AI Opportunities for Business: Stop Looking for Use Cases. Start Looking for Friction
- 11 minutes ago
- 8 min read
One of the first questions businesses tend to ask when they start taking AI seriously is:
“What could we use AI for?”
It sounds sensible however it is often the wrong place to start.
It encourages people to begin with the technology and then search the business for somewhere to put it. That is how you end up with an AI chatbot nobody needed, an agent looking for a job, or a collection of disconnected experiments that are interesting but don’t materially improve how the business operates.
There is a simpler way to find useful AI opportunities. Look for friction.
Look for the work people repeatedly complain about. Look for information they have to hunt down. Look for things copied between systems, emails that have to be summarised, documents that need to be checked, reports that need to be assembled manually and processes that spend more time waiting for information than actually doing anything with it. This is usually where the useful opportunities are hiding.
The best AI opportunities for business often start with ordinary operational friction rather than a new AI capability.
AI opportunities often look like ordinary operational problems
A good candidate for AI rarely announces itself as an “AI use case”.
“Every Monday morning, I have to pull this information together from four different places.”
“Someone has to read all of these emails and work out which ones need action.”
“We enter the customer details here, then someone enters basically the same information into another system.”
“Sarah knows how to do it.”
None of those statements contains the word AI. All of them tell you something useful about how work actually happens.
They expose repetition, information retrieval, interpretation, hand-offs, duplicated effort or dependence on knowledge that exists primarily in someone’s head. Which are much better starting points than sitting in a meeting trying to brainstorm ten things your business could do with the latest AI model.
AI opportunities for business often look like ordinary operational problems
As enterprise AI adoption has matured, the emphasis has started shifting from simply giving people access to AI towards incorporating it into repeatable workflows. OpenAI’s 2025 enterprise research reported that weekly use of Custom GPTs and Projects had increased approximately 19 times year-to-date. In recent months covered by the report, around 20% of Enterprise messages were processed through one of these structured environments.
The important part isn’t the particular OpenAI products. It is the pattern.
There is a big operational difference between “Our staff have access to AI” and “When this job occurs, we have a defined process that uses AI for these specific steps, accesses these approved sources, produces this output and leaves these decisions with a person.”
The second one is an operating capability. The first one is access to a tool.
McKinsey’s 2025 State of AI research found workflow redesign had the largest effect of 25 attributes tested on an organisation’s likelihood of reporting EBIT impact from generative AI. Yet only 21% of respondents reporting gen-AI use said their organisations had fundamentally redesigned at least some workflows.
For a small business, that doesn’t mean launching a major transformation program. It means starting with the work.
Start with a friction map
You don’t need an AI strategy workshop to begin finding opportunities. For a week, pay attention to where people lose time or momentum. Look specifically for these patterns.
1. Repetition
What does someone do repeatedly that follows roughly the same pattern? Preparing a weekly report. Categorising enquiries. Drafting a standard response. Reviewing a form. Updating a customer record. Turning meeting notes into actions.
Repetition doesn’t automatically mean AI is the answer. Traditional automation may be cheaper, safer and more reliable. But repetition tells you where to look.
2. Searching
Where do people repeatedly have to find information before they can do their job? It might be buried in SharePoint, Google Drive, email, a CRM, an accounting platform, project folders or someone’s memory.
The problem here may not initially be “we need AI”. It may simply be: We have an information retrieval problem.
AI can sometimes provide a much better interface to that information, particularly when people need to search by meaning rather than an exact filename or keyword. But the underlying information still needs to be accessible, trustworthy and appropriately governed.
3. Copying and re-keying
Look for information moving manually from one system to another. A customer fills in a form. Someone reads the form and enters the details into a CRM. Later, someone copies some of those details into an accounting system. Then another person copies them into a project document.
This is an important category because the solution may not involve AI at all. If the data is structured and the rules are deterministic, an API integration or conventional automation may be the better solution.
The point of looking for friction isn’t to prove that AI should be used. It is to find a problem worth fixing.
4. Checking
What does your team have to read, compare or inspect before deciding whether something is okay? Documents against requirements. Invoices against purchase orders. New enquiries against qualification criteria. Contracts against standard clauses. Reports against expected numbers.
AI can be particularly useful when the checking involves language, context or interpretation rather than a simple exact match. But if the consequence of getting the check wrong is significant, AI should usually assist the decision rather than quietly become the decisionmaker.
5. Summarising
Businesses produce an extraordinary amount of information that somebody later has to compress. Meetings become action lists. Long email threads become status updates. Customer notes become handover summaries. Documents become management briefs. Research becomes recommendations.
Current language models are good at this kind of work, but even here the useful question isn’t simply “Can AI summarise this?” Ask: “What happens after the summary?” A useful implementation connects the summary to the next part of the workflow.
6. Waiting
Some of the most expensive friction doesn’t involve anyone actively doing anything. Work simply stops. Someone is waiting for information, a document, somebody to notice an email, a manager to review something, or data to be collected before a decision can be made.
This is why looking at the whole workflow matters. Making one task 80% faster doesn’t achieve much if the work immediately sits in somebody else’s inbox for two days.
7. Hand-offs
Every time work passes between people or systems, context can disappear. Sales hands a new customer to delivery. Delivery asks finance to create something. Finance needs information from sales. A customer emails one employee, who forwards it to another employee with an explanation.
The friction isn’t necessarily the task itself. It is reconstructing the context required for the next person to continue. AI can help assemble that context, but only if it has access to the right information.
8. Information chasing
Almost every business has some version of this: “Have they sent that through?”, “Did anyone reply to this?”, “Where are we up to with that quote?”, “Has the customer approved it?”, “Who is waiting on whom?”
Individually, these interruptions look trivial. Collectively, they consume attention.
A useful AI system might identify outstanding actions, prepare follow-ups, surface exceptions or tell someone what actually requires attention. The goal isn’t necessarily to remove the person. It is to stop the person acting as the workflow’s search engine.
Not every friction point needs AI
This may be the most important part of the exercise. Once you’ve found a frustrating process, don’t immediately build an agent. Work out what kind of problem you actually have.
It is useful to separate possible interventions into four categories.
Remove it. Sometimes the best solution is to stop doing the work entirely.
Automate it conventionally. If the inputs are structured and the rules are clear, normal software, integrations or workflow automation will often be more predictable than an AI model.
Assist it with AI. AI searches, summarises, drafts, prepares or recommends while a person remains responsible for the decision.
Delegate defined parts to AI. For bounded, well-understood tasks with appropriate controls, an AI system or agent can perform more of the workflow itself.
Anthropic’s Economic Index provides an interesting real-world reference point. Its initial analysis found more Claude usage associated with augmentation of human work than automation: 57% versus 43%.
AI adoption doesn’t have to mean finding jobs for autonomous agents. Sometimes the best result is simply removing 20 minutes of frustrating work from something a person does every day.
Turn the friction into a specific task
Once you’ve identified a friction point, make it concrete. Instead of: “We spend too much time on customer enquiries.” Write: “Every morning, someone reads approximately 30 incoming enquiries, determines what each customer is asking for, identifies which ones require an urgent response and sends the relevant enquiries to the appropriate person.”. Now you have something you can evaluate.
Ask:
What information comes in?
What judgement is required?
What systems contain relevant context?
What happens next?
What would happen if the system got it wrong?
Which steps require a person?
Which steps could be deterministic?
Which steps genuinely benefit from language or reasoning?
How would we know whether the new process is better?
This is where an AI opportunity starts becoming an operational design problem. And that is a much better problem to have.
Find the smallest, useful intervention
There is another temptation once you identify a good workflow. You immediately redesign the whole thing. Don’t.
Suppose your business receives enquiries through email. Someone currently:
reads the email
identifies the customer
works out what they need
looks up previous interactions
decides who should handle it
drafts a response
updates the CRM
sends the email
It would be easy to draw an architecture diagram in which an autonomous AI agent performs all eight steps. It would also introduce a lot of risk and complexity at once.
The first useful intervention might simply be: AI reads the enquiry, retrieves relevant context and prepares a suggested classification and draft response for a person to review. That’s enough to find out whether the idea actually improves the work. If it does, you can expand from there.
Measure the friction, not the AI
Before changing the process, establish what the problem costs today. It doesn’t need to be a sophisticated ROI model.
minutes spent per item
number of items each week
time spent waiting
number of hand-offs
corrections or rework
response time
backlog
missed follow-ups
how often someone has to ask another person for information
Then implement the smallest, useful change and measure again. The metric isn’t: “We deployed an AI agent.” It is: “This process used to take 90 minutes every Monday. It now takes 25.” Or: “Customers used to wait six hours for enquiries to be triaged. They now wait 20 minutes.”
Or perhaps: “We tested AI here and discovered a simple integration solved the problem better.” That last result is still a success. You found and fixed the operational problem without forcing AI into it.
The AI opportunity is usually hiding inside the work
There is enormous pressure on businesses to “do something with AI”. That pressure makes it easy to start with models, agents and products.
But your business doesn’t exist to deploy AI. It exists to serve customers, deliver work, make decisions and operate effectively. So rather than asking everyone “Where could we use AI?”, try asking: “What part of your job is unnecessarily difficult?”
Then listen for the clues: “I have to do this every week.” “I can never find…” “I have to copy this into…” “I always need to check…” “I spend ages summarising…” “We’re waiting for…” “I have to ask them…” “Only Sarah knows…”
Those aren’t complaints to work around. They are a map of how your business actually operates. And somewhere on that map is probably a much better AI opportunity than the one you would have invented in a brainstorming session.
Cypher Agency helps businesses understand where AI can create practical value, then designs and builds the systems, integrations and controls needed to make it work reliably in the real world.
Comments