The next step for SaaS is not new screens but less typing: how we fill our blog, website content and project backlog through AI with a person guiding, verifying and responsible for the result, and what stays human.
Software today is inputs and outputs. Forms, tables, charts, a search box, a save button. Someone types the data in, someone else reads it out. The screens are fine. The typing is the part whose time is up. I believe the natural next step for this kind of service is that the AI does the filling and the person guides it, checks the result and stays responsible for it. For the last few months I have run our blog, our website content and a full project backlog exactly that way: I tell an AI what I want, it does the work, I check it and approve it. The only things I still type by hand are the ones too sensitive to hand over.
This is not the end of anything. It is evolution. Within a year I expect everything in this space to be moving in this direction, and some products to be there already.
This is not a vision document. It is how we already work.
The blog. I give the AI the facts, the angle and the audience, usually by voice. It drafts the post and shows me a preview. If I like it, I say publish, and it presses the one button it has access to. It cannot touch the database; everything behind that button is our own code. It can also attach images. Coding agents upload them directly. The chat tools cannot upload files at all, so the agent gives me an upload link, I drop the image in, say "done", and it attaches it. That is the division of labour in one gesture: I do the single step the machine cannot, and nothing else.
The website content. When we rebuilt our website, every page needed real text: services, cases, products, the company story. I did not want to type it into the admin screens, and I did not want anyone writing straight into the database and breaking formats and tags. So we gave an agent a controlled way into the content system, fed it everything about the company, and let it draft the whole site. I read it, corrected it, approved it. The agent did the filling; I did the judging.
The project backlog. For the same website, I connected the agent to our design tool and to our work tracker, and asked it to go through the design page by page and create the tasks. I read every task, fixed what was wrong, and only then let it post them. The full story is here. Nobody typed a work item.
In each case the pattern is the same: the person guides, the AI fills, the person verifies, the person approves.
Because the AI will do exactly what you asked and also something you did not.
For one piece of content I gave the AI our real internal numbers so it would understand the subject and write properly, and told it not to disclose them. The draft came back with every number in it. Nothing went out, because the draft is a draft and I read drafts. I removed the numbers, tightened the instruction and ran it again.
Another time it simply misunderstood what I wanted to say, confidently, and in good English. Same fix: read, correct, rerun.
People ask whether I would still read every draft at twenty posts a month. Yes, and so should anyone in my position. That reading is not a cost that AI added; it is the work that always had to be done. Without credible oversight, AI work is exactly as bad as unsupervised human work. A person is effective only when they are responsible for something. If nobody owns the result, the result is poor, AI or no AI. What changes with AI is what you are responsible for, not whether you are.
AI changes what you are responsible for, not whether you are.
Passwords, keys, salaries, client names, personal data. None of it goes to an AI provider.
That is not the same as "AI cannot touch it". The next thing we are building is invoice generation, and invoices are full of client names and amounts. That work will be done by a model running on our own hardware, so the data never leaves the building. Ordinary content goes to the big providers because they are better and faster. Sensitive data goes to a local model or stays with a person. The question is never "AI or not", it is "which AI, where, with access to what".
Privacy is a real concern here. It is also a solved one, if you are willing to run some of the work locally.
Setup and strategy. Deciding what the system is for, what the AI may touch, which approvals exist and who gives them. That is first time work, done by people, and it is where most of the thinking happens. After that, people keep deciding during the work which accesses to widen and which to keep closed.
Approvals, where you have set them. Sensitive and critical decisions. And the step the machine cannot do, like dropping a file into an upload box.
The structure stays too. The forms, the screens, the tables are not going anywhere, and a person must always be able to do the work by hand when needed. What moves to the AI is the repeatable, same type, boring work that takes too long to enter: the part that was never really the job.
The AI cannot be left alone. There has to be a person in the loop, and that person has to own the outcome. But they now do the parts that need a person, and far more of them than before, because the filling, clicking and searching are gone.
I dictate most of my instructions to the AI, including the notes this post came from. For me voice is much easier: I can give more detail and more of the thinking, much faster than I type. That is a personal preference, not a rule. Some people type fast. The point is that the interface to software becomes language, spoken or written, and the forms disappear behind it.
Because it is the natural next step for this type of service, and a lot of companies and SaaS products are already working on it. Today a product is a set of inputs and outputs. Next, the inputs are a conversation, spoken or typed, and the outputs are available both to people and to their AI, with full access to the data and statistics so the AI can give proper feedback and not just fill fields. Manual entry stays available, the way you can still do a calculation by hand. It just stops being the default.
I do not expect every product to be there in a year. I expect every product to be moving there, and some to have arrived.
The objections I hear are mistakes, privacy and risk, which is why most of this post is about verification, local models and what stays human.
The end state is simple to describe. AI does the work. People oversee it. People can do far more with AI, and AI is far more useful with people watching. We are already working that way, and I expect most software to catch up soon.
If you want to see how we set this up for other companies, start with AI solutions and agents and private and local AI.
Our engineers can help map this to your stack.
