The conversation around artificial intelligence often begins in the wrong place.
It begins with a model, a feature, or a demonstration. Someone asks what the technology can do, and the organization starts searching for places to put it. The result may look impressive in a presentation, yet still struggle to earn a meaningful role in the way customers or teams actually work.
A more useful starting point is simpler:
What valuable job needs to be done better?
That question changes the nature of the work. It shifts attention away from novelty and toward value, context, adoption, and responsibility. It also creates a standard against which an AI idea can be judged before significant time and capital are committed.
Begin with friction, not features
Strong AI opportunities tend to emerge where people repeatedly encounter one of four conditions:
- Too much information to review confidently
- Repetitive work that still requires judgment
- Decisions that would improve with better context
- Experiences that need to adapt without becoming inconsistent
These are not automatically AI problems. Some should be solved through clearer content, better service design, conventional automation, improved data, or a simpler product flow.
The first responsibility is therefore not to justify AI. It is to identify the source of friction accurately.
If a customer cannot find an answer because the information architecture is poor, a conversational interface may only place a polished layer over disorganized knowledge. If a team repeats a task because two systems do not communicate, a reliable integration may create more value than a probabilistic agent. If a decision has no reliable data behind it, adding a model can increase confidence without increasing truth.
The right solution begins with the right diagnosis.
Four questions that sharpen an AI opportunity
Before moving into concepts or prototypes, we ask four questions.
1. Where is the value?
What becomes meaningfully better if this works?
The answer should be concrete. A customer reaches clarity sooner. A specialist spends less time searching and more time deciding. A team can respond consistently across a larger volume of work. A service becomes accessible in a context where the existing interface fails.
“Using AI” is not an outcome. The opportunity needs a beneficiary and a change worth creating.
2. What evidence can the system use?
An intelligent experience is only as useful as the information, rules, and context available to it.
What knowledge can be trusted? Where does it live? Who owns it? How current is it? What should the system never infer? Which decisions require a human source of authority?
These questions are not background technicalities. They shape the experience itself. A system with limited evidence should communicate uncertainty differently from one operating within a controlled knowledge domain.
3. What happens when it is wrong?
Every AI concept has a failure mode. The important question is whether that failure is noticeable, recoverable, and proportionate to the value being created.
A poor recommendation in a low-stakes discovery tool may create inconvenience. A poor recommendation within a consequential financial, health, legal, or operational decision can create real harm.
The response should influence the product architecture from the beginning: when to ask for confirmation, when to show sources, when to constrain the system, when to escalate, and when not to use AI at all.
4. Why will someone adopt it?
People do not adopt a new system because the technology is sophisticated. They adopt it when it fits naturally into the way they understand and perform a task.
Adoption depends on trust, timing, control, transparency, and the perceived cost of changing behaviour. An AI tool that asks someone to abandon a familiar workflow must provide significantly more value than one that assists them inside it.
The experience should make the role of intelligence understandable. Users need to know what the system is doing, what it needs from them, and where their own judgment remains essential.
Choose the right level of intelligence
Not every opportunity requires an autonomous agent.
Sometimes the best answer is a narrow automation that performs a repeatable task reliably. Sometimes it is a copilot that helps a person explore, draft, compare, or decide. Sometimes it is a retrieval experience that makes trusted knowledge easier to access. In specific contexts, a more agentic system may be appropriate—but only when its authority, boundaries, and oversight are explicit.
The ambition of the technology should match the maturity of the problem.
Starting narrowly is not a lack of vision. It is often the fastest way to learn what creates value, where trust breaks down, and what the system needs before it can take on more responsibility.
Prototype the workflow, not only the interface
A convincing interface can hide an unresolved operating model.
For that reason, an AI prototype should demonstrate more than a polished conversation. It should explore the complete journey:
- What initiates the experience?
- Which information is available?
- What does the system produce?
- How is uncertainty communicated?
- Where does a person review or intervene?
- What happens when evidence is missing?
- How is feedback captured?
- Who remains accountable for the final action?
This is where design, engineering, content, data, and governance become inseparable. The experience is not simply the visible interface. It is the relationship between the model, the information around it, the person using it, and the organization responsible for it.
Define what “good” means before scaling
AI experiences need evaluation criteria that reflect their real purpose.
Technical performance matters, but so do relevance, clarity, consistency, usefulness, recoverability, and user confidence. A system may produce a plausible response and still fail if the user cannot understand why it should be trusted or what to do next.
Evaluation should therefore combine multiple forms of evidence: controlled tests, expert review, realistic scenarios, user observation, operational feedback, and product behaviour after release.
The goal is not to prove that the system is perfect. It is to understand where it performs reliably, where it requires support, and whether the value justifies the remaining uncertainty.
Governance is part of the experience
Responsible AI is often discussed as a policy layer added after the product has been conceived. In practice, it is also a design material.
Source visibility, permissions, escalation, feedback, consent, human review, and data handling all affect how the experience feels. When designed well, safeguards do not merely reduce risk. They can make the product easier to understand and more deserving of trust.
This is especially important when an organization is introducing AI to customers or employees for the first time. The experience establishes an expectation for how intelligence will operate across the wider brand.
Intelligence should earn its place
The most valuable AI products are rarely the ones that display the most intelligence. They are the ones that apply it with precision.
They begin with a meaningful job. They use evidence deliberately. They make failure visible and recoverable. They respect the judgment of the people they are designed to support. And they create a result that is valuable enough to become part of real behaviour.
That is the standard AI should meet before it becomes part of a product, service, or organization: not whether it can be added, but whether it has earned a useful role.