Gemini can now read your CRM and your books
Buy this instead of hiring
Somebody is halfway through a quote in a Doc when they stop, open the CRM to see what this customer was quoted last time, then the accounting system to check whether the last invoice was paid, then come back and finish the paragraph. Nobody counts those trips; each costs two minutes.
That interruption is two lookups and a judgment wearing one coat. What did we agree to, and what have they paid, are just reading a screen and carrying a number back in your head. Only what to do about it needs somebody who knows the job.
Google has moved the lookups to where the writing happens. Its September 15, 2026 announcement says Gemini in Workspace can now interact with seven outside systems through Model Context Protocol connectors — Asana, Atlassian Rovo, HubSpot, Mailchimp, QuickBooks, Monday and Salesforce — reached from the Gemini side panel in Docs, Sheets and Slides, and in Google Chat. Google lists it as available now for Rapid Release and Scheduled Release domains, on by default for anyone with Gemini for Workspace access, with an administrator choosing which connectors are enabled.
Google's post describes accessing information. Whether a connector also writes back depends on the vendor: HubSpot documents its Gemini connector as read-only for contacts, companies and deals, while Google's Marketplace listings for Salesforce and QuickBooks each describe taking actions directly in Gemini. Check the one you care about. Where it is read-only, the trip is one-way rather than gone. And nothing says a single question can draw on two connectors at once, which is the question an owner actually asks: the deal in the CRM against the payment history in accounting.
The office manager who is the company's human index will not trust it, and should not at first. Watch for the spreadsheet they keep anyway. And a connector pointed at a neglected CRM answers confidently and wrongly, so fix what is in the system before making it easier to ask.
One route per cart, and the sorting job disappeared
Reorganize and the job disappears
It is four in the afternoon and somebody on your dock is walking between carts of finished goods, working out which go on which truck. They are good at it. It is an hour of their day, every day.
Viwinco, a high-volume maker of custom windows and doors, had exactly this: finished goods came off the line as a mix of dates and runs, so the loading department had to find and reorganize products, which occasionally led to a missed delivery. The source is George Taninecz's Lean Enterprise Institute write-up of September 22, 2026; LEI coaches Viwinco, so treat it as a partner case study.
The fix was not a better sorting method. They made the cart come off the line already right: "one route per cart, period," in LEI senior coach Karen Gaudet's words, color-coded by day of delivery and labeled with which of roughly 75 weekly milk runs takes it.
The sorting did not get faster. It stopped existing, because the mixed cart it existed to fix is no longer produced.
The same principle turns up on the line itself: a first time study found more than half of process time went on finding the next part among more than 40 pieces of work-in-process, later cut to six between stations. The article does not say which change came first.
It transfers anywhere output is sorted after it is made, because it was not made in the order it ships — picking to stock then re-picking to route, or batching parts then hunting out one job's set.
The catch is that the sequence has to be known while the work is still being made, which pulls the schedule earlier and removes slack the dock was quietly absorbing. The dock lead will resist, and they are right to, because sorting chaos fast is a real skill and it is why they are valued. They will keep a staging area just in case. And if you ship twice a week to one customer, there is nothing here to sequence.
Knowing which customers actually make you money
Nothing off the shelf fits
Every owner has a feeling about which customer is the problem. It is usually the one who calls most. The feeling is sometimes right, it is never evidence, and the gap between those two is where bad pricing decisions get made.
"Which customers actually make us money" sounds like one question. It is three.
The first is what we billed them. That is in the accounting system, it is clean, and it takes a morning.
The second is what it cost to serve them. That is not in the accounting system. It is spread across the job system, the timesheets and the card statement, and those do not share a customer key. The same company is a contact in the CRM, a customer record in accounting, and four job numbers in operations, and nothing anywhere states they are the same company.
The third is what it cost that nobody billed. The callback handled as a favor. The second trip because the site was not ready. The hour a week spent on that account's paperwork because they want everything a particular way. That is usually the whole difference between the customer who looks profitable and the customer who is, and it is the least recorded thing in the business.
How far the tooling gets
Xero is a fair measure of the state of the art. In a post dated September 16, 2026 it says its ChatGPT plugin — the piece first announced at Xerocon US — is now live, with answers built on live Xero data rather than an export. The Claude side, built on a connector it launched in May, is mixed on tense: the post announces a release and says Claude can now retrieve Xero data, but lists the named workflows — business pulse, overdue invoice chasing, cash flow snapshot, month-end preparation — as ones it will support. Treat the list as intent until you have seen one run. Accountants and bookkeepers can use the integrations today, one client organization at a time, with multi-client switching described as coming soon and a Xero and Anthropic webinar on October 27 to explain them. The only accounts of any of this are Xero's and Anthropic's.
For the first question that is useful. For the second and third it does nothing, and not because it is underbuilt. Every answer comes from Xero. Revenue lives in Xero. Cost mostly does not.
What would have to be built
A record that joins the same customer across the accounting system, the CRM and whatever runs the actual work, matching them despite different IDs, a trading name in one place and a legal entity in another. That is the unglamorous half and it is most of the work.
Then an allocation policy, written down. Which overhead a customer carries. Whether a warranty return counts against the original job. Whether a salesperson's time is a customer cost or a cost of being in business. These are not settings. They are your accounting policy, and two reasonable people will answer them differently.
Then a way to capture the unbilled work, which does not mean a new timesheet nobody fills in. The extra trips and callbacks already leave a trace in a dispatch record, a calendar entry, a truck's mileage; the job is tying that trace back to the customer.
What comes out is one page per customer: billed, cost, and the difference, with the few things driving it named.
Nobody sells this, and the reason is not market size. The joining logic is specific to how your records happened to get entered over fifteen years. The allocation rules are yours. A product that shipped with those decisions already made would be wrong for almost everyone who bought it.
Who should not do this
If your job costing is not already honest, do not. If hours get booked to whatever job is open in the app rather than the one being worked, this produces a confident, precise, wrong answer about which customer to fire, and unlike most bad reports, people act on this one. Fix the booking discipline first; it is worth more than anything built on top of it.
If you have twelve customers, do it by hand once. A week with a spreadsheet will teach you more than a system would, and you will have the allocation arguments while they are still cheap.
And write the allocation policy before building anything. That document is the real deliverable. Plenty of businesses that start this find the argument about what overhead a customer carries was the whole insight, and stop there having already taken the value.
The resistance comes from whoever owns the relationship. When the number says their best account is marginal, they will dispute the allocation, and they will often be right, because the first version always charges overhead crudely. The argument is not the danger. What happens quietly afterward is: they stop logging the extra visits, and next quarter the customer looks fine again.
If one of these is your Thursday afternoon, the ten agent roles are in the Library.