Guide
Software or AI: Four Tests We Apply Before Recommending Either
Much of what is proposed as AI would be cheaper and more dependable as conventional software. We put each proposal through four tests on your own cases before committing budget.
4-minute read2 November 2026
In Brief
- Conventional software suits work whose rules can be written down and whose answer must be right every time. AI suits unstructured, variable inputs where a measured error rate is acceptable or can be caught.
- We see four tests: can the rules be written down, what does an error cost, how variable are the inputs, and can you measure the result. Most strong designs combine the two.
- Thinqmesh tests each initiative against your own cases, where the answer is already known, and recommends conventional software in writing when the evidence favours it.
Boards now ask where AI is being used. It is a reasonable question, and it can distort priorities, because it invites teams to look for places to apply AI when the better start is a problem worth solving. A meaningful share of what is proposed as AI would be cheaper, faster and more dependable as ordinary software. We help you tell the difference before budget is committed.
The Two Have Different Economics
Conventional software, once correct, stays correct for a given input. An AI system gives a statistically good answer, which needs measuring, monitoring and a plan for the cases it gets wrong. That overhead pays for itself only when the problem requires it.
Choosing wrongly costs money in both directions. AI applied to work with clear rules adds a running cost and an error rate that software would not have had. Software applied to work with endlessly varied inputs adds rules that someone must keep rewriting. The strongest designs give each the step that suits it: AI reads the invoice, the email or the scanned form, and software validates, applies the rules, routes the approval and keeps the record.
Four Tests We Put Every Proposal Through
Each is answered with the people who do the work, using real cases, and the result is a recommendation you can act on.
Can the Rules Be Written Down?
If a competent employee could write out the steps and the exceptions, the work suits conventional software.
Tax calculation, payroll, approval routing and most reporting follow rules. A report that no one has built is often called an AI opportunity when the real gap is that the data was never brought together. Where the rules resist being written, because the work is reading language or recognising a pattern, AI becomes a candidate.
Get started
- Ask the person who does the task to write the steps and the exceptions.
- Note any step they cannot put into words.
What Does an Error Cost?
Statutory filings, payments and safety decisions need the answer right every time.
AI may assist with these, and the controls around it are designed before the system is, with a person as the final authority. Where a mistake is inexpensive and easily corrected, such as a first sort of incoming requests, a measured error rate can be perfectly acceptable. The question is which mistakes can be tolerated and who catches the others.
Get started
- List the mistakes in this process that would reach a customer or move money.
- Name who would catch each one.
How Variable Are the Inputs?
The same form every day needs no model. Documents in hundreds of layouts may.
Variability is the clearest sign that fixed rules will become expensive to maintain. A process that receives documents from hundreds of suppliers in hundreds of layouts is a fair candidate for a model that can generalise. A process fed from a structured form is better served by software.
Get started
- Collect twenty recent inputs and count how many different layouts they use.
- Mark the fields that change from one to the next.
Can You Measure It?
An AI system enters production only when real cases with known answers exist to score it.
We build a test set from your real cases, where the correct answer is already known, and produce a measured score on your own files with the mistakes named. If the cases cannot be assembled, the right first step is the data work that would make measurement possible.
Get started
- Gather fifty past cases where you know the right answer.
- Agree who will check the system's answers against them.
How Thinqmesh Solves This
We begin with the operational questions: which process justifies the investment, how accuracy is established, and who remains accountable. A defined scope on an established platform is delivered in four weeks.
- Week 1
A Test Set and a Measured Score
We build a test set from your real cases and produce a measured accuracy score on your own files, with the mistakes named.
- Week 2
The Right Tool for Each Step
Each step is assigned to software or AI by the four tests. Where the evidence favours conventional software, we say so in writing.
- Week 3
One Job Built From Start to Finish
One job is built end to end on your own system, with a person checking the work wherever a wrong answer costs money.
- Week 4
Running, With Accuracy in View
The job goes live, with ongoing checks that keep accuracy in view after launch and a way for your team's own AI tools to ask the system questions.
You finish with one job automated on your own system, a score on your own files, a person checking where errors are costly, and checks that keep accuracy in view.
