Skip to content

Strategy & Systems Design

AI opportunity analysis: a model only where a rule can't decide

Most steps in a business don't need AI. If the decision can be written as a rule, ordinary software applies it every time, for free, and can explain itself. A model earns its place where the input is language, photos or documents and no rule can decide. Where a wrong answer is costly, a person checks it. And where nobody has decided the rule yet, the business decides it first.
Talk through where AI might helpSee six candidates sorted ↓
now sorting: invoices over $2,500 wait for approvalnow sorting: who gets priority in a heatwavenow sorting: sort the office inboxnow sorting: read a furnace label's model numbernow sorting: “is my repair under warranty?”yesnonoyesnoyesCan a written rule decide it?Is the input language, photos or documents?Is a wrong answer costly?A rule: ordinary softwareinvoices over $2,500 wait for approvalthe rule is already writtenDecide the rule firstwho gets priority in a heatwavenobody has decided yetAI, checked by rulessort the office inboxa misfiled email is easy to moveAI drafts, a person decidesread a furnace label's model numberthe technician confirms“is my repair under warranty?”the office checks the draft

Six candidates, sorted

The steps worth changing from the automation analysis, each asked whether it needs a model at all.
Invoices over $2,500 wait for approvala written rule: the amount, or a discountordinary software
Reminders the day beforethe appointment timeordinary automation
Who gets priority in a heatwavea business choice nobody has made yetthe owner writes the rule; software applies it
Sorting the office inboxfree-text emails; a misfile is easy to moveAI sorts, rules check, unsure ones stay with the office lead
Reading a furnace label's model numbera photo; a wrong part means a second visitAI reads it; the technician confirms before parts are ordered
Warranty questionsthe customer's words and the warranty terms; a wrong answer costs trustAI drafts from the terms; the office checks before sending

Signs AI is the wrong tool

Any one of these, and we'll recommend something simpler.

  1. 01

    The rule can be written down

    One sentence describes it

    Software applies it the same way every time, costs nothing per decision and can show why.

  2. 02

    Nobody has decided the policy

    A business choice, not a reading task

    A model will guess at a choice the business hasn't made. Decide it, then write it as a rule.

  3. 03

    The input is already structured

    Amounts, dates, statuses

    There's nothing to read, only things to compare.

  4. 04

    It must be exactly right, with nobody checking

    No review, high cost

    Models are sometimes wrong. If nobody can review the result and a mistake is expensive, that's too risky.

  5. 05

    There's no way to tell if it's working

    No examples of right answers

    Without real examples to test against, nobody will know when it starts getting things wrong.

What each AI use needs before it's built

A short brief for every step that gets a model. If a line can't be filled in, it isn't ready.

Sort the office inbox

Input
each new email: sender, subject and text
Output
one of: booking, billing question, complaint, other
Checked by
rules: an email from accounting is never a booking
When unsure
left in the inbox, marked for the office lead
Measured by
a weekly sample checked by the office lead
Owner
the office lead

Where the analysis sits

In

Steps worth changing
from the automation analysis
Real examples
emails, photos, questions
Cost of a wrong answer
money, visits, trust
Verdict and brief
for every step

Out

Ordinary software
the rules
AI steps
with checks and a fallback
A person
where answers cost

Worth doing before anyone builds an AI feature

Good fit

  • You've been told to “add AI” and want to know where it would help.
  • Your staff read and sort a lot of emails, photos or documents.
  • You want AI only where it earns its place, with someone accountable for it.

Not the right fit

  • Every decision you have in mind can be written as a rule. Software does those better.
  • You want a chatbot because others have one, without a job for it to do.

Where this capability is used

Next to AI opportunity analysis

Questions about AI opportunity analysis

Do you always recommend AI?

No. Most steps we look at are better served by rules or ordinary automation. AI goes where the input is language, photos or documents and no rule can decide.

How do you know an AI step works?

Before building, we collect real examples with the right answers. They become the test it has to pass, and the measure we keep checking after launch.

What if it gets something wrong?

Every AI step has a check after it and a path for when it's unsure. Where a wrong answer is costly, a person decides.

What about AI agents?

Only where the system has to choose among steps nobody can list in advance. Most business tasks work better as one structured step with checks.