Build it with AI
Build an inventory workflow around stock, movement, exceptions, and replenishment.
Create a focused system for inventory records, locations, counts, thresholds, movement, and operational alerts.
Introduction
What an inventory system has to hold.
Most teams end up with an inventory system the same way: a stock spreadsheet updated after the fact, and a physical count that contradicts it. A spreadsheet holds while one person does all the receiving and picking. It breaks with the second person, because the sheet records the count at a moment and movement happens continuously between those moments.
The recorded quantity and the shelf quantity diverge continuously, and only a stock take reveals by how much. There is no record of movement as it happens, no reorder point that accounts for lead time and variability, and no distinction between stock that is present and stock that is available after commitments.
What follows covers building an inventory system: the records it holds (items, quantities, locations, movements, reorder points, and supplier lead times), the systems it reads (Google Sheets and Airtable), and what it does not fix.
The problem
Stock on the sheet, stock on the shelf, and the gap between.
Spreadsheets model a snapshot and have no concept of a transaction, which is why the count is right on the day it was taken and drifting by the next. Full ERP inventory modules model everything and impose a process most small operations cannot staff.
The records are items, quantities, locations, movements, reorder points, and supplier lead times, and the authoritative copy of most of them already lives in Google Sheets or Airtable. The sheet says fourteen, the shelf has eleven, and nobody knows whether three were sold, damaged, or never received — because the only record is the count itself.
The cost is not the inconvenience: an order is accepted for stock that is not there.
You're likely here because
- Stock accuracy is only known immediately after a count
- The recorded quantity and the shelf quantity diverge continuously, and only a stock take reveals by how much.
- When it is wrong, an order is accepted for stock that is not there
What gets built
Launch builds it, Grow operates it.
Built in Launch
- • Inventory table
- • Location views
- • Threshold alerts
Operated through Grow
- • Notifications
- • Supplier follow-up
- • Action queues
Systems it reads
- • Google Sheets
- • Airtable
- • Slack
The record model
What an inventory system has to record.
- Movement transaction
- Receipts, picks, adjustments, and returns as events. Quantity is what the events imply; storing quantity instead means a variance can be corrected and never explained.
- On-hand and available, separately
- Committed stock counted as available is how the same unit gets sold twice, which is the error customers notice and remember.
- Location
- Because stock that exists in the wrong place is functionally absent, and a single total hides it completely.
- Item class by cost of error
- Uniform inventory discipline across every item is the most common reason these projects are abandoned. The class decides how much rigour each item gets.
- Reorder point with lead time and variability
- A fixed threshold set two years ago is the difference between a buffer and a stockout, and demand variability is the term usually omitted.
- Adjustment reason and owner
- Unattributed adjustments are how shrinkage becomes invisible, and they are the default in every spreadsheet-based system.
- Cycle count variance and tolerance
- Per item class, so an investigation is triggered by rule rather than by whoever is looking on the day.
How it runs
From a periodic count to recorded movement.
Step 01
Describe what an inventory system has to do
Model movement rather than quantity. Receipts, picks, adjustments, and returns are the events; the quantity is what those events imply, and that inversion is the whole design.
Step 02
Connect the systems of record
Sales channels supply commitments, the accounting system supplies purchases, and the operational sheet supplies history. Reading them turns a count into a reconciliation.
Step 03
Build the operating surface
Item records with locations, movement transactions, reorder points computed from lead time and demand variability, and alerts on the exceptions.
Step 04
Start narrow
Movement recording plus reorder points on the twenty items that actually matter — the ones whose stockout costs a sale or halts work. The long tail can stay on the sheet.
Step 05
Route the exceptions
A count variance beyond tolerance goes to a named owner as an investigation rather than being written off silently into an adjustment.
Step 06
Measure stockouts on items that were above reorder point in the system
Track inventory accuracy by cycle count and the number of stockouts on the items that matter. Both are countable, and accuracy is the one that predicts the other.
Implementation path
Getting inventory accurate enough to trust.
- 01
Classify items by the cost of being wrong before doing anything else. Uniform inventory discipline across every item is the reason most inventory projects are abandoned.
- 02
Baseline accuracy with a cycle count on the top items, and count stockouts over the last quarter. The first number is usually lower than anyone expects.
- 03
Record movement rather than periodic counts. A system that models transactions can reconcile; a system that stores quantities can only be corrected.
- 04
Introduce cycle counting on a rolling schedule instead of an annual full count. Continuous small counts find the process problem while it is still identifiable.
- 05
Build the narrowest useful version first: items below reorder point with the supplier lead time attached, so ordering is a decision rather than a reaction.
- 06
Classifying items by the cost of being wrong is a half-day and shapes everything else. Movement recording plus reorder points for the top twenty items is two to three weeks. Rolling cycle counts start immediately and are the mechanism by which accuracy becomes known rather than assumed. Barcode scanning is a later step and arrives sooner than most teams plan for.
- 07
Once movement recording is habitual on the top items, add commitments read from the sales channels so available quantity is real. Lot, serial, or expiry tracking follows only where an obligation requires it.
Controls
Controls that matter.
Control 01
Adjustments recorded with a reason and an owner, since unattributed adjustments are how shrinkage becomes invisible
Control 02
Available quantity distinguished from on-hand, because committed stock sold twice is the most expensive inventory error and the easiest to make
Control 03
Cycle count variance tolerances defined per item class, so an investigation is triggered by rule rather than by someone’s judgement on the day
Examples
Three surprises that stop being surprises.
The item sold twice
Distinguishing available from on-hand prevents the same unit being committed to two orders, which is the error customers notice and remember.
The reorder that was too late
A reorder point computed from lead time and demand variability, rather than a fixed threshold set two years ago, is what makes the difference between a buffer and a stockout.
The annual count
Rolling cycle counts turn one large annual variance that cannot be explained into small frequent ones that can, which is how a process problem is actually found.
How it goes wrong
Three ways inventory accuracy erodes.
The system stores quantities, and every variance ends as an adjustment nobody can explain.
Model movement, not quantity. A count tells you what is there; a transaction tells you why, and only the second stops the variance recurring.
Reorder points are computed from a naive average and under-buy through every seasonal peak.
Include lead time and demand variability, and model seasonality explicitly where it exists. An average across a seasonal series under-buys exactly when being short is most expensive.
The same discipline is applied to every item, and the process is abandoned within two months.
Rigour proportional to the cost of error. Ninety-eight percent accuracy on the items that halt work is worth far more than the same figure across the whole catalogue.
Limitations and considerations
What a system cannot count for you.
- A system cannot count the shelf. Accuracy depends on movement being recorded at the moment it happens, and the discipline to do that is an operational change rather than a software feature.
- Reorder points computed from historical demand fail on seasonal and promotional patterns unless those are modelled explicitly. A naive average will under-buy exactly when it matters.
- Stock that carries lot, serial, expiry, or regulatory tracking has obligations a general system should not assume. Verify the requirement before relying on a general model for a regulated item.
- If one person does all receiving and picking, a spreadsheet is honest and adequate. If stock carries regulated lot or expiry obligations, verify what the regime requires before relying on a general model for it.
- Connector coverage varies: Google Sheets, Airtable, Slack are representative rather than guaranteed, and the fields exposed depend on your workspace permissions.
FAQ
Build an inventory system with AI: common questions.
Do we need barcode scanning?
Not to start, and eventually yes for accuracy at volume. Manual movement recording is workable at low throughput; the point at which it stops being workable arrives sooner than most teams plan for.
How accurate does inventory need to be?
It depends entirely on the cost of being wrong per item, which is why classification comes first. Ninety-eight percent on the items that halt work matters far more than the same figure across the whole catalogue.
Why record movements rather than counts?
Because a count tells you what is there and a movement tells you why. Without transactions, a variance can only be corrected; with them, it can be explained, and the explanation is what stops it recurring.
Can it connect to our sales channels?
Reading commitments from sales channels is what makes available quantity meaningful, and it is usually the highest-value connection. Whether write-back is available depends on the channel and on your workspace permissions.
What should the first version contain?
Items below reorder point with the supplier lead time attached, so ordering is a decision rather than a reaction. Everything else waits until that one is genuinely used.
How will we know whether it worked?
Measure stockouts on items that were above reorder point in the system against the baseline taken before anything changed.
Start with ARIA
Ask ARIA to build it.
Describe the website, application, workflow, or operating surface you need. ARIA plans, connects, builds, tests, and keeps refining it — inside the permissions you set.
- ARIA acts only through the systems and permissions you connect.
- Connections use scoped credentials you can change or revoke.
- Actions are recorded, and consequential ones can require approval.
Start here
Build an inventory system around the process you actually run.
Model movement rather than quantity, classify items by the cost of being wrong, and start cycle counting the twenty that matter.