From Filing Cabinet to SaaS to AI: Why Software Is Finally Doing the Work

Every wave of business software has done the same thing: take a filing cabinet and put it online. AI agents are the first wave that can open the cabinet and do the work inside it, but only for the businesses that build the Brain first.

In our first article we argued that the model is almost irrelevant if your organisational Brain is broken. In our second article we argued that the CLI, not the dashboard, is how agents actually get anything done. This article connects both of those ideas by going back further, to where enterprise software actually started, because the history explains exactly why this moment is different from every software cycle that came before it.

The First Filing Cabinet Moment

In the 1960s, airlines kept every seat reservation in a physical filing cabinet. A booking agent would search paper cards by hand to check whether a seat was free. IBM and American Airlines built a joint project to fix this called SABRE (Semi-Automated Business Research Environment), one of the first large-scale reservation databases. It took the filing cabinet and put it on a screen.

That was the beginning of enterprise software as we know it, and the pattern it set has repeated in almost every category since. HR had filing cabinets full of employee records. That became PeopleSoft, and later Workday. Legal had filing cabinets full of case law and contracts. That became LexisNexis. Accounting had filing cabinets full of ledgers and invoices. That became QuickBooks, and later NetSuite.

What SaaS Actually Automated

Here is the uncomfortable part. None of that software actually reduced the amount of human labour required to run those departments. The same number of people work in HR today, per company of the same size, as did in 1950. We just replaced the locked filing cabinet with an IT department and a security team making sure nobody breaks into the digital one.

What all of this software did brilliantly was storage and retrieval. It made information faster to find, easier to edit, and simpler to share. What none of it did was the work itself.

Take a real example. A dentist can open QuickBooks and see a list of overdue invoices in about four seconds. QuickBooks will not call the patient and ask them to pay. It will not follow up when they don’t answer. It will not check the claim against the patient’s insurance policy, work out what’s actually owed, and close the loop. A human still has to do every part of that job. The software just made the list of jobs easier to see.

This is why, despite forty years of enterprise software, small and medium businesses still run on people doing manual, repetitive, unglamorous labour. The filing cabinet got a lot nicer. The work inside it never got smaller.

Why This Cycle Is Actually Different

In a recent conversation on the a16z podcast, Steijn Pelle, co-founder of Lassie (an AI company that runs the back office for dental practices), put the entire shift into a single line: “the incumbent was named Betty, and she quit two weeks ago.”

That line matters more than it sounds like it should. In most software categories, a startup has to worry about an existing software company copying its idea once it proves the market exists. That is the classic startup fear: you build something good, an incumbent with more resources and more customers copies it, and you lose. It has played out again and again, from digital video recorders to email search tools.

But in a huge number of small business categories, there is no software incumbent to compete with at all. Ask who the market-leading software company is for running a dental practice’s back office, or a physio clinic’s admin, or a small plumbing business’s scheduling and invoicing. There isn’t one. The incumbent was a person. She did the books, chased the invoices, ran payroll, filed the insurance claims, and kept the whole place running, usually while wearing far more hats than the job title implied. Eventually she got tired of it and left, and the owner ended up doing her job themselves at midnight, on top of their own.

That is the actual opportunity AI agents represent for small and medium business, and it has nothing to do with a smarter model. It is the first time software can pick up the labour itself, not just the paperwork about the labour.

Labour Was Always the Hard Part, Not Storage

This is the point worth sitting with. Every prior wave of software treated the problem as a storage and retrieval problem, because that was the only problem software could actually solve. AI agents are the first technology that can treat it as a labour problem instead.

Read the claim. Work out what’s owed under the policy. Fill in the form correctly. Chase the payment. Follow up when nobody replies. Update the record when it’s resolved. That is not a smarter filing cabinet. That is the work Betty used to do, done without waiting on a person to sit down and do it.

This same shift shows up in smaller, less obvious ways too. One of the most tedious parts of running any services business is what happens after you send an invoice: the follow-up questions, the scope disagreements, the client who wants to negotiate a job that’s already finished. Every one of those conversations currently costs a real person real time, and time is the actual product a services business sells. An agent that has the full context of the job, the scope, the hours logged, and the history of the client relationship can have that exact conversation instantly and correctly, as many times as it needs to, without ever getting tired of explaining the same line item twice.

Why the Brain Has to Come First

None of this works by accident, and it is not the model getting smarter that makes it possible. An agent can only resolve an invoice dispute correctly if it already has access to the scope document, the hours logged, the email where the client agreed to the extra work, and the history of every similar dispute before this one. That is the Brain we described in our first article: the denormalised memory, the context layer, the skill registry, all built and kept current before the agent is ever asked a real question.

Skip that step and an agent does not resolve the dispute. It guesses, confidently and often wrongly, which is a worse outcome than not answering at all. This is the actual reason most AI pilots inside businesses stall out. The business bought the outcome without doing the unglamorous work of building the thing that makes the outcome possible.

Build the Brain first, though, and the payoff is direct: accuracy. The same task that an agent would otherwise guess at, it now gets right, consistently, because it is working from the actual facts of your business instead of a blank slate. That is the real difference the Brain makes — not automation that’s merely possible, but automation accurate enough to actually trust with a real client, a real invoice, a real claim.

Why the Harness Is the Hands

Having the right Brain solves what the agent knows. It does not solve what the agent can actually do. This is where our second article comes in. What actually lets an agent act on anything — the tools it’s allowed to call, the permissions it’s been given, the commands it can run — is what’s called the harness. An agent stuck clicking through a dashboard, screenshotting the screen to find a button, and simulating a mouse click is slow, expensive, and fragile. A harness built around a CLI, where the agent reads and writes plain text directly, lets it take the action the moment it decides what needs to happen.

Put the two together and you get the actual shift this article is describing. The Brain tells the agent what is true and what has happened before. The harness lets the agent act on it directly. Neither one on its own gets you past being a nicer filing cabinet.

Filing Cabinet, SaaS, AI Agent

Three eras, one underlying pattern.

The filing cabinet stored information and required a person to act on every piece of it.

SaaS moved that storage online and made it faster to search, but still required a person to act on every piece of it.

The AI agent, built on the right Brain and given real execution through a CLI-based harness, is the first version of this pattern where the software can act on the information itself.

Most businesses buying AI right now are still shopping for a nicer filing cabinet, a chatbot that summarises the invoice for a human to go and process manually. That is not this shift. The businesses that actually benefit are the ones building the Brain and the execution layer together, so the software stops being a place to look things up and starts being something that does the job Betty used to do.

What This Means for You

If you are running or advising a small or medium business, the signal to watch for is not a new model release. It is the moment a role in your business is described almost entirely as “wears a lot of hats.” That role is where a person is currently doing the job an under-built filing cabinet never could. It is also exactly where an agent, built on the right Brain and given real execution, can take the busywork off that person’s plate so they spend their time on the parts of the job they actually enjoy, which tends to make them a lot more likely to stick around.

The next model release will not fix that. Building the Brain, and giving it hands, will.

Frequently Asked Questions

Why didn’t SaaS software reduce the amount of labour needed to run a business?

SaaS software solved storage and retrieval. It made records easier to find, edit, and share, but it still required a human to act on every piece of information: chasing a payment, filing a claim, following up with a client. The same number of people were needed to do the actual work, just with better tools for looking at it.

What is the “Betty” problem in AI for small business?

It refers to the observation that in many small business categories, there is no software incumbent to compete with, only a person doing the job by hand. When that person leaves, the business has no system to fall back on. AI agents that can do the actual labour, not just store information about it, are the first real alternative to hiring another person into that role.

Why do AI agents need a “Brain” before they can automate real work?

An agent can only resolve a task correctly, like an invoice dispute or an insurance claim, if it has access to the right context: the scope document, the transaction history, the relevant policy. Without that memory layer built and kept current, the agent guesses instead of answering correctly, which is worse than doing nothing.

Why does an AI agent need a CLI instead of a dashboard to actually do the work?

A dashboard is built for a human to click through. An agent forced to use one has to screenshot the screen and simulate clicks, which is slow and expensive. A CLI lets the agent read and write plain text directly, so it can take an action the moment it decides one is needed.

How is this different from a chatbot that summarises information for a human?

A chatbot that summarises an invoice or a claim for a human to still process manually is a faster filing cabinet, not a labour replacement. The shift described in this article is software that completes the task itself: reading the claim, filling in the form, chasing the payment, and closing the loop without a person actioning each step.