Software is solved. SaaS is dead. What is left?

Two claims

The first claim is that software is solved. The newest AI models can take an ambitious description and build working software from it, with little help. One widely followed reviewer put it this way after putting the latest one to work:

Matthew Berman, who tests these models for a living, is careful to add that he does not mean every application is perfect. He means the ability to build is no longer scarce. If a capable team can produce in weeks what used to take years, then code by itself stops being a reason to choose one vendor over another.

The second claim is that SaaS is dead. It is usually pinned on Satya Nadella, who never said those words. What he did say, in a December 2024 conversation, was more precise and more interesting. Talking about business applications, he said:

His point was that the logic in those applications is moving up, into agents. A year later he described Microsoft’s own future the same way: a tools business turning into “an infrastructure business in support of agents doing work”. Andrej Karpathy has made the same observation from the builder’s side: software used to have two kinds of user, people working through screens and programs working through APIs, and agents are now a third.

What actually dies

Put the two claims together and one thing is clearly in trouble: the screen as the product. For forty years we sold software by the seat. A person logged in, clicked and typed, and the work lived in whichever application they had open. A loan officer today touches about ten of those applications to close one loan.

An agent does not need any of them. It needs to know the facts and to be allowed to act. When it does need a person, to review a number or approve a step, it can draw the screen for that moment and throw it away afterward. The permanent, hand-built interface becomes the exception.

So yes, something is dying. It is the idea that the interface is where the value lives.

The end of the point solution

If both claims hold, the first casualty is the point solution: the product that does one job well and leaves you to connect it to everything else. I think most of them are going away, and I want to be precise about which.

A point solution used to be protected by how hard it was to build. That protection is gone. A texting tool, a document portal, a lead dashboard: each is now a few weeks of work for a capable team. That makes it a feature, and features end up inside whichever product already holds the loan.

The second problem is worse. Every separate tool is another place where part of the loan lives, and an agent cannot reason across ten places. Lenders put up with fragments while people did the stitching. They will stop paying for them once fragments are what hold their AI back.

The ones that survive own something gated: a credit bureau relationship, an underwriting engine, investor pricing. They do not disappear. They stop being screens your staff log in to and become services that the system of work calls. The vendor I would not want to be is the one in the middle, with a pleasant interface and nothing scarce behind it.

Cheap software cuts the other way too, and I should say so. There will be more small tools than ever, because anyone can make one. That is exactly why each of them will be worth less.

What does not die

Three things survive, and they get more valuable as the screens fade.

1. The record

   A complete, current, trustworthy account of the loan: the people, their goals, the application, the documents, the conditions and every conversation. An agent is only as good as what it can see.

2. The tools

   The actions an agent is able to take: pull credit, run underwriting, price the loan, request a document, send a disclosure, message the borrower. Each one has to work every time, against real systems.

3. The rules

   Who may do what, what must be redacted, what needs a person’s approval, and what gets written down. In a consumer app, a sloppy agent is an annoyance. In mortgage, there is a regulator, an investor and a family’s largest financial decision behind every action.

This is also why the obituary is premature. Jensen Huang has called the idea that AI will replace software tools “the most illogical thing in the world”, arguing that agents will use the tools that exist. When Klarna was widely reported to have swapped its software vendors for AI, its chief executive corrected the story himself: “we did not replace SaaS with an LLM”.

Nadella goes one step further than I would. He expects the back ends to become interchangeable once the logic moves into agents. In general business software he may be right. In a regulated industry I think the record is much harder to swap than that, because it carries the audit trail, the permissions and years of decisions an examiner can ask about.

What a prompt cannot get you

Here is the part the “software is solved” argument tends to skip. Anyone can now describe a loan application and have working software by the weekend. People call it vibe coding, and it is real. What nobody can describe into existence is access.

A mortgage runs on a handful of gated systems. A credit report needs a permissible purpose under federal law and an agreement with a bureau or a reseller. Automated underwriting means being let into Fannie Mae’s and Freddie Mac’s systems. Pricing engines are licensed lender by lender. The loan origination system opens through a partner program, on its terms. Each gate has a contract behind it, and usually a certification and a security review as well.

Software that cannot pass those gates is a demo. It can show a handsome application and a clever chat window, and it cannot pull credit, get a finding or lock a rate. The scarce thing is no longer the code. It is the trusted connections, and the work of tying them to one record, so that the credit pull, the underwriting run and the price all describe the same loan.

Access is one kind of scarce. Operating the thing is another. A texting screen takes an afternoon. Texting borrowers lawfully takes carrier registration, consent that is captured and honored, opt-outs enforced on every message, and each call, text and email filed against the right loan. An app takes a weekend. The same live loan on the loan officer’s phone, the borrower’s and the real estate agent’s, each seeing only what they should, does not. We have built both, and neither came quickly.

This is where being an Ascend company matters to us. Credit and verifications come from inside the family. Underwriting runs through Fannie Mae’s Desktop Underwriter today. The loan stays in step with Encompass. None of that came from a prompt.

Why mortgage is different

Mortgage is one of the hardest places to make this work, which is exactly why it matters here. The loan origination system keeps the loan compliant, and it should stay the system of record. Around it sit the CRM, the point of sale, the pricing engine, the texting tool, the document portal and a shared spreadsheet or two. The loan’s real story is scattered across all of them.

Point an AI at one of those tools and it sees one corner of the loan and has nowhere to act. That is why so much mortgage AI disappoints. The fix is not a smarter model. It is a single record, tools the model can call, and controls a compliance team can sign off on. In a regulated industry, that combination is hard to assemble and hard to copy. It is the moat.

It is also what we mean by a Mortgage OS. It sits on top of the LOS, holds the record of how the loan is actually worked, and exposes the tools and the rules to whoever is doing the work, human or agent.

How it arrives

I do not know how fast this happens. I am fairly confident about the order.

1. Now

   Screens, with an analyst beside you

   People still click through the loan. An AI analyst reads the whole file, answers questions with citations and drafts the next message. This is where most of the industry is, and where Luna AI works today.

2. Next

   Talk, don’t click

   A loan officer says “pull credit and price a 30-year fixed against a 7-year ARM” on a call or in a message. The agent does it, shows its work and waits for approval where approval is required. The screen appears only when a person needs to look.

3. Then

   Agents talk to agents

   A borrower’s own assistant, which already knows their finances and goals, asks three lenders for terms. It does not fill in three web forms. It sends one structured request, and each lender’s agent answers from its own record under its own rules.

4. Later

   Agents with a body

   The conversation about buying a home happens at the kitchen table, with a machine that can speak, listen and act. When the family settles on a plan, it hands the request to the lenders’ systems. Nothing about the record or the rules changes. Only the front door does.

Notice what stays constant through all four stages: the record, the tools and the rules. A lender that gets those right today is ready for each stage as it comes. A lender that buys another screen is buying something with a shorter life than its contract.

What I would do as a lender today

1. 1

   Get the loan onto one record

   An agent that sees one corner of a loan gives one corner of an answer. Before buying any more AI, ask where the borrower’s goal, the documents, the conditions and every conversation live, and whether they live together.

2. 2

   Insist on a receipt for every action

   If an AI pulls credit, sends a disclosure or changes a field, you should be able to show an examiner who asked, what it saw, what it did and why. If a vendor cannot show you that today, autonomy will only make it worse.

3. 3

   Ask which doors it can already open

   Before any pilot, ask which gated systems the product is connected to in production: credit, underwriting, pricing, your LOS. A tool that is about to integrate is a tool that cannot yet act.

4. 4

   Ask what is left when nobody logs in

   Put this question to every vendor, including us: when my staff stop clicking through your screens, what in your product is still valuable? A good answer names the data, the actions and the controls. A weak one describes the interface.

Where that leaves us

inflooens is software, so I should answer my own question. We expect our screens to matter less every year, and we are building for that. Luna AI already reads the whole loan and acts through the same tools our users do, inside each user’s permissions, with identifying data redacted before any model sees it and a receipt for every action. The next steps are Luna working alongside a live conversation, and then doing the work a lender chooses to hand over.

Software may be solved. SaaS, as we have known it, may well be dying. The record, the tools and the rules are not going anywhere. That is where we have put our effort.