Please Don't Paste Unreviewed AI Output

TL;DR

AI is a great way to draft, research, and accelerate your thinking. It is not a way to outsource your judgment. Don't paste a generated answer into a chat, ticket, email, or customer thread when you can't vouch for it. If you don't understand the domain well enough to spot an error, the recipient usually can't either — so now there are two people scratching their heads, and one of them came to you for the answer.

The rule: Do the verification offline. Then answer in your own words, say what you confirmed and how, cite where it came from, and be explicit about what you're still unsure of.

What "slop" actually means

Slop isn't "content that used AI." Slop is content where nobody did the thinking. Telltale signs:

Pasted verbatim, formatting and all — bold headers and a tidy numbered list for a one-line question.
Confidently answers a question the sender couldn't answer themselves and can't defend in a follow-up.
Generic: no product version, no tenant, no SKU, no reference to the actual error in the actual environment.
Cites docs that don't exist, settings that aren't there, or cmdlets that were renamed three releases ago.
Answers a similar question rather than the one that was asked.
Five paragraphs where "yes, but only on v2.14 and later" would have been the whole answer.

The cost is not just a wrong answer. It's trust. Someone who catches one hallucinated setting will re-check everything you say afterward — including the parts you got right.

Chat etiquette

Please don't do this:

14:02 customer: Conditional Access policy isn't applying to our guest users. Why?
14:02 you:      [pastes the question into an AI tool, copies the answer back]
14:03 you:      Great question! There are several reasons a Conditional Access
                policy may not apply to guest users:
                **1. Policy scope** - Ensure the policy targets "All guest and
                external users"...
                **2. Licensing** - Conditional Access requires Entra ID P1...
                **3. Exclusions** - Check for excluded groups...
                Let me know if you'd like me to elaborate on any of these!
14:05 customer: We checked all three. Still not applying.
14:06 you:      [has no idea what to say next]

You bought yourself sixty seconds and spent the customer's afternoon. You also told them, in effect, that they could have typed the question into a chatbot themselves and skipped the meeting.

Please do this instead:

14:02 customer: Conditional Access policy isn't applying to our guest users. Why?
14:03 you:      Good question - I want to check this against your actual config
                rather than guess. Can you send the policy name and a sign-in log
                entry for an affected guest? I'll dig in and come back by 16:00.
15:40 you:      Found it. The policy targets "All users," which does not pick up
                B2B guests unless "All guest and external users" is selected
                explicitly. Documented here: [link]. I reproduced it in a test
                tenant to confirm.
                Two things I have NOT verified: how this interacts with your
                cross-tenant access settings, and the B2B direct-connect case.
                Want me to test those next?

Slower by ninety minutes, worth many times more. Note what the second version does: it commits to a time, states what was confirmed and how, cites a source, and names the open questions instead of papering over them.

Say "I don't know" out loud

"I don't know, let me find out" is a professional answer. A fabricated answer delivered in a confident tone is not. Nobody ever lost credibility by scoping their own uncertainty; plenty of people have lost it by being fluent and wrong.

you: Honestly, I don't know off the top of my head, and I don't want to
     guess on something that touches your production auth path.
     Give me until tomorrow morning and I'll come back with a real answer.

Customer-facing threads

In a customer conversation, you are the brand. A pasted generated answer signals one of two things, and both are bad: either you didn't engage with their problem, or you can't tell the difference between a plausible answer and a correct one.

Don't:

Paste generated content into a customer thread without reading every line and standing behind it.
Put the customer in the position of fact-checking you.
Answer a question you don't understand in order to avoid the discomfort of saying so.
Send generic guidance that ignores the environment, version, or licensing they already told you about.
Relay a hallucinated setting, cmdlet, portal path, or doc link. One of these undoes a lot of goodwill.

Do:

Use AI to generate a hypothesis list, then go test the hypotheses yourself.
Reproduce it, or check it against official documentation, before you assert it.
Open every link you intend to send. Run every command you intend to recommend.
Rewrite it in your own voice, at the length the question actually deserves.
Separate verified from likely from worth testing, and label them as such.
Escalate or loop in a specialist instead of improvising. That reads as strength, not weakness.

The test: If the customer replies "where did you get that?" or "we tried that, what now?" — can you answer on the spot, without going back to the AI? If not, it isn't ready to send.

A reasonable way to use the tool

1. Ask the AI. Get the draft, the hypotheses, the starting points.
2. Verify every factual claim against docs, code, or a live repro.
3. Delete everything you could not confirm.
4. Rewrite what's left in your own words, at your own length.
5. Add what the AI could not know: their environment, their history,
   their constraints, and your actual recommendation.
6. State your confidence and your open questions.
7. Send.

Steps 2 through 6 are the job. Step 1 just got cheaper.

Before you hit send

Did I read every single line of this?
Can I defend every factual claim without reopening the AI?
Did I open the links and run the commands?
Is it specific to their environment, version, and constraints?
Have I said clearly what is verified and what is not?
Is it as short as the question deserves?
Does it sound like me?
If this turns out to be wrong, am I comfortable owning it?

Any "no" means it isn't ready. Any "no" you'd rather not think about definitely means it isn't ready.

Linking someone here should be a nudge, not a public correction. Prefer a direct message over a reply in the thread, and prefer "here's how we handle this" over "you did this wrong."

you: Hey - quick note, just between us. That last reply read as straight
     AI output, and the customer pushed back on one of the steps.
     Worth a look: aka.ms/nosloppypasta. Happy to help verify the next one.