AXIOM FORGE / FIELD NOTES

AI Automation · Illustrated edition

Did the supplier promise delivery, or only dispatch?

Perplexity’s Decider v1.3 can highlight exact words in a supplier quote. A café opening shows why the buyer still needs to read the whole promise.

A café owner holding a furniture quote looks across a room of tables with just one sample chair.
Illustrative scenario: the tables are ready. The buyer needs to know when people can sit at them.

The short answer

Use Locate to show a buyer the words behind a proposed field, not to turn every date into a commitment. Save the exact text version, display the surrounding sentence, and leave arrival unconfirmed unless the supplier actually promised it.

A date can be copied perfectly and still answer the wrong question.

In this article

01 / 04

A Friday with two meanings

Imagine a café owner ordering chairs for an opening. One furniture maker writes, “If the deposit clears by Tuesday, we can dispatch on Friday.” Another writes, “Delivery to your café is confirmed for Friday.” On a purchasing dashboard both might become a green cell marked Friday. Only one supplier said the chairs would be there.

If you can set the terms of submission, give the supplier separate fields for dispatch, arrival and conditions. But a small purchasing team may have to work with the emails and quotes its suppliers already send. That is where a new extraction feature from Perplexity deserves a look.

02 / 04

An answer you can point to

On 12 November 2026, Perplexity announced Decider v1.3 and Decider Flash v1.3. Earlier Decider models could select a label or assign a score. The new Locate question asks for words already present in the supplied text. A match includes the wording, the positions where it starts and ends, and a score. A request can return several matches or none. This is an analysis of the release and model card; we did not run the models.

That makes a different interface possible. Instead of showing buyers a generated explanation for “Friday,” let them click the proposed date and see the supplier’s sentence. The service is available through Perplexity’s Decisions API; the company also publishes model weights. The larger model’s reference setup calls for roughly 49 GiB just for BF16 weights, plus working memory. Running it yourself is a separate infrastructure decision, not a prerequisite for trying the workflow.

One boundary matters for procurement: the documented Locate request takes text, not images. A character range in extracted text is not a rectangle on the original PDF. For scanned quotations, budget separately for text recognition and a way to map the result back to the page. If a faint scan loses the word “not,” a perfect match against that transcript will faithfully reproduce the mistake.

03 / 04

Keep the verb and the condition

A furniture maker straps a pallet of bentwood chairs beside a waiting truck.
A promise to load the truck is not a promise that the chairs will reach the café that day.

Ask for the wording that commits the supplier to arrival at the destination. If the model highlights only “Friday,” the reviewer cannot distinguish either offer above. Show the clause around the match, including its condition. In the first offer, the buyer could record a conditional dispatch date and ask for a delivery estimate. There is no arrival promise to approve.

In the second offer, the buyer can verify the sentence and confirm the arrival field. If the software needs a calendar date, resolve which Friday in a separate step with the buyer. Do not replace the supplier’s words with the resolved date. Someone checking the order later needs both.

There are two different checks here. Can the application reproduce the highlighted characters from the saved input? And do those words support this particular field? The first can be checked mechanically. The second needs examples judged by a buyer who understands what the supplier is promising. The model’s score is not a substitute for either check.

04 / 04

Give the supplier a rehearsal, not a clean sample

A buyer compares two chair quotations, following a highlighted line with a clear ruler.
The highlight belongs to one version of the quote. Keep the older offer readable, not silently replaced.
Quote presentedWhat the buyer should see
Dispatch Friday if the deposit clears.Condition and dispatch wording visible; arrival unconfirmed.
Two identical dates: one for deposit, one for delivery.The arrival highlight opens at the delivery clause, not the first matching date.
A new offer withdraws the original Friday delivery.The earlier approval stays attached to the old quote. The active order asks for review.
No delivery date appears anywhere.No invented date. The buyer can request one without treating an empty result as proof no promise exists.

For the rehearsal, ask the vendor to retain each input exactly as supplied to the model. Give it two versions of the same offer, then add a line above the highlighted sentence in the new version. Open the original approval again. If the highlight jumps to another line, the application has saved positions without saving the text they refer to.

For a first release, keep the tool beside the purchasing decision, not inside the rule that places orders. Compare its suggestions with the team’s normal review on a set of previously resolved quotes. Count unsupported arrival promises separately from missed ones: the former can put an opening at risk, while the latter usually cost a buyer another email. Agree which mistakes your team can live with before commissioning automation beyond highlighting.

Sources and editorial scope

This is Axiom Forge’s procurement and product-delivery guidance, not a guaranteed project plan. The example is illustrative. Sources checked on 16 Nov 2026.

From uncertainty to a clear scope

Know what you are commissioning.

Send the workflow, audience and constraints you already know. We’ll review the context privately and discuss a suitable next step. Please leave out passwords and confidential customer data.

Write to us

Email us — we reply in English, Russian or Armenian.

Our time now:

Analytics preferences

We use Google Analytics and Google Ads to understand which pages help visitors and to measure enquiries. Form contents and personal details are never sent to them.