Not Every Problem Needs AI. Here’s Why.
As organizations race to adopt AI, I've noticed that we are becoming so enthusiastic about what AI can do that we're starting to assume AI should be part of every solution. Every product needs an agent. Every workflow needs intelligence. Every experience needs a conversational layer.
But not every problem needs AI.
In fact, using AI where it isn't needed can make a product more complicated, less predictable, and significantly more expensive to operate. Just because a model can perform a task doesn't mean it is the right technology for that task.
I think we need to start the conversation somewhere else.
Instead of asking, "Where can we use AI?" we should ask, "What kind of intelligence does this task actually require?"
When you map work that way, my observation has been that it starts to fall into three distinct categories:
Execution, where tasks don't require reasoning at all.
Delegation, when tasks where reasoning can safely happen on the user's behalf, and
Collaboration, when and where tasks where AI and human judgment are more valuable together.
Each requires a very different experience.
1. When intelligence isn’t required.
Let's start with the category we're overlooking the most: work that doesn't need AI at all.
There are countless interactions where the user already knows what they want to do, the available choices are known, and the outcome is predictable.
Transfer $500 from checking to savings. Select seat 14A on a flight. Choose an available appointment time. Change a project's status from "In Progress" to "Complete."
There is nothing for AI to reason about in these moments. The person has already made the decision. The system simply needs to help them execute it.
Could we add AI?
Of course. Instead of selecting 14A from a seat map, we could ask someone to type, "Put me in seat 14A." But we've taken a single click and turned it into an inference problem. We've added computational cost, latency, uncertainty, and another opportunity for the system to misunderstand something already clear.
These experiences need great UX, not artificial intelligence.
The interface should be designed specifically for the task, making the available information and actions clear and helping people complete the work as efficiently as possible. Traditional interface patterns and deterministic logic are incredibly good at this.
AI should have to earn its place in an experience.
If the intent is explicit, the rules are known, and the outcome is deterministic, the simplest solution may still be the best one.
2. When the system should do the work.
The second category is almost the opposite.
Some tasks don't require much user input because the system can reliably do the work on their behalf.
Think about a financial system automatically reconciling transactions that meet clearly established criteria.
A security system identifying routine threats and blocking them according to predefined policies. An enterprise system categorizing incoming requests and routing them to the appropriate team. Or a healthcare system continuously monitoring large amounts of information and surfacing the handful of changes that require someone's attention.
This category requires another important decision. Does the automation actually require AI?
Sometimes the answer is no. If X always results in Y, simple rules and traditional automation may be more reliable, less expensive, easier to test, and easier to govern.
AI becomes useful when the inputs are less structured, or the system needs to interpret, classify, summarize, recognize patterns, or make decisions where deterministic rules aren't sufficient.
But the UX challenge here isn't necessarily helping someone perform the task. The system is already doing that.
The experience needs to help people understand what happened.
If a system is automatically reconciling hundreds of transactions, the user probably doesn't need to watch every transaction happen.
However, they need to know that 487 were successfully reconciled, 12 require review, and three couldn't be resolved. They need to understand why something was flagged, have confidence that the automation is behaving correctly, and be able to intervene when it isn't.
The interface shifts from execution to oversight.
Design becomes about visibility, trust, exceptions, tuning, and control.
What did the system do? Why did it do it? What requires my attention? How do I correct it? How do I change the rules?
Automation shouldn't mean invisibility.
3. When humans and AI should think together.
The third category is the most interesting because neither traditional interfaces nor fully autonomous AI are enough.
There is a growing class of work where AI can dramatically expand what a person is capable of doing, but we don't necessarily want AI making the final decision.
Here are four scenarios:
Think about a physician evaluating a complex patient. AI might synthesize years of medical history, surface relevant changes in labs, identify possible interactions, and suggest diagnoses to consider. But the physician brings clinical experience, knowledge of the patient, judgment, and ultimately accountability for the decision.
Or consider a financial advisor helping someone plan for retirement. AI can analyze thousands of scenarios, identify patterns, model different assumptions, and surface options. The advisor understands the person's goals, family circumstances, tolerance for risk, and the nuances that aren't necessarily represented in the data.
Meanwhile, a designer might use AI to analyze research, explore dozens of product concepts, generate working prototypes, or identify patterns across customer feedback. But the designer still needs to determine whether the right problem is being solved, evaluate the quality of the experience, understand the context, and decide which direction is worth pursuing.
Finally, consider an engineer investigating an incident might have AI analyze logs, correlate events across systems, and suggest likely causes. The engineer applies experience and system knowledge to determine which hypothesis is credible and what action is safe to take.
These aren't automated workflows. They're assisted workflows.
And I think they require a new interaction model altogether.
Designing AI as a collaborator.
A collaborator is very different from an automation.
It needs context about what the user is working on. It needs to understand what has already happened. It should be able to surface something they may have missed without constant interruption. It should offer recommendations without making it difficult for the user to disagree.
Most importantly, the user needs to be able to judge what they are given.
If AI recommends a diagnosis, design direction, investment strategy, or technical fix, simply showing me an answer isn't enough. The user may need to understand what information it used, what evidence supports the recommendation, how confident the system is, and what information might be missing.
The experience needs to support a loop between human and machine.
AI analyzes or generates. The human evaluates. The human redirects. AI responds. The human applies judgment and ultimately decides.
That's much richer than putting a chat box next to an existing application.
Chat may sometimes be part of the experience, but AI collaboration can also happen through recommendations embedded in workflows, comparisons, generated artifacts, simulations, contextual suggestions, editable plans, or interfaces we haven't invented yet.
This is where I think some of the most interesting UX work of the next decade will happen.
Don't confuse AI with the interface.
One of the mistakes we're making right now is conflating AI with conversational interfaces.
If someone is transferring $500 between accounts, they probably don't need a conversation. Give them a great interface and let me complete the transaction.
If AI is reconciling thousands of transactions, a user likely doesn’t need to chat with it about every one. Provide visibility into what happened, surface the exceptions, and let them intervene.
If someone is trying to make a complex financial decision, that's different. Now AI might help them explore scenarios, understand tradeoffs, and consider possibilities they hadn't thought about.
Same underlying technology. Three completely different relationships between the human and the system.
The interaction model should follow the work, not the technology.
There is an economic reason to get this right.
There is another reason organizations need to make these distinctions carefully: AI isn't free.
Every time we introduce model inference where deterministic software could have solved the problem, we're potentially adding computational cost, latency, complexity, testing requirements, governance, and unpredictable behavior.
Multiply those decisions across millions of interactions and the economics become significant.
The smartest AI product may therefore be the one that knows where not to use AI.
Use deterministic software when the answer is deterministic. Automate when the system can safely perform work without human involvement. Use AI when interpretation and reasoning actually create value.
Bring humans and AI together when it produces a better outcome than either could produce alone.
That isn't being conservative about AI. It's designing intelligently.
Map the intelligence before you design the experience.
I think this should become an early step in designing AI-powered products.
Map the workflow before choosing the technology.
For every part of the experience, ask what kind of intelligence the task actually requires. Is the intent already known and the outcome deterministic? Can the work happen safely in the background? Does that automation actually require AI, or would simple logic work better? Does the task require interpretation? And where does human judgment add something AI cannot?
Those answers should determine the interaction model.
No reasoning required. Design a great interface and use deterministic logic.
Reasoning can be delegated. Automate the work and design for visibility, trust, exceptions, and human intervention.
Reasoning benefits from collaboration. Create an assisted experience where AI capability and human judgment work together.
We shouldn't start with AI and look for places to put it.
Start with people, the work they're trying to accomplish, and the outcome we're trying to create.
Then use the simplest technology that can solve each part of the problem well.
Sometimes that's AI. Sometimes it isn't.
Knowing the difference may be one of the most important design skills of the AI era.