---
title: Eight trade-offs before you build an AI system | Fractals Collective
description: Eight choices anyone building an AI system has to make, each set out as a straight comparison: what you get on one side, what you give up on the other.
source: https://www.fractals.sg/decisions/
---

[fractals](https://www.fractals.sg/)

# Eight trade-offs before you build an AI system

A working document from Fractals Collective

Every one of these is a straight choice between two things you can have, and you cannot have both. What changes is the price, and the price is not always money.

Read each pair by asking one question: **what do I give up if I take that side?** There are no right answers. There are only costs you are willing to carry.

01

## Where your data lives

Buying a building, against renting an office. One is yours to keep and yours to fix.

Your own machine

It runs on hardware you control, in a room you can point to.

You getConfidentiality you can defend in front of a regulator, and nobody else holding your records.

You give upThe upfront bill, and your evenings as the person who looks after it.

Someone else’s servers

Rented infrastructure from a company whose job is running it.

You getWorking this week, and room to grow without buying anything.

You give upYour data sits on their machines, and their outage becomes your outage.

vs

**The middle option** — keep the sensitive records on your own machine and rent everything else. You get most of the privacy and most of the convenience, and you now run two systems.

**Ask yourself** — Whose information is in this, and what would they say if they knew where it was going?

02

## Whose intelligence you use

Booking the best chef in the city for one night, against hiring a cook you keep on staff.

Rented intelligence

You rent access to the most capable models available today, over the internet.

You getThe best answers on day one, and no model to look after.

You give upThe price and the behavior are theirs to change. Your system can get worse without you touching it.

A model you run

You download one and run it on your own hardware.

You getA fixed version that never shifts under you, and a flat cost once you are running at volume.

You give upCapability. It is the good cook, not the famous one, and you have to host it.

vs

**The middle option** — the rented one for hard reasoning, your own for routine and sensitive work. Two systems, two bills, and a rule about which task goes where.

**Ask yourself** — If the answers changed next week without anyone telling you, what would break?

03

## How hard it is to leave

A fitted kitchen, against furniture you can move when you change your mind.

All in with one supplier

Everything built around a single provider so it works out of the box.

You getThe fastest setup, and one company to call when something breaks.

You give upLeverage. If their price doubles next year, moving house is a rebuild.

Able to swap

Built so the engine underneath can be replaced with little disruption.

You getThe option to leave, which is the thing that keeps a price honest.

You give upMore work at the start, and the job of keeping it genuinely swappable.

vs

**Ask yourself** — If they doubled the price next year, what would you do?

04

## When it works

A taxi you hail, a bus on a timetable, or a night watchman on duty.

When you ask

Nothing happens until someone types something.

You getNo bill while it is idle, and a person in the loop by default.

You give upAnything that has to happen without you, which is most of what makes these systems worth having.

All the time

It watches for something and reacts when it sees it.

You getWork that continues while you sleep, which is the whole promise.

You give upQuiet. Something is always running, so something is always able to go wrong.

vs

**The middle option** — a schedule. It does its one job every morning at seven. Most of the value of always-on, at a fraction of the risk, and it fails the same way every day until you notice.

**Ask yourself** — What has to happen while you are not looking?

05

## How much it decides

A recipe, against a chef. The recipe is dependable and blind. The chef can improvise and will one day surprise you.

It follows a script

The same steps, in the same order, every single time.

You getOutput you can check line by line, and a result you can defend.

You give upAnything messy. It breaks the moment reality stops matching the script.

It works it out

You give it a goal and it chooses the path to get there.

You getOpen-ended work actually handled, including the cases you did not think of.

You give upThe ability to pre-approve every step, because you will not see them coming.

vs

**Ask yourself** — When it gets something wrong, how will you know?

06

## Who signs off

A junior who drafts the letter for you, against one who signs it and sends it.

You approve

It prepares the email, the record or the action, then waits for a person.

You getNothing reaches a customer without you having seen it.

You give upMinutes every day, and the speed that made the idea attractive.

It acts

It sends, files and replies without asking.

You getWork that finishes while you are somewhere else.

You give upThe gap between a mistake and its consequences, which drops to zero.

vs

**The middle option** — automatic where a mistake is cheap, approved where it is not. It only works if you write the line between the two down, because the system will not know where it is.

**Ask yourself** — What is the worst thing it could send without you seeing it?

07

## How you get it built

Off the peg, against a tailored suit. One is on you by Friday and fits well enough. The other fits properly and takes weeks.

Buy it ready-made

An existing tool, configured around your workflow.

You getSomething running in days, at a price you can see before you start.

You give upPer-seat fees that grow with use, and a ceiling on anything it does not already do.

Build your own

Made for your exact workflow, and owned outright by you.

You getNo platform markup, and something that changes the day your business does.

You give upA larger bill at the start, and a thing that only you can look after.

vs

**Ask yourself** — Who will still be looking after this in two years?

08

## Whether it remembers

A colleague who reads the file fresh each time, against one who remembers the last conversation you had.

It forgets

Every task starts clean, with only what you hand it.

You getThe cheapest and cleanest version, with no stale information to manage.

You give upAnything that needs yesterday, and anything that takes more than one sitting.

It remembers

It keeps client history and picks up a job where it left off.

You getA system that becomes more useful the longer it runs with you.

You give upData to maintain, and the risk of a system that is confidently wrong about something it learned last month.

vs

**Ask yourself** — What should it forget?

## Where these choices turn into each other

Each pair is fine on its own. The damage happens in the combinations, and three of them account for most of the wreckage.

**All the time, acts on its own, mistakes are cheap.** You have not built what you think you have built. You will hear about the failure from a customer before you hear about it from the system.

**Yours, built for you, nobody maintaining it.** Four good months, then it stops matching how the business actually works.

**Rented, drafts only, mistakes are cheap.** Perfectly safe and worth almost nothing. This is the most common outcome, and the reason most people conclude that AI does not work when what they tried was too small to tell.

The question underneath all eight is smaller than any of them. What is the one repetitive thing you want to stop doing? Not a category, one task, in a sentence. If you cannot say it that plainly, no answer on this page will help yet.

**Mark your own eight and see what you have agreed to**

Pick a side on each row. You get back the total bill, which is every concession you just made, written in one place.

Where the data lives

OursRented

Whose intelligence

RentedOur own

How hard to leave

Locked inSwappable

When it works

When askedScheduledAlways

How much it decides

ScriptWorks it out

Who signs off

You approveTieredIt acts

How it gets built

Ready-madeBuilt for you

Whether it remembers

ForgetsRemembers

What you have agreed to

If you want to work through where these land for you, that is what [the session](https://www.fractals.sg/ai-sessions/) is for: 45 minutes, US$39, and you leave with your own eight marked and the first job planned.

Fractals Collective

Made deliberately small in Singapore.

[fractals.sg](https://www.fractals.sg/)[alan@fractals.sg](mailto:alan@fractals.sg)[LinkedIn](https://www.linkedin.com/in/alansoon/)
