From AI Tools to Better Workflows: How to Identify Where AI Can Actually Add Value

AI adoption doesn’t start with choosing a tool. It starts with understanding how work gets done.

Many organisations already have access to ChatGPT, Gemini, Copilot or other AI tools. Yet having more AI available does not necessarily mean work is getting better. Teams may still be spending hours on repetitive tasks, moving information between systems, reviewing documents or producing the same type of output again and again.

The question, then, is not simply:

“Where can we use AI?”

A better question is:

“Which parts of our workflows are worth redesigning with AI?”

That shift in perspective is where tools such as AI Specialists, also commonly called Skills, become useful.

A chatbot is useful when you have a specific question or problem. You provide an instruction, the AI responds, and you decide what to do next.

A Specialist works differently. It is configured to handle a specific, recurring type of work by following a defined process. Once it has been set up and trained, the human involvement required for each run can be significantly lower.

Think of it less as another chatbot and more as a capability embedded within a workflow.

For example, instead of asking an AI tool every time you need to create a set of assessment questions, you could configure a Specialist with the relevant context, examples, process and expected output. You trigger it when the task arises, review the result and improve the Specialist over time.

But there is an important step before building one:

Understand the workflow first.

When AI takes on more responsibility, you need to understand how the whole piece of work flows and where AI can actually contribute.

A useful way to map a workflow is through four elements:

• Goal: What value does the workflow deliver?
• Trigger: What starts the workflow?
• Steps: What activities need to happen?
• Deliverable: What does the workflow ultimately produce?

Consider a simple example: creating a new skills assessment for a learning platform.

The goal might be to make a new assessment available to users. The workflow could include defining learning objectives, creating a maturity model, writing assessment questions, creating test questions, updating the database and publishing the assessment.

A Specialist does not necessarily need to take over the entire workflow.

In fact, it probably shouldn't.

The opportunity is to identify the specific step where AI can create the most value without compromising quality or ownership.

Not every task is suitable for a Specialist.

A good candidate tends to have two characteristics: it is feasible to automate and valuable enough to justify the effort.

1. Start with feasibility

Ask four questions:

Is the task repetitive?

Does the basic process stay similar each time, or is every case completely different?

Do you have enough data or examples?

Can you show the AI previous inputs and good outputs? A Specialist needs a knowledge base to learn from.

Are the inputs and outputs structured?

Tasks with predictable inputs and clearly defined outputs are generally easier to delegate.

Is the process stable?

If the underlying process changes every week, maintaining the Specialist may become more work than doing the task manually.

Together, these questions help determine whether the task is technically and operationally feasible.

2. Then consider value

A feasible workflow is not automatically worth automating.

Consider:

Frequency: How often does the task happen?

Time taken: How much time does it consume?

Logic vs. intuition: Is the task based on a repeatable logic that AI can reproduce, or does it depend heavily on human judgement and experience?

This last distinction matters.

A task can be technically feasible for AI but still have limited value if a human needs to rewrite most of the output anyway.

One useful way to prioritise workflows is to score feasibility and value separately.

For example, feasibility could be scored out of 20 and value out of 15.

But don't simply add the two scores together and choose the highest number.

Feasibility should act as a gatekeeper.

You might decide that a workflow needs at least 14/20 in feasibility before you consider building a Specialist. You can also establish minimum requirements for specific criteria. For example, if there are no examples or data available, the workflow may not be a good candidate regardless of how valuable it is.

Among the workflows that pass the feasibility threshold, use the value score to decide where to start.

This creates four broad categories:

High value + High feasibility: Best candidates — start here.
High value + Low feasibility: Future candidates — improve feasibility first.
Low value + High feasibility: Good practice — useful for learning.
Low value + Low feasibility: Don’t do — keep doing it manually.

This distinction prevents a common mistake: automating something simply because AI can do it.

The goal is not maximum automation.

The goal is better work.

Once you have selected a workflow, don't stop there.

Break it down into its individual steps and evaluate those steps as well.

For example:

Create a new assessment

→ Define learning objectives
→ Create maturity model
→ Create assessment questions
→ Write test questions
→ Update database
→ Publish assessment

Perhaps defining the learning objectives is highly dependent on expert judgement, while writing test questions is repetitive, structured and easy to evaluate.

That changes the answer.

Instead of asking AI to “create a new assessment”, you might give the Specialist responsibility for writing the first version of the test questions.

This narrower scope makes the system easier to train, easier to evaluate and easier to trust.

This is where AI adoption becomes a question of governance, not just productivity.

Think of trust as a dial.

At one end, there are tasks you simply should not delegate. They may depend on your organisation's unique expertise, involve significant risk or represent a core part of the value you provide.

In the middle are human-in-the-loop workflows. AI does the work, but a person reviews the result before anything important happens.

At the other end are tasks that can be fully delegated because the process is predictable, the output is easy to validate and there is an appropriate safety net if something goes wrong.

Most organisational work will sit somewhere between these extremes.

And that's okay.

Imagine an AI Specialist that generates 40 assessment questions.

The workflow could look like this:

1. The relevant learning objectives and assessment structure are defined.
2. The Specialist generates the first batch of questions.
3. A human reviews every question and its possible answers.
4. The human provides detailed feedback.
5. The Specialist revises the output.
6. The human approves the final version.
7. The feedback is used to improve the Specialist for future runs.

The human is still involved, but their role has changed.

Instead of spending four hours creating every question from scratch, they spend their time evaluating, correcting and making decisions.

That distinction is important.

The objective isn't to remove humans from the workflow. It's to remove unnecessary human effort from the workflow.

A Specialist doesn't become useful simply because you have given it an instruction.

Like a new employee, it needs context.

A well-designed Specialist typically includes five components:

Persona: What role is it performing?

Context: What knowledge, process documentation and examples does it need?

Capabilities: Which tools or systems does it need access to?

Input and output: What triggers the work, and what exactly should be delivered?

Edge cases: What should it do when information is missing or something doesn't go according to plan?

This is also why the first version of a Specialist is rarely the final version.

Training takes time. Outputs need to be evaluated. Feedback needs to be incorporated. The process itself may need to be refined.

The work doesn't end when the Specialist is created.

It evolves as you learn where it performs well—and where it doesn't.

A useful test is to compare the work before and after introducing the Specialist.

In one example, producing 40 test questions manually took approximately four hours. With a Specialist, the process took around 30 minutes, with the potential to reduce that further.

The AI generated the initial questions in minutes. The remaining time was spent on human review and feedback.

That's still roughly 3.5 hours saved for every 40 questions.

But there is an important lesson here:

A Specialist making you faster doesn't mean you should stop checking its work.

If the output affects customers, employees, products or business decisions, speed should not come at the expense of quality.

The right measure is not simply:

“How much work did AI do?”

It is:

“How much unnecessary effort did we remove while maintaining the level of quality and control we need?”

AI Specialists are not valuable because they are another type of AI tool.

They are valuable because they can change how recurring work gets done.

That requires a different starting point.

Don't begin with:

“What can we automate?”

Begin with:

“Which workflows consume significant time, and where is human involvement genuinely necessary?”

Then assess feasibility. Prioritise value. Break workflows into steps. Decide how much trust is appropriate. Keep humans involved where judgement matters. And continuously improve the system based on real results.

The organisations that get the most from AI won't necessarily be the ones with the most AI tools.

They will be the ones that understand how work flows through their organisation—and deliberately redesign that work around the strengths of both humans and AI.

AI adoption is not just about adding intelligence to existing processes. It's about creating better ways of working.