The useful AI apps do one job. They take something a person already has, such as a photo of a page of notes, and turn it into something that person needs, such as a set of questions to test themselves on. Everything else is a demo. If an app can’t name the single job it does, it isn’t ready to be built.
The useful AI apps automate one step. Everything else is a demo with a chat box attached.
What an AI-powered app is
An AI-powered app is ordinary software with a language or vision model doing one step inside it. The model handles the part that used to need a person: reading handwriting, summarising a document, turning a topic into questions, or sorting messy input into a structure. The rest of the app is normal code.
That framing matters, because it keeps the cost and the risk in proportion. You’re paying for one step to be automated, not for an app that thinks.
Who it’s for
This suits businesses and projects with a repetitive transformation job:
- turning documents, photos or recordings into structured, searchable content
- generating practice material, summaries or briefs from a source
- taking free text from customers and sorting it into something usable
- anything where a person currently retypes information from one format into another
The problems it usually solves
Information exists but it’s in the wrong form. A page of handwritten notes, a recorded session, a PDF. All of it useful, none of it searchable or testable until someone does the work.
The value disappears when the event ends. A teaching session, a workshop, a client call. Without something turning it into notes and questions, it’s gone.
Nobody has time to make the practice material. This is the whole argument for the revision apps I’ve built. Reading is easy to do and easy to fake. Being tested is the part that works, and it’s the part nobody has time to prepare.
What’s included
- Scoping the one job the AI does, and confirming it’s genuinely worth automating
- A prototype using real inputs of yours, so you can judge the output quality before committing
- The build, with the AI step properly contained and the rest as normal software
- Cost controls, including per-user or per-day limits so usage can’t run away
- Fallback behaviour for when the model is slow or unavailable
- Handover, with the prompts and settings documented so they can be changed later
What you get
- A working app that does the job on your own material
- Honest timings, measured on the live build rather than estimated
- Usage limits you control
- A record of what the AI is actually asked, so its behaviour can be adjusted rather than guessed at
A worked example
I built a GCSE revision app for a member of my family, which is live and used daily. It covers eight subjects and works in three numbered steps per topic: learn it, practise it, tick it off.
The AI does four jobs: transcribing a photo of schoolwork, turning that into notes, generating practice questions, and generating questions from a topic. On the live build those take roughly 8, 12, 9 and 17 seconds. Those are measured timings, not estimates, and they’re the kind of number worth knowing before you design a screen, because 17 seconds needs a different interface from 2 seconds.
Two decisions from that build are worth repeating on any project of this kind. First, everything starts at “not started”. An earlier version pre-filled progress from an existing checklist, and it was removed, because only progress made in the app should count. Second, usage limits were set at 80 AI actions per account per day to keep running costs predictable.
A second version of the same idea exists as a clickable prototype for a school, where a teacher-run session becomes searchable notes, flashcards and questions instead of disappearing when it ends. It’s tested across 19 journeys at phone and tablet size, and it hasn’t been deployed.
Limitations
AI output needs checking, and an app has to be designed on that basis rather than assuming the model is right. Anything factual should show its source or be reviewable. Costs are usage-based, so they rise with adoption, which is why limits go in from the start rather than after the first surprising bill.
What this won’t solve
An AI step won’t rescue a process that doesn’t make sense manually. If a person couldn’t do the job well from the same input, a model won’t either. It also can’t take responsibility for accuracy in a regulated or safety-critical setting, where a human check has to stay in the loop.
Talk to the person doing the work
There is no account manager here and no team to be handed to. If you get in touch, it is me who reads it, me who looks at your site, and me who does the work if we go ahead.
Get in touch, or start with a free assessment.
Engagement and price
There are no list prices here, and that is deliberate. What a piece of work costs depends on the state of the site, how much of it there is and what you actually need, none of which I know before looking at it.
So it starts with a free analysis. I look at your site, your Search Console and who you are really competing with, then come back with what I would do first and what it would cost to do. No obligation attached, and if the honest answer is that you do not need me yet, I will say so.
How the work gets done
Prototypes run on your real inputs before any build is agreed. Where the app has public pages, they’re checked through the technical SEO checks so they’re indexed properly. This service sits under app and tool development. The AI revision app and Tower Revision case studies cover both builds, and generative AI for marketing covers using AI in content rather than in software.
FAQ
What does the AI actually do in these apps?
One defined step, such as reading handwriting from a photo, summarising a document, or turning a topic into practice questions. The rest of the app is ordinary software. Keeping the AI’s job narrow is what makes the cost predictable and the output quality something you can judge.
How much does it cost to run?
It depends on usage, because AI is billed per request rather than a flat fee. That’s why I build in limits, such as a cap on actions per user per day, from the start. The prototype stage gives a realistic per-user figure before you commit to anything.
Is AI output reliable enough to use?
For transformation jobs such as summarising or generating questions from a source, it’s good enough to be genuinely useful. For anything factual or regulated, the app needs to show its source and keep a human check in the loop. I’d design it that way rather than assume the model is right.
Can you use my own data safely?
Yes, and how the data is handled is part of the scoping rather than an afterthought. That covers what gets sent to the model, what’s stored, and for how long. Some processing can happen entirely on the user’s own device, which avoids sending anything at all.
Talk it through
Get in touch with the repetitive job you’d like to automate, and I’ll tell you whether an AI step is the right way to do it.