When to Use AI and When Not To

One of the least-discussed aspects of using AI effectively is knowing when not to use it. Most writing about AI — including a lot of my own — focuses on the capabilities: what it can do, how to prompt it, how to build workflows around it.

But judgment about when AI is the right tool is as valuable as skill at using AI when it is. Bad AI use isn't just wasted effort — it can actively make things worse by introducing errors, false confidence, and over-complex solutions to simple problems.

Here's the framework I actually use.

The Four Quadrants of AI Use#

I think about AI use across two dimensions: task clarity (how well-defined is the task?) and consequence of error (how bad is it if the AI gets it wrong?).

High clarity, low consequence: Use AI aggressively. These are the bread-and-butter AI use cases — data transformation, first drafts of routine communications, summarization, classification, enrichment. The task is well-defined enough for the AI to understand, and errors are easy to catch and fix. This is where you should be delegating heavily and automating broadly.

High clarity, high consequence: Use AI with verification. The AI can do the work, but you need to verify the output before acting on it. SQL queries on production data, legal language in contracts, specific numerical claims in financial documents. Use the AI to accelerate, not to replace judgment.

Low clarity, low consequence: Use AI as a thinking partner. When you're not sure exactly what you want but the stakes are low, AI is excellent for exploring options, generating alternatives, and helping you clarify your own thinking. "Help me brainstorm approaches to this problem" is better than "do this thing" when the thing is undefined.

Low clarity, high consequence: Don't use AI as the primary driver. Strategic decisions, relationship-sensitive communications, novel situations where you haven't established the AI's accuracy in similar contexts — these are areas where human judgment should lead. AI can provide research, context, and options, but shouldn't be driving.

Most people apply AI well in the first quadrant. The others require more deliberate thinking.

Specific Situations Where AI Works Well#

Repetitive processing. Any task that involves doing the same type of thing many times — enriching records, classifying emails, extracting data from documents, formatting text consistently — is AI's strongest domain. The AI applies a consistent logic at scale, which is exactly what you want for these tasks.

First drafts. AI is excellent at getting from "blank page" to "something to react to." For emails, proposals, documentation, reports, social posts — the AI draft gives you a starting point that's usually 70-80% of the way there. You edit and improve rather than create from scratch.

Research and synthesis. Gathering information on a topic, summarizing long documents, comparing options, identifying patterns in data. AI does this faster than humans and often surfaces things that would take significant manual work to find.

CRM and database queries. Natural language queries against structured data are an AI strength. "Who haven't I contacted in 30 days" is faster through an AI interface than through filter menus, and more flexible than a pre-built report.

Structuring thinking. Giving an AI a messy, unstructured brain dump and asking it to organize, clarify, and identify gaps is a surprisingly high-value use. The AI doesn't generate new ideas; it helps you see the structure of your existing ones.

Specific Situations Where AI Fails#

Novel judgment calls. Situations you haven't encountered before, where there's no established pattern to apply, where the relevant considerations are subtle and contextual. AI works by pattern-matching against past data. Novel situations are where its patterns don't apply.

Relationship-sensitive communication. A message where the exact phrasing, tone, and emotional register matter enormously — a difficult negotiation, a termination conversation, a personal apology. AI can draft these, but the output often lacks the precise human awareness these situations require. Use AI for rough structuring, not final language.

When you don't know what you want. AI can help you explore options when you're uncertain, but it can't figure out your actual goals for you. If you don't have a clear enough sense of the desired outcome to evaluate AI output, you'll either accept bad output or endlessly iterate without converging.

When accuracy is critical and verification is hard. Medical diagnoses, legal interpretations, specific numerical data in high-stakes documents. AI can hallucinate plausibly-sounding wrong answers. If you can't verify the output efficiently, don't use AI as the primary source.

Simple tasks that don't need it. Sending a two-line email to a colleague you know well. Remembering to buy groceries. Deciding what to have for lunch. The cognitive overhead of engaging with an AI for truly simple tasks exceeds the value. Know the floor below which AI isn't worth it.

The Verification Budget#

One of the most useful practical frameworks I've developed: the "verification budget."

For any AI-assisted task, ask: how long will it take me to verify the AI's output? If that verification time is more than half of what it would have taken to do the task manually, the AI might not be saving you much.

This is especially relevant for high-consequence tasks. Verifying an AI-drafted contract might take as long as writing parts of it yourself. Verifying an AI-generated financial analysis might require you to rebuild the analysis mentally anyway.

The verification budget tells you where AI is genuinely saving time vs. where it's shifting the work from "doing" to "checking" without much net benefit.

The Context-Sensitivity Calibration#

AI output quality degrades sharply when you leave the domain of its training data — when you're operating in a context that's unusually specific, recent, or niche.

Questions about your specific company, your specific industry niche, your specific customers, your specific product — the AI has less relevant training data and is more likely to make confident generalizations that don't apply to your situation.

For DenchClaw users, this is why having good local context data matters so much. When the AI is querying your DuckDB database with your actual contacts and your actual pipeline, it's working from specific, accurate data about your situation. When it's generating generic advice about "typical sales cycles," it's working from training data that may not match your reality.

The practical implication: the more domain-specific the task, the more important it is to give the AI your specific context rather than asking for general-case advice.

Building Your Personal AI Judgment#

The meta-skill underlying all of this is AI judgment — knowing when you're in a "use AI aggressively" situation vs. when you're in a "lead with human judgment, use AI for support" situation.

This judgment develops through experience and reflection. I try to notice, after any significant AI-assisted task:

• Did the AI output need heavy editing?
• Did I catch any significant errors?
• Did the AI understand what I actually wanted, or did I have to re-prompt significantly?

Over time, this builds an accurate personal calibration: the situations where the AI is reliably excellent, the situations where it needs significant supervision, and the situations where I'm better off without it.

The goal isn't maximum AI use. The goal is maximum value from work done, with AI as one tool among many. Sometimes the best tool is the AI. Sometimes it's your own judgment. Knowing which is which is the whole game.

Frequently Asked Questions#

How do I know if an AI output is good enough to use?#

Define your acceptance criteria before generating the output. "I'll use this email draft if the tone is appropriate and all factual claims are correct" gives you a clear evaluation standard. Evaluate against your criteria, not against "does this feel impressive?"

When should I never use AI without review?#

Any output that will be: sent to an external party, published publicly, used as input to another high-stakes process, or cited as a factual claim. These categories warrant human review regardless of how confident the AI seems.

Is it possible to over-use AI?#

Yes. Over-reliance on AI for tasks that benefit from direct human attention, relationship sensitivity, or creative judgment can lead to outputs that feel generic, miss important context, or erode the personal quality of your work. Use AI to multiply your effectiveness, not to replace your judgment.

How do I explain to my team when to use AI vs. not?#

Use the four-quadrant framework: high clarity + low consequence = use freely; high consequence = verify; unclear goals + high consequence = lead with human judgment. Make a few specific examples in your team's context concrete and it becomes intuitive quickly.

Ready to try DenchClaw? Install in one command: npx denchclaw. Full setup guide →