For sixty years we taught people where the buttons are, and the next sixty will teach software what people want, which leaves one job for the person: checking.
That is the argument, and it runs against the tidy version going around. Intent-based systems let you state a goal and leave the software to find the path. The design press concludes that the interface therefore disappears. For business software we think the conclusion is wrong, and the error changes how a company should prepare.
The conclusion is not unreasonable, because it comes from two good sources. The Nielsen Norman Group calls intent-based interaction the first genuinely new interaction paradigm in roughly 60 years. An essay from Die Produktmacher separates command-based computing, where you specify every step, from intent-based computing, where the system chooses the steps, and names the price: users have to audit output instead of executing a task. Take that last clause seriously and the disappearance story falls apart. If the user audits, the user needs something to audit with.
Think about what clicking through a form actually gave you. Nothing fancy: you knew what happened because you did it. When an agent raises the purchase order, nobody on your team knows what it read, what it assumed or which rule it applied, unless a screen says so. The entry screen goes, and a review screen has to take its place. That screen shows the proposed action, the records the agent relied on, the rule behind it and a way to reverse the result. Almost no ERP screen built for typing shows any of those four things. So the interface that survives in operations software is the evidence screen, and it will matter more than the one it replaces.
Most firms will meet this inside software they already own. Gartner projects that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. Read that as a delivery channel: your accounting, CRM and ERP vendors will switch agents on whether or not you wrote an AI strategy. The same firm expects more than 40% of agentic AI projects to be canceled by the end of 2027, and names unclear business value and inadequate risk controls among the reasons. Risk controls are a review problem. Nobody can control what they cannot inspect.
Search behavior leans the same way, though only as a rough signal. In Google Ads data for US English, measured on 11 October 2026, "prompt injection" averages 5,400 searches a month and is rising, with a spike to 9,900 in August 2026. "Shadow AI" is at 3,600, up from 2,400 a year earlier. "Context engineering", the craft of feeding a model the right inputs, fell from 6,600 a month in autumn 2025 to 3,600 by summer 2026. One country and one language prove little. The pattern still fits: after the first excitement, people search less for how to make the model do things and more for how it goes wrong.
There is a second reason the review screen gets harder than it sounds, and it has nothing to do with the agent. The old interface was hiding the state of your operations. A person at a keyboard repairs things without noticing. They know which report to trust, they quietly correct the stale stock figure, they remember that the customer with two spellings is one customer. The screen looked fine because the human behind it was patching.
An agent does not patch. It reads what is there. Connected, consistent records produce a reliable outcome, while a dozen disconnected spreadsheets produce a confident, plausible, wrong one. And the reviewer, comparing the agent's proposal against those same records, inherits the problem: a check against dirty data checks nothing. Once nobody sits in front of the screen, the state of your operations is what customers and staff experience. Our piece on observability as the gate between pilot and production covers how you notice when it goes wrong.
What would we do first? Give agents read access before write access, since diagnosis and summaries can run with little supervision long before anything may move money. Design the review view before you build the agent: for every action that touches money or permissions, draw on one page what the reviewer will see and how they reverse it. If that page cannot be drawn, the agent is not ready, whatever the demo looked like. Then fix the records the reviewer will compare against, and take one high-volume, low-risk workflow through a full quarter before adding another.
You may disagree and hold that the agent should simply act and the screen should go. Test that view with one question. When the agent is wrong on a Thursday, who signs off on the correction, and what do they look at?
Not sure where your operations stand? Our free Operations & AI Readiness Assessment takes about three minutes and returns a benchmarked report with your priority gap. Our AI services start from the same place, mapping what an agent will touch before any automation goes live. To talk it through, get in touch.
