Most AI rollouts skip the one step that decides everything — making the work visible before redesigning it. The Missing Middle Mapper reads an audit of a single workflow, surfaces the judgment hiding inside each step, then tells you where AI fits, where it doesn't, and what to fix first.
Open the folder in Claude Code · upload your audit · get a map.
The problem
Two failure modes. One root cause.
Failure mode 01
Tried AI. Nothing improved.
You gave people tools, encouraged experimentation, maybe ran a pilot. Usage went up. Impact didn't. The AI made the work faster but not better — or faster and worse at the same time. Now there's skepticism, and no one can explain what went wrong.
Failure mode 02
Haven't really started.
You know AI matters. You've heard the mandate. But the gap between "use AI more" and knowing where it actually fits in the work feels enormous. The learning curve looks steep, the starting point isn't clear — so nothing has moved.
These feel like different problems. They're the same one: both skipped the step of making the work visible before changing it. AI can't reliably improve what the organization can't see — the judgment calls, the undocumented handoffs, the quality bars that live only in people's heads. That's where the value is. And it's invisible.
What it does
It reads the work. Then it makes the call.
Feed it an audit of one workflow. For every step, the Mapper surfaces the hidden judgment layer, then classifies the step into one of four verdicts — and tells you how confident it is and why. It doesn't hand the decision back to you. It decides, names the risks, and tells you what to fix before you automate anything.
Automate
Repetitive, rule-based, recoverable if wrong. Hand it off entirely.
Augment
High-value work where AI sharpens human judgment without replacing it.
Keep Human
Taste, context, relationships, real stakes. Not safely delegated.
Stop / Rethink
A process that runs on habit or ambiguity. Fix it before automating it.
When a step looks simple from the outside but carries years of expertise underneath, the Mapper won't classify it confidently until that judgment is on the page. When a workflow is structurally broken, it says so — and refuses to help you automate the wrong thing.
How it works
Audit in. Opportunity map out.
1
Fill the audit
Describe one workflow honestly: what each step produces, where it breaks down, what judgment is applied, and what expertise isn't written down anywhere. A guided template walks you through every question — and the deeper the audit, the more confident the map.
2
Run it in Claude Code
Open the folder as a workspace, upload or paste your completed audit.
> Produce an AI opportunity map from this workflow audit.
3
Get your map
An interactive HTML map you can open in a browser and share with your team — every step classified, reasoned, and risk-checked. If the audit is thin anywhere, a gap report lands beside it telling you exactly what to go find out, and why the classification depends on it.
See it
One step from a real map.
Internal content approval workflow/Step 02 of 04/Click to expand the judgment layer
02
Editorial Review
Editor reviews drafts for brand, tone, accuracy, and strategic alignment
AugmentKeep HumanProvisional
Expand+
Hidden judgment layer
What actually happens here
The editor applies different treatment by content type — blog, customer email, and social each get a different level of review — but the criteria for each are not documented
A style guide exists but doesn't cover everything; "sounds right for the brand" is the operative standard
The editor tracks what the exec team cares about this week — strategic messaging priorities that change and aren't formally communicated down
The comms lead reviews sensitive pieces, though what qualifies as "sensitive" is also undocumented
Where it breaks down
Drafts get sent back multiple times with feedback that isn't always specific enough to act on
Output consistency depends on who's reviewing — the editor's judgment isn't systematically the same across all content types
During launches, review volume exceeds capacity — the editor spends 20–25 hours per week on review alone, and more during spikes
No documented quality standard exists for AI to check against; the criteria live in the editor's and comms lead's heads
This classification is provisional — see gap report for what's missing.
Augment
Pre-review triage: before the editor opens a draft, AI checks it against the existing style guide and flags structural issues — grammar, prohibited language, missing CTAs, overly long sentences, format problems by content type. This handles the mechanical layer of review and surfaces specific issues with locations, so the editor's attention can go to judgment rather than hunting. A style guide already exists; AI can operationalize it immediately, even before deeper criteria are documented.
The style guide AI checks against will gradually become outdated if it isn't owned and maintained. Assign a specific owner and a review cadence — otherwise AI will confidently flag things the team stopped caring about and miss things they started caring about. Also watch for anchoring: if AI flags issues before the editor reads the draft, the editor's judgment may be unconsciously shaped by what AI found first. Consider having editors note their preliminary read before reviewing the AI triage output.
Keep Human
Final editorial judgment: brand voice beyond what's in the style guide, exec-priority alignment, tone calibration by audience and moment, and the call on what goes back vs. what goes through. This judgment layer is real — the editor can identify what's wrong but often can't fully articulate the criteria in advance. Until the tacit standards that drive these calls are externalized and documented, AI cannot reliably assist with them.
The editorial judgment that makes this step work lives primarily with the editor and comms lead. If either is unavailable during a launch week, the workflow stalls. The judgment isn't documented, can't be delegated, and can't be backed up. Even if AI augmentation doesn't expand here, the team should invest in knowledge codification — not to automate, but so the judgment isn't fragile.
Also included in every map
A confidence rating on every callConfident, Provisional, or Conditional — never false certainty.
The skills your team needsWhat capabilities each classification actually requires to execute.
A measurement baselineCurrent state → target, so you can prove AI improved the work.
You applied AI without making the work visible first. The map shows what you skipped, the gap report tells you what to surface, and the risk callouts name the problems you're probably already seeing — over-reliance, quality erosion, inconsistent practice.
→ See what you skipped.
Haven't started
The audit is your starting point.
You don't need AI fluency before you begin. Audit one priority workflow. The map tells you where AI can help, which tells you what to learn — so fluency becomes just-in-time instead of front-loaded and overwhelming.
→ Learn the right thing first.
Leading transformation
A repeatable diagnostic.
Run it across your priority workflows. The maps show where investment creates leverage, the skills sections tell you what to build, and the gap reports reveal where work isn't visible enough yet — the thing standing between your fluency numbers and real impact.
→ See where to invest next.
Get started
Map one workflow this week.
It's open source and runs in Claude Code. Three example audits are included — clean, partial, and broken — so you can watch the reasoning before you run your own.
Download the repo and fill out audit-template.md — it walks through every question with help text on what a good answer looks like. Drop the completed file into Claude Code and run.
Answer the questions step by step and download a properly formatted .md file, ready to feed the operator. Good for a first audit, or a team that wants more structure.