Prompt sprawl is the new shadow IT.

Ask a team lead where their best prompts live and you'll get a tour of the modern shadow stack: a personal Notion page, a pinned Slack DM to self, a text file called prompts-final-v3.txt, a browser bookmark to an old ChatGPT conversation that still has "the good version" in it.

These are not toy artifacts. In most AI-adopting organizations, some of these prompts now sit in the production path of real work — the account manager's renewal-email prompt, the legal team's contract-summary prompt, the support lead's escalation-triage prompt. Work product that reaches customers is being shaped, daily, by text files nobody else has read.

IT has seen this movie. Unsanctioned SaaS, personal Dropboxes, macros in a spreadsheet that quietly became load-bearing. The pattern always has the same three acts: the tool is adopted because it works, it spreads because it's invisible, and it surfaces as a problem the day something breaks or somebody leaves. Prompts are in act two.

Prompts are business logic now

The category error is treating prompts as personal productivity — like keyboard shortcuts or an email signature. Look at what a prompt actually is in functional terms: it has defined inputs (the variables the team fills in), defined outputs (the artifact it produces), failure modes (the edge cases where it generates something wrong), and versions (v3 exists because v2 had a problem someone found the hard way).

Inputs, outputs, failure modes, versions. That's software. More precisely, it's business logic — the encoding of how your organization does a piece of work. A prompt that determines what goes into a customer-facing renewal email is doing the same job as the template engine in your CRM, except the template engine went through procurement, has an owner, and gets reviewed when it changes.

The prompt got none of that, because it never crossed a boundary where controls live. No repository, no deploy, no ticket. It went from someone's experimentation directly into daily production use, and every improvement since has happened in the same private channel.

Three ways this bites

Departure risk. The prompt's author resigns, and the prompt resigns with them — locked in a personal workspace, or simply undiscoverable. The team's output quality drops in a way nobody can name, because the asset that was lost never appeared on any asset list. Institutional knowledge used to walk out the door slowly, in expertise. Now it walks out in a single unshared document.

Divergence. Five account teams each have their own version of "the proposal prompt," forked from a screenshot at different points in its evolution. Customers now receive materially different quality depending on which team they landed with. When leadership decides the messaging must change — new positioning, new legal disclaimer — there is no mechanism to change it everywhere, because there is no everywhere. There are five private copies, three of which nobody will remember to update.

The governance blind spot. This is the one that should worry security and compliance. Prompts contain instructions about what data to include, what tone to strike, what claims to make. A prompt that tells the model to paste in the full customer record for context, or to include aggressive claims a regulator would frown at, has never been reviewed by anyone whose job is to catch that. Customer-facing output is being generated under instructions that exist outside every review process the organization runs.

Amnesty first, then management

The wrong response is a policy banning unofficial prompts. That's how shadow IT is fertilized, not fixed — the prompts work, so they'll simply go quieter. The right response is the one that eventually worked for SaaS: make the sanctioned path better than the shadow one.

1. Run an amnesty, not an audit. Ask teams to bring their working prompts into a shared library, no questions asked about why finance has been running unreviewed customer-communication prompts for a year. The goal is visibility. Blame kills visibility.
2. Give prompts owners and versions. Each shared prompt gets a named owner and a change history. Not heavy process — just enough that "which version are we on, and what changed" has an answer.
3. Review the ones that matter. Not every prompt needs governance. The ones producing customer-facing or compliance-adjacent output do. A lightweight review — what data does it pull in, what does it instruct the model to claim — catches the worst problems in minutes per prompt.
4. Make the library the easiest place to work from. Fill-in variables, one-click use, team sharing. If using the managed copy is faster than hunting for the DM, the shadow copies wither on their own.

The uncomfortable truth about shadow IT was always that it mapped real demand the official stack had failed to meet. Prompt sprawl is the same signal: your organization has already rebuilt chunks of its business logic on top of language models, bottom-up, without telling you. You can read that as a threat, or as the most honest adoption data you'll ever get — a map of exactly where AI is already doing real work. Either way, the logic is running. The only question is whether it runs inside your controls or outside them.