One Home per Fact: How to Stop AI From Spreading Bad Data

Bad data rarely arrives with a warning. It looks like a normal customer record, project status, or renewal date. Trouble starts when the same fact lives in two places and nobody knows which copy is right.

The risk grows when an AI agent can read and write across tools. A person might notice that a CRM says “active” while a project board says “paused.” An agent may use whichever appears first, then carry the mistake into an email or database.

Before connecting an agent, give every important fact one official home. Other systems may reference it, but only one should define it.

Start with facts, not apps

Listing apps does not answer the key question: where does each fact belong?

Build a short register for information that can change a decision or trigger an action.

Important fact	Official home	Owner	Other tools that may read it
Customer lifecycle stage	CRM	Sales lead	Email, reporting
Contract renewal date	Contract system	Account lead	CRM, calendar
Invoice payment status	Accounting system	Finance lead	CRM, reporting
Delivery status	Project system	Delivery lead	CRM, client updates
Support priority	Support desk	Support lead	Project system

Focus first on facts that affect money, promises, access, deadlines, or customer communication. A harmless duplicate note can wait. A conflicting renewal date cannot.

Choose the home and owner

Choose the official home with four questions:

1. Who is accountable for accuracy? Name a role, not “everyone.”
2. Where is the fact created or approved? That system is often the best candidate.
3. Which system keeps the useful history? Payment, consent, and status changes may need an audit trail.
4. What depends on the fact? The home must be suitable for those decisions.

Write the rule plainly: “The CRM is the official home for lifecycle stage. The project board may show that stage, but it cannot change it.”

The owner defines valid values, edit rights, exceptions, and conflict decisions. They approve new workflows that write to the field. Assign a backup and review date so the rule does not quietly expire.

Set sync rules before connecting an agent

A sync is not a neutral pipe. It creates authority unless you limit it. For every shared fact, document:

• the source and destination;
• whether access is read-only or can write;
• which fields move;
• what starts the sync;
• how quickly the destination should update;
• what happens when data is missing or invalid;
• where errors are logged and who receives an alert.

Prefer one-way movement from the official home. Two-way sync creates conflicts when both copies change.

If another tool only needs context, show a link or record ID instead of copying the value. When a duplicate is necessary for search or reporting, label it as a mirror and block edits there.

If you later connect apps and tools to an agent, grant only the access needed for the agreed flow. Connection is the last step, not the first.

Decide how conflicts stop

Never let the agent improvise when systems disagree. Choose a conflict rule in advance.

For a low-risk field, the official home can win automatically. Log the rejected value, source, time, and reason for later review.

For a consequential field, stop and ask for review. A mismatched payment status should not trigger a collection email. Do not settle conflicting consent with “latest timestamp wins.” Delayed imports can make stale information look new.

Use three outcomes:

• Accept: the official value is clear and valid.
• Quarantine: preserve the conflicting input without using it.
• Escalate: send the mismatch to the owner before dependent work continues.

The safe default is no outbound message, irreversible update, or financial action while a consequential conflict remains unresolved.

Protect the data around the workflow

A clean source of truth can still be badly governed. Review permissions, privacy, and retention alongside the sync design.

Start with least privilege. If the agent needs a company name and delivery status, do not grant access to private notes, payment details, or an entire drive. Separate read from write access. Restrict destructive actions and changes to consent, access, billing, and contracts.

Define how long copies, prompts, logs, quarantined records, and exports remain available. Remove them when their purpose ends, subject to legal and contractual duties. Check that deletion in the official home removes copies elsewhere.

Keep sensitive values out of broad error alerts. “Record 1842 has a status conflict” is usually safer than pasting the full customer record into a group channel.

Run a small, reversible test

Pick one important fact, one source, and one destination. Use test records or a small internal segment, and keep the old workflow available.

For two weeks, let the agent read from the official home and update only a non-sensitive mirror or draft. Review anything that could reach a customer or change a consequential record. Seed safe conflicts to confirm the workflow stops, logs the mismatch, and alerts the right person.

Measure:

• percentage of tested records matching the official home;
• conflicts detected and actions correctly stopped;
• time from alert to resolution;
• unauthorized or out-of-scope write attempts;
• mirrors that miss the agreed update window.

Expand only after the owner reviews the evidence. If the test fails, remove write access, clear the test mirrors, and return to the prior process.

Watch for predictable failure modes

Naming a source of truth without removing edit access elsewhere leaves a second home.

Treating “most recently updated” as a universal rule also fails. Bulk imports can disguise old information.

Derived facts also need a home. If “customer health” is calculated from support, billing, and delivery data, document the formula, inputs, refresh schedule, and owner.

Watch for silent failures, renamed fields, expired credentials, departed owners, and tests without conflicts.

The candid limitation

One home per fact will not make the fact true. The official system can still contain a typo, an outdated decision, or a value entered by the wrong person. This method reduces ambiguity about where correction belongs. It does not remove the need for review, evidence, and accountable judgment.

The principle was prompted by Corey Ganim’s X post about his Audit, Optimize, Automate framework and his YouTube walkthrough. In that source-reported client example, customer information lived in both ClickUp and a CRM. The duplicate ClickUp copy was removed because mismatched records invite errors. That experience is useful inspiration, not independent proof that the same design is right for every team.

Make the rule visible

Keep a one-page record of the fact, official home, owner, readers and writers, sync direction, conflict response, retention, and review date.

Then connect one bounded process and watch what happens. The goal is not to move more data. It is to ensure every important fact has one place where it can be checked and corrected before an agent carries it further.

Audit one workflow before you automate it.