Fix Memory Before You Automate Anything

Most failed rollouts are bad memory and unlabeled data, not a weak model. Teams blame the vendor. The real issue is quieter: the company never decided which numbers are true, so the software invents ones that sound right.

Working memory vs durable facts

Working memory is what you hold for this conversation or this task. Durable facts are what the business stands behind next quarter.

Example: a sales lead says "we did about 18% margin on that job" in a chat. That is working memory. Useful for a question. Dangerous as policy. The durable fact lives in the CURRENT pack (or your equivalent): labeled, dated, owned, and ready for a banker to open without a scavenger hunt.

When you automate, you must tell the system which is which. If you do not, the chat will promote vibes into facts. Then your agent emails a customer a number that never lived in finance.

Provenance or it is gossip

Provenance means: where did this number come from, who owns it, and when was it last refreshed?

Owners often run on tribal knowledge. That works until you hire faster, sell the company, or ask a tool to act without asking you. Then tribal knowledge becomes expensive fiction.

Simple labels beat clever retrieval:

– Source file or system – As-of date – Owner – Status (draft, CURRENT, archived)

If a figure lacks those tags, treat it as discussion material. Never as an input to an irreversible action.

Do not let a chat invent company numbers

This is the failure mode we see most. Someone asks a helper for "last year's EBITDA" or "average job margin." The tool answers fluently. Nobody checks. The number enters a deck, a loan package, or a customer quote.

Fluency is not accuracy. A system trained to be helpful will fill gaps. Your job is to refuse gaps.

Rule we like: if the answer is not pulled from a labeled source of truth, the answer must say so. Better: the tool refuses and asks for the CURRENT pack path.

That feels slower. It is cheaper than a corrected 10-K narrative in buyer diligence.

Parallel to Icon's CURRENT pack rule

At Icon Business Advisors, OneDrive CURRENT packs are the numbers source of truth. HubSpot is the task book. The point is not software brand loyalty. The point is one place that means "this is what we stand behind today."

When packs are CURRENT, automation gets safer. Agents draft from known files. People approve. Shadow sheets die. When packs are stale or unlabeled, automation just scales the mess.

Same idea for CRM fields. If deal stage, close date, and owner are garbage, no assistant will magically produce a clean pipeline. Memory first. Tools second.

A pre-automation audit you can finish this week

Pick the workflow you want to automate. List every number or record it needs. For each one, write:

• Where does the durable version live?
• Who updates it?
• How fresh must it be?
• What happens if the field is blank?

If you cannot answer those four, stop. You are not ready for agents. You are ready for a data cleanup sprint.

Also kill shadow systems. Two "final" forecast files means your software will pick one at random and sound confident either way.

What "good" looks like in practice

Good is boring:

– One CURRENT pack for the metrics that matter this period – HubSpot (or your CRM) as the living task and deal record – Helpers that read labeled sources and draft – Humans who approve irreversible steps – Logs that show which source fed which action

Bad is a chat history treated as the company brain, plus an agent with send rights.

You can buy stronger models later. You cannot buy back trust after a wrong number went external.

The AI pitch will keep getting louder. Memory discipline is still the unpaid work that decides whether automation helps or just moves errors faster.

Before your next automation project, which company number still has no labeled owner and no as-of date?