Now onboarding a limited number of teams Request access
This field note is not published yet. See what is live on the blog.
FIELD NOTES

Why we will not let AI send a quote

We automate a great deal of correspondence. Not this. The line is not drawn at how confident the system is, it is drawn at what the message does, and a price does something the others do not.

It would be easy to build. The quote data is in the record, the audit has already run, the draft grades well. Sending it automatically is a smaller technical step than most of what already runs.

We do not, and the reason is not caution about model quality. It is that a price is a different category of object from a reply.

A clarification can be corrected. A price has been seen, and cannot be unseen.

The asymmetry

Consider what recovery looks like for each kind of message.

MessageIf it is wrongRecovery
A clarification to a buyerYou send a correction the same dayMild embarrassment
A follow-up to a supplierYou clarify in the next messageNothing
A priceThe buyer has already read it and formed an expectationYou are negotiating from a number you never meant to offer

The recovery column is what decides which messages can be automated.

Once a number reaches a customer it becomes an anchor. Withdrawing it costs credibility even when you are entitled to. And in a competitive process, a price you did not intend to give is not corrected so much as inherited.

Why grading is not enough here

Our own gate blocks anything under 8 out of 10, and quote drafts routinely score well. That is exactly why the score is the wrong basis for this decision.

A grade measures whether the draft is supported by the record. It cannot measure whether the commercial position is one you want to take this quarter, against this competitor, on this account, given a relationship the record knows nothing about.

That is not a gap in the model. It is information that was never in the system, and no amount of context recovers a judgement that has not been made yet.

Where the line sits

  • Routine correspondence can send automatically once it passes the gate. Acknowledgements, clarifications, supplier questions, follow-ups. High volume, low commitment, recoverable if wrong.
  • Quotes, proposals and submissions wait for a person. Your click is the approval, it is logged under your name, and the rule is enforced on the server rather than in the interface.
  • Submission is stricter again. Your password, an exact preview of the message and attachment, an explicit confirm. It is the one action that cannot be recalled, so it gets the most friction, deliberately.

The friction is the feature

There is commercial pressure to move that line. A vendor who promises fully automatic quoting is promising a bigger number in a demo, and the demo is where these decisions get made.

But the value of automating correspondence is that it clears the volume so a person can spend their attention on the few messages that set a price. Automating the price too does not extend the benefit. It removes the reason the benefit existed.

Your click is the approval

Celestix AI drafts quotes and proposals from the complete record, at your margin, graded before you see them. It will not send one. That is not a setting.

Field notes are written from work we do on live deals. All figures and documents shown are sample data. No customer, supplier, or buyer is identified.