Don't Be the Relay: The Meat-Proxy Norm for AI Rollouts

When you roll AI-enabled workflows out across a revenue organization — in my case, Slack plus Agentforce, still in progress — the failure mode you end up watching for isn't the one the vendor risk deck warns about. It isn't hallucination, and it isn't the model being wrong. It's a person being absent: someone pastes the model's answer into a thread, unread, with their name attached, and moves on.

The internet recently handed that behavior a name. In an August essay titled "Don't be a meat proxy", software engineer Niklas Gruhn described the experience of asking a colleague a question and getting back an AI's response verbatim — long, polished, and obviously never read by the person who sent it. Simon Willison amplified it the same week, and the term spread from there. The etymology does the work: "meat" is the old cyberpunk word for the human body, "proxy" is the networking word for a pass-through intermediary. Your body showed up to the exchange. Your judgment didn't.

I think the term deserves a place in every AI rollout plan, because it names the human-layer failure precisely — and rollouts have plans for every layer except this one.

The Cost Doesn't Disappear. It Moves.

Here's the mechanism, stripped of the mockery. Every piece of model output carries a verification cost: someone has to check that the numbers are real, the claims hold, and the answer actually fits the question. When you relay output unread, that cost doesn't go away. It transfers — silently — to whoever receives it. And it transfers with interest, because your name on the message reads as endorsement. The recipient reasonably assumes a human stands behind these words, so they either extend trust the words haven't earned or burn time re-verifying something you implied was already checked.

I wrote in AI on Governed Data that an agent doesn't average your bad data — it speaks it fluently, which is exactly what makes the failure dangerous. The meat proxy is the same property one layer up. Fluent output looks verified. A confident, well-structured answer gives the recipient no visible signal that nobody checked it, the way a hesitant human answer would. The data-layer version of this failure erodes trust in the tool. The human-layer version erodes trust in you — and in a revenue organization, where the output lands in front of sellers, executives, and sometimes customers, that trust is the thing the whole rollout runs on.

Verbatim Isn't the Sin

One boundary worth drawing carefully, because the norm fails if it's stated too broadly: quoting the model verbatim is not the problem. If someone asks for the exact command, the config snippet, or the SOQL query, forwarding it untouched is often precisely the right move. The same goes for pasting a stack trace or a generated report the recipient asked to see raw.

The sin is narrower: passing something along without having opened the envelope. Verbatim-and-verified is a fine deliverable. Verbatim-and-unread is the meat proxy. The distinction matters for rollout policy, because a rule that reads "never paste AI output" gets ignored within a week — it bans obviously useful behavior. A rule that says "never pass along what you haven't verified" bans only the failure.

The Norm: Verify and Rewrite, or Hand Off the Prompt

Gruhn's essay prescribes the fix, and it maps cleanly onto how I frame the expectation for teams adopting these tools. There are exactly two acceptable moves with model output, and both are honest ones:

Verify it, then write the response in your own words. Read the output, check it against what you know, discard what's wrong, and restate what survives as your own answer. The rewrite isn't busywork — it's the certificate. Writing it yourself is hard to fake without actually understanding the content, which is precisely why it works as evidence that the verification happened. This is the move when the answer is going out under your name.

Or hand off the question honestly. "I asked the model and got something that looks useful — here's the prompt." That's not laziness; it's accurate labeling. You're pointing at the tool instead of relaying for it, and the recipient gets to decide how much verification the output deserves, with full knowledge of where it came from. This is the move when you don't have the time or the context to verify — which is a legitimate state to be in, as long as you say so.

The one move that's off the table is the middle path: the model's words, your name, nobody's judgment. Stated as a principle: AI output is an input to your judgment, never a deliverable with your name on it. Everything else about the rollout — which tools, which workflows, which guardrails — is negotiable by use case. This part isn't.

This Site Runs the Same Rule

The norm isn't hypothetical for me; it's in this site's colophon. The About page discloses that these posts are drafted with AI assistance — and that every claim and number is mine and verified. That second clause is the entire point. The disclosure without the verification would just be labeled meat-proxying.

The drafting of this very post made the case for the rule. The research brief I started from arrived — AI-generated — with roughly ten citations. On checking, two were real and primary: Gruhn's essay and Willison's post. The rest were unreachable or led to content farms that had themselves aggregated the story, some of them likely machine-written. Those citations looked exactly as authoritative as the real ones. An article about unverified AI relay, sourced from unverified AI relay, would have been a genre of irony I'd rather read about than publish.

Usage Is Not Adoption

Here's why I treat this as an adoption problem rather than an etiquette problem. Every rollout dashboard measures usage: seats active, prompts issued, messages generated. A team of meat proxies scores brilliantly on all of it — the tool is being used constantly. What the dashboard can't see is that judgment has left the loop, which means the organization is now producing output no one stands behind, at a volume no one can verify after the fact.

I've written before about the gap between go-live and adoption: shipped and used are different claims, and used and operated with judgment are different again. The meat-proxy norm is where that lesson lands for AI. The rollout succeeds not when the usage numbers climb, but when the humans in the loop are still doing the one job the model can't: standing behind the words.

Prompt the model all you like. Just don't be the relay.