AI AUTOMATION · BEYOND $1M | breadcrumb: Home > Learn > AI automation > Beyond $1M
Beyond $1M: systems, data, governance and the human in the loop
Past $1M, "can we automate this?" matters less than "who owns it, what may it do alone and how would we know if it went wrong?" Brands that answer those three questions in writing, and keep a human in the loop for the costly decisions, can automate more with less risk.
Quick answer
Above $1M, ecommerce AI automation needs governance and a human in the loop: a register of every automation with a named owner, one source of truth for each type of data, written approval thresholds for refunds, prices and purchase orders, least-privilege access, a log of every automated action and a weekly review. People approve anything that costs money, changes a price or touches an account.
This page is part of our free AI automation guide, which follows the path from first automations to governance.
What changes about automation once a brand passes $1M?
At a few hundred orders a week, one founder can notice when an automation misbehaves. Past $1M you typically have more SKUs, two or three channels, a 3PL, a support team and dozens of workflows built by different people over time. The risks move:
| Below $1M | Above $1M |
|---|---|
| "Is it built?" | "Who owns it, and is it still right?" |
| A handful of rules in one tool | Workflows across channels, the 3PL, the helpdesk and finance |
| The founder spots errors | Errors repeat across thousands of orders before anyone notices |
| One login per tool | Many users, contractors and AI tools with access to accounts |
| Speed matters most | Auditability matters as much as speed |
None of this means slowing down. It means writing things down.
What systems does a $1M+ brand need under its automations?
Think in layers. Each layer has one job and one owner.
| Layer | Job | Typical tools | Owner |
|---|---|---|---|
| Channels | Sell and take orders | Amazon Seller Central, Walmart Seller Center, Shopify | Channel lead |
| Catalog | One record of product data | Shopify, a PIM or a master sheet | Catalog owner |
| Stock and fulfillment | Where units are and what is sellable | 3PL or WMS, FBA, WFS | Ops lead |
| Orders | Collect, check and route orders | Order hub or OMS | Ops lead |
| Customer | Conversations and policy | Helpdesk | Support lead |
| Money | Fees, payouts, landed cost, margin | Accounting, payout reports, cost sheet | Finance owner |
| Automation | Rules and connectors between layers | Shopify Flow, Zapier, Make, n8n, integration platforms | Automation owner |
| AI | Drafts, classification, agents within limits | Platform assistants, helpdesk AI, approved AI tools | Automation owner plus each layer's owner |
The AI layer sits on top, not underneath. It reads from the layers below. It never becomes the system of record for stock, prices or money.
What data rules keep automation honest?
Most automation failures at this size are data failures. Five rules prevent the common ones:
- 01One definition per metric.
Write down what "available stock", "contribution margin" and "late order" mean in your business, and use the same definition in every report and rule.
- 02One system of record per data type.
Others read; they never override. (The table is in the Scale guide.)
- 03Daily reconciliation.
Orders per channel versus orders at the warehouse; stock per channel versus put-away; payouts versus orders. Mismatches go to an owner.
- 04Stop on disagreement.
When two systems disagree, the automation pauses and asks. It never picks a number.
- 05Dated changes.
Fee changes, tariff changes and rule changes get a date and a source in a change log, and the rules that depend on them are updated the same week.
What goes in an AI operations policy?
A one-page policy, signed by the owner and kept where the team can see it, does more than any tool. It covers:
| Section | What it says |
|---|---|
| Register | Every automation and AI tool, what it does, which systems it touches and its named owner |
| Approval thresholds | The dollar and percentage limits above which a person must approve (see the matrix below) |
| Access | Named users only, least privilege, MFA, no shared logins, never the owner password; no tool gets bank, payout or legal-entity permissions |
| Official connections only | AI tools connect through each platform's official APIs, apps or plugins, where the agent identifies itself |
| Logging | Every automated action is logged with time, input, output and result |
| Testing | New automations run in shadow mode (drafts only) before they act |
| Change control | Who may change a rule, and how changes are recorded |
| Incident response | Who can pause any automation, how customers and marketplaces are told and how the fix is checked |
| Review | Weekly log review; monthly sample audit; quarterly review of the whole register |
| Customer honesty | What your AI may say about policies, how you disclose automated replies and a ban on fake or incentivized reviews |
On access, the platforms are clear. An Amazon post on the Seller Forums tells sellers: "Do not share your password; use User Permissions to provide access instead," and to "Always add external service providers as Authorized Partners, not Secondary Users" (Amazon Seller Forums, "User Permissions Guide for Amazon Sellers," posted by the Amazon account Ezra_Amazon; checked Oct 20, 2026). Shopify says: "Asking a merchant for their password, or using a merchant's credentials to access their store, is prohibited," and recommends collaborator accounts (Shopify Help Center, Working on client stores; checked Oct 20, 2026). On Walmart, keep the Admin role with the business owner: Walmart says Admin access includes "Reviewing, signing and accepting binding legal agreements in Seller Center on behalf of the seller" and to limit it to people "who can accept binding terms on behalf of your company" (Walmart Marketplace Learn, updated Jul 2, 2025; checked Oct 20, 2026).
On official connections: Amazon sued Perplexity on Nov 4, 2025 (Bloomberg, Nov 4, 2025), arguing that its Comet shopping agent was disguised as a regular Google Chrome browser session, and in Mar 2026 a federal judge granted Amazon a preliminary injunction blocking Comet's agent from password-protected parts of Amazon's site (GeekWire, Mar 10, 2026; both checked Oct 20, 2026). Meanwhile Amazon opened a seller plugin for outside AI assistants where "Sellers choose which types of data the plugin can reach and approve each action before it's carried out" (GeekWire, Sep 23, 2026; launched first in the US). The lesson: use the door the platform built.
On reviews: the FTC rule on consumer reviews and testimonials bans fake and AI-generated reviews, buying reviews conditioned on sentiment and review suppression (16 CFR Part 465, effective Oct 21, 2024). Never point an AI tool at review generation.
What does an approval matrix look like?
Here is a sample matrix. Your limits will differ; the point is that they are written.
| Action | AI or rule acts alone | Ops lead approves | Owner approves |
|---|---|---|---|
| Refund | Up to $25, inside policy | $25 to $250 | Over $250 or outside policy |
| Price change | Inside the written floor and ceiling | Up to 10% outside the band, with margin shown | Below floor, or any change to MAP-protected items |
| Purchase order | Draft only | Up to the monthly budget | Over budget or a new supplier |
| Ad budget | Bids under the set maximum | Budget changes up to 20% | New campaigns over a set spend |
| Marketplace submissions (appeals, policy replies) | Assemble documents | Drafts and checks | Approves before sending |
| Customer promise outside policy | Never | Replies personally | Sets new policy |
(Illustrative · sample figures. Set your own thresholds.)
Human in the loop, on the loop or out of the loop?
There are three ways a person can relate to an automation. Pick one per workflow, on purpose.
| Pattern | What it means | Use it for |
|---|---|---|
| In the loop | The AI drafts; a person approves before anything happens | Refunds above limits, prices outside rules, appeals, POs, new customer-facing replies |
| On the loop | The automation acts; a person watches the log and can pause it | Repricing inside a band, routine order-status replies, stock sync |
| Out of the loop | The automation acts; reviewed in the weekly log only | Tags, internal alerts, drafts that nobody receives |
Amazon's own Seller Assistant follows the same idea: sellers choose whether it only surfaces insights, drafts actions for approval or acts on its own within limits they set (Amazon Accelerate 2026 recap, Sep 29, 2026).
One caution from operators: people stop paying attention to approvals they see a hundred times a day. Keep approvals for decisions that matter, sample the rest in a weekly review and rotate who reviews.
Where automation stops
Let AI do the busywork. Three actions always wait for a person:
- 01Price changes outside the limits you set.
A pricing mistake can sell your stock at a loss in minutes.
- 02Refunds above a set amount.
Small ones can be automatic. Big ones get a second look.
- 03Anything touching account health.
Appeals and policy replies are written and checked by people.
We add three more to every build: account appeals and reinstatement, policy decisions (what your store promises customers) and any situation nobody has written a rule for. Everything else: automate it, log it and review it every week.
Five guardrails sit under every workflow: limits on every action, approval for anything risky, only the access each tool needs, a log of every step and one person who owns the outcome. What your AI says, you said.
How do you know your automations are still working?
Track automation health like any other operation. Five measures, reviewed weekly:
| Measure | What it tells you |
|---|---|
| Exception rate | How often the automation hands off to a person; a sudden rise means something upstream changed |
| Override rate | How often a person changes what the AI drafted; high means the rule or prompt needs work |
| Errors found in review | Mistakes caught in the weekly sample, by type |
| Time to clear exceptions | Whether the hand-offs are being handled, or piling up |
| Drift incidents | Times a platform, fee or policy change broke a rule |
Who does what?
| Role | Owns |
|---|---|
| Owner or founder | The policy, the approval thresholds and decisions above them |
| Ops lead | The register, the exceptions queue and the weekly review |
| Workflow owner (one per automation) | Its rule, its log and fixing it when it drifts |
| Channel leads | Marketplace submissions and account health |
| Support lead | What the AI may say to customers |
If nobody on your team has the time to own the register and the exceptions queue, that is the gap to fill first. You can hire an ops lead, bring in Ops Desk to run daily operations and the weekly review from $1,499 a month, or add a named automation and data analyst through our Dedicated Ops Team, a full-time person who works only on your business and is managed by us. What matters is that the owner is a named person with the time to do the job, not a shared inbox.
At Ecom Relays, every Relay Module we build has a named owner on our side and one on yours, a written spec and acceptance tests, and lives in your accounts. Anything failing acceptance is fixed free for 60 days, and upkeep (Module Care) is included on any monthly plan.
A note on AI claims: the FTC has warned businesses not to exaggerate what their AI can do (FTC business guidance, "Keep your AI claims in check," Feb 27, 2023). We describe each tool from its own documentation, label vendor figures as vendor estimates and never say that AI will produce sales for you. AI here means software that drafts, sorts, checks and alerts. People still own the decisions.
Problems and fixes at this stage
| Problem, in the buyer's words | Cause | Fix | Relay Module or service | How to measure |
|---|---|---|---|---|
| "Nobody knows what half our automations do." | No register or owners | Register every automation with an owner; retire the orphans | 14-Day Ops Foundation; Ops Desk | Automations with a named owner, as a share of all |
| "An AI tool changed prices we never approved." | No floor, ceiling or approval step | Written price band; approval outside it; log | RELAY-02 Margin Guard | Price changes outside the band per month (target zero) |
| "We find account health problems when it is already too late." | No daily check | Daily checks, same-day action, appeal preparation by people | RELAY-05 Account Health Watch | Days from warning to action |
| "Our numbers live in spreadsheets nobody trusts." | No single definition or source | Defined metrics from source data | RELAY-01 Profit and Cash Pulse | Hours a week spent reconciling reports |
Talk it through with a real person
Book a free 30-minute session with an ops lead. Bring one workflow you want off your plate, and we will tell you plainly whether to automate it, how and where it should stop. A real person replies within one business hour, 9am to 6pm US Central, Monday to Friday. No pitch deck.
A written report within 48 hours of working read-only access. You keep it whether or not you hire us.
Free
Ask a real expert. Free.
A real person replies within one business hour, 9am to 6pm US Central, Monday to Friday.
Request received.
That did not send.
You can reach an ops lead directly, and nothing you typed is lost.
Do we need a formal AI policy at our size?
If more than one person builds or uses automations, yes. One page is enough: the register, approval thresholds, access rules, logging, testing, incident response and review cadence. It protects the account and makes it easier to hand work between people.
Can AI tools get our seller account suspended?
They can contribute, the same way a person breaking policy would: unsupported listing claims, repricing that trips pricing rules, review requests outside platform programs, shared logins. Use official connections, least-privilege access, written rules against each platform's policy and a person approving every marketplace submission.
What approval thresholds should we set?
Start from your own margins and policy. Many brands set a small refund limit for automatic approval, a price band per SKU tied to its margin floor and a rule that every purchase order needs a person. Write the limits down and review them quarterly.
What is the difference between human in the loop and human on the loop?
In the loop: a person approves before the action happens. On the loop: the automation acts and a person monitors and can pause it. Use in the loop for anything costly or hard to undo, and on the loop for routine actions inside tight limits.
How do we test a new AI workflow safely?
Run it in shadow mode first: it drafts, a person does the real work and you compare. When the drafts are consistently right over a few weeks, let it act inside limits, with the log reviewed weekly. Count the differences between the drafts and what your team actually did each week, and widen the limits only while that override rate stays low.
Should AI ever submit appeals or policy replies to Amazon or Walmart?
AI can gather documents and draft, but a person writes the final version, checks every claim and submits it. These submissions affect your account directly and cannot be taken back. Keep the person who submits named in your AI operations policy.
Who should own the automations?
Each one needs a named owner who understands the rule and checks the log. The ops lead owns the register and exceptions; the founder owns the policy and the thresholds. Write the owner's name in the register next to each automation, and change it the day that person changes role.
Will governance slow us down?
It slows down the first build slightly. It speeds up everything after, because people trust the automation enough to let it run, and problems are caught in a weekly review instead of in a marketplace warning. The extra work is a one-page policy, a register and a 30-minute weekly review, which is less time than cleaning up one bad run of automated prices or refunds.
Sources14 sources
- 01help.shopify.comchecked Oct 20, 2026Primary
- 02help.shopify.comchecked Oct 20, 2026Primary
- 03zapier.comchecked Oct 20, 2026Primary
- 04make.comchecked Oct 20, 2026Primary
- 05n8n.iochecked Oct 20, 2026Primary
- 06sellingpartners.aboutamazon.comchecked Oct 20, 2026Primary
- 07geekwire.comchecked Oct 20, 2026Primary
- 08sellercentral-europe.amazon.comchecked Oct 20, 2026Primary
- 09marketplacelearn.walmart.comchecked Oct 20, 2026Primary
- 10marketplacelearn.walmart.comchecked Oct 20, 2026Primary
- 11geekwire.comchecked Oct 20, 2026Primary
- 12bloomberg.comchecked Oct 20, 2026Primary
- 13federalregister.govchecked Oct 20, 2026Primary
- 14ftc.govchecked Oct 20, 2026Primary
Free and ungated. No sign-up wall. Ecom Relays is not a course or coaching program. We do not sell stores and we never promise income.
Next step
Ask a real expert. Free.
A real person replies within one business hour, 9am to 6pm US Central, Monday to Friday.