How do I build an AI automation roadmap for my business

You build an AI automation roadmap by inventorying the work that already burns hours, scoring each item by blast radius (what a failure costs) and hours saved (what a success returns), then shipping one recoverable path with a production spine — not by starting with the shiniest agent. Blast radius is the veto. Hours are the sort. Agents are a tool you add after the path is boring.

This spoke sits under the Production n8n handbook. Across 600+ automations built and 500+ live, plus 20,000+ hours on agentic systems and 35,000+ hours saved for clients, the expensive weeks were almost never “we needed a smarter model.” They were a ranked queue that started with a public bot while intake was still copy-paste. The rail is n8n. The ranking function is not.

I will not invent a studio-wide ROI, payback month, or “X% of tasks automated” figure for your business. Those numbers only exist after you measure your week.

The short answer

• Inventory first. One page of work that already happens: trigger, destination, weekly hours, failure, owner.
• Score two axes. Blast radius (money, customer contact, deletes, PII) is a gate. Weekly hours saved is the sort — measured hours, not a multiplier.
• Ship one path. Happy path plus exception path, idempotency before writes, error workflow with a named owner.
• Promote autonomy later. Approval gates stay on until the errors are understood. A clean demo week is not a promotion.
• Agents last. If the path is rule-shaped, it is a workflow. If it needs judgment you cannot encode, it stays manual until you can.

Ranking you used	What you actually bought
Shiniest agent	A demo and a mute channel
Loudest stakeholder	Political debt, not hours
Cheapest connector	A sync nobody owns
Blast radius + hours	A queue you can defend on Tuesday

Shiny is not a ranking function.

What is an AI automation roadmap, and what is it not?

A roadmap is a ranked queue of business paths with owners, recovery rules, and a ship order. It is not a slide of tools, a backlog of “AI ideas,” or a twelve-canvas program with no first promote.

Artifact	Exists when	Fake version
Inventory	Trigger, system of record, weekly hours, failure, owner on one page	A Notion graveyard of “automate invoicing someday”
Scoreboard	Blast class + hours + recoverability, dated	A 2x2 from a workshop with no numbers
First path	One happy path, one exception path, autonomy level written	“Phase 1: AI layer”
Spine	Webhook verify, idempotency, schema, DLQ, named owner	Green Execute on a sample JSON
Cadence	Weekly failure glance, monthly kill/promote	A quarterly “AI steering” meeting

Decision list:

1. If it does not name a trigger and a destination, it is a wish.
2. If it does not name weekly hours, it is not ranked — it is guessed.
3. If it does not name recovery, it is not production — it is a demo.
4. If the first item is an agent that talks to customers, you inverted the queue.

NIST’s AI Risk Management Framework is voluntary, and as of 2026 NIST says version 1.0 is under revision. The part that still helps here is the loop: Map, Measure, Manage — not “stand up an agent, then invent governance.” A roadmap that skips Map is a shopping list.

A backlog of workflows is still a backlog. Sequence is the product.

Why rank by blast radius and hours saved instead of agents?

Agents look like progress in a demo. Blast radius and hours decide whether Tuesday is quieter. Rank the work the way an operator would defend it after a failure, not the way a vendor would demo it.

Ranking method	Sorts by	Typical first item	Typical Tuesday
Agent-first	Novelty, model, “autonomy”	Public chatbot or “AI employee”	Wrong send with your logo
Stakeholder-first	Who yelled	Founder’s pet report	Sales muted the real routing bot
Connector-first	What the tool already has	A sync of a sheet nobody uses	Drift, no owner
Blast + hours	Failure cost, then weekly load	Intake, routing, or a reversible internal path	One path actually quieter

Blast radius answers: if this fires twice, or fires wrong, how bad is the afternoon? Hours saved answers: if this fires once and correctly, how much calendar comes back? You need both. Hours without a recovery path is how you automate a mess. Recovery without hours is how you spend a month on a two-hour-a-week reminder.

What “agent-first” usually hides:

• The bot cannot write a correct CRM row, so humans still copy-paste
• Customer email is on by default because the demo needed a send
• Nobody named an owner, so the mute is the monitoring
• The workflow is a god canvas because “the agent will decide”

If inbound leads go cold, that is a routing problem — treat it as a ranked path, not a personality. See lead routing automations that sales will not mute. If kickoff after a “yes” is still a scramble, intake beats a chatbot. Agents draft. Rules and humans decide until the error rate earns more rope.

How do you inventory the week before you rank anything?

You cannot rank a week you have not watched. Sit the work for five operating days, or pull it from the people who already do it. Do not interview for “AI opportunities.” Interview for copy-paste, chase, and re-key.

Capture this row per candidate. If a cell is blank, the item is not on the scoreboard yet.

Field	What to write	Disqualify if
Name	{domain}-{outcome} e.g. sales-lead-route	“The AI layer”
Trigger	Form, webhook, inbox, schedule, human paste	“Whenever we remember”
System of record	One CRM / sheet / billing object	Dual-write “for now”
Weekly hours	Clocked or estimated from a time study	“A lot” / “4x productivity”
Failure mode	Duplicate, miss, wrong owner, wrong send	“It should be fine”
Recovery	Minutes with a runbook, or days of archaeology	Blank
Owner	Named human + backup	“Ops” / “engineering”
Autonomy	Manual / gated / auto	“The model decides refunds”

Procedure:

1. List every handoff that happens more than twice a week.
2. Strike anything that changes definition every sprint.
3. Strike anything whose success is judgment you will not encode (pricing exceptions, legal nuance, sensitive replies).
4. Keep the rest. You now have an inventory, not a roadmap.
5. Date the page. An undated inventory is a mood.

• Time study exists (five days of real work, or a week of “who did this”)
• Each row has one trigger and one destination
• Weekly hours are a number you can say out loud
• Failure and recovery are written, not implied
• Owner would still own it next quarter

Hours you cannot name are hours you will not save.

How do you score blast radius without a fake dollar model?

Blast radius is a class, not a spreadsheet of invented dollars. You do not need a Monte Carlo. You need to know whether a double-fire is an undo, a ticket, or a lawyer.

Class	Examples	Default until proven
Irreversible	Charges, refunds, invoices sent, customer email at volume, deletes, CRM merges	Gate. Do not auto.
Customer-visible	Wrong owner ping, wrong confirmation, wrong SLA claim	Quiet notify; log the reason
Internal-only	Digest, reminder, draft in a queue, shadow write to staging	Fine to auto if recovery is minutes
Silent miss	Trigger never fired, cron died, webhook still on the test URL	Heartbeat, not hope

Stripe’s webhook docs are blunt: endpoints occasionally receive the same event more than once, and you must make processing idempotent. That is not a Stripe quirk. It is how HTTP callbacks behave when the 200 is slow. Any roadmap item that writes money or CRM on a webhook inherits that fact. If you cannot say what a second delivery does, the blast class is “unknown,” which is worse than “high.”

Score the class before you score the hours:

1. Name the side effects: create, update, send, charge, delete.
2. Mark each one reversible (undo in minutes) or not.
3. If any irreversible effect is on the path, the item cannot sit in “quick win” no matter how many hours it claims.
4. If the only effect is an internal draft, do not pretend it is a transformation. It can still be item one.

Question	If yes	If no
Can a duplicate create a second invoice or email?	Irreversible until idempotent + gated	Still check silent miss
Can a bad payload create an empty CRM row?	Schema gate, or it is not ready	You still need a DLQ
Can sales mute this in a week?	Notify is too loud or too dumb	Keep notify boring
Can the original builder take Friday off?	Need a pause + replay note	Do not ship

A fake dollar model is how a two-hour-a-week report outranks a five-hour intake path because someone typed $50k in a cell. If you do not have finance-grade cost of failure, use the class. Class is honest. Invented NPV is not.

How do you score hours saved without a multiplier you cannot defend?

Hours saved are clocked load minus remaining exceptions, not “AI will 10x the team.” The handbook heuristic still holds: nodes are cheap, ownership is not. Measure your week. I will not paste a benchmark from a vendor PDF and call it your number.

Weekly hours on the path	Failure recovery	Default
Under ~2	High (money, customer, legal)	Leave manual or semi-manual
Under ~2	Low	Maybe a reminder, not a workflow
Over ~5	Recoverable in minutes	Automate the happy path; park exceptions
Over ~5	Recovery takes days	Do not automate until recovery is short
“We think it is a lot”	Unknown	Time-study week. Do not rank yet.

Those cuts are a heuristic, not a result you can put in a case study. If your week is four hours of chase and one hour of judgment, the automation can eat the chase. It cannot eat the judgment until you write rules.

How to get a number that survives a skeptic:

1. Watch five operating days, or pull last week’s tickets/inbox/CRM activity.
2. Count occurrences × minutes, including the “quick” re-keys people forget to log.
3. Subtract exception rate you will leave manual (do not assume 0%).
4. Write hours_saved ≈ (occurrences × minutes × happy_path_share) / 60.
5. If you cannot fill occurrences, you do not have hours. You have a vibe.

Input	Honest	Dishonest
Occurrences	37 form submits last week	“Leads, constantly”
Minutes	Timed three real ones	“Five minutes, probably”
Happy-path share	80% after you listed exceptions	100% because the demo was clean
Multiplier	None	“AI productivity 4x”
After-hours	Counted if someone actually works them	Ignored so the number looks bigger

Do not annualize a sloppy week into a six-figure “savings” slide. Annualization of a guess is still a guess. Rank on the weekly number. Revisit after the first path has a month of executions.

The 35,000+ hours saved for clients figure is an aggregate across a book of work, not a promise that your first workflow returns a slice of it. Your scoreboard is local.

Copy this time-study sheet for one path. Fill it from last week, not from a brainstorm.

Day	Occurrences	Minutes each (timed)	Exceptions left manual	Notes
Mon
Tue
Wed
Thu
Fri
Week total	sum	median of timed samples	count	hours ≈ total_min × happy_path_share / 60

If Friday is blank because “we were busy,” you did not measure. Extend the study. Do not fill the blanks with optimism.

What scoreboard actually ranks the queue?

Put blast class in the left column so it can veto. Sort the survivors by weekly hours, then by how short recovery already is. Do not average the two axes into a single “priority score” that lets a high-blast, high-hour money path sneak in as item one without a gate.

Rank	Item (examples, not client results)	Blast class	Weekly hours (heuristic)	Recoverability	Ship posture
1	Intake: form → one CRM row → confirmation	Customer-visible if send is wrong	Often high	Minutes if gated + idempotent	First path, gate the send
2	Incomplete-intake nudge (one, not a drip)	Customer-visible	Medium	Minutes if complete-flag is real	Same canvas or immediate follow
3	Qualified inbound → owner + one alert	Customer-visible if notify is loud	High if leads sit	Minutes if deduped	After intake, or first if you are losing the close
4	Internal status digest / “where is this?”	Internal	Medium	Minutes	Fine once 1–3 are live
5	Reporting / board pack from systems of record	Internal	Low–medium	Easy	Do not skip 1–3 for this
6	Customer-facing agent / public chatbot	High	Demo hours, not ops hours	Days when it lies	Last, if ever
7	Autopilot invoicing / refunds	Irreversible	Can look high	Days when double-applied	Not until spine + gate + owner

Worked ranking rules:

1. Veto. Irreversible + recovery longer than minutes → park. Do not “just ship it behind monitoring.”
2. Sort. Among items that pass the veto, highest weekly hours first.
3. Tie-break. Prefer the path whose exception rate you already know.
4. Cap WIP. One production path in flight. A second canvas is how you get two half-spines.
5. Re-rank monthly. Hours move. Mute rate is a re-rank signal.

• Veto column filled before the sort
• No item with blank weekly hours in the top three
• No agent in the top three unless the path is already rule-shaped and gated
• WIP = 1 named {domain}-{outcome}-{env} workflow
• Date on the scoreboard matches this month

If two items both pass, and one is lead routing while the other is a weekly PDF nobody reads, routing wins. Hours and blast agree. The PDF can wait.

What ships first, what waits, and what never gets a canvas?

First is the path that deletes calendar and fails small. Never is the path that fails large while the definition of done is still a meeting.

Bucket	Put here	Do not put here
Ship now	High hours, recoverable in minutes, one system of record, named owner	Anything whose trigger is still “we will figure out the form”
Ship gated	High hours, irreversible step exists, recovery can be an approval	Autopilot “to learn faster”
Wait	Medium hours, dirty schema, access not issued, exception rate unknown	“We’ll map properties during promote”
Never (yet)	Process changes every sprint; judgment-heavy; no owner; hours unnamed	A chatbot to hide a missing CRM

Decision list for the #1 slot:

1. Every new “yes” creates a scramble → intake / onboarding first.
2. Qualified inbound sits untouched → routing first.
3. Both hurt → onboarding first if you already close work; routing first if you are losing the close.
4. Neither hurts and you still want an agent → you do not have a roadmap problem. You have a boredom problem. Do not spend money on boredom.

What never belongs on v1 of the company roadmap:

• A public chatbot with no system of record
• Autopilot refunds
• Dual-write to two live CRMs
• “Migrate all of Zapier this quarter”
• Queue mode, Redis, and workers before the first path has an idempotency key — n8n queue mode is a concurrency fix, not a strategy

Parked is not failed. Parked is how you keep the first path from drowning.

How do you implement the first ranked path in n8n?

Implement one ranked path as infrastructure, not as a company-shaped canvas. The nodes change. The spine does not. Full treatment lives in the handbook; this is the roadmap cut.

Step	Done when	Skip this and you shipped a demo
1. One-pager	Happy path + exception path + autonomy level + owner	“We’ll discover in the build”
2. Identity	Provider event ID named (not the n8n execution ID)	Duplicates become two CRM rows
3. Verify	Signature / auth on the Webhook node using the raw body	Open write API
4. Claim	Atomic idempotency before irreversible nodes	Second invoice
5. Schema	Required fields checked; null / type drift → review	Empty records, green run
6. Gate	Money / customer send / delete behind approval until measured	Autonomy as a launch party
7. Error path	Shared Error Trigger workflow, execution link, named owner	“Workflow failed” in a muted channel
8. Promote	Production URL published; vendor has that URL, not the 120-second test URL	Traffic dies when you close the editor

n8n facts that change the calendar (not folklore):

• You cannot test the Error Trigger by clicking Execute Workflow. It fires on an automatic failure. Force one with Stop And Error on a staging workflow, then trigger it for real.
• n8n’s webhook development docs split test vs production URLs. Mixing them is a production bug that looks like “the workflow never runs.”
• Error-workflow executions do not count toward the execution quota. Skipping the handler to “save runs” is how you save a mute.

Procedure for the first ranked path:

1. Name the workflow {domain}-{outcome}-prod (and a matching staging copy).
2. Build verify → claim → validate → write. Reverse that order and you will double-apply.
3. Attach the shared error workflow under Settings → Error workflow.
4. Keep irreversible nodes on a gate. A gated v1 is still production.
5. Replay a dirty payload and a duplicate webhook on purpose before you call it live.
6. Write a half-page runbook: pause, replay DLQ, who pages.
7. Only then discuss a second item on the scoreboard.

• Vendor points at the production webhook path
• Duplicate delivery returns 200 without a second side effect
• Poison payload lands in a review table with the original body
• Owner can pause without the original builder
• WIP is still one path

Map scoreboard fields onto n8n artifacts so the canvas cannot drift from the rank:

Scoreboard field	n8n artifact	If missing
Trigger	Webhook / Form / Schedule node using the production path	You are still on Listen for test event
Identity	Key from provider event ID, claimed before writes	Execution ID will double-apply
System of record	One write node family (CRM or sheet, not both live)	Dual-write archaeology
Autonomy	Approval node or a hard threshold in front of send/charge/delete	Demo autonomy
Owner	Sticky + error-workflow mention + runbook link	Mute-worthy “Workflow failed”
Recovery	DLQ table + replay branch	Slack screenshot as the archive

Do not implement the roadmap as one god workflow. Smaller workflows fail smaller. The scoreboard is allowed to have twelve rows. The canvas is not.

What breaks a roadmap once it touches production?

The failure mode is not “n8n was down.” It is a ranked queue that pretended ranking was finished when the slide was.

Failure	What it looks like	Cost	What you do instead
Agent-first queue	Chatbot ships; intake still handmade	Hours not saved; brand risk on sends	Re-rank; freeze the bot; ship the record
Invented ROI	Annualized guess becomes the KPI	You cannot tell if the path worked	Weekly hours on that path, DLQ age, mute rate
God canvas	Intake + billing + reporting on one graph	One schema drift takes the company down	Split by trust boundary
Test URL in production	Vendor still has /webhook-test/	Silent miss all weekend	Production URL + heartbeat
No idempotency	Provider retries; two invoices	Afternoon of reconcile	Claim the provider event ID first
Personal OAuth	Founder changes a password at 2am	DLQ storm	Service account; pause on auth drift
Unbounded WIP	Four “quick” workflows, zero spines	Four demos, one owner, no vacation	WIP = 1 until the first path is done
Ranking by politics	Loudest exec’s dashboard is item one	Sales mutes the actual routing bot	Hours + blast, dated, public to the operators

Concrete Tuesday: you ranked “AI SDR” first because the deck needed a hero. The form still dumps into a shared inbox. A duplicate webhook creates two HubSpot contacts. The SDR agent emails both. Sales mutes the channel. The roadmap now “needs better prompts.” It needed a scoreboard.

Signal	Means the roadmap is lying	Means it is working
Mute rate	Notify is wrong or loud	Routing/intake alerts stay unread-able and rare
DLQ age	Nobody owns replay	Items clear inside a published window
Duplicate side effects	Key was the execution ID or missing	Second delivery is a no-op
“Never ran”	No heartbeat; test URL; disabled workflow	Canary missed → ticket
New agent request	Scoreboard skipped	Agent shows up only after the path is boring

Retries without classification are how a down API becomes a self-inflicted outage. Bound them. Poison goes to a dead-letter path. Partial apply does not get a blind replay of the whole flow.

If the original builder cannot take a week off, you do not have a roadmap. You have a dependency on one person’s memory.

When should you hire vs DIY this roadmap?

DIY the ranking and the first reversible path when you will still own the canvas next quarter. Hire when the first ranked item can email a client, merge a CRM row, or charge a card — or when you already have a pile of green demos and no spine.

Situation	DIY	Buy the $500 Automation Audit
Inventory + scoreboard	Yes. A week of watching your own work	If nobody will tell the truth about hours
Reversible internal digest	Yes, if you will attach an error workflow	If you want a second pair of eyes on WIP
Form → CRM → confirmation	Maybe, if send stays gated and duplicates are claimed	If the form already double-fires or the CRM merge rules are tribal
Lead routing with SLA	Only if notify is quiet and ownership is mapped	When sales already muted a bot, or territories fight
Money / refunds / mass email	No	Yes. Do not DIY a blast-radius you cannot reverse in an hour
“We need an agent first”	No. Re-rank	Yes. That sentence is the audit

Hire vs DIY is not “can we click nodes.” It is “can we survive a double-fire.” The how long production automation takes clocks — scope, integrations, hardening — apply the moment you leave a personal sandbox. Access delays usually outlast canvas work. If IT is still issuing a service account, that is the timeline, not the n8n tutorial.

• You can name weekly hours without a multiplier
• You can name recovery in minutes
• A backup human exists
• Staging credentials are not production
• You are willing to keep a gate on for a watch window

If three of those are false, you are not behind on AI. You are behind on ownership. The audit exists for that. It is not a twelve-week transformation program.

How long until the first ranked path is production, not a slide?

The roadmap is not “done” when the scoreboard looks pretty. The first ranked path is done when duplicates do not double-apply, poison lands in review, and a backup human can pause it. Demo time is not that clock.

Phase	What it is	What it is not
Inventory + scoreboard	Days, if people will sit the work	A substitute for a spine
Happy-path canvas	Often an afternoon	Production
Access + schema	Often the long pole	“We’ll remap on promote”
Spine + staging proof	Days to weeks	Optional polish
Watch window	Scheduled after promote	“We’ll keep an eye on Slack”

From the production-clock post: a two-week promise is only real when access exists, the path is one, side effects are reversible or gated, and done is written. I will not quote a studio-wide “days to ROI.” Early proof is operational: the veto held, the first path has an owner, and last week’s hours on that path moved. Revenue attribution can wait until you are not lying about the hours.

Compress scope. Do not compress recovery.

Compress	How	Do not compress
Scoreboard width	Ten honest rows beat forty wishes	Skipping the time study
First path	One object, exceptions manual	“Also sync the old Zapier pile”
Autonomy	Gate on; learn from real traffic	Autopilot to hit a slide date
Meetings	Written one-pager, one signer	Standing alignment with no owner

If the person quoting the date cannot name scope, integrations, and hardening, the quote is a wish. Happy path is not the clock. Recovery is.

What does a 90-day roadmap look like without invented ROI?

Ninety days is enough to inventory, ship one production path, and decide whether a second path has earned a canvas. It is not enough to “automate the company.” Anyone selling you a 40-workflow transformation in a quarter is selling a backlog.

Window	Outcome	Explicit non-goals
Days 1–14	Dated inventory + scoreboard; WIP = 1 chosen	No new agent, no billing autopilot
Days 15–45	First path live: spine, gate, watch window, runbook	No second production canvas
Days 46–75	Harden: DLQ hygiene, mute review, heartbeat, credential split	Queue mode unless concurrency is the actual pain
Days 76–90	Re-rank with real hours; promote a gate only if errors are understood; pick path two or kill path one	Invented annual savings slide

Success for the 90 days (local, measurable, not a vendor KPI):

• Scoreboard dated this quarter, veto column filled
• One {domain}-{outcome}-prod workflow with production URL
• Forced duplicate did not double-apply
• Error alert includes execution link and owner
• Weekly hours on that path re-measured (same method as baseline)
• DLQ items older than your published window = 0, or a ticket exists
• No public agent shipped unless it was already item one after the veto (it should not be)

What 90 days is allowed to fail at: covering every department, replacing judgment, showing a company-wide ROI number. If you need a number for a board, report hours returned on the shipped path and incidents that did not become customer-facing. That is a receipt. A 4x multiplier is not.

Kill criteria at day 90:

1. Hours on the path did not move and the exceptions ate the happy path → the process was not rule-shaped. Park it.
2. Mute rate spiked → notify/routing is wrong. Fix ranking of alerts, not prompts.
3. Owner left and nobody can pause → you shipped a person, not a workflow. Write the runbook or turn it off.
4. Leadership still wants the chatbot first → stop building. The ranking function was rejected. More canvas will not fix a political veto.

How do you keep the roadmap honest after the first ship?

A scoreboard that never re-ranks becomes a shrine. Re-measure hours. Re-read mute and DLQ. Promote autonomy only after a stretch of understood errors. Demote when a vendor changes shape or the owner disappears.

Cadence	Look at	Do not look at
Daily (async)	Failure notifications on the live path	A vanity “automations run” count
Weekly	Hours still spent on the path; DLQ age; schema-fail spikes	New agent ideas from Twitter
Monthly	Autonomy gates; kill/keep; WIP still 1 or 2	Annualized savings from month one
Quarterly	Secret rotation; workflows that no longer earn keep; scoreboard rewrite	A new platform because the last one was “not AI enough”

Operating checklist:

• Baseline hours method is written (so month two uses the same clock)
• Heartbeat exists for “never ran,” not only for “ran and failed”
• Vendor changelog subscribed for every write destination
• Last-known-good export is newer than the last promote
• Agents, if any, sit below the fold of the scoreboard until the path is boring

When you add path two, copy the spine. Do not copy the canvas. The second item inherits verify / claim / validate / error / owner, not a fork of a 80-node monster.

If leadership asks “where is the AI,” point at the ranking function. The AI, when it shows up, drafts inside a gated step on a path that already worked. That is the roadmap. The other thing is a demo calendar.

FAQ

How do I build an AI automation roadmap for my business?

Inventory the work that already burns hours, score each item by blast radius then weekly hours saved, and ship one recoverable path with a production spine in n8n. Blast radius vetoes money, mass customer send, and deletes until they are gated. Hours sort the survivors. Do not start with the shiniest agent.

How do I measure whether an AI automation roadmap is working?

Re-measure weekly hours on the shipped path with the same method as baseline, plus DLQ age, mute rate, duplicate side effects, and whether a backup human can pause the workflow. A company-wide ROI slide is not a measurement. If hop-level “AI activity” is up and hours on the path are flat, you ranked theater.

What usually fails first when teams try this?

Agent-first ranking and invented hours. The chatbot ships, intake stays handmade, and a duplicate webhook double-writes because nobody claimed a provider event ID. Politics is a close second: the loudest dashboard becomes item one and sales mutes the actual routing bot.

How long does this take to show results?

You should see operational control when the first ranked path has a spine, an owner, and a forced-duplicate that does nothing — often inside a focused stretch once access exists, not once the slide exists. I will not invent a studio-wide days-to-ROI figure. Early proof is hours on that path and incidents that stayed internal.

What should I skip if I only have a week?

Skip new agents, billing autopilot, queue mode, and “automate the company.” Do the inventory, fill blast + hours on ten rows, pick WIP = 1, and either ship a reversible internal path with an error workflow or write the one-pager if credentials are still personal. A week of ranking beats a week of canvas.

When is this not worth doing yet?

When you cannot name weekly hours, the process changes every sprint, success is judgment you will not encode, or nobody will own the graph next quarter. A roadmap without an owner is a backlog. Fix the definition of done and the time study first. Then rank.

CTA

Rank the week. Then build the first path that survives a double-fire.

Explore the automation lane, then book a $500 Automation Audit. Bring the inventory, not a bot demo. We will tell you what to sequence first — and what not to automate yet.