Field notes / AI
The useful question is not where AI can be added. It is where it can remove repetitive thinking without removing human judgement.
Igniter Studio · 2 Oct 2026 · 5 min read
There is no shortage of things you can add AI to.
Inbox assistants. Chatbots. Content generation. Document analysis. Sales summaries. Reporting. Search. Customer support. Internal knowledge bases.
The harder question is:
Which of those things are actually worth doing?
For a small business, the best AI projects are rarely the most dramatic ones.
They are usually focused improvements to work that already happens repeatedly.
A person reads something.
Extracts information.
Compares it with something else.
Writes a summary.
Classifies it.
Routes it.
Turns it into a first draft.
AI becomes useful when it removes some of that repetitive cognitive work without pretending human judgement is no longer required.
A common AI project starts backwards.
Someone discovers a new model or tool and asks:
What can we build with this?
A better starting point is:
Where does the business repeatedly spend time turning messy information into a decision or action?
That is a much more useful question.
Imagine a business receives twenty detailed enquiries each week.
Someone reads every message, identifies the service requested, checks the location, looks for urgency, decides whether it is a good fit and then writes a response.
That process already exists.
AI might help structure the information, classify the enquiry and prepare a response draft.
The person still decides what happens.
That is very different from adding a chatbot because “we should have AI”.
Traditional automation works extremely well when the rules are clear.
If payment status equals paid, send receipt.
If form field equals Perth, assign Western Australia.
If date is seven days away, send reminder.
AI becomes interesting when the input is less predictable.
For example:
These tasks previously required a person because the information could not easily be reduced to a fixed set of rules.
That is where modern AI can be genuinely useful.
There is an important difference between:
AI recommends
and
AI decides
That difference matters.
Suppose an investment research system reads company announcements and highlights potential risks.
Useful.
Suppose it automatically sells an asset based on its own interpretation.
Very different risk profile.
The same principle applies in less dramatic situations.
AI can prepare a customer response, but perhaps a person should approve it.
AI can score an enquiry, but perhaps it should not automatically reject somebody.
AI can summarise a contract, but the summary should not be treated as legal advice.
Good AI products make the human boundary clear.
The goal is not maximum automation.
The goal is appropriate automation.
A lot of useful AI should barely feel like “using AI”.
Imagine an internal dashboard with a button:
Summarise this month's customer feedback
Behind that button, the system might collect hundreds of messages, group themes, detect repeated complaints and prepare a concise report.
The user does not need a chatbot.
They need an answer.
This is often better than placing a large empty chat box in the middle of the interface and expecting people to figure out what to ask.
Chat is useful when the task is genuinely exploratory.
When the task is predictable, purpose-built interfaces are often better.
AI output is heavily influenced by the information available to it.
A generic model asked:
Reply to this customer.
knows almost nothing.
A system that also knows:
can produce something much more useful.
This is why a good AI implementation is not just an API call.
The surrounding software matters.
Which data is retrieved?
What is trusted?
What is optional?
What should never be sent?
What happens when information conflicts?
What is stored?
What requires approval?
The intelligence of the system is partly determined by these decisions.
One underrated feature is uncertainty.
A poorly designed AI feature feels compelled to produce an answer.
A better one can say:
There is not enough evidence to determine this reliably.
Or:
These two documents contain conflicting values.
Or:
This requires human review.
That behaviour is particularly important when the system works with business-critical information.
Confidence should not be manufactured by polished language.
Before adding AI to a workflow, ask:
Does this task happen often enough to matter?
Does it involve interpreting messy or variable information?
Would a useful draft, summary, classification or recommendation save time?
Can a person review important outputs where necessary?
Can we measure whether the result is actually better?
If most of those answers are yes, there may be a good use case.
If the main reason is simply “AI would look cool here”, probably not.
That is not a criticism.
The most valuable AI feature in a business might be the thing that quietly saves twenty minutes every morning.
Or turns a messy weekly process into a two-minute review.
Or makes a large archive searchable.
Or prepares the first draft somebody used to write from scratch.
Those improvements do not require a humanoid assistant floating around the screen.
They require understanding the work, choosing the right boundary between software and human judgement, and building something people actually want to use.
That is where AI becomes useful.
Not when it is everywhere.
When it is in the right place.
Your next chapter starts here
Tell me what you’re working on. I’ll help you find a practical way forward.
Let’s talk about it