The AI button in the corner is not the next generation of business apps

There’s a version of the AI business app that’s everywhere right now, and it’s a chat box wedged into the top corner of software that worked fine without it. It got added because a competitor added one, or because “AI-powered” needed to go on the pricing page, or because the board asked what the AI strategy was and a button was easier than an answer. It answers questions nobody was asking and sits there like an ornament.

That’s not the next generation of anything. It’s AI theatre, and it’s already wearing thin. The apps actually worth building don’t bolt a model onto the corner. They put AI down inside the workflow, at the specific points where the work currently grinds, and everywhere else they leave the ordinary software alone because ordinary software was never the problem. The skill is knowing which is which.

Forms that fill themselves from what you already have

Look at how much of any business app is a person retyping information that already exists, sitting right there in a document, an email, or a photo. The customer sent a PDF, and someone’s copying fields off it into a form. That’s the drudgery ordinary software could never fix, because a plain form can’t read a PDF.

This is where AI changes the app. It reads the source, drafts the fields, flags what’s missing, and hands the person a filled form to confirm instead of a blank one to transcribe. Onboarding, compliance, quoting, applications, inspections, support intake, anywhere data gets re-keyed off something the business already holds. The person goes from data-entry clerk to checker, which is faster and less error-prone, and nobody had to open a chat box to get there. That’s AI in the workflow, not on top of it.

Put numbers on it. An equipment hire company in regional Queensland takes bookings that arrive as emailed PDFs, photos of paper forms, and half-complete web enquiries. An intake officer retypes about forty of them a day at four or five minutes each, call it three hours of typing daily, plus the transposed digits and skipped fields that come free with re-keying. An intake screen that reads the email, drafts the booking and asks the officer to confirm turns those forty transcriptions into forty quick checks: under an hour, fewer errors, and the officer back on the phones where she’s worth more. Nobody using that screen would call it a chatbot. It just feels like the app finally doing its job.

Asking the database a question in plain words

The second real shift is search that works how people think. A staff member should be able to type “which jobs are waiting on supplier approval” or “show me contracts expiring next month for high-value customers” and get an answer straight from the records, not go and build a report.

The catch, and it’s the part the gimmick version skips, is that this only works on top of structured data and real permissions. The AI makes the asking natural. The clean data model underneath makes the answer true. And the answer has to link straight to the actual records, not hand back a confident paragraph you then have to go and verify, because a summary you can’t check is slower than no summary at all.

Agents on a short leash, or not at all

Some apps will run small agents that do a bounded job: prepare the weekly report, chase the info that hasn’t come back, draft a response, reconcile records that drifted apart. Useful, when they’re kept on a short leash with narrow permissions and a clear log of what they did.

The failure mode is the agent working in the dark, quietly changing things nobody can see or explain later. That’s not a productivity feature, it’s an incident generator. If a user can’t see what the agent changed, why it changed it, and what source it acted on, the agent shouldn’t have the keys. Autonomy without a visible trail is exactly the thing that makes people stop trusting the software.

Less screen clutter, but only if the workflow was sound

When an app can infer context and tee up the likely next action, it can lean less on giant forms and dense menus, and put the actual decision in front of the person with the evidence sitting beside it. That’s a real improvement in how the thing feels to use.

But design still has to do its job. AI won’t rescue a workflow that was confusing to begin with, it’ll just add a fluent layer over the confusion. If nobody understood the process before, an assistant explaining it in complete sentences doesn’t fix the process, it decorates it. Sort the workflow first.

The plumbing matters more, not less

None of this removes the need for clean data models, validation, permissions, APIs and audit logs. It raises the stakes on all of them. Point a model at bad records and you get bad assistance delivered with total confidence, which is worse than obviously-broken software because people believe it. A good next-generation app is smarter because the plumbing underneath is better connected and better governed, not because it grew a model on top.

You probably don’t need a new app to get this

The market pitch is that AI apps are a new generation, so you should buy new or rebuild. Usually you shouldn’t. If your current system keeps clean records behind an API, the highest-value AI features can be added to what you already run: an intake screen that drafts from documents, plain-language search over your records, a summarised history sitting beside the customer record. That’s AI integration work measured in weeks, not a rebuild measured in quarters, and it lands inside software your staff already know how to use.

The exception is the ageing system with no API, no coherent data model and no safe way in. There, the AI feature isn’t the real project; modernising the foundations is, and the AI capability is part of the payoff for doing that work. Be suspicious of anyone offering to skip the step, because a model pointed at a decade of inconsistent records doesn’t clean them up. It summarises the mess fluently and calls it an answer.

What to ask when someone demos the AI feature

Whether it’s a vendor’s product or a dev shop’s proposal, the demo of the AI feature will be smooth, because demos always are. The useful questions start after the applause. Ask what data the model read to produce the answer on screen, and whether it respects the permissions of the person logged in or just the demo account’s. Ask to see it run on a record with half the fields missing and a typo in the customer name, because that’s what your data looks like. Ask what happens when it gets something wrong: who notices, where the log lives, how a user flags it. And ask what it costs at your volume, because AI features increasingly ship with their own meter, and a per-call price that looks cute on ten demo records is a real line item at forty thousand a month.

None of that needs a technical background. It just needs you to hold the feature to the same standard as the rest of the software: it exists to move work off your people, and if it can’t be shown doing that with messy inputs, it’s decoration.

So the honest test for any business app selling itself as AI-powered is simple: is the AI down in the workflow doing the messy, written, judgement-heavy work that plain software always fumbled, or is it a chat box in the corner that exists so the feature list has the word on it? Put AI where the inputs are messy, where answers need context, where staff burn hours preparing work. Leave the ordinary rules and forms alone where they’re already clear. Get that line right and you get a useful app. Get it wrong and you get a button the market expected and users ignore. If you’re weighing up where AI actually fits in an app you’re building, tell us the workflow and we’ll point at the spots worth it.