Gemini can write it up, but not from your numbers
Stop producing the unread write-up
Somebody just came off a forty-minute call. Three things have to exist: a summary document, an email to the two people who missed it, and a calendar hold. That is twenty minutes of moving between apps, four times a day.
Since early September 2026, Gemini can be prompted from inside Gmail, Docs, Drive, Slides, or Chat and will generate a formatted doc, sheet, or deck in the background, draft or send the email, find open times and book the meeting, and log reminders into Tasks. Before, you had to be inside each app to get its help. Rollout began September 2, 2026, on Business and Enterprise Standard and Plus among other tiers, in English only. Only Google's own account of this was available; nothing here has been independently confirmed.
But writing up a meeting is not one job, it is three. What was said is clerical and recoverable. Who needs to know is routing. What the numbers are is neither, and that is where the post goes quiet. Google says a document the user cannot open stays shut to Gemini too. About the accounting system and the field software it says nothing at all. The job cost and the open balance still arrive by hand.
Notice where the safety net sits. Google describes a review card before mail goes out or a meeting is booked, and describes docs and decks as generated in the background. Whether those get a checkpoint it does not say. The item most likely to be quietly wrong is the one Google says least about.
And if the write-up is a document nobody has opened in three months, stop producing it and see who complains.
The schedule nobody in the field was reading
Reorganize and the job disappears
Somebody in your office rebuilds the schedule every week. They open the software, drag the bars, print it, email it out. On the job the foreman tells his crew what they are doing that morning, and nobody opens the attachment. When something changes, the question goes back to the office, waits, and returns as a revision.
The plan is doing two jobs badly. It is meant to coordinate the people who decide things, and it is meant to tell the crew where to be. Detail helps the first a little and the second not at all.
The Pentagon renovation is the case. Two years before the 2001 attacks, the Department of Defense had already issued the design-build RFP: six pages, where federal proposals of that kind run to thousands. That left the solution to the builders. After the attacks they stood up a co-located office within a week and put every decision maker in it, every trade and the owner, settling things together and tracking cost and schedule in the room. Southland regional president Tim Michael told Facilities Dive that rather than a scheduling effort carrying some fifteen thousand line items for every wedge of the building, they “got it down to a single 11-by-17 sheet,” and that the crews could read off it where they were due each day and each week.
The scheduling work did not get faster. It stopped existing, because what it was for, carrying decisions between people not in the same room, was being done another way.
Michael names the limit himself: not every job can put every decision maker into one trailer on site. Where you cannot, the smaller sheet on its own will not save you.
The same part under five different numbers
Nothing off the shelf fits
A purchase order arrives from your largest customer. Six line items, and on four the part number is theirs, not yours. Somebody who has been there eleven years recognizes three on sight and hunts for the fourth: a spreadsheet, then last year's order, then an email to the buyer. The order sits.
That is not a filing problem. One item travels under a different name everywhere it goes, and nothing in the building holds all the names at once.
What Shopify shipped
Shopify's changelog entry of September 8, 2026, says a variant could previously hold exactly one barcode, forcing merchants to park the others in a spreadsheet. It can now carry up to twenty, each typed as UPC, EAN, ISBN, GTIN, ASIN, or a custom code, validated against that standard's format as it is entered. Only Shopify's own account of this was available; nothing here has been independently confirmed.
Where it stops is the interesting part. Shopify says the channels still read whichever barcode sits first, and that labels, feeds, and apps carry on unchanged. The lookup problem is solved; the translation problem is not, which is the shape of the gap in most businesses, online or not.
Matching a part number is four questions
Watch what the veteran actually does. It is not one act of recall.
First: have we seen this number before? Clerical, a lookup against a record, and the only part software has always been good at.
Second: is it even well-formed, the kind of number it claims to be? A UPC carries a check digit; a customer's internal code has a house style. That is a rule, and rules can be written down.
Third: is a number that looks right actually right? A correctly formed UPC belonging to a different product passes every format check ever written. This is the question that costs money, because it fails silently and downstream, in a shipment.
Fourth: what do you do with a number nobody has seen? New item, renamed old one, or the customer's typo? That is judgment, it depends on knowing this customer, and it is what lands on a person.
Most attempts solve the first two, declare victory, and leave the third and fourth where they were.
What would have to be built
A cross-reference record: one row per physical item, holding every identifier it travels under, yours, the manufacturer's, each customer's, each supplier catalog's, with the source and date each was learned.
Then the part that earns its keep. When a new catalog or customer part list arrives, it proposes matches rather than making them: this number probably means that item, here is why, please confirm. A confirmation becomes a new row. The record improves every time somebody uses it, which is the only way such a mapping ever gets built. Nobody enters four thousand cross-references on a Tuesday.
Nobody sells you this because the mapping is yours. Your accounting or ERP system carries one part number per item by design. Supplier catalogs arrive as spreadsheets in each vendor's own format. The customer's number for your item exists only on their purchase orders. Tools that handle this properly are built and priced for distributors with a data team.
The hard part, and who will fight you
The hard part is the third question, and it is not technical. Being confident enough to let a match through unattended is a different standard from being useful enough to suggest one. Set the threshold high and everything queues, which is what you have now. Set it low and you ship the wrong part with a tidy audit trail saying you meant to.
Your inside sales or purchasing lead will resist, and she is right. The mapping in her head beats anything you build in the first year, and she carries the exceptions: the customer whose engineering department renumbered everything in 2019 and still orders under the old codes. She will not argue. She will use the system for the easy ones and keep her spreadsheet for the hard ones, and you will not discover it for eight months, by which point her spreadsheet is the real system and yours is decoration. Give her the confirming seat, so the record learns her judgment instead of competing with it.
The honest no: below a few hundred active parts this does not pay. The spreadsheet is cheaper and the veteran is faster. Fix your own part list first: if it carries duplicates and items you stopped selling in 2021, a cross-reference will formalize the mess and make it look authoritative.
If one of these is your Thursday afternoon, the ten agent roles are in the Library.