It is easy to find things AI could do in a business.

It can draft emails, summarize documents, answer questions, organize information, prepare reports, classify requests, and help employees think through decisions.

But “AI can do this” does not mean “this is worth pursuing.”

A useful AI opportunity needs to improve something that matters. It should make work faster, easier, more consistent, less expensive, less frustrating, or more valuable without creating an unreasonable amount of new cost, risk, or complexity.

That means the better starting question is not:

Where can we use AI?

It is:

Which part of the business is worth improving, and could AI play a useful role?

That small change in framing can prevent a business from spending time and money on impressive technology that does not produce a meaningful result.

01Business frictionFind work worth improving.
02WorkflowMap the complete work.
03Readiness checksTest value, information, risk, and adoption.
04Solution levelChoose the least complex response.
05Bounded pilotAnswer one real decision.
A worthwhile AI opportunity moves from business friction to a bounded test—not from a tool to an implementation.

Start with friction, not features

Do not begin by browsing AI products or collecting lists of popular use cases.

Begin with work that is not going as well as it should.

Look for places where:

  • customers wait too long;
  • employees repeat the same work;
  • important information is difficult to find;
  • quality depends too heavily on one person;
  • simple requests take too much attention;
  • work is regularly delayed, corrected, or redone;
  • valuable tasks are being neglected because routine work consumes the available time.

The first goal is simply to identify a real problem.

“Use AI in customer service” is too broad.

A more useful description would be:

Our team spends a large part of each week answering routine order-status questions, which leaves less time for unusual or urgent customer problems.

That gives you something concrete to investigate. It identifies the work, the burden, and the result you would like to improve.

It also leaves room for an honest answer. The right solution might be AI, but it might also be better notifications, clearer policies, improved self-service, conventional automation, or a simpler process.

Look at the whole workflow

A common mistake is to focus on one task without looking at the work around it.

For example, a business may decide that AI should “write customer quotes.”

But writing the final quote may be only one small part of the process. The complete workflow might include:

  1. receiving the customer’s request;
  2. figuring out what the customer actually needs;
  3. checking specifications and availability;
  4. applying pricing rules;
  5. identifying unusual terms or risks;
  6. preparing the quote;
  7. getting approval;
  8. sending it;
  9. recording it in the correct system;
  10. following up.

AI might write a polished quote and still fail to improve the workflow.

It may not have the correct pricing information. It may miss an approval rule. Employees may spend more time checking its work than they previously spent writing the quote themselves.

This is why the workflow matters more than the isolated task.

Before considering a solution, understand:

  • what starts the work;
  • who is involved;
  • what information they need;
  • which decisions they make;
  • what systems they use;
  • what result they produce;
  • what happens after that;
  • what exceptions are common;
  • who remains responsible.

Business value comes from improving the completed work, not from producing an impressive AI output in the middle of it.

Make sure the problem is worth solving

Not every inconvenience needs a new system.

A worthwhile opportunity usually involves work that is frequent, expensive, slow, risky, frustrating, or important to growth.

Ask:

  • How often does this happen?
  • How much attention does it consume?
  • Where does the work slow down?
  • How much correction or rework is needed?
  • What happens when the result is late or wrong?
  • Does this affect customers, revenue, cost, capacity, quality, or risk?
  • Is improving it important enough to work on now?

You do not need a perfect financial estimate at this stage.

You do need enough information to show that the problem is real.

A task that takes two hours per year is unlikely to justify a new AI workflow. A process that delays customers every day, consumes scarce expertise, or limits how much work the company can handle may deserve serious attention.

Repetitive work is not automatically a good AI opportunity

Repetition is often a useful clue, but it is not a decision.

A repeated task may still be:

  • too small to justify changing;
  • caused by a broken process;
  • better handled by ordinary automation;
  • dependent on poor information;
  • too risky to hand off;
  • easy to remove entirely.

Imagine that employees repeatedly answer questions about delivery dates.

An AI assistant could draft those answers. But the stronger solution might be to send customers better shipping updates so they do not need to ask in the first place.

Or imagine that employees repeatedly copy data from one system to another. That may call for an integration or automation rule, not generative AI.

Before asking how AI could perform the work, ask:

Why does this work exist, and should it continue in its current form?

Sometimes the best AI decision is to improve the business without using AI.

“No AI required” is a valid result

A good evaluation should consider simpler solutions first.

The workflow may improve through:

  • clearer instructions;
  • better training;
  • improved templates;
  • cleaner data;
  • better-organized information;
  • a form, rule, formula, or checklist;
  • configuration in existing software;
  • conventional workflow automation;
  • removing unnecessary steps.

AI is most useful when the work requires flexible language, interpretation, summarization, classification, comparison, generation, or judgment support that simpler tools cannot provide well enough.

This does not mean a business must exhaust every other option before experimenting with AI.

It means AI should have a clear purpose.

A conclusion of “no AI required” can save money, reduce complexity, and solve the problem faster. That is a successful result, not a missed opportunity.

Check whether the right information is available

AI cannot perform a workflow well if it does not have the information needed to do the work correctly.

Ask what a capable employee would need.

That might include:

  • current customer or product information;
  • policies;
  • approved examples;
  • pricing rules;
  • templates;
  • previous decisions;
  • exception criteria;
  • information from other systems.

Then ask whether that information is:

  • available;
  • accurate;
  • current;
  • organized;
  • accessible at the right time;
  • appropriate and authorized for the intended use.

A system may understand a customer’s question but still give the wrong answer because it was not given the company’s current return policy.

It may draft a convincing quote using an outdated price list.

It may produce a complete-sounding response even though important information is missing.

If the information is not ready, the opportunity may still be worthwhile—but the business may need to prepare first.

That preparation might include organizing documents, improving data quality, documenting decisions, or collecting good examples.

Decide what role AI should play

“Use AI in this workflow” is not specific enough.

You need to decide how much responsibility the system should have.

AI assistance

AI prepares a draft, summary, classification, recommendation, or analysis. A person reviews the result and takes the final action.

Partial automation

AI completes a defined part of the process. A person reviews important decisions, exceptions, or higher-risk cases.

Full automation

The system completes the workflow or takes action without regular human approval.

These are very different designs.

Full automation is not automatically the most advanced or valuable outcome. For many businesses, human-reviewed assistance may be the best long-term solution.

The right level depends on:

  • how serious an error would be;
  • whether errors are easy to notice;
  • whether an action can be reversed;
  • how much judgment the work requires;
  • how often unusual cases appear;
  • how much review is practical;
  • who remains accountable.

Drafting an internal meeting summary is different from changing a financial record, making an employment recommendation, or sending a legally important customer message.

The goal should not be to remove a person simply because it is technically possible.

The goal is to decide carefully which responsibilities belong to the system and which remain with people.

Define what a good result looks like

A business cannot tell whether AI is working if “good” has not been defined.

Goals such as these are too vague:

  • write a good response;
  • produce an accurate summary;
  • identify the important information;
  • improve productivity.

A stronger definition is more observable.

For a customer-response workflow, a good result might need to:

  • identify the customer’s actual problem;
  • use the correct policy;
  • avoid inventing customer or order information;
  • include required language;
  • use an appropriate tone;
  • escalate unusual cases;
  • require only a reasonable amount of editing.

The exact requirements will depend on the workflow.

For document extraction, you may need every answer to be traceable to its source.

For classification, certain mistakes may matter much more than others.

For drafting, the important measure may not be whether the first version sounds polished. It may be whether the draft is accurate, complete, and genuinely reduces the reviewer’s work.

You should also test more than easy examples.

Include:

  • ordinary cases;
  • difficult cases;
  • missing information;
  • unclear requests;
  • conflicting information;
  • important exceptions;
  • cases where the correct answer is to stop and ask for help.

If you cannot explain what an acceptable result looks like, you are not yet ready to decide whether the system works.

This follows a broader principle in NIST’s AI Risk Management Framework: risk management and evaluation should reflect the system’s intended context and continue across its lifecycle. NIST’s AI Resource Center also provides resources for testing, evaluation, verification, and validation rather than treating general model capability as sufficient evidence. NIST AI RMF 1.0 · NIST AI Resource Center

Estimate value without pretending to know the future

Early AI ideas often come with very confident savings estimates.

A business may calculate that a task takes ten hours per week, multiply that by an hourly wage, and treat the result as guaranteed value.

That can be a useful starting point, but it leaves out several important questions:

  • How much time will actually be saved?
  • How much review will still be required?
  • Will people use the system consistently?
  • What will happen with the time that is freed?
  • Will the solution reduce overtime, hiring, delay, or rework?
  • What will setup, training, maintenance, and correction cost?
  • What happens when the workflow changes?

Saving five hours does not automatically create five hours of financial value.

The time must be used productively, increase capacity, avoid another cost, improve service, or enable work that was not previously completed.

At this stage, directional estimates are usually more honest than precise forecasts.

Look at:

  • current volume;
  • current effort;
  • delays;
  • rework;
  • consequences of mistakes;
  • expected review effort;
  • implementation difficulty;
  • recurring cost;
  • maintenance needs;
  • plausible business benefit.

A useful conclusion might be:

This opportunity appears valuable enough to test, but we need evidence about quality and review time before estimating a reliable return.

That is more useful than an impressive number built on assumptions.

Let risk shape the solution

Risk does not always mean “do not use AI.”

It means the system may need stronger limits, oversight, testing, or expertise.

Consider whether the workflow affects:

  • financial records;
  • contracts;
  • legal rights;
  • employment;
  • health or safety;
  • regulated activity;
  • sensitive information;
  • customer commitments;
  • irreversible actions;
  • the company’s reputation.

Then ask:

  • Can a mistake be caught before it causes harm?
  • Can the action be reversed?
  • Can the system stop when information is missing?
  • Is a qualified person available to review important cases?
  • Can the business understand how the result was produced?
  • Can the system be disabled or replaced safely?

Risk should influence the design before the system is put into use.

A lower-risk internal drafting assistant may be suitable for a small test.

A system that makes consequential decisions or takes action on behalf of the business may require expert review before any operational use.

The OECD AI Principles similarly emphasize context-appropriate human oversight, traceability, accountability, and ongoing risk management. They also recognize that different uses create different levels of risk and therefore need proportionate safeguards. OECD AI Principle: Human-centred values and fairness · OECD AI Principle: Accountability · OECD AI Principle: Robustness, security and safety

Include the people who will use it

A technically capable system can still fail because it does not fit the people or the process.

Ask:

  • Who will use the result?
  • Where will it appear?
  • What will the user do next?
  • Does the new process remove work or add more steps?
  • Who handles uncertain cases?
  • Are responsibilities clear?
  • What training will be needed?
  • Who will maintain the workflow?
  • Will employees understand when not to trust the output?

An AI tool may save writing time but require employees to copy information between several systems.

A recommendation may be useful but ignored because no one knows who owns the final decision.

A workflow may perform well in a test but feel too confusing or unreliable for daily use.

These are not problems to solve after implementation.

They are part of deciding whether the opportunity is viable in the first place.

Adoption is part of the design.

A team maps business friction, information, risk, human responsibilities, and review before choosing an appropriate AI option.
Opportunity selection is a workflow-design decision involving the people who understand the work—not merely a search for more AI tools.

Choose the simplest solution that could work

A worthwhile AI opportunity does not automatically require custom software.

It helps to think in levels.

Level 0 — No AI

Improve the process, policy, documentation, training, search, rules, software configuration, or conventional automation.

Level 1 — Standard AI assistance

Use an existing AI product with human review.

Level 2 — Configured AI workflow

Add repeatable instructions, templates, examples, business context, and quality checks.

Level 3 — Integrated AI workflow

Connect AI to business systems or partially automate parts of the workflow with clear human oversight.

Level 4 and beyond — Expert assessment

Consider custom applications, advanced agent workflows, larger integrations, shared governance, or broader organizational capabilities.

These levels are not a ladder that every business needs to climb.

Level 1 may be the right permanent answer.

Level 0 may produce the strongest return.

A more complex system is justified only when it creates enough additional value, reliability, or control to earn its ongoing cost.

Every extra integration, model, tool, review step, and data connection creates something that must be understood, secured, tested, and maintained.

Complexity is not automatically bad. It simply needs a reason to exist.

Use a pilot to answer a real question

A promising opportunity should lead to a small, focused test—not an immediate implementation project.

A useful pilot begins with a decision that needs evidence.

For example:

Can a standard AI product produce acceptable first drafts for routine customer questions while reducing employee writing time without increasing policy errors?

That is a useful question because it defines:

  • the workflow;
  • the role of AI;
  • the expected benefit;
  • an important quality requirement;
  • the evidence needed to decide what happens next.

A weaker question would be:

Can we build an AI customer-service agent?

That may produce an impressive demonstration without telling the business whether it should use it.

A sensible first test might involve:

  1. Document the current workflow.
  2. Measure current volume, time, delay, and rework.
  3. Collect representative examples.
  4. Define an acceptable result.
  5. Identify important risks.
  6. Test the simplest plausible solution.
  7. Compare it with the current process.
  8. Record failures and reviewer effort.
  9. Decide whether to stop, prepare further, test again, or expand.

The purpose of a pilot is to reduce uncertainty before committing to complexity.

NIST’s Generative AI Profile applies the same lifecycle perspective specifically to generative AI, covering design, development, use, and evaluation rather than treating a successful demonstration as proof that a workflow is ready for normal operation. NIST AI RMF: Generative AI Profile

A practical check before moving forward

A promising AI opportunity usually has most of the following:

A meaningful problem

The workflow has a recurring or important problem worth addressing.

A defined workflow

You understand the trigger, inputs, people, decisions, output, systems, exceptions, and next action.

A clear role for AI

AI has a specific contribution that simpler methods may not provide as effectively.

Usable information

The necessary information is available, accurate enough, current, accessible, and appropriate to use.

A result you can evaluate

You can explain what acceptable performance means and which mistakes matter.

Manageable risk

The consequences are understood, and appropriate review or controls are possible.

People who can use it

The intended users can fit the system into their work and understand their responsibilities.

Plausible economics

The potential benefit is large enough to justify testing, setup, review, operation, and maintenance.

A proportionate solution

You can begin with the simplest level likely to create enough value.

A bounded next decision

The test will help you decide whether to stop, prepare, redesign, or continue.

The opportunity does not need to be perfect.

Unknowns are normal. The question is whether those unknowns can be investigated safely and affordably.

Choose the opportunity, not the technology

A worthwhile AI opportunity is not simply a task that looks impressive in a demonstration.

It is a business workflow where:

  • the problem matters;
  • the intended improvement is clear;
  • the right information exists;
  • quality can be evaluated;
  • risk can be managed;
  • people can use the result;
  • the economics make sense;
  • the required complexity is reasonable.

That process may lead to AI assistance, partial automation, a custom system, conventional software, a better process, or no change at all.

Any of those may be the correct result.

Gecko Road’s position is simple:

Start with meaningful business friction. Understand the complete workflow. Choose the least complex response that could reliably improve it. Then test the assumptions that matter before committing to a larger solution.

The goal is not to find as many places as possible to use AI.

The goal is to identify the few opportunities that are genuinely likely to improve the business.

Sources

  1. NIST Artificial Intelligence Risk Management Framework 1.0
  2. NIST AI Resource Center
  3. NIST Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
  4. OECD AI Principle: Human-centred values and fairness
  5. OECD AI Principle: Robustness, security and safety
  6. OECD AI Principle: Accountability