Syntic
All posts

Published October 4, 2026 in Software

Software Is Moving From Interfaces to Intent

By Syntic AI

Every interface ever shipped is a translation service. A person has a goal, such as "send the Berlin team the revised forecast," and the software has a vocabulary of menus, fields, buttons and commands. The user's job has always been to translate one into the other.

Language models change who does the translating. A user can now describe the outcome, and the system works out the steps. It is tempting to conclude that screens, menus and forms are on their way out.

The research on how people actually work with computers points somewhere more interesting. Intent removes the hardest part of operating software, which is knowing how to do the thing. It makes a different part harder, which is knowing whether the thing was done right. The products that win this transition will be the ones that use language to capture intent and keep a visible, manipulable interface for checking, correcting and approving the result.

Every interface was a translation layer

Jakob Nielsen, writing for Nielsen Norman Group in June 2023, described the history of computing in three paradigms. In batch processing, from around 1945, users submitted a complete set of instructions and waited, often overnight, for the result. Command-based interaction, which he dates to time-sharing in 1964, introduced turn-taking: the user issues a command, the computer responds, the user issues another.

The command line, the graphical interface and the touchscreen all live inside that second paradigm. They differ in how commands are expressed. Typed syntax demanded that users memorize a vocabulary. The GUI made the vocabulary visible as icons and menus. Touch put the user's finger on the object itself. But in each case, the user decided what to do, step by step.

Nielsen called the third paradigm "intent-based outcome specification." The user states what they want, not how to get it, and the system plans the steps. In his phrase, this "completely reverses the locus of control." For sixty years the person drove and the computer followed. Now the computer drives towards a destination the person names.

That reversal is the real story. Natural-language interfaces are not new; researchers built them for databases in the 1970s. What is new is that the system behind the text box can plan, call tools and complete multi-step work, a change we trace in how AI is moving from answering to doing.

What intent removes, and what it adds

The most useful lens for this shift comes from a 1985 paper by Edwin Hutchins, James Hollan and Donald Norman on direct manipulation. They described two gulfs that every interface must bridge. The gulf of execution separates a user's goal from the actions the system requires. The gulf of evaluation separates what the system did from the user's ability to tell whether the goal was met.

Ben Shneiderman had coined "direct manipulation" two years earlier to describe interfaces with visible objects, rapid reversible actions and immediate feedback. Its genius was that it narrowed both gulfs at once. You drag the file to the folder, and you see it land.

Intent-based interaction collapses the gulf of execution almost entirely. Nobody needs to know which submenu holds the pivot table settings. But it widens the gulf of evaluation. When a system takes twelve steps on your behalf, you did not watch any of them. Nielsen named this problem in the same 2023 essay: users often "don't know how something was done," which makes errors hard to find.

This argument is older than large language models. In 1997, Shneiderman and Pattie Maes debated "direct manipulation vs. interface agents," published in ACM Interactions. Maes argued that people facing more information and more tasks than they could handle would need software agents to delegate to. Shneiderman argued for control and predictability, and against modelling human-computer interaction on human-human interaction. Both turned out to be right. Delegation is now practical, and the need for predictability has not gone away.

The blank prompt problem

A GUI shows you what is possible. A text box shows you nothing. That is the central usability cost of intent, and the research on it is consistent.

Nielsen calls it the articulation barrier. Citing literacy research that classifies roughly half the population of rich countries such as the United States and Germany as low-literacy, he argues that writing clear prose instructions is harder still, so the share of people who struggle to articulate what they want is likely higher.

Even capable users struggle. In "Why Johnny Can't Prompt" (CHI 2023), J.D. Zamfirescu-Pereira and colleagues watched non-experts design prompts for a chatbot. Participants worked opportunistically rather than systematically, generalized from single successes or failures, and brought expectations from talking to people that the model did not meet.

Hariharan Subramonyam and colleagues at Stanford gave the problem a name at CHI 2024: the gulf of envisioning. Before a user can even write a prompt, they face three gaps. Can the model do this task at all? How should I instruct it? How will I judge whether the output is right?

Microsoft researchers anticipated this in their 2019 "Guidelines for Human-AI Interaction." Of their 18 guidelines, the first two are to make clear what the system can do and how well it can do it. An empty prompt field does neither. Discoverability was the GUI's quiet achievement; menus are a catalogue of capability. Remove them and users must guess.

Where the GUI still wins

Pure conversation is poor at four things that graphical interfaces do well.

Precision

Some intents are easy to state and hard to state exactly. "Make the chart look cleaner" is a fine prompt. "Move the legend four pixels left" is a slider. When a value is continuous, or when the target is a specific cell, word or region, pointing beats describing. Nielsen Norman Group's 2023 usability study of chatbot use, with eight professionals working in ChatGPT-3.5, documented "accordion editing," where users repeatedly asked the AI to expand or shrink its output, and "apple picking," where they referenced fragments of earlier answers to get the response they wanted. The recommended fixes were graphical: let users edit text directly, change just a portion of a response, and select a location by pointing rather than describing it.

Habit and speed

Repeated tasks do not need a sentence each time. In a 2024 study, Florian Lehmann and Daniel Buschek at the University of Bayreuth designed a conversational writing assistant based on a survey of 121 people. In practice, participants "regressed to using short command-like prompts." When the researchers then offered those commands as toolbar shortcuts alongside free prompting, users switched dynamically between fixed functions and open-ended requests. People want the button for the thing they do daily and the text box for the thing they have never done.

Direct manipulation

Spatial, visual and structural work, such as arranging a layout, trimming a clip or reordering a pipeline, is still best done with hands on objects. Language can propose the arrangement. Fine adjustment wants a canvas.

Auditability

A conversation is a poor record of state. If an assistant changed forty records across three systems, a scrolling transcript is the wrong place to review it. A table of what changed, with before and after values, is the right one. The more consequential the action, the more the evaluation side of the interaction needs structure.

Hybrid is the destination, not the compromise

The most promising designs do not choose between intent and interface. They use each where it is strongest.

Generative interfaces

One direction lets the model build the interface itself. In November 2025, Google Research described generative UI, in which a model produces a complete interactive experience for a prompt, and shipped it in experiments in the Gemini app and in AI Mode in Search. In the accompanying paper, human raters overwhelmingly preferred these generated interfaces to standard markdown answers. The authors were careful about limits. Pages designed by human experts were still preferred overall, generated interfaces were at least comparable to them in about half of cases, generation could take a minute or more, and outputs were occasionally inaccurate.

Generative UI matters because it turns intent back into something users can see and manipulate. You ask a question in language and get controls, comparisons and visual state in return. The gulf of evaluation narrows again.

Structured confirmation

The second pattern is less glamorous and more important. The system turns a fuzzy request into a concrete, inspectable plan or artifact, and the user approves it before anything irreversible happens.

Eric Horvitz set out the logic in "Principles of Mixed-Initiative User Interfaces" (CHI 1999). Automated action under uncertainty about a user's goals has costs as well as benefits. A system should weigh the expected value of acting, and when it is unsure, it should resolve the uncertainty through efficient dialogue while respecting the cost of interrupting the user. In modern terms: ask when it matters, act when it is cheap to undo, and show your work when it is not.

Syntic for Excel is one example of this pattern. A user describes what they need in plain English; Syntic writes the formula, explains what it does, and the user reviews it before it is inserted into the workbook. The request is intent. The output lands in the most precise direct-manipulation surface in business software: a spreadsheet cell.

What changes for SaaS product teams

If users increasingly arrive with a goal rather than a click path, several long-standing assumptions in software design stop holding.

The data model becomes the product surface

Navigation used to be a large part of a product's design. In an intent-first product, much of it is replaced by a system planning over the product's objects and actions. What matters is whether those objects are well named, whether actions have clear preconditions and whether everything the system does can be expressed as a reviewable change. A product with a clean model of its own nouns and verbs is easier to drive by intent. A product whose logic lives in its screens is not.

Design effort moves from the path to the proof

The hardest screens to design will be review screens: plans, diffs, previews, change summaries and confirmations. These are where the widened gulf of evaluation gets closed. Teams that spent years tuning onboarding funnels will spend the next few years on how to show a user, in five seconds, what the system is about to do and what it just did.

Empty states need a catalogue

The blank prompt is a design failure, not a neutral starting point. Suggested tasks drawn from the user's real data, visible examples of what the system can do and honest signals about what it cannot do are the intent-era equivalent of a menu bar.

Success metrics change

Clicks, sessions and time in app measured engagement with an interface. When the system does the clicking, those numbers can fall while value rises. Better measures are whether the stated goal was completed, how often users had to correct or rephrase, and how often they reversed what the system did.

Pricing follows the work

Seat-based pricing assumes value scales with the number of people operating the software. Intent weakens that link: one person can direct far more work, and some work happens with no person in the interface at all. Pricing is already moving. Intercom lists its Fin AI agent from $0.99 per outcome, alongside per-seat plans for its helpdesk. Salesforce introduced Agentforce at $2 per conversation and has since added Flex Credits at $0.10 per action. These models are still settling, and none is a clean answer, but they point the same way: charging for work done rather than access granted. The shift also interacts with falling model costs, which we examine in what happens as the cost of intelligence falls.

The same tension shows up on the builder's side, when applications are described into existence; we cover that in from prompt to product.

The interface becomes the receipt

The graphical interface is not disappearing. Its job is changing. For forty years, the GUI was primarily how users told software what to do. In intent-based software, it increasingly becomes how software shows users what it did, what it plans to do next, and where they can step in.

That is a more demanding role. A receipt has to be accurate, legible and complete, and it has to make the important exceptions impossible to miss. Products that treat the screen as an afterthought to the chat box will produce outcomes users cannot verify. Products that treat it as the place where trust is earned will make intent usable for people who never learned the menus, and safe for the people who did.

Sources