好文集ChatGPT Plus¥119

A CRM Row Was Built for Reports

A CRM row was built for reports. It holds a stage, an amount, an owner, and a close date. A manager can filter it. A dashboard can sum it. For twenty years that was the whole job, and the row did it well.

An AI agent has a different job. Ask it "why is this deal stuck?" and a close date tells it nothing. It needs the story: what the buyer said on the last call, who went quiet, which champion changed roles, and what moved in the past three weeks. None of that lives in a row. It lives in emails, call transcripts, meeting notes, and the gaps between them.

Most CRMs answer this by adding columns. A notes field. Then an AI summary field. Then a sentiment score, a health score, and ten more. Every new column is a schema migration, a form nobody fills in, and one more number sales ops has to defend. The table grows and nobody trusts it.

The failure is structural, not a matter of effort. A column holds one value of one type. Understanding a deal is not one value. Forcing a narrative into cells either flattens it until it is useless or fragments it until nobody can read it back.

Why retrieval at question time is not enough

The obvious fix is to skip storage and let the agent read everything when the question arrives. Pull the emails, the transcripts, the activity log, and let the model sort it out.

That works for one question. It falls apart as a system. Every question re-pays the full cost of reading months of history. Latency grows with the age of the account, not the difficulty of the question. Two agents asked the same thing do the work twice and can land on different answers, because each reconstructed the state from scratch.

Worse, nothing compounds. The hundredth question about an account should be cheaper and better informed than the first. If every answer starts from raw history, it never is.

Understanding should be written down when the event happens, not reconstructed when the question arrives. An email lands once. Process it once, update the state, and let every later question start from the result.

Two layers, not more columns

The fix is not another column, and it is not a vector index bolted on the side. It is a second layer in the data model itself.

Keep the fields as they are: typed, validated, filterable. Filters, reports, routing rules, and forecasts read them. They do that job well, and nothing about agents makes them worse at it.

Beside them, give each record a semantic layer: keyed, unstructured values written for a model to read. The key states the intent. The value carries the full story. An executive summary of the account. A timeline of the engagement. The transcript of the last call. A note on who has gone quiet and since when. A model opening the record finds the understanding already assembled, under keys that say what each piece is for.

The layer only works if it can be trusted, and trust comes from freshness. A summary that lags its record by a week is worse than no summary, because someone will act on it. So the semantic layer cannot be a nightly batch job or a cache that expires when it feels like it. When an email lands or a meeting ends, every value that depends on it has to move. Staleness is a bug, not a degraded state to tolerate.

And the layers have to stay apart. A structured field must be correct: a filter that returns the wrong rows is broken. A semantic value must be complete: a summary that drops the buyer's objection is broken. Those are different bars, tested differently and maintained differently. One column asked to do both jobs does neither well, which is exactly what happened to the notes field.

What it changes

An agent starts with understanding instead of a blank page. It does not read 40 emails to learn the state of a deal; that state was written down the day the 40th email arrived. Each agent builds on what the last one established, the way a new account executive inherits a well-kept file instead of a cold start. The work compounds.

The tradeoff is real, and nobody should pretend otherwise: analytics on unstructured text is hard. The rule that keeps it honest is promotion. When a semantic value needs a filter, promote that one value into a typed field. Promotion is cheap, because the information already exists; it just needs a type. The fields stay the source of truth for reports. The semantic layer stays the source of truth for reasoning. Neither pretends to be the other.

Rows are not going away. Reports still matter, forecasts still matter, and every system that reads fields today will keep reading them. But the next generation of software reads a record the way a colleague would, and a colleague has never once been helped by a close date alone.

全文完
查看原文