Don’t put AI in everything you build
As you look to embed AI thinking and practices across your business, its important to know when AI is a sensible part of your solution and when it is not.
In the stampede to sprinkle some AI magic on to every process you can quickly end up creating over-engineered solutions that are expensive, slow and inaccurate.
I use this matrix to explain the point.
On the horizontal axis you have two approaches to building tools and products - you can use AI tools like v0, Cursor, Lovable, Bolt, Windsurf and Replit to build your new tools, or you can not and go slower.
On the vertical axis you have two approaches to running your tools and products - you can use AI as part of the solution, or you can not and have a ‘traditional’ deterministic application or script.
Bottom left (Extinct) - Not using AI to build or to run. If you take this approach you are falling behind. It will take you too long and cost too much to build anything and what you do build will not provide the value that your users and customers deserve and expect.
Top right (Category Crushers) - these are the companies winning right now, using AI to build products at a rapid rate and delivering AI functionality that their users demand - you can put v0, Cursor, Lovable, Bolt, Windsurf and Replit all in this bucket - they are seeing revenue growth that has never been seen before.
Top left (Missed Opportunity) - this is the unrealised potential, the companies that are bolting on an AI-widget to their core product because it is the buzzword of the day, but they are not embracing Claude, v0, Cursor in their own development practices. A two year roadmap is too slow. Move faster.
Bottom right (Quick Wins) - for most non-tech companies, and non-tech teams within companies this is the biggest opportunity to get quick wins. Here we are using AI to build tools and mini products for our own teams quickly, but we are not looking to use AI in the finished solution.
Let me give you a quick example.
A finance team I am working with have a very manual reconciliation process that they need to complete at the end of every financial period.
It requires them to export three large reports from their ERP Netsuite.
One of their team needs to manually manipulate the files - aligning data, and then matching data across the three tables using a PO Number as a unique key.
Having matched the data, they then need to create a fairly complex pivot table and report on each supplier across different financial periods so that they can provide data required by their auditors.
This very manual process is ripe for automation - but is it ripe for AI?
The process itself is very deterministic. What I mean by that is there are predefined steps that need to be repeated. There is no creativity required in the process - put the same three files in 100 times and we should get exactly the same output.
Therefore what we need to create is a more traditional automation or workflow, as opposed to an AI powered creative solution.
But that doesn’t mean we can’t use AI to get there.
In this situation, the finance team don’t have any development experience, nor do they have access to development resource at a price that would make sense for automating this simple process.
But with AI we can shortcut this.
Here’s the process we followed.
Document the process step by step as if you were telling an intern
We wrote it in human language in explicit detail, “In Excel 1 row 7 and 8 contains the column headings. Column row 7 has Date, Number, Payee…, Row 8 has Type, Account, Memo. We need to merge these two rows, and every subsequent set of two rows so that each record is on one row with these column names….”
We broke each step down into its constituent parts so that a human would not get confused or stuck.
We explained every acronym, term or field name so there was no confusion
This created a ten page document, which if we had provided it to an intern, they could now take on the manual task from the finance team.
Remember: if an intern couldn’t follow your instructions and give you the result you want, it is highly unlikely that an AI model would be able to either.
Ask AI how to approach the task
Now we have the ‘instruction manual’ we can turn to our AI co-worker (Claude, ChatGPT or your preferred model) to ask for their suggestions on how to approach it.
I use the example of crazy golf in these situations.
If you’ve played crazy golf you will know there is always an obstacle like a windmill or a wall in the way.
If you aim straight at the pin you will fail - you need to aim away from the pin at a wall or other obstacle to bounce off and towards the pin.
This is how I approach my AI-co-worker.
If you ask AI “Here are the instructions for this task. I am now going to upload these files and I want you to do the job” then this is like aiming straight at the pin.
The AI will try to do a good job, but for a complex task will likely get stuck, make a mistake, and even if it gets it right, you don’t have a repeatable process for the next time you need to do this task.
So instead, ask a different question.
“I have documented a task that I need to regularly complete. I am not a developer, but I would like your help suggesting ways that I could approach this to create a repeatable tool I could use to complete the task. Here is the task {add in your document}”
Provide as much detail about your needs:
Is it just you, or would others use this?
How often would you use it?
What existing systems or platforms do you use?
Now you’ll get some valuable suggestions for how you (as a non-developer) can solve this challenge.
Once you have determined an appropriate path forward ask your AI co-worker to provide a step by step “project plan” for putting the solution into action. Remind it that you are not a developer and if it suggests anything you don’t understand, just ask it to explain.
Build your solution
Depending on the complexity of your request and what tools you have access to, you can now try and build your solution.
There are a number of tools that can help you on your way here:
are three good starting points.
I won’t go into the detail of using these tools in this article, but they are each simple enough that if you have your ‘instruction manual’ from the first step, and you have your “project plan” from the second step, then you will be able to provide these inputs and you will be amazed at how far you get in just an hour of discussion with these platforms.
For our finance controller, they now have a simple interface to drop in their three files and in under 5 seconds out pops their reconciled, cleaned, pivoted file.
No AI is used in the ongoing running of this process. AI just helped us to get there.
Not only does the finance team have a great solution, they now know that the path to solving the many other manual, frustrating, error prone processes is simpler than they could have imagined.
For clients I work with, we keep this matrix in mind, and are hunting down those quick wins where we can use AI to improve the experience and effectiveness of their teams with AI, but not necessarily using AI in the final solution.
Get Started
Whenever you are ready there are three ways I can be helpful:
YouTube Channel: Subscribe for regular updates on AI developments
AI Inspiration Briefing: Show your people the path forward in this 60 minute live session
Kowalah: AI-powered buying platform for technology buying groups