TL;DR
The saving is not that a machine read the request. It is that a person reviews three lines instead of forty — which only happens if something tells them which three, and per-line uncertainty is that mechanism.
How should an AI quoting system decide which lines a human must check?
Start from the honest version of the promise. Nobody buys a quoting system so a machine can read an email. They buy it because a person currently checks forty lines and would rather check three.
That is the entire saving. It is not the parsing, not the matching, not the pricing — those are the work that makes the saving possible, and none of them is the saving itself.
And the promise only holds if something tells the reviewer which three. If they have to re-verify all forty to find out which ones to trust, nothing was saved and the afternoon comes straight back.
For anyone specifying the review screen, which is where this whole category lives or dies. 7 minute read.
The saving is the review, not the reading
Worth stating plainly because it inverts how these systems usually get evaluated.
A demonstration that impresses shows a messy request going in and a tidy quote coming out. That demonstrates extraction. Extraction is the part that works — chapter 0 covers how thoroughly, and Microsoft ships a capable version of it inside Business Central.
The question that decides whether anyone saves time is the one no demonstration answers: when this is wrong, will I know which line to look at?
- Reads the request — The typing
- Matches every line — More typing
- Tells you which lines it is unsure about — The re-reading, which is the actual cost
Per-line, or the promise collapses
The mechanism is uncertainty attached to individual lines, not to the document.
A document-level score is close to useless for this purpose. "This quote is eighty per cent confident" tells a reviewer nothing actionable: they still do not know whether the doubt is spread thinly across forty lines or concentrated in three, and the only safe response to an unlocated doubt is to check everything.
Per-line uncertainty is not a feature to add later. It is the thing being bought, and it has a corollary that chapter 7 argues in full: the score has to be about the same thing the reader thinks it is about. A confidence about an item, displayed at the end of a row, silently becomes a claim about the quantity and the price as well.
Two rows, both matched, both priced. One was found by item number with nothing else it could have been. The other edged a near-identical catalogue entry from a different supplier. A reviewer who cannot tell those apart has to treat them the same, and the only safe way to treat them the same is to check both.
Scale the attention to what a line is worth
A refinement that costs little and matters commercially.
Not every uncertain line deserves equal attention. A doubtful match on a line worth a few units of packing material and a doubtful match on the largest line of the order are the same epistemic problem and very different business problems.
So confidence thresholds should move with line value. Be more willing to accept a marginal match on something trivial; be less willing on the line that carries the order. This is not a machine-learning decision — it is a policy decision, and it should be visible and adjustable rather than buried.
It also protects against the failure mode where a reviewer, faced with a long list of flags, starts clearing them by reflex. A flag that fires too often stops being read, which is the same argument chapter 3 makes about noise in the instructions panel.
Show every line by default, and let filters be a working aid
There is a strong pull toward collapsing the settled lines. It makes the point — look how much we did, here is what is left — and it looks like focus.
It is the wrong default, for a reason that has nothing to do with usability: the accepted lines go out with the reviewer's name on them too. Somebody is about to send this document to a customer. A reviewer who cannot see the whole quote cannot check the whole quote, and collapsing puts a click between them and the thing they are taking responsibility for.
So the default view is all lines, with a counter making the point that most of them are settled.
Filters are for working, not for hiding. A needs you view is genuinely useful for clearing open work without settled rows in the way — and it must not be where the reviewer lives. The distinction is whether the reviewer can see the document they are signing, and the default decides that.
A row that says "needs you" has to give you something to press
Small rule, disproportionate effect.
Every open line carries its own controls, on the row, always visible, never on hover:
- Change it — wrong item, wrong quantity, wrong unit. The reviewer's correction beats every heuristic, and it should be remembered for this customer.
- Ask about it — the question goes into the thread against this line, and the answer comes back against this line.
- Source it — the line is real but not on the shelf, which is chapter 9.
A control reachable only by hovering is a control somebody misses, and a control in a toolbar requires knowing the toolbar exists. Keyboard shortcuts are excellent for the person who has used the system for a year and irrelevant on day one.
And nothing sticky in the middle of the page. A floating bar slides over the panel that has to be read in order to keep visible the one that does not.
How Business Central surfaces work, and where the difference sits
Microsoft's Sales Order Agent [GA] has a considered review model, and it is worth describing accurately rather than as a straw man.
Work arrives in a Tasks pane. Microsoft: "Tasks that require attention—like reviewing an incoming or outgoing email—are at the top of the list." Steps needing a person are listed under Needs Attention, and the reviewer opens a timeline, selects Review, and either Confirms or Stops.
The agent asks for help in named situations — "Reviewing and confirming incoming email requests", "Reviewing outgoing emails", and "Providing assistance to Sales Order Agent when it needs to get unblocked—for example, to provide missing data."
And it explains itself. Business Central "can also display the reasoning used by the agent and citations that led to a suggested value." That is genuine transparency and more than many systems offer.
The difference this chapter is drawing is narrow and specific. Those surfaces are organised around tasks and steps. What the four pages do not describe is a per-line ranking — a signal on each row saying this one is least safe, start here. At the availability step Microsoft states plainly: "You can't change the items and quantities at this point."
*Explanation answers why did it say that. A ranking answers which one should I check first.* A reviewer with fifteen lines and no ranking reads fifteen lines.
About Cognilium Cognilium builds AI systems that work in tandem with Microsoft Dynamics 365 — the decisions the ERP records but does not make. Business Central and Finance & Operations, on your own governed stack. https://cognilium.ai · https://www.linkedin.com/company/37180269/
Want to see what a per-line review screen looks like against a real request of yours? Book a 15-minute call and bring one. No deck.
Sources
- Process sales quotes and orders with Sales Order Agent
- Sales Order Agent overview — Business Central
- FAQ for Sales Order Agent — Business Central
- Item availability in Sales Order Agent (preview)
Sources
Share this article
The workspace these articles describe — one queue, per-line confidence, supplier choice and the write-back — as a product for Business Central distributors.
