Stop Asking Can AI Do This

TL;DR

• “Can AI do this?” is usually the weakest question in the room.
• The better question is whether AI should do the work, under what constraints, and with what review path.
• Strong AI opportunities have value, manageable risk, clear evaluation, and workflow fit.
• Some workflows need LLMs. Some need rules. Some need human judgment. Some need nothing new.

Abstract

Generative AI can do enough things that capability alone no longer tells you much.

It can summarize documents, draft emails, classify tickets, answer questions, write code, inspect images, generate images, extract fields, and simulate conversation. A team can spend weeks proving that a model can participate in a workflow without proving that the workflow is better.

The useful question is not only:

Can AI do this?

The useful question is:

Should AI do this, and what must be true for users to trust it?

This post gives a practical decision guide for answering that question.

Table of Contents

• The Wrong Question
• The Better Question Set
• The Four-Part AI Fit Test
• When AI Is a Bad Fit
• When AI Is a Strong Fit
• Decision Matrix
• Summary
• Next Steps

The Wrong Question

“Can AI do this?” feels practical. It is not useless. Early in exploration, you do need to know whether a model can produce a plausible output.

But the question becomes dangerous when it stands alone.

AI can often produce something that looks like the work. It can write a contract summary, answer a policy question, generate a chart explanation, describe an image, or draft a support reply. The output may be fluent enough to make the team believe the hard part is over.

Usually, the hard part has just started.

The model’s ability to produce an answer does not tell you:

• whether the answer is correct
• whether the user can verify it
• whether mistakes are recoverable
• whether the workflow is faster
• whether the output is worth its cost
• whether the organization can accept the risk

That is why many AI demos feel convincing and then stall. The team answered the capability question but not the product question.

The Better Question Set

Replace “Can AI do this?” with a better set of questions.

Should AI do this?

Some tasks are technically possible but operationally foolish. If the cost of a wrong answer is high and review is weak, the workflow may not be ready.

Is it better than the current workflow?

Better can mean faster, cheaper, safer, more complete, more accessible, or more creative. If AI does not improve a meaningful dimension, it is decoration.

Can we trust the output?

Trust does not mean the model is always right. It means the system has ways to ground, inspect, review, correct, and improve its behavior.

Can we recover from mistakes?

Wrong outputs are not rare enough to ignore. A useful AI product has a recovery path.

Can users tell when it is wrong?

If users cannot evaluate the answer, automation may create false confidence.

Does the system improve with feedback?

If failures disappear into chat logs or support tickets, the product will keep repeating them.

These questions connect directly to the developer trust ledger. Trust is earned through evidence, constraints, and repeated behavior, not claims.

The Four-Part AI Fit Test

Use four filters before building.

1. Value

What does AI improve?

Look for a concrete gain:

• less time
• lower cost
• better coverage
• better personalization
• more accessible workflow
• better synthesis across messy information

If the value is “it feels modern,” stop.

2. Risk

What happens when the system is wrong?

Risk includes financial loss, user harm, privacy exposure, legal exposure, brand damage, and workflow disruption. High risk does not always mean no AI, but it does change the design. You may need human review, strict scope, audit logs, or abstention behavior.

3. Evaluation

How will you know whether it works?

A product team needs more than vibes. It needs test cases, rubrics, comparisons, human review, automated checks, or production feedback. If quality cannot be measured or reviewed, improvement becomes guesswork.

4. Workflow Fit

Where does the AI output go?

An answer that lives in a chat window may not be enough. The output may need to become a ticket, report, design asset, code change, dashboard note, or reviewed recommendation.

Good AI products fit into work. They do not ask users to invent new work around them.

When AI Is A Bad Fit

AI is often a bad fit when:

• the task is deterministic and rules would work
• the risk is high and review is weak
• the data is unavailable or low quality
• the user cannot judge the output
• latency or cost makes the workflow worse
• the organization wants automation but not accountability

For example, an LLM is probably not needed to route three known form types. Rules may be cheaper, faster, and easier to debug.

An LLM may also be a poor fit for final approval in a high-stakes compliance workflow. It may be useful as an assistant that drafts, summarizes, or flags issues, but not as the decision-maker.

The point is not to be anti-AI. The point is to put AI where it changes the work.

When AI Is A Strong Fit

AI is usually stronger when the workflow involves:

• messy language
• long documents
• repeated drafting
• summarization
• semantic search
• multimodal interpretation
• classification with review
• synthesis across many sources
• exploration where there are many acceptable outputs

Support triage can be a good fit because AI can summarize the issue, suggest a category, retrieve relevant docs, and prepare a draft response for human review.

Internal research can be a good fit because AI can help search, summarize, compare, and cite information across a large corpus.

Visual inspection can be a good fit when a vision-language model helps identify regions of interest, describe images, or support a human expert.

The strong pattern is not “replace the human.” The strong pattern is often “reduce the load around the human decision.”

Decision Matrix

Use this matrix before starting a build.

Use AI

Use AI when the task benefits from generation, retrieval, synthesis, language understanding, or multimodal interpretation, and the risk is manageable.

Use AI With Human Review

Use AI with review when the output can help but should not be final. This is common in support, legal, healthcare, finance, hiring, education, and safety-sensitive workflows.

Use Traditional Automation

Use rules, scripts, forms, or traditional ML when the task is deterministic, structured, and easy to verify.

Do Not Automate

Do not automate when the workflow is rare, poorly understood, high-risk, or valuable precisely because of human judgment.

We can help you decide where AI belongs and where it does not. The right answer is often smaller, clearer, and more useful than the first demo idea.

Summary

“Can AI do this?” is a starting question, not a product strategy.

The stronger questions are about value, risk, evaluation, and workflow fit. They force the team to decide whether AI makes the work better, whether users can trust it, and whether the system can improve.

The best AI teams are not more excited. They are more selective.

Next Steps

Take one AI idea and classify it:

• Use AI.
• Use AI with human review.
• Use traditional automation.
• Do not automate.

Then write down why. If the reason is only capability, keep working.

Need help turning an AI idea into a working system? We build, break, and explain AI. Then we hand it back working.