Almost every CPQ vendor now has an AI assistant, and most of them are genuinely useful. But generating a document and remembering what the document committed you to are different problems, and only one of them just got easier.
Almost every quoting vendor has shipped an AI assistant in the last eighteen months. Most of the demos are genuinely impressive, and most of them are showing you the easy half of the problem.
Drafting is a real win. Describing a deal in a sentence and getting a structured quote back — right products, sensible quantities, the ramp laid out — removes work that nobody enjoyed and nobody was good at. So does summarising a forty-page master agreement into the five terms that differ from your standard paper.
Finding things is a win too. "Which customers have a clause letting them exit at twelve months" is a question that used to mean reading contracts. Now it is a question you can ask.
None of that is hype. If your reps are still building quotes in spreadsheets, an AI assistant will genuinely give you hours back.
Here is the line that matters: AI makes it faster to produce a document. It does not make your system remember what the document said.
Ask any of these assistants what a given customer is paying right now, after two amendments and a partial cancellation, and it will answer from whatever is stored. If the upgrade in March was never written down as an event, the model cannot retrieve it. It will either say it does not know, or — worse — produce a confident number derived from the stale opportunity amount.
That second failure is the one to watch. A spreadsheet that is wrong looks wrong. A fluent answer that is wrong looks like an answer.
The pitch behind every AI layer is conversational access to your own data. It is a good pitch, and it rests on an assumption nobody says out loud: that the data describes what happened.
For most revenue questions, it does not. The CRM holds what deals closed at. The billing system holds what was charged. The actual history — which thing changed, when, and why — lives in signed PDFs and email threads, which is to say it does not live anywhere queryable at all.
Point a language model at that and you get fluent answers over an incomplete record. The fluency is new. The incompleteness is not.
AI raises the value of boring data discipline rather than replacing it.
When analysis was slow and manual, bad records were an annoyance — someone spent a week and got to an answer anyway. When analysis is instant and conversational, bad records become a liability, because the wrong answer now arrives in seconds, in a confident sentence, in front of a board.
The companies that will get the most out of AI in revenue operations are not the ones adopting it fastest. They are the ones whose systems already record what happened, in a form a query can reach.
"Show me a customer who upgraded mid-term, then cancelled part of it." Not a fresh deal. A messy one. Watch whether the assistant can state what that customer pays today.
"Where is that stored?" If the answer involves re-reading a document or recalculating from the original opportunity, the record does not exist — it is being inferred.
"Can you show me the same figure as of six months ago?" Point-in-time history is the hardest thing to fake and the fastest way to tell a ledger from a lookup.
None of this is an argument against AI in quoting. It is an argument that the interesting work is underneath it — and that a confident answer over a thin record is a worse place to be than a slow answer over a complete one.
Thirty minutes in your Salesforce sandbox. We install it, configure your pricing, and quote a real opportunity end to end.