How to Build a Really Bad AI Agent
Here is what you want to do if you want to make a really bad AI agent for yourself. I am not being entirely sarcastic — I watch businesses do every one of these.
Step one: choose something you hardly ever do
Preparing your tax. Auditing the business for a new compliance regime. Pick one of those and just go ahead and point an agent at it.
Whatever you do, do not choose anything repeatable. Do not choose anything that happens in the normal running of your business.
Step two: make sure there is no process document
The knowledge should live in one person's head. You want to be talking to Bob from the department so you can extract what he knows and feed it into the agent as you go.
Step three: do not write down your thinking
This matters most on problems of classification. Just let the AI figure out what "good" means.
Say you are building a marketing agent. Ask it to work out the best market for you to enter, and give it no information about your business or what you actually do.
Step four: choose a process with no margin for error
This might be getting the AI to control the brakes on your car. It might mean opening up all of your banks to the agent so it can control your business finances.
Step five: let it run by itself
No human looking over it. Then congratulations — go take the day off, make yourself a nice cuppa, or knock off and go down to the pub.
Just kidding. Please do not do this.
Those five things are a recipe for ruining your experience with AI. Go down that route and you will transform your business into one that is no longer operating.
Now invert all five
Take a process that is repeatable. Something that runs in the normal course of business, not once a year.
Make sure the process is already documented. If it is not, that is the real first job — and if you do not do it, people like me end up doing it for you anyway.
Write down your thinking behind the process. This is the step almost everybody skips. If you have a classification problem — say you are working out which leads are worth pursuing — you need to understand what actually makes a good lead. Is it someone already in your industry? Someone you have connected with in the past?
Include your exact reasoning, so that when the AI thinks about these things it knows which thoughts to pick up on and what it is supposed to be doing. That reasoning is usually the part missing from documentation, and putting it in is an easy way to make your agents ten times better at the work.
Leave a little margin for error. AI is never going to get everything one hundred per cent right. There is a cycle involved in perfecting what you are asking, and the thinking step above is what powers that cycle. But you have to start with a process that can tolerate a wrong answer on the way through.
Keep a human over the output. Have someone check the result — even if an agent does the first pass of that checking — to make sure the output is right.
Skip those steps and you will, boom, have a bad AI agent.
What this means for choosing an AI partner
When you are looking at AI partners, this is the work you want to be doing with them. You do not want to just be hands off. That is going to be being a bad business owner.
If you are struggling with where an agent actually fits, come and have a chat with us. We start with people and process before we go anywhere near the technology.