There are plenty of places where AI is useful.
It can read messy inputs, summarise long conversations, classify requests, draft first versions, extract details from documents, and help staff deal with volume without losing context.
But there are also plenty of places where AI is the wrong answer.
Not because the technology is weak. Often it is perfectly capable in a narrow sense.
The problem is that the business is asking AI to compensate for decisions, structure, or discipline that do not exist yet.
That is where AI projects become expensive, confusing, and hard to trust. The team expects automation. What they get is a faster version of the same ambiguity.
Do Not Use AI To Avoid A Process Decision
A common pattern is easy to spot.
The business has a workflow that depends on informal judgement. Different people handle the same case in different ways. Exceptions are not written down. Ownership is unclear. Nobody is quite sure when work should move from one stage to the next.
Then someone asks whether AI can decide what happens.
Sometimes it can assist.
But if the business has not defined the rules, AI is being asked to invent them.
That is not automation. That is delegation of an unresolved operating decision.
Before using AI, ask a blunt question: could a competent new hire follow this process from written guidance?
If the answer is no, AI will probably struggle too. It may still produce an answer, but the answer will depend on patterns, prompts, examples, and assumptions rather than a clear business rule.
That is a poor foundation for anything operationally important.
Do Not Use AI Where The Input Is Bad For Boring Reasons
AI can handle messy language. It is good at turning unstructured text into useful structure.
That does not mean it should be used to clean up every bad input.
If customer records are incomplete because nobody owns the CRM fields, fix ownership.
If job notes are vague because staff do not know what detail is expected, fix the capture process.
If product names are inconsistent because there is no controlled list, fix the data model.
If every order arrives differently because the sales process promises anything to get the deal over the line, fix the commercial handoff.
AI can help at the edges. It can suggest values, flag missing information, standardise language, or route uncertain cases to a queue.
But if the source of the bad input is a weak process, AI becomes a permanent correction layer.
Sometimes the cheaper fix is not smarter technology. It is a better field, a required checklist, a clearer handoff, or one agreed source of truth.
Do Not Use AI When A Simple Rule Would Do
There is no virtue in using AI for work that a rule can handle cleanly.
If an invoice is overdue by more than seven days, send a reminder.
If a lead has no next action, create a task.
If a support ticket contains a known category, assign it to the right team.
If a required field is missing, block the stage change.
Those are not intelligence problems. They are workflow problems.
Rules are easier to test, explain, monitor, and maintain. They fail in more predictable ways. They are also easier for staff to understand, which matters when the system starts affecting their day.
AI is useful when rules become brittle because the input is varied or judgement-heavy.
Use AI where the work genuinely involves interpretation. Do not use it to make a straightforward process look more sophisticated.
Do Not Use AI Where Accountability Must Be Crystal Clear
Some decisions need a clear owner.
Pricing exceptions. Credit decisions. Contract commitments. Compliance-sensitive responses. Employment decisions. Anything that materially affects a customer, supplier, employee, or legal position.
AI can support these workflows. It can summarise, prepare, compare, check for missing information, or draft a recommendation.
But the business still needs a named human decision-maker and a clear audit trail.
The worst version is a vague approval process where AI generates something, staff assume the system knows best, and nobody is sure who is responsible when the output causes a problem.
If accountability matters, design the human review properly before adding AI.
Where AI Does Make Sense
AI is strongest when three conditions are present.
First, the task involves messy information: emails, calls, PDFs, notes, forms, support conversations, or documents that do not arrive in a neat structure.
Second, the business can define what good output looks like: the fields to extract, the summary style, the category set, the confidence threshold, the escalation path.
Third, there is enough volume or repetition to justify the effort of building, testing, and maintaining the workflow.
That is where targeted AI automation can be genuinely useful.
Not as a strategy in itself. As a practical tool inside a well-understood process.
Examples include extracting order requirements from inbound emails, summarising sales calls into CRM updates, triaging support requests, checking documents for missing information, or turning long internal notes into consistent handover summaries.
Even then, the important work is the workflow around it: validation, fallback, review, monitoring, ownership, and the decision about where AI should stop.
The Practical Test
Before starting an AI project, ask five questions.
- What exact operational problem are we trying to remove?
- What decision, handoff, or rule is currently unclear?
- Could a simple workflow change solve most of it?
- What should happen when the AI is uncertain or wrong?
- Who owns the outcome after the system goes live?
If those questions are hard to answer, the project is not ready for AI yet.
That does not mean doing nothing.
It means doing the useful groundwork first: map the workflow, remove ambiguity, clean up the source of truth, define the exceptions, and identify the smallest place where AI would actually reduce friction.
At Intelligent Marmalade, that is the point of an operational friction review.
The aim is not to find places to use AI.
The aim is to find where work is slow, duplicated, unclear, or unnecessarily dependent on individual memory, then choose the most robust fix.
Sometimes that fix is AI.
Sometimes it is an integration, a small internal tool, a cleaner process, a better data model, or a decision that leadership has avoided making.
The valuable question is not "where can we add AI?"
It is "what is the real constraint, and what is the simplest reliable way to remove it?"