The Can vs Should Matrix: Deciding When to Use AI at Work
Somewhere in the middle of my Technology and AI course, a student always asks a version of the same question: just because I can use AI for this, should I. That gap, between what is technically possible and what is actually appropriate, is exactly what the Can versus Should matrix is built to help you navigate.
The Two Axes
Can (does the tool technically produce a usable result for this task) and Should (is it appropriate given the stakes, the audience, and what is actually being represented as yours). High on both is the easy case. The dangerous quadrant is high Can, low Should: tasks the tool handles competently, where using it anyway creates a real problem, whether that is accuracy, trust, or misrepresentation of effort.
The Can versus Should matrix, filled in
Everyday work tasks placed by what the tool can do and what it should do.
도구가 할 수 있는 일과 해도 되는 일을 기준으로 업무를 배치한 예.
Do it yourself
The task is fine to delegate, but the tool is not good enough yet. A niche legal translation.
맡겨도 되지만 아직 도구가 부족하다.
Use it, and say so
Routine drafts, data cleaning, meeting summaries. Check, then disclose if asked.
쓰되 검토하고, 물으면 밝힌다.
Skip the tool
Neither capable nor appropriate. A condolence note to a colleague.
할 수도 없고 해서도 안 되는 일.
The danger zone
The output looks good, but it should be yours. A performance review, a graded essay.
결과물은 좋아 보여도 본인이 해야 하는 일.
A CareerComms framework.
The bottom right cell is the one worth memorising. Everything in it looks like a productivity win, which is why it is dangerous: the output is good enough that nobody questions it, and the problem only appears when someone finds out how it was made. I wrote about the learning version of this cell in the danger zone, where perfect grammar hides the absence of understanding.
Where the Should Question Actually Comes From
Should is rarely about the tool’s technical capability. It is almost always about three things: whether the output is being presented as more personally authored than it is, whether the stakes of being wrong are high enough that unverified AI output is a real risk, and whether the task itself is meant to demonstrate your own thinking, like a graded assignment or an original piece of analysis someone is relying on.
Korea has already turned part of this into law. The AI Basic Act, which took effect on January 22, 2026, includes mandatory labeling requirements for certain generative AI outputs, a legal version of the same Should question: even where a tool technically can produce something, the law now requires disclosing that it did (Cooley’s overview of the AI Basic Act).
The official explanation is worth reading in the original. The government’s policy briefing on the Act taking effect on January 22 says generative AI outputs, and deepfakes in particular, must be marked in a way users can clearly recognise, with invisible digital watermarks allowed for things like animation and webtoons. It also says the government will hold off on fact finding investigations and fines for at least one year while companies adjust.
Here is the part that matters for individuals. A later explainer on the same policy briefing site, relaying answers from the government’s AI Basic Act support desk, says the labeling duty falls on businesses that provide and operate generative AI services, not on users who make their own content with them. In other words, the law answers the Should question for the platform. For your own report, slide deck or cover letter, the Should question is still yours.
A Practical Test
Before using AI for a specific task, ask honestly: if the person receiving this knew exactly how it was produced, would they feel misled. If yes, that is a Should problem, regardless of whether the Can answer is a confident yes. This is the same honesty standard behind professional trust generally: it is not about the tool, it is about whether your representation of your own work is accurate.
A Worked Example: Three Tasks at a Pangyo Game Company
Here is how the matrix plays out in one real kind of week. Hana, 28, is a localization project manager at a game company in Pangyo. Her team uses an AI translation tool every day, and her manager has told everyone to use it as much as possible. In one week she has three tasks on her list, and the tool can handle all three.
Three tasks from one week, run through the matrix
An illustrative week for a localization PM at a game company in Pangyo.
판교 게임회사 현지화 PM의 한 주에서 고른 세 가지 업무 예시.
Swipe for more columns · 옆으로 밀어 보기
| Can 할 수 있나 | Should 해도 되나 | Her call 그녀의 결정 | |
|---|---|---|---|
| Patch notes in four languages 패치노트 4개 언어 번역 | High. Fast, mostly accurate. | High. Players expect speed, not authorship. | Use it, then native reviewers check. |
| Review of a junior translator 후배 번역가 평가서 | High. It writes fluent reviews. | Low. Her judgment is the product. | Write it herself, tool for spelling only. |
| Market analysis for the director 본부장 보고용 시장 분석 | Medium. Good structure, shaky numbers. | Depends on what she claims. | Use it for structure, say so in the email. |
A CareerComms example.
The tool performed well on all three. Her decisions differed because the Should answer differed. Patch notes are judged on speed and accuracy, so the tool belongs there with a human check. The review of her junior translator is judged on her own judgment, so a fluent AI draft would misrepresent exactly the thing the review is supposed to contain. The market analysis sits in between, which is why the fix was not to avoid the tool but to say clearly what it did. If your team is still arguing about where its lines are, the AI conversation nobody is having at work is a good place to start, and The Offload Line turns the discussion into a task by task list.
This Changes by Context, Not by Rule
The same task can be a clear yes in one context and a clear no in another. Using AI to draft a first pass of routine internal communication is usually fine. Using it to produce a final analysis you are presenting as entirely your own independent judgment is a different situation entirely, even though the tool performs identically in both cases. The matrix forces you to evaluate context every time, rather than applying one blanket rule.
Saying How It Was Made, in One Line
Most Should problems in the middle of the matrix can be solved with one honest sentence. You do not need a formal disclosure policy to tell the reader which parts came from a tool and which parts came from you.
One disclosure line, from vague to accurate
An illustrative cover email for an analysis shared with a director.
본부장에게 분석 자료를 보낼 때 쓰는 메일 첫 줄 예시.
Attached is my analysis of Q3 retention in Southeast Asia.
True in a way, but the reader will assume every part is hers.
틀린 말은 아니지만, 전부 직접 했다고 읽힌다.
Attached is my Q3 retention analysis. An AI tool cleaned the data and drafted the tables; the conclusions are mine.
One sentence, and the reader knows what to trust and what to check.
한 문장으로 무엇을 믿고 무엇을 확인할지 알려 준다.
A CareerComms example.
The second version costs Hana nothing and protects her if the numbers turn out to be wrong, because the director now knows the tables came from a tool and the conclusions came from her. That is the same principle behind AI will write your email, but it will not save your reputation: the tool can carry the words, but you carry the accountability. When something does go wrong, who benefits when AI gets it wrong is the question to ask next. Before you send anything that sits in the middle of the matrix, read it once through Copy Check and once through the eyes of the person receiving it.
If you want help thinking through where your own AI use sits on this matrix, this is exactly the kind of question we work through with clients and students. Work with us.
Frequently Asked Questions
What is the Can versus Should matrix for AI use?
A two axis decision framework: Can asks whether AI technically produces a usable result, and Should asks whether using it is appropriate given the stakes, audience, and what is being represented as your own work.
When is it not appropriate to use AI even if it works well?
When the output is being presented as more personally authored than it actually is, when the stakes of an error are high and the output is unverified, or when the task itself exists specifically to demonstrate your own thinking.
How do I know if a specific use of AI crosses a line?
Ask honestly whether the person receiving the work would feel misled if they knew exactly how it was produced. If the answer is yes, that is a Should problem, even if the tool handled the task competently.
Does the same AI use case apply the same way in every situation?
No. The same task can be appropriate in one context, like a routine internal draft, and inappropriate in another, like a final analysis presented as your own independent judgment, even though the tool performs identically both times.
CareerComms. “The Can vs Should Matrix: Deciding When to Use AI at Work.” CareerComms, 2026, https://careercomms.com/can-vs-should-ai-matrix/.
Quotation with attribution and a link is welcome. Republication without permission is not.
STAY SHARP
Get career frameworks, not career advice.
Practical tools and structured thinking for your next career move. New posts, tools, and resources — no fluff, no filler.
Get updates →Try it now, no signup
The Offload Line
Two questions decide what you hand to the machine and what stays with you. Plot your real tasks and read the shape.
The Briefing, Twice A Month
One idea, one tool, one thing happening in Seoul. No filler, unsubscribe in one click.
