Education Centre Guides & Checklists
Building custom software in the age of AI: a guide for business owners
AI has made software dramatically faster and cheaper to build. It hasn’t made it easier to decide what to build, or to run it well once it exists. This guide covers what has changed, how a modern build fits together, and a six-step way to think about it before anyone quotes.
At a glance
This Promptwise Advisory guide by Adam Agresta explains how to approach custom software now that AI can write much of the code: what changed (Software 3.0), the six parts of a modern build, why a demo is not a product, the autonomy slider for AI, Retool’s 2026 build vs buy findings, and six steps to take before anyone quotes.
- The six parts
- People, screens, workflows, data, the AI layer, connections
- Key idea
- A demo only has to work sometimes; a product has to work every time
- Start with
- The whole business, not the software
- Data source
- Retool 2026 Build vs. Buy Report, 817 builders
What changed
Software 3.0, in plain English
In a well-known June 2025 talk, AI researcher Andrej Karpathy described three eras of software. First, people wrote every line of code by hand. Then, models were trained to do specific jobs. Now, a lot of software is built by describing what you want to an AI model in plain language.
For a business owner, that means two things. A first version of an app can now take days rather than months. And the hard part has moved: the value is in knowing exactly which problem to solve, what the software is allowed to do, and how a person checks its work.
What businesses are doing
Custom builds are no longer just for big companies
Retool surveyed 817 people who build software inside their organisations in late 2025. The sample leans towards Retool’s own customers, but the direction is clear.
Source: Retool, 2026 Build vs. Buy Report.
Under the hood
The six parts of a modern build
Whether it’s an internal dashboard or a customer app, every build has the same six parts. The four in the top row are the core of the app. The AI layer and the connections run across all of them.
The people
Who uses it, and what each person is allowed to see and do.
- Customers
- Your team
- Managers
- Partners
The screens
Where people use it: a web app, a mobile app, a portal or simply a chat window.
- Web app
- Mobile app
- Portal
- Chat
The workflows
Your business rules: what happens, in what order, and who approves it.
- Bookings
- Follow-ups
- Approvals
- Reporting
The data
The single source of truth: customers, jobs, payments and files, owned by you.
- Records
- Files
- History
- Permissions
The AI layer
Models that read, write, sort and decide, inside guardrails you set, with a person checking anything sensitive.
- Drafting
- Sorting
- Answering
- Human check
The connections
Links to the tools you already pay for, so nobody copies data between tabs.
- CRM
- Accounting
- Calendar
- Payments
The trap to avoid
A demo is not a product
Karpathy’s warning applies to every business build: a demo only has to work sometimes, while a product has to work every time. AI closes the first gap fast. The second still takes care.
A demo
- Works on the happy path
- One user, sample data
- Nobody checks the AI’s answers
- Built in days
A product
- Handles mistakes, edge cases and bad data
- Real users, permissions and security
- A person checks anything sensitive
- Owned, maintained and measured
Keep AI on a leash
Decide how much the AI does on its own
Karpathy calls it an autonomy slider. AI is brilliant at some tasks and makes odd mistakes on others, so start with it suggesting or drafting, and only move it to the right once it has proved itself on that task.
1
Suggest
AI suggests, a person does the work.
Lowest risk
2
Draft
AI drafts, a person checks and sends.
Most common start
3
Act and report
AI acts on routine work and reports back.
Once it has proved itself
4
Act alone
AI handles it end to end.
Only for low-risk, well-tested tasks
The Promptwise way
Six steps before anyone builds
This is how I approach it with clients. It starts with the whole business, not with the software.
- 01
Map the whole business first
Sales, marketing, operations and service. The best automation opportunities usually sit between teams, not inside one tool.
- 02
Find the hours and the leaks
Where is repeatable, manual work eating time, and where are leads, money or customers slipping through? Put a number on each one.
- 03
Buy, connect, then build
Take the cheapest path that works. Buy if a proven tool fits, connect what you have, and only build what nothing else does.
- 04
Prototype small and prove it
AI makes a first version quick and cheap. Use it to test the idea with real users, and budget for the gap between a demo and a product.
- 05
Decide how much the AI does alone
Start with AI suggesting or drafting and a person checking. Give it more independence only once it has earned it.
- 06
Own it and measure it
The code, data and accounts in your name, a plan for maintenance, and a clear measure of hours saved or revenue won.
The thinking behind the build
Eight questions to answer before anyone quotes
AI can write code quickly, but it can’t decide these for you. Answer them first and any quote you get will be sharper and fairer.
01
Who will use it?
Customers, staff, managers, partners? Each group needs its own screens and permissions, and each one adds testing.
02
What data does it hold, and who owns it?
Customer records, bookings, payments, files. Decide what’s stored, for how long, and make sure it’s yours, not the builder’s.
03
How do people log in?
Email and password, Google or Microsoft sign-in, two-factor. The more sensitive the data, the stronger this needs to be.
04
What does it need to connect to?
Your CRM, accounting, calendar, payments or email. Ready-made connectors are quick. Anything else needs custom work.
05
Where does AI fit, and where doesn’t it?
Which jobs AI does, what it’s allowed to see, who checks its work, and what each AI task costs to run.
06
Privacy and compliance
Where the data is hosted, what the Australian Privacy Act asks of you, and any rules specific to your industry.
07
What will it cost to run?
Hosting, AI usage and third-party subscriptions are separate from the build, so ask for them upfront.
08
Who maintains it?
Software needs looking after. Agree who fixes, updates and improves it after launch, and what that costs.
What makes a build simple or complex
Complexity comes from a handful of things, and most of them are visible before a line of code is written. That’s why builds are quoted only after scoping.
Keeps it simpler
- One type of user
- A few screens and a simple form or dashboard
- One or two ready-made connectors
- No payments, or a simple payment link
- New data, nothing to migrate
Adds complexity
- Several user types with different permissions
- Real-time updates, syncing or offline use
- Several integrations, some without connectors
- Money moving between parties, refunds or subscriptions
- AI making judgements that need checking
- Sensitive data, compliance or industry rules
- Years of old data to clean and move across
- Native mobile apps in the App Store and Google Play
Watch out for
Three risks that come with building faster
Tools nobody knows about
In Retool’s survey, 60% of builders had built something outside IT’s oversight in the past year. Keep a list of what exists, who owns it and what data it touches.
The maintenance bill
Time saved can be eaten by fixing, updating and securing a growing pile of small apps. Fewer, connected builds beat many disconnected ones.
Code nobody checked
Only 8% of builders use AI-written code as is. Make sure someone experienced reviews and tests it before it touches customer data.
Before you sign
Questions to ask any builder
Whoever builds it, these answers protect you. A good builder will answer every one of them in writing before starting.
- Will I own the code, the accounts and the data?
- Where will my data be hosted?
- What will it cost to run each month, and what makes that go up?
- How are changes made after launch, and what do they cost?
- How is security tested before we go live?
- Who reviews the AI-written code?
- What happens if we stop working together?
- Can I see a working prototype before committing to the full build?
How Promptwise helps
My scoping isn’t just about software. I look at the whole business first, then work out where AI and automation should and shouldn’t go.
Step 1
Scope the whole business
Sales, marketing, operations and service: how work flows today, where time goes and where revenue leaks.
Step 2
Flag what to automate
The areas that could and should use AI or automation, ranked by the hours or revenue at stake, and the ones to leave alone.
Step 3
Introduce the right builder
One of my trusted AI and build contacts, with a clear written brief. Any referral fee is disclosed in writing first.
Step 4
Stay involved, or build it myself
I can check the work against the plan by the hour, and in some cases I can set up the solution myself.
Free · 3 minutes
Find out what’s worth building
The AI and custom builds check shows where your team is losing hours, and whether a build would pay off.
FAQ
Frequently asked questions
How has AI changed building custom software?
AI can now write much of the code from plain-language instructions, so a first version can take days rather than months. The hard part has moved to deciding exactly what to build, what the AI is allowed to do and how a person checks its work.
What are the parts of a modern custom build?
Six parts: the people who use it, the screens they use, the workflows and business rules, the data it holds, an AI layer with guardrails, and connections to the tools the business already uses.
Are businesses replacing software subscriptions with custom builds?
Increasingly. In Retool’s 2026 Build vs. Buy Report, 35% of teams had replaced at least one SaaS tool with a custom build and 78% expected to build more in 2026.
What is the autonomy slider?
A way to decide how much AI does on its own: from suggesting, to drafting for a person to check, to acting and reporting back, to acting alone on low-risk, well-tested tasks.
Does Promptwise Advisory build custom software?
Adam Agresta scopes the whole business, flags where AI and automation should help, then introduces a trusted AI or build contact. In some cases he sets up the solution himself. Any referral fee is disclosed in writing.
Thinking about a custom build?
Start with a free 30-minute chat, over coffee in Sydney or on Google Meet. We’ll work out whether you need to build at all.