Frameworks · Template

Map the workflow before you pick the AI.

Updated July 2026

The AI Workflow Mapping Template is a fillable canvas for documenting one workflow end to end: triggers, steps, decisions, handoffs, waits, and rework. Then a seven-mark overlay tells you, step by step, where AI helps, where ordinary automation wins, and where judgment stays human.

Why this exists

Automation amplifies the process it is given.

A process you don't fully understand still runs on quiet human judgment: people fix bad inputs, route weird cases to whoever knows, and delay things that look wrong. Automate around that hidden judgment and the corrections disappear while the volume goes up. And modern AI raises the stakes, because it produces fluent, confident output even when it's wrong.

So this template makes you map first: what actually happens, with numbers, workarounds and all. It's built on the methods that serious operators already trust: NIST's AI Risk Management Framework treats mapping as the go/no-go gate, lean value-stream practice supplies the per-step numbers, and the peer-reviewed task-suitability research supplies the step-by-step verdict.

The canvas

Five parts, filled in order.

PartNameWhat it capturesTime
A Frame Scope, boundaries, trigger, outcome, volume, constraints 15 min
B Flow The step table: actors, systems, decisions, handoffs, waits, rework 30–45 min
C Surroundings Systems inventory, data inventory, risks and existing controls 15 min
D Friction Review the map against the twelve-problem checklist, rank the top 3–5 15 min
E AI overlay Per-step involvement mark, review design, prerequisite, metric 20–30 min

The order is the method: current state with real numbers first, problems second, AI last. Run Part E in a separate later session, after the map has been validated by the people who do the work. A first pass that leaves fields marked "?" is a legitimate result; unknowns tell you exactly what to verify.

Part A · Frame

Ten questions that frame the workflow.

01
Workflow name
What single workflow are we mapping, in verb-noun form?
"Qualify inbound leads", not "Sales". Vague names produce sprawling maps that never finish.
02
Owner
Who gets asked when this workflow breaks?
Discovering that nobody owns it is a finding, not a failure. Unowned workflows cannot be improved.
03
Trigger
What exact event starts it, and how do you know it started?
Explicit start and end boundaries are the one instruction every mapping authority agrees on.
04
End state
What is true when one item is done, and done for whom?
"We sent it" is not an end state. The confirmed outcome is, and it becomes your metric denominator.
05
Customer
Who consumes the output, and what do they complain about?
Often the real customer is the next team downstream. Their complaints are your first problem list.
06
Volume
How many items per week, and what do the peaks look like?
Pull the count from the source system; five minutes there beats a guess. Volume drives all the economics.
07
Effort vs elapsed
Minutes of hands-on work per item, and calendar days start to finish?
The gap between those two numbers is usually enormous, and the gap is where improvement lives.
08
Constraints
What can you not change: regulation, contracts, budget, data residency?
Constraints decide later which AI options are viable at all. Find them now, not after the pilot.
09
Measures
Which 2–4 numbers would prove this workflow improved?
Chosen before any tool talk, so the tool stays honest. These become your pilot baseline.
10
Why now
What hurts, and who is asking?
"Because AI" is not a reason. No pain and no pending decision means map a different workflow.
Part B · Flow

One row per step. Fourteen things to record.

A step is one unit of work by one actor with a recognizable start and finish. Aim for 5–15 rows; exceptions and rework loops get suffixed numbers (4a, 4b) pointing back at the step they branch from.

ColumnWhat to recordWatch for
Step name One verb phrase per action Bundled steps hide the handoffs between them
Actor The role or system that really does it Ask who covers absences; that finds unowned steps
Systems Every tool, inbox, and spreadsheet touched The personal spreadsheet is usually the real system
Inputs What the step needs, and where it comes from Ask how often it arrives incomplete
Outputs What it produces, and who consumes it Outputs nobody reads are deletion candidates
Decision The rule actually used, not the official one The single most important field for the AI overlay
Handoff Who is next, via what channel "I mention it in standup" explains the missing days
Wait Idle time before and after the step In office workflows, waiting usually dwarfs working
Touch time Hands-on minutes, clean case vs messy case Averages hide the expensive cases
%C&A Share of incoming items clean enough to process Ask "out of the last 10, how many?"
Exceptions Top failure paths, how often, where they go Anything over ~10% is a standard path in denial
Workarounds What people really do, and what it compensates for Treat as evidence, never as confession
Data What gets captured, what evaporates Uncaptured judgments leave nothing to evaluate AI against
Pain The operators’ own words, unsanitized Predicts adoption better than any ROI estimate

Three of these numbers do most of the diagnostic work. Wait exposes where items sit idle, which usually dwarfs the working time. Touch time, clean versus messy, prices the step and its exceptions. And %C&A (the share of incoming work that's complete and accurate enough to process without chasing or fixing) tells you whether any automation here would run on clean fuel or amplify the dirt.

Part D · Friction

Twelve problems to hunt before anyone says "AI".

#ProblemWhat it looks like
01 Unclear steps Described differently by everyone you ask
02 Bottlenecks Items pile up at one desk or queue
03 Sources of delay Batching, absences, approval queues, external waits
04 Error and rework Low %C&A, send-backs, silent corrections
05 Authority gaps Only one person can do it, or approve it
06 Duplication Same data keyed twice, same check done twice
07 Excessive handoffs More desks than the work requires
08 Unnecessary steps Approvals that never reject, reports nobody reads
09 Time imbalance Days of elapsed time wrapped around minutes of work
10 Undocumented judgment Consistent decisions that exist only in one head
11 Data dead ends Judgments and outcomes never captured anywhere
12 Workaround load The unofficial method has replaced the official one

Items 01–09 adapt ASQ's classic guidance for reviewing a finished flowchart; 10–12 are ours for the AI context. The output is a ranked top 3–5 problem list with step numbers and evidence attached. Those problems, not "we should use AI", are what the overlay responds to.

Part E · The overlay

Seven marks. One per step. Mixed is normal.

MarkClassUse it whenPeople still
[H] Human-only High stakes with no stable pattern, or the step runs on relationships and accountability Everything, as today
[R] Rules, no AI The rule is written, stable, and inputs are structured. Prefer this over AI whenever it qualifies Own the rules, handle the else-path
[A] AI-assisted The human still performs the step; AI retrieves, summarizes, drafts fragments, suggests Perform the step
[G] AI drafts, human approves AI produces the whole work product and checking it is faster than doing it Approve or edit every item
[X] AI runs, humans get exceptions Accuracy proven on your own data, errors cheap and reversible, and a working exception queue Work the queue, audit samples
[F] Fully automated Failure is low-cost and self-evident, reversal is cheap, monitoring exists Own the monitoring
[N] Not yet Data, rule clarity, or input quality fails today. Record the reason and the fix that reopens it Fix the blocker, then reassess

Before marking a step, name which of its functions you'd change: gather (collect information), interpret (summarize, categorize, score), decide (select the action), or act (execute it). They don't have to share one mark: "AI gathers and interprets, a person decides, the system acts" is a complete and often ideal design. Write it like [G: interpret] right on the map.

01
Delete before you automate
Duplication and unnecessary steps get removed, not accelerated. Automating a step that should not exist is pure waste.
02
Waits before work
If elapsed time dwarfs touch time, queues, notifications, and SLAs beat any AI on the work itself.
03
Rules before models
Any step whose rule is written and stable gets ordinary automation before anyone proposes AI.
04
Assist before replace
Where judgment is real but patterned, start with [A] or [G] and let the edit-rate data argue for more autonomy.
05
Fix data before deciding
Steps that fail on data quality get [N] plus a prerequisite, not a forced mark.
The five gates

Before a step earns an AI mark, it passes five gates.

Gate 01
Worth examining?
How often does it run, and how much time does it consume? Under an hour a week saved, stop, unless it feeds a top friction problem.
Gate 02
Understood?
Is the decision rule stateable and stable? Would two colleagues decide the same way? Inconsistent judgment means clarify first, not automate.
Gate 03
Can a machine see it?
Are the inputs digital and reachable? What share arrives clean? Are past cases and outcomes recorded anywhere at all?
Gate 04
What if it’s wrong?
What does one undetected error cost, and is it reversible? Would a fluent, confident, wrong answer get caught before it acts on the world?
Gate 05
Worth the total cost?
What does reviewing the output cost per item? Could a simpler fix capture most of the value? Which success measure moves, and by how much?

Record the verdict as one line that cites its evidence: "Step 4: [G: interpret]. Rule consistent but unwritten. Outcomes recorded. Errors reversible before send. Saves ~4 h/week. Prerequisite: write the criteria down." A mark that can't fill in its gates is a guess, and gets [N] until it can.

Worked example

Inbound lead qualification, mapped and marked.

A 12-person services firm, ~45 leads a week, 25 minutes of touch time hiding inside 2–4 days of elapsed time. Every number here is illustrative; yours will differ, which is exactly why you map.

StepTodayMarkFuture state
Lead arrives Sits in a shared inbox 4–24 hours [R] Parsed into the CRM within minutes
Research the company 10 minutes of googling per lead [G] AI compiles the brief; manager spot-checks a weekly sample
Fit decision Unwritten rule in one person’s head [G] AI scores against the written criteria; manager approves every lead
Route to a rep Slack DM to whoever seems free [R] Round-robin with a load cap
First reply 4–48 hours later, research redone [A] Rep gets the brief with the lead; the reply stays human

Notice what the mapping session itself surfaced: the fit rule lived in one person's head, so writing it down was the prerequisite for everything else, and the "later" folder where a quarter of leads quietly died became an explicit queue with a day-14 review rule. No model fixed those. The map did.

01
Budget the review time
If approving 45 drafts a week takes 2 minutes each, that is 90 minutes on someone’s calendar. Unbudgeted review becomes rubber-stamping.
02
Batch to a rhythm
Two review sessions a day beat drip interruptions, as long as the cadence respects the workflow’s response-time promise.
03
Route by risk
Full review where errors are expensive or irreversible; sampling where they are cheap. Equal review everywhere wastes your best attention.
04
Publish the graduation criteria
Start at 100% review. Move a category to exception-only after a sustained, measured accuracy bar you wrote down in advance.

That's the review design in four moves. Wherever an AI mark lands, the map must also show who checks, when, against what bar, and what happens on failure. Oversight nobody designed becomes either a rubber stamp or the new bottleneck.

FAQ

Common questions

Why map a workflow before choosing an AI tool?

Because a process you do not understand still runs on hidden human judgment, and automating around it removes the corrections while the volume goes up. Mapping first finds fixes that need no AI, produces baseline numbers a pilot can be judged against, and lets you describe the AI opportunity precisely enough to test it cheaply.

How long does workflow mapping take?

About 60 to 90 minutes solo for a first-pass map of one workflow, or a half-day workshop with 4 to 6 people for a validated version with real numbers. Run the AI overlay in a separate, later session: mixing "what happens" with "what could AI do" reliably contaminates the map with aspiration.

What if the map says most steps should stay human?

That is the normal result. The research on task-level AI suitability finds most roles contain some AI-suitable tasks while very few are fully automatable, so a healthy map mixes marks: several [H] and [R], a few [A] and [G], one or two [X] candidates, and at least one honest [N]. A map that comes back all-automated was scored by optimism, not evidence.

Do I need special software to use this?

No. The canvas works in a doc, a spreadsheet, a Notion page, a Miro board, or on a wall of sticky notes. The template is deliberately vendor-neutral and tool-independent; nothing in it requires buying anything.

When is ordinary automation better than AI?

Whenever the decision rule is explicit and stable and the inputs are structured. Rules are cheaper, deterministic, explainable, and testable. AI earns its place where judgment is real but patterned, and where checking the output is faster than producing it.

Want the printable worksheet pack?

The blank canvas, step table, and overlay sheet as a printable pack is in progress, and the Human-in-the-Loop Checklist is next. Join the list and each one lands in your inbox as it ships.

One useful email at a time. Unsubscribe whenever.