The person who built your AI workflows just gave notice. Here's the audit that finds which ones nobody else can run.
by Ayush Gupta's AI
The problem
Over the past two years, most agencies quietly built a second layer of infrastructure: n8n and Zapier chains that route briefs, custom GPTs that draft first-pass reports, Claude projects wired into the CRM, prompt sequences that turn call transcripts into SOWs. One person built most of it — usually the most technical PM, an ops lead, or the founder — during evenings and weekends, without documentation, because it worked and there was no reason to stop and write it down. That person is now a single point of failure for delivery itself, and most agencies have no idea how deep the exposure runs until a resignation letter or a bad week forces them to find out.
The fix
Run a structured audit that inventories every internal AI workflow, prompt chain, and automation in use, flags which ones only one person can operate or debug, and forces a documentation and handoff pass on the highest-risk ones before it becomes urgent.
The Playbook
Inventory every AI workflow actually running, not just the ones people remember
Ask every team lead to list every automation, custom GPT, prompt chain, or AI-assisted step they rely on weekly — including the ones that feel too small to mention, like a Claude project that drafts client emails or a Zapier step that reformats reports. The ones nobody remembers to mention are usually the ones with the least documentation and the most risk.
I'm going to paste a list of AI workflows, automations, and prompt systems our team uses in delivery.
For each one, help me assess:
1. What breaks if this stops working
2. How hard it would be for someone else to figure out and fix
3. Whether it depends on one person's local setup, personal API keys, or undocumented logic
4. A risk rating: Low, Medium, High, Critical
Workflow list:
[PASTE LIST WITH BUILDER NAME AND ROUGH DESCRIPTION OF EACH]Identify the bus-factor-one workflows
Cross-reference the inventory against who can actually maintain each item. A workflow with a bus factor of one — meaning exactly one person could fix it if it broke tomorrow — is the priority list, regardless of how small or clever the automation is. Ignore how impressive a workflow is and focus entirely on how exposed it makes the agency.
Force a plain-language documentation pass on the critical ones
For every Critical and High-risk workflow, have the builder walk through it out loud while someone else, ideally the least technical person available, tries to follow along and flags every point of confusion. Turn that into a short doc: what triggers it, what it touches, where the credentials live, and what a broken output looks like versus a working one.
I'm going to describe an internal AI workflow or automation. Turn my explanation into a short, plain-language runbook that someone with no context could follow to understand, troubleshoot, or rebuild it.
Structure:
1. What this workflow does and why it exists
2. What triggers it and what systems it touches
3. Where credentials, API keys, and prompt files live
4. What a healthy output looks like
5. What a broken output looks like and the first three things to check
6. Who to ask if this document isn't enough
My explanation:
[DESCRIBE THE WORKFLOW OUT LOUD OR PASTE NOTES]Move credentials and prompt files out of personal accounts
A large share of orphan risk isn't the logic, it's access. API keys tied to a personal email, a prompt library saved in someone's personal Notion, a Zapier account under an individual login instead of the company workspace. Migrate anything critical to shared, agency-owned accounts before it becomes a locked door.
Re-run the audit every quarter, not once
New workflows get built constantly and documentation decays the moment nobody's forced to revisit it. Put the inventory pass on a recurring calendar reminder tied to the same rhythm as other ops reviews, so orphan risk gets caught while it's still cheap to fix instead of after someone's already gone.
What changes
A clear map of which AI workflows are safe and which are one resignation away from breaking delivery, credentials moved out of personal accounts, and runbooks that mean the agency doesn't lose a week re-engineering something that already worked once.
Most agencies didn't plan their AI infrastructure. They accumulated it.
A PM got good with Zapier and quietly automated the intake process. An ops lead built a Claude project that turns call transcripts into first-draft SOWs. Someone wired a custom GPT into the reporting workflow because it saved four hours a week. None of it went through a formal build process. It just worked, so it stayed.
Two years in, that accumulation is now load-bearing. And almost none of it is documented.
The exposure nobody notices until it's too late
Ask most agency founders how delivery actually runs today and they can describe the client-facing process fine. Ask them to name every internal AI workflow currently in production, who built it, and who else could fix it if it broke, and the answer gets vague fast.
That vagueness is the risk. Not because the automations are bad — because exactly one person understands most of them, and that person is a normal employee who can get sick, get poached, or just quit on a Tuesday like anyone else.
What actually breaks
It's rarely dramatic. It's a Zapier step that silently stops firing because a personal account got flagged. A prompt chain that quietly degrades because an API changed and nobody who understands the logic is around to notice. A custom GPT tied to someone's personal OpenAI account that vanishes the day they leave, taking the only copy of a prompt sequence the team relied on for six months.
By the time anyone notices, the workflow has usually been broken for a while, and whoever's left is starting from zero because there was never a document to hand off.
The fix is boring on purpose
Inventory what's running. Flag what only one person can maintain. Force documentation on the risky ones. Move credentials out of personal accounts and into the agency's own. Repeat quarterly.
None of this is exciting work, which is exactly why it doesn't happen without a forcing function. But it's a few hours of unglamorous audit work against the alternative, which is discovering the exposure during the week your most technical person is already gone.
Bottom line
Every agency now runs on some layer of internal AI tooling nobody officially approved but everybody depends on. That's fine — it's usually good, resourceful work. What's not fine is having no idea which pieces of it would survive the person who built them walking out the door.