When the digital business diverges from the physical one
Why dependable automation starts where the facts are observed

- By
- Cam Myers
- Content type
- Point of View
- Reading time
- 6 min read
- Status
- Author’s perspective
- Topics
- Operating Models
In brief
- Applications agreeing on a record does not establish that the physical fact is correct.
- Meaning, identity, timing, and traceable events determine whether a record can support a promise.
- Distinguish missing observations from information that already exists but needs better interpretation.
The argument
The screen says eighty units. The warehouse has fifty-three.
Every system connected to that record can agree perfectly. Purchasing, advertising, sales, and the owner's dashboard can all repeat the same number. Their agreement does not put the missing twenty-seven units on the shelf.
This is a difficult problem because technical consistency looks reassuring. The integrations are running. The reports reconcile. Nobody sees an error message. Yet the company is making decisions about a business that exists in its records and differs from the one employees are operating.
AI makes the distinction more consequential. Software can analyze an inaccurate record and produce a persuasive recommendation. It can also move that recommendation into action quickly enough for the original error to affect many customers before someone checks the shelf.
The decision underneathWhich observation makes the promise true?
Separate a recorded fact from an observed one
Inventory records are claims about physical objects. Someone received, counted, identified, moved, damaged, reserved, sold, or returned something, and the business captured that event with some degree of accuracy.
The record becomes trustworthy through that chain of observations and corrections. Connecting more applications to it improves availability. It does not establish that the observations were right.
Consider an illustrative midsized building materials supplier in the northeastern United States buying products by the pallet and selling by the piece. Receiving enters ten pallets. A later system interprets the number as ten pieces. Another employee corrects the quantity without preserving the conversion or the reason for the change. Sales records may now appear reasonable while the inventory history no longer explains what happened.
An accurate quantity also needs a clear meaning. Does it include damaged stock? Is it physically present but already promised? Is it available at the location from which the customer will collect it? A company can answer every arithmetic question correctly and still promise goods it cannot supply.
The first design decision is therefore what the record means in the operation. A field called available quantity should have an agreed relationship to physical stock, commitments, condition, location, and the timing of updates. A shared label is insufficient when departments use different definitions.
Identity carries the history
The company must also know whether two records describe the same thing.
A product may have a supplier number, an internal inventory number, a barcode, and several descriptions. Similar products may differ only by size, handing, finish, or pack quantity. If those distinctions matter to fulfillment, a match based on a similar name can be commercially wrong even when it looks plausible.
The opposite problem is equally costly. The same product appears under several identifiers, splitting its sales history and making each record look like a slow seller. Purchasing then makes a decision from incomplete demand.
AI can help propose matches or interpret messy descriptions. The system still needs rules for accepting those matches and preserving uncertainty. A likely match is not automatically an acceptable basis for combining financial histories or promising a specific physical item.
That distinction should depend on the action. A tentative match may be useful for exploratory analysis. The same match may need verification before stock is committed or a customer's order is changed.
The cost of a wrong identity determines how much assurance is worth buying.
A current snapshot cannot explain every change
Suppose today's quantity is lower than yesterday's. The difference could reflect a sale, damage, a transfer, a counting correction, or an import problem. A series of daily totals shows that the number changed. It may not establish why it changed.
When the company needs historical inventory, product profitability, or time-to-sell analysis, the events behind the totals become important. Otherwise a correction made today can be mistaken for a sale, or a late record can distort the period in which an activity actually occurred.
Preserving events does not require capturing every conceivable detail. Capture enough to explain the changes relevant to decisions: what moved, how much, when the business event occurred, when it was recorded, and the basis for important corrections.
Keep the original record where practical and make corrections traceable. Quietly overwriting history may clean up a dashboard while destroying the evidence needed to understand the discrepancy.
A model can estimate an unobserved quantity, but the company should know when it is looking at an estimate. A reconstructed inventory balance should not acquire the authority of a physical count merely because it appears in the same column.
Spend observation effort where it matters
Perfect information everywhere would be expensive and often unnecessary. The sensible goal is sufficient confidence for the decision being made.
Errors in an inexpensive, slow-moving item have different consequences from errors in a scarce part that can halt a valuable installation. Focus verification where a wrong record can cause meaningful loss, repeated customer failures, or expensive decisions.
Freshness matters in the same way. A monthly count may support one planning decision while being unsuitable for a promise made to a customer this afternoon. Data quality is partly a relationship between the information and the use someone intends to make of it.
Discrepancies should also lead to a named correction process. Who checks the physical stock? Who resolves the unit conversion? Who can change the authoritative record? Without those responsibilities, a report of discrepancies becomes another queue nobody owns.
Correcting a number addresses today's symptom. Investigating recurring differences can reveal a receiving problem, an ambiguous product definition, or a transfer that is never recorded. The more durable improvement changes the point where reality enters the system.
Ask which information is actually missing
Not every data problem calls for a company-wide cleanup before useful work can begin. The information needed for one outcome may already exist outside the field people habitually use.
A broad internal category might be inadequate for advertising classification while the product description contains the size, material, and intended use needed to classify it. A different route can use that information without waiting for an overhaul of the internal category structure. The alternative still needs to be checked, but the original dependency may be avoidable.
Physical quantity is a different problem. A description that identifies the product does not establish how many usable units remain in the warehouse. Better interpretation cannot recover an event nobody observed or recorded with sufficient reliability.
This distinction changes the investment. Sometimes the company needs a better way to interpret information it already has. Sometimes it needs a new observation, such as a count or a recorded transfer. Treating both as generic data cleanup conceals the actual design choice and can produce either unnecessary work or unjustified confidence.
For a business planning automation, I would begin with one important promise: an available item, a delivery date, a rental booking. Trace the information behind it back to the observation that makes it true. Examine where the claim can become stale, lose its meaning, or be attached to the wrong object.
Then decide how the business will detect and correct those failures before expanding the automation.
The most valuable data work may happen at the receiving desk or in the yard. That is where the company establishes whether its digital account of the business deserves to be trusted.
For the leadership team
In practice
- Trace an important customer promise back to the observation that supports it.
- Define stock availability, product identity, update timing, and correction ownership.
- Spend verification effort in proportion to the consequences of a wrong record.
Questions worth asking
- What evidence makes this record trustworthy for the action we intend to take?
- Do we need a better interpretation, a new observation, or a corrected operating definition?
Weighing this decision now? Tell us what you are considering and we will say plainly whether we can help.
Discuss your business

