.

We may find that a certain layer deserves to become a platform.

Nothing obliges us now.

The mistake would be choosing the business model too early merely because SaaS has a more modern vocabulary.

The profession already exists.

The demand already exists.

The budget already exists.

Companies already pay to access and structure capital.

Perhaps the opportunity is not convincing the market to buy a new software category.

Perhaps it is delivering an old service with a new cost structure.

That is less revolutionary in PowerPoint.

Economically, it can be much better.

It also fits curiously naturally with what I have been looking for since 2009.

When I heard that nothing makes more money than selling money, I imagined the secret lay in the merchandise.

Then I discovered the merchandise is not the center.

The value appears in the capacity to find, select, structure, and coordinate.

I spent years working as a loan broker without having to own the money that passed through the operation.

Now the question is another.

If I managed to develop a machine to do that work better, why would I sell the machine instead of using it?

Perhaps I am finally reaching the point where technology stops being the company and becomes the company's competitive advantage.

There are old precedents for that.

The most important systems of some financial institutions were never a product.

They were simply the reason those institutions operated better than their competitors.

Software only looks obligatorily like a separate category because we spent twenty years being taught to turn everything into SaaS.

Not every engine has to appear in the shop window.

Sometimes the most valuable thing in a house is precisely what the client does not see.

A good restaurant does not sell its kitchen.

It sells dinner.

A good asset manager does not sell its investment process.

It sells management.

Perhaps a good financial house should not sell its origination system.

It should sell well-organized capital.

That hypothesis is becoming hard to abandon.

And it changes the kind of institution I imagine building.

Not a platform originators enter to administer their deals.

A boutique companies enter to solve capital problems.

The technology stays behind.

It knows the history.

It organizes the operation.

It maps the market.

It learns.

The businessman meets few people.

Those people have a much larger machine behind them.

That can produce a rare combination: the experience of a boutique and the productivity of software.

It is the first time I can clearly see how I want the two to meet.

Perhaps the definition is simple.

Service on the outside.

Software on the inside.

Human judgment in front.

Automation behind.

And one obsession: making a small team able to operate capital like a much larger institution, without acquiring the slowness that usually accompanies much larger institutions.

If we manage that, perhaps the real product is never the software.

Perhaps it is Seferu itself.

I have not yet decided to use that name publicly for this thesis.

But I have begun to understand what I am building.

And, for the first time, I think selling the engine may be the least interesting way to monetize it.

Leo Bentier