For three years the advice on getting more out of AI has been the same, and most of us have given it or received it at some point: learn to write a better prompt. Be specific, give it context, show it an example, ask it to think step by step. Whole training budgets have gone on this, and I’ve handed out a fair amount of the advice myself. It was sound advice for the tool we had, because the tool we had was a text box, and a text box rewards people who are good with words. Most people aren’t, or at least not in the particular way a text box demands, and no amount of training was ever going to change that at the scale a business needs.

In the second week of March, OpenAI and Anthropic each shipped something that changes the shape of the problem. On 10 March, ChatGPT gained interactive visual explanations for more than seventy maths and science concepts, for every logged-in user on every plan: ask about Pythagoras and you get a triangle whose sides you can drag, with the hypotenuse recalculating as you go. Two days later, Claude began drawing charts, diagrams and small interactive widgets inline in ordinary conversations, on by default, on every plan. Ask how compound interest works and you get a curve to play with rather than a paragraph to parse; ask about the periodic table and you get one you can click.

I think those two releases matter more than most of the model launches on either side of them, because for the first time the mainstream assistants are working in the medium most of us are better at. What’s becoming obsolete is the prompt box as the only way in, and prompt-writing as the skill we train everyone for. And if that’s right, the prompting barrier that training has struggled to remove may simply stop mattering for most of the people in your business.

Ten years of trying to get rid of the menu

Back in 2016 I was giving talks to clients about conversational interfaces, and one slide in the deck read “no menus or UI to navigate — just ask”. I’ve written about that presentation before. The timing was six years out and the technology path was wrong, but the underlying idea held up completely: people want less friction between themselves and the computer.

What I didn’t see then is that “just ask” only removes the friction for people who are comfortable asking. A blank box with a blinking cursor is one of the least friendly interfaces ever built for a large part of the population. We took the menu away and replaced it with an essay question.

That’s the prompting skills gap in a sentence, and I’ve spent a good deal of the last two years treating it as a people problem. Show people what the tool can do, embed it in a real workflow, give them permission to experiment: all of that still works, and I’d still recommend it. But some of the gap was in the medium rather than the people. We were asking accountants, designers, warehouse managers and sales directors to describe in well-formed prose the things they’d normally have sketched on a whiteboard or pointed at on a screen. Then we were surprised when a lot of them typed “summarise this” and left it at that.

What a diagram does that a paragraph can’t

Take a question most owner-managers ask at some point in the year: when does the cash get tight? Put that to a text box and you get four paragraphs about receivables, payment terms and a dip in the third quarter, and unless you’re already good at holding numbers in your head you’ll read it twice and still not be sure. Put it to a tool that answers with a chart of the year, and you see the dip before you’ve read a word. You click it. You ask what’s driving it. You drag the big supplier payment back a fortnight and watch the line change. Three questions in, you’ve learned something about your business, and you never composed a prompt worth the name. The visual output is doing the prompting for you.

Two things are happening in that example. The picture is telling you what your next question is, which is a feedback loop that text struggles to produce, because in a text conversation every follow-up has to be composed from scratch, and composing is the step most people find hard. And the question itself is a spatial one, about how something moves across time, which is the kind of thing sentences describe badly however good the prompter is. Describing a cash flow, a customer journey or a dependency between three systems in prose is a translation with losses at both ends: you lose something turning the shape into words, and the model loses something turning your words back into a shape. When the tool draws the thing and lets you correct the drawing, most of that loss goes. Today’s versions are still mostly explanatory, a curve here and a clickable table there, but that’s the direction they point.

The third effect follows from the first two, and it’s the one I’d watch most closely as a business owner: a visual, explorable interface lowers the floor. The person who would never have written a three-paragraph prompt will happily drag a slider and see what changes, and having seen it, will ask a question they’d never have thought to type. OpenAI cited research suggesting that manipulating variables and seeing the effect builds understanding better than instruction does, and that finding doesn’t stop applying when the learner runs a warehouse and the concept is her own stock levels.

There’s a rule of thumb in that. Text still suits work that’s made of language, such as a contract clause, a letter or a policy. The visual route earns its place when the task is about position, relationship, sequence, scale or change, and a surprising share of the questions asked in a small business are of that second kind.

Test the interface, not just the model

All of which changes what we should be doing when we evaluate an AI tool. Most evaluations I’ve seen, including some I’ve run, amount to a bake-off between models: same prompt, several tools, compare the answers. That was a reasonable test when the box was the only way in. It’s the wrong test now, because the differences between the models matter very little if the person who’s meant to use the tool can’t get past the interface. The best model behind a text box will lose, inside your business, to a good-enough model behind a workspace people can see and touch. A few things to try instead:

  1. Give it to the person who’d never use it: Hand the tool to the least confident member of the team rather than the enthusiast, ask them to get an answer to a real question from their own work, and watch where they stall. If they stall at the blank box, the model behind it doesn’t matter.
  2. Ask a spatial question: Pick something from your business that lives naturally as a shape, such as a process, a plan or a set of relationships, and see whether the tool draws it or describes it. Then see whether you can change it by pointing rather than by re-explaining.
  3. Count the follow-ups: In a good visual workspace, the second and third questions come from the output itself. If every follow-up has to be typed from nothing, you’re still sitting the essay exam.
  4. Check where the output lives: Claude’s inline visuals, for instance, are deliberately temporary, built to help you understand something in the moment and gone when the conversation moves on, and they’re distinct from the permanent artefacts you can share and come back to. Both are useful, but for work you need to know which you’re getting.

None of this makes a clear brief worthless: when we hand an agent a whole task, the brief becomes the work, and that’s a real skill for the people who direct agents. For everyone else, we spent three years asking people to get better at typing, and AI has now started getting better at meeting them where they are. That leaves us free to spend the training budget on the part that was always harder: knowing which questions are worth asking.