GPT-6 Intelligent UI Will Replace Your Software's Menus

By James Daily

GPT-6 intelligent UI is the first frontier-model pitch aimed at the interface rather than the answer sitting behind it. If the concept survives contact with real products, the model stops filling in a form your software vendor designed and starts designing the form itself, which moves power away from whoever wrote your software and toward whoever is doing the work. That matters more to a small business than any benchmark number, and it arrives with a lock-in problem attached.

What has actually been put on the table

The claim circulating is narrow and specific: GPT-6, the next entry in the GPT line, is being paired with something described as an intelligent UI for everyone. Read that literally and it stops being a model announcement with a nicer chat box attached. It is a proposal that the model generates the interface for each task, deciding which fields appear, which controls vanish, and what order the work happens in. No benchmarks, pricing, context limits or release date have been published alongside the phrase, so the shape of the idea is reported and the performance of it is not.

The practical question is not whether GPT-6 beats whatever Google or Anthropic ships next on some evaluation. It is whether the screen your staff stare at for eight hours a day stops being a fixed artefact that ships twice a year and becomes something a model composes on the fly, per person, per task.

Every software purchase a business makes is really an interface purchase. You are not paying for the database underneath, you are paying for the arrangement of buttons that lets a person get through a task without breaking anything. The menu structure in your CRM or your booking system encodes the vendor's assumptions about how your industry works, and every hour of staff training you have ever paid for is an hour spent learning somebody else's arrangement. That is why a model that builds the arrangement itself belongs to a different category from a model that answers questions about it.

A mild version of this already exists and is easy to miss. Salesforce and HubSpot have spent years bolting natural language assistants onto their object models, and reporting tools let you type a question instead of dragging fields into a pivot table. Those are still fixed interfaces with a text box in front of them. An intelligent UI implies the reverse: a small set of primitives underneath, with the screen assembled per user and per task at the moment it is needed. The vendor ships behaviour, not layouts.

The phrase for everyone carries the actual weight. If the same model generates the interface, the receptionist, the bookkeeper and the operations manager can work from the same data and each get a different front door into it. Today that costs money: three roles, three pieces of software, three sets of licences, three integration projects. The promise is one data layer with generated surfaces on top. That only holds if the underlying records are clean and correctly permissioned, which is precisely where most small firms are weakest.

An interface generated per task also changes what the model needs access to. A fixed dashboard can be built against a database view with narrow permissions. A generated interface needs the model to understand the schema, the role of the person in front of it, and the limits of what that person may see and do. The permissions layer stops being an IT detail and becomes the core of the product. Anyone who has tried to give a chatbot access to a live CRM understands how quickly that gets difficult.

Frontier labs have shown generated interfaces in demos before, and a demo is not a product. What makes this one worth attention is that the concept is attached to a named model rather than floated as research. The models are already capable of assembling a working form. The remaining difficulty is not intelligence. It is permissions, audit trails and the unglamorous business of knowing which customer record is the real one.

What changes for a business that buys software

The first change is what you are actually buying. Today a licence buys a seat in someone's interface, and that interface is the moat. If screens are generated, you are no longer buying screens. You are buying access to your own data plus the permissions model wrapped around it, and switching costs drop because nobody has to relearn where the buttons are. That should change how you negotiate. Multi-year seat commitments signed to lock in a workflow become a bad trade.

Training is the hidden line item in every software decision. A new CRM rarely fails because the licence was too expensive. It fails because six weeks of onboarding turns into six months and the sales team keeps the old spreadsheet running in parallel. A generated interface compresses that, because the screen can match the way a specific person already works instead of demanding they adopt the vendor's process. For a Costa del Sol agency hiring multilingual staff mid-season, the difference is measured in weeks of productivity rather than licence fees.

The second change is uncomfortable. If the interface is generated per user and per task, there is no fixed screen to screenshot for an audit. Regulated businesses, from clinics to law firms to anyone handling payment data, rely on a process looking the same every time it runs. A generated interface can still be logged, but it has to be logged deliberately, at the level of actions rather than the level of windows. Most vendors will not build that until a buyer demands it, and most buyers will not demand it until an inspector asks.

Lock-in does not disappear. It moves. The old lock-in was retraining cost, and it was soft. The new lock-in is your data model and your permission graph sitting inside a vendor's system and tuned by their model, which is a harder thing to walk away from. If this plays out, the question at renewal stops being how much per seat and becomes how portable the schema is and who owns the access rules.

The most interesting consequence is smaller than any of that. Custom internal tools used to sit behind a budget line only large firms could justify: a developer, a designer, three months, twenty thousand euros and a maintenance contract. A generated interface over a clean spreadsheet plus a single model call is not that project anymore. Small firms will start owning tools that were previously out of reach, a quoting screen for one estimator, a scheduling view for one coordinator, a claims tracker for one team. That is the pattern behind most of our AI workflow automation work: the constraint moves from build cost to data quality.

Not everything benefits. High-frequency, muscle-memory work gets worse when the layout keeps changing, and a warehouse scanner or a till will never be generated on the fly. The businesses that gain are the ones where tasks are varied, low volume and awkward, which describes most professional services, most agencies and most property businesses. The ones that lose are vertical software vendors whose entire product was a fixed workflow nobody could easily leave.

Who this lands on first

This lands first on firms between roughly five and fifty people in professional services: agencies, property businesses, clinics, logistics brokers, accountancy practices. They carry enough volume to feel the cost of rekeying data between systems, they have no IT department to build internal tools, and they buy software on seats because that is the only model anyone has offered them. They are also the group most exposed to the promise, since the promise is aimed at the thing they currently pay for.

Take a Costa del Sol estate agency. The conversion variable on an international enquiry is response speed, not brochure quality, and the tooling usually runs through a CRM, WhatsApp and a spreadsheet because nobody built anything better. A generated interface would not fix a slow agent. It would remove the navigation between an enquiry arriving and someone replying, and in this market that gap is the whole game.

The losers are less obvious. Vertical software companies whose moat is a workflow nobody can export will feel this first, because their customers will start asking why the screen cannot simply be generated. Businesses that signed three year contracts to lock in a platform are now committed to a fixed interface in a market drifting toward generated ones. And any firm that spent two years buying tools without cleaning the data underneath will find that a generated interface makes the mess visible faster than any dashboard ever did. The same applies to teams already running custom AI agents against a live CRM: the interface was never the hard part, the access rules were.

What to do now

One action, and it does not involve buying anything. Sit down and list the five workflows where your team rekeys information from one system into another. For each one, check whether the source record is clean and whether the permissions around it are written down anywhere. That inventory is the asset a generated interface will consume. If it is a mess, the model will build you a fast, confident interface on top of a mess and you will discover it at the worst possible moment.

Doing this before the products ship has a second benefit: it makes you a competent buyer. When a vendor arrives with an intelligent UI pitch, the person who can ask which schema it reads from, how actions are logged and what happens to the permission model is negotiating from a different table than the person asking about price per seat. For the same reason, contracts should get shorter over the next eighteen months, not longer. A three year deal buys certainty about a screen that may not exist in its current form by 2028.

Treat GPT-6 intelligent UI as a direction of travel rather than a date in the diary. The models improve on someone else's schedule, and the announcements will keep arriving whether or not your business is ready for them. Your data and your access rules improve on yours, and that is the half of the problem you can actually control this quarter.

The prediction is straightforward. The first wave of GPT-6 intelligent UI products will look impressive in a demo and disappoint in production, not because the model is weak but because permissions and messy records are the real work. The second wave will quietly remove a category of internal tooling projects that small firms never started. The businesses that benefit will not be the ones that moved fastest. They will be the ones whose data was already clean enough that a generated interface had something true to work with, and that advantage is being decided now, in spreadsheets nobody has tidied.

If you want that inventory built with you rather than alone, our AI consulting services start with your workflows, not with a tool.