A model can understand inventory and still misunderstand your inventory report. It does not know that your team measures slow movers from the last sale, excludes customer-held shipments from late deliveries or closes the reporting day at store close.
That missing handover is business context.
At a glance
- Tell the agent how your business defines the work, not just what to produce.
- Separate stable rules from live data and keep a source, date and owner with each rule.
- Turn a wrong answer into a context correction and a check for the same mistake.
Five things an experienced colleague already knows
| Kind of context | What the agent needs to know | A useful example |
|---|---|---|
| Identity | Which thing are we discussing? | Which branch or supplier a name refers to |
| Process | How do we do this here? | The steps for the weekly stock review |
| Decision | What was agreed, and why? | Customer-held shipments are excluded from late orders |
| Relationship | Who owns the result? | The inventory controller owns recount requests |
| Time | Which version is current? | The policy took effect on 1 September |
Connecting documents helps, but it does not automatically settle these questions. Two old policies can both look relevant. A spreadsheet can contain the right figures under a definition the team no longer uses.
Knowledge explains the rule, a skill explains the work
In Hubzoid, reference knowledge holds definitions, policies and decisions. A skill describes how to perform a repeatable task. Keeping them separate makes both easier to maintain.
For example, a returns policy says when a refund is allowed. The returns-checking skill says how to find the sale, apply the current policy and prepare an answer. The skill refers to the policy instead of repeating it.
When the refund window changes, there is one rule to update. Someone should still check the procedure and its outputs after the change.
A business rule should answer a real question
Consider an illustrative trade discount request: an 18 percent discount on an order worth 9,000 in the business's currency. The approved policy says the sales manager reviews discounts above 10 percent and up to 20 percent.
A useful answer names the sales manager as the reviewer, cites the policy and says the discount is awaiting approval. It does not turn the lookup into an approval.
The note needs enough detail to support that answer:
- the bands and how the discount is calculated
- any order-value exception
- the policy's effective date and source
- the role that owns the rule and its backup
If a workflow must enforce those bands, use code for the comparison. The model can explain the result. A prose rule alone does not enforce a decision.
Keep changing facts connected to their source
Today's balance or stock count belongs in the system that owns it. A tool can fetch the current record, while the Hub explains what it means.
Give the answer a reporting period and source timestamp. If the export is missing, report the missing source. A plausible substitute can be worse than an incomplete answer.
Fix the handover before changing the model
When an answer is wrong, ask what the agent lacked. Was the rule missing, the term ambiguous, the source stale or the exception unwritten?
Correct the shared note, then add a small behavioral check. For the late-order example, ask about a customer-held shipment and check that the answer excludes it. Hubzoid calls these checks evals.
Run the check after later edits. It helps catch a known mistake, though passing a few checks does not prove every answer will be correct.
Next step: list the three definitions your team explains most often, then put them in the Hub's knowledge.
Watch Knowledge and skills, episode 2 (opens in a new tab) for the same distinction in the fictional Wren & Page bookshop.
