Notebook · Signal 4

Decide the authoritative system first

When two tools overlap—CRM vs. helpdesk vs. spreadsheet—automating both directions without a declared winner creates conflicting customer truth. Fix the architecture decision first; only then does AI amplification become safe. Models are very good at sounding coherent; they are not inherently good at knowing which of your databases lied last Tuesday.

Symptoms of unresolved authority

You will see duplicate tickets, mismatched billing addresses, “correct” AI summaries built on stale fields, and endless reconciliation spreadsheets that exist only because two systems disagree. The fix is rarely “smarter AI.” It is usually a boring rule: for each field class, one writer, controlled reads everywhere else, and explicit exceptions documented where they must exist.

How this ties to privacy and procurement

Data mapping depends on knowing authoritative sources—see respecting customer data and the full stack coherence pillar. Growing SMBs often hit this pain when integrations multiply faster than governance; the right time to draw the diagram is before you add another AI connector, not after the first customer complaint escalates to legal.

Exceptions: where dual writers sneak back in

Real businesses have legitimate exceptions: acquisitions, grandfathered contracts, regional subsidiaries with different billing systems. The failure mode is ungoverned exception—everyone knows “finance is messy for EU,” but nothing encodes that in integrations or prompts. Document exceptions with an owner, a review cadence, and explicit customer-facing rules (“EU invoices always originate from system X”). AI should not infer exceptions from folklore; it should read structured rules or hand off to humans when the case matches an exception code.

Conflict resolution when two systems disagree

Even with good intentions, conflicts happen: sync lag, manual overrides, partial outages. Define precedence explicitly—timestamp wins, CRM wins, support wins for ticket state—and log which rule fired. Without precedence, models reconcile creatively. With precedence, you may still have ugly edge cases, but they are explainable in a postmortem. Pair technical rules with workflow design so humans see conflicts as structured tasks, not detective work buried inside model output.

The diagram as a social contract