From prompts to systems — how AI work is really delivered

The gap nobody names

Most people who say they use AI mean this: open a chat window, type a request, paste the answer somewhere. That is a real skill and it is now a common one. It is also the layer of work AI absorbs most easily — data entry and encoding, basic writing and rewriting, copy-pasting between systems, simple research and summarising, formatting and sorting. That is the work sitting in the $4–7 per hour band, and the pressure on it is not letting up.

Delivery is a different activity. Delivery means a business had a problem on Monday, and by the time you are done there is a working system that solves it, that the right people can log into, that runs when nobody is watching, and that the owner can still operate six months from now. That difference is what separates work that gets repriced from work that does not. What follows is what actually sits inside that sentence.

It starts with a business problem, not a tool

Clients rarely describe problems. They describe solutions they have already imagined. "We need an app for our tenants." "Can you build us a dashboard." Take that at face value and you will build the wrong thing beautifully.

The first job is finding out what is actually broken. Consider a familiar shape of request in property management: the client asks for a tenant app. Underneath it, repair requests arrive by Viber, by email and by phone call, three staff each keep their own list, and nobody can answer a simple question, which repairs are still open right now. The app is not the problem. The missing single record is.

Useful questions are boring ones. Who does this today, step by step. What happens when it goes wrong. What does the owner count at the end of the month, and where does that number come from. Who chases whom.

An LLM helps here. ChatGPT, Claude, Gemini or Perplexity will draft your discovery questions, tidy your notes into a summary and suggest angles you missed. What it cannot do is sit in the room, notice the hesitation when you ask who approves payments, and follow it. That noticing is yours.

Modelling the data comes before building

Every business runs on a handful of things it keeps track of. In the property example: properties, units, tenants, repair requests, contractors, invoices. Data modelling means writing those down as tables, deciding what each one stores, and being explicit about how they connect. One property has many units. One unit has one current tenant and many past ones. A repair request belongs to a unit and to an assigned contractor.

This is the part clients never see and the part that decides everything else. Screens sit on top of the data model. Automations read and write it. Access rules depend on it. Get it wrong and you find out in week six, when a report cannot be built because the information was never separated properly.

Airtable and Softr Database are the usual places to do this without writing code, because you can model the data and set the access rules in the same tool. The JobTayo foundation covers data at the level of the decision rather than the software — a spreadsheet, a proper database, and what changes the moment several people need the same record. The deeper Softr work is a specialisation, and that is still being built.

Deciding who is allowed to see what

Access rules are not a settings toggle you flip at the end. They are a design decision, made alongside the data model, and they are one of the clearest markers of someone who delivers rather than demonstrates.

Work it through role by role. A tenant signs in and sees their own repair requests and nothing else, not their neighbour's and not the building's. A contractor sees the jobs assigned to them, with the address and contact for those jobs only. An office admin sees every request across the portfolio but not the payroll table. The owner sees everything.

Write that grid down and confirm it with the client in plain words before you build it. "Contractors will be able to see tenant names and mobile numbers for their own jobs. Is that what you want." The conversation costs far less than the alternative.

Two sibling guides go deeper here: backend security when you are not a developer and what you are responsible for with client data.

Building screens people can actually use

Now, and only now, the interface. Softr covers most of what businesses need: a client-facing portal, a staff dashboard, a form that writes into the right table with the right permissions attached. Glide is built around the phone. Bubble goes further and asks more of you in return. A public marketing website is a different job again to the system sitting behind it, and it is worth keeping the two apart in your head.

The test of a screen is not how it looks in a demo. It is whether a busy person can finish the task on a phone, one-handed, while something else is going on. That usually means fewer fields, not more. One obvious next action per screen. Labels in the words the business already uses, not the words your tool uses by default.

A screen with everything on it is easier to build and harder to use.

Automating the flow, including the failures

Automation is what turns a tidy database into something that saves real hours. Make.com covers the ordinary connective work: a form submission creating a record, a status change sending a notification, a weekly summary landing in someone's inbox. n8n handles workflows with branching logic, several conditions and steps that depend on each other. AI agents come in where a step needs judgement expressed in language, such as reading an incoming message and routing it to the right category, or drafting a first-pass reply for a person to approve.

What separates a delivered automation from a demo is failure handling. What happens when the API is down. When a required field arrives empty. When the same form is submitted twice. Every step needs an answer, and someone needs to be told when that answer is triggered.

Checking AI's work before the client sees it

AI output is a draft. It is fluent, it is fast, and it is confidently wrong often enough that shipping it unread is a professional risk you are taking on the client's behalf.

Have a fixed check before anything reaches a client. Does it match the source record, field by field, rather than merely sounding plausible. Is it in the client's voice and terminology. Does it contain anything the recipient should not see. Would you sign your name under it.

Better still, build the check into the system. A draft status that requires a human to approve before an email sends. A review queue for anything an agent classified with low confidence. A log of what was generated and by whom, so a mistake can be traced rather than argued about.

JobTayo assesses this rather than assuming it. Each foundation course ends in a quiz you have to pass before the next one opens: 67% to pass, one sitting, unlimited retakes after 24 hours with the questions reshuffled, best score counts. The specialisations end in a live assessment with a person instead of a quiz, and those are still being built. The gate is there so that what you learn actually holds.

Handing it over so it survives without you

A project is not finished when it works on your screen. It is finished when someone else can run it.

Handover is short and written, and it covers five things: how the data is structured, who has what access, what runs automatically and when, what to do when something breaks, and how to add or remove a user. A five-minute screen recording alongside it does more good than twenty pages nobody opens. Name the person on the client side who owns the system, and walk them through it while they click, not while you click.

This is also how repeat work happens. A system somebody understands gets used and extended. One nobody understands gets abandoned, and the tools get blamed.

AI accelerates the work, it does not own the result

Notice where AI appeared in that sequence. It drafted discovery questions. It sped up the writing. It handled classification and first-pass replies inside a workflow. It did not decide what the business actually needed, how the data should be shaped, who was allowed to see what, or whether the output was fit to send. A person was accountable at every one of those points, and that person is the one the client is paying.

That is the whole shift, and it is why the work is teachable. Mindflows is an agency that teaches rather than a course company: 100+ custom apps delivered for 400+ businesses across 14+ countries and 15+ industries, with 3,000+ people using those systems, and the standing of #1 Softr Premium Partner worldwide. The curriculum comes from those projects, including the parts that went wrong.

In practical terms it is a subscription at ₱999 a month, self-paced, with three foundation courses taken in order. Every tool named in this guide has a free plan, and no coding or degree is required. The whole curriculum is published before you pay for any of it, so you can read what is in it and decide.

Where to go next

• What an AI business app actually is — the thing you are delivering, described concretely
• Backend security when you are not a developer — access rules in more depth
• The skills that outlast the tools — what stays valuable when the software changes
• See the full curriculum