The 400-Line County Fleet Bid Took Four Days to Price. Claude Cowork Hands the Packet Back Finished.

You tried this in 2024 and it invented three part numbers. Fair enough.

The objection is legitimate and most buyers in this trade have earned it. Somebody pasted a bid tab into a chat window, asked it to price the lines, and got back a tidy spreadsheet containing two superseded numbers, one part that does not exist, and a core charge on a line that has no core. It took longer to check than to price by hand. The buyer closed the tab and has not opened one since.

What shipped in 2026 is not a better chat window. It is a different arrangement of the work. You state a goal, it connects to your own files and systems, it breaks the job into steps, it works on its own for hours, and it hands back a finished thing — a priced bid tab, a completed form set, a stocking proposal. Anthropic shipped Claude Cowork on 12 January 2026, aimed squarely at people who are not technical, and expanded it to mobile and web in July. OpenAI shipped ChatGPT Work on 9 July 2026 running on GPT-5.6, doing the same shape of job. That is the change: the output is work product, not conversation.

The job: 400 lines of county fleet bid, due Thursday at 2 p.m.

Every distributor with a commercial side knows this packet. The county fleet department issues an invitation to bid for automotive and light-truck parts. The pricing tab has three or four hundred line items — filters, brake friction, batteries, belts, hoses, lamps, wiper blades — some listed by OE number, some by a competitor’s part number, some by a description a fleet mechanic typed in 2016 and nobody has touched since. You have to cross every one of them to a line you actually stock, price it at a percentage off list or a fixed net, state a core charge where one applies, and commit to a delivery window and a fill rate.

Then there is the rest of the packet, which is the part that makes people miss deadlines: the signed bid form, the non-collusion affidavit, the W-9, the certificate of insurance naming the county as additional insured, three references, the addenda acknowledgments that came out eight days ago, and sometimes a bid bond. Miss one page and a technically perfect price gets thrown out as non-responsive.

Four days of your best category manager, minimum. Which is why most distributors your size look at nine of these a year and bid four. The other five are not a decision anybody made. The calendar made it.

flowchart TD
  A["Buyer states the goal: price the county bid, packet ready Thursday 2 p.m."] --> B["Agent opens the bid tab, your price file and supplier cost sheets"]
  B --> C["Crosses OE and competitor numbers to lines you stock, flags no-bids"]
  C --> D["Returns a draft: priced tab, cover letter, forms checklist"]
  D --> E{"Buyer spot-checks 25 lines and the margin floor"}
  E -->|Kick it back| C
  E -->|Clean| F["Buyer signs, notarizes the affidavit, submits"]
  F --> G["Result and pricing logged so next year starts from it"]

What is actually different from the 2024 attempt

Three things, and none of them is that the writing got better.

First, it works from your files, not from the internet. It reads the price file you exported out of Epicor, your supplier cost sheets, last year’s bid and what you won it at. A cross-reference it cannot support from your own data is supposed to come back flagged rather than filled in. That is the failure mode that burned everybody in 2024, and it is the one to test hardest.

Second, it can hold the whole thing at once. Claude Opus 4.6 brought a million-word working memory, which in practical terms means you hand it the entire bid document — all sixty pages of terms, the pricing tab, and the addenda — instead of feeding it in slices and hoping it remembers page nine when it gets to page fifty-one. Agent Teams, still a research preview, lets several of them split one job in parallel: one on friction, one on filtration, one on electrical.

Third, it runs for hours without you. You start it Tuesday morning and go run your branch. That is the piece that turns four days into an afternoon of review, and it is the reason the fifth, sixth and seventh bid become possible at all.

What the owner has to change about how work gets assigned

This is the part vendors skip and it is the part that decides whether it works in your building. You cannot hand this a task list. Task lists are what you gave a temp. What it needs is a goal with the finish line defined and the rules written down — and those rules currently live in your category manager’s head.

Write them down. Which lines you never bid, because the freight kills you. Your margin floor by category, and where it is different for a fleet account with a two-hour delivery commitment. Which brands you are program-committed to. How you handle a line where the county specified an OE number and you intend to bid an aftermarket equivalent, which is a bid protest waiting to happen if you do not say so on the sheet. Whether tariffed lines get quoted with a surcharge line or baked in. That document is two pages and it is the single highest-value thing an owner can produce this quarter, because it is also what you hand the next buyer you hire.

The second change is the job description. Your category manager stops being the person who types the bid and becomes the person who reviews it, sets the rules, and makes the judgment calls on the fifteen lines that are genuinely arguable. Some buyers love that change. Some find it uncomfortable, because typing the bid felt like the job. Say it out loud rather than letting them work it out from the new workflow.

The arithmetic: it is a win-rate story, not an hours story

Everyone frames this as labor saved. That is the smaller half. Here is the way it actually pays.

Stated assumptions, not a promise: your win rate may fall as you bid marginal accounts, and a fleet contract carries delivery obligations you have to actually staff. The labor saving is real but modest — roughly three days of a category manager per bid, call it twelve days a year at whatever you load him at. The contract you were never able to bid on is the number that matters.

Where it hands you something wrong, and what that costs in a public bid

A submitted bid is binding. That is not true of a quote to a walk-in, and it changes how much checking is enough. If it crosses a heavy-duty friction number to a light-duty line and you win at that price, you are supplying it at that price all year.

The specific things to check by hand every single time: superseded and interchange numbers, because catalog application data is imperfect and always has been; core charges, which get dropped or invented; freight and hazmat handling on batteries and aerosols; anything touched by Section 301 tariffs where your cost moved after the price file was exported; and California-specific lines where an emissions part needs a CARB EO number. Then the forms — the affidavit needs a wet signature and often a notary, the insurance certificate has to name the right entity, and no software signs anything.

Practical rule: spot-check twenty-five lines chosen across categories, plus every line where the margin lands below your floor, plus every no-bid it flagged. If those come back clean the rest usually is. If two of twenty-five are wrong, send it back and find out why before you trust the other three hundred.

Frequently asked questions

Which of the two should a distributor start with?

Whichever your team will actually open. Claude Cowork has had six months of use by non-technical staff and expanded to web and mobile in July; ChatGPT Work is newer, from 9 July 2026, and built for the same kind of user. Both connect to your files. Run the same bid through both for one cycle and let your category manager tell you which one produced less rework.

Do I have to move off Epicor or change systems?

No. The work happens against exported files — your price file, your movement report, your cost sheets. That is a feature, not a limitation: it means nothing is writing back into your system without a person doing it.

What else in a parts business fits this shape of job?

Anything with a deadline, a form set, and a lot of lines. Quarterly supplier price file loads. The annual stock adjustment submission. National account line review packets before the fall buying meetings. Warranty and defective claim packets to manufacturers, which is real money most distributors leave uncollected because nobody has the hours to chase it.

Will my 58-year-old category manager use it?

He will if the first job you give it is one he hates. Do not start with something he is proud of. Start with the claims backlog or the price file reformatting, get a win, and let him bring you the next candidate himself.

Pick one bid and run it twice

Take the next fleet bid on your desk. Have your category manager price it the way he always does. In parallel, state the goal and let it produce a packet from the same files. Compare them line by line and count the disagreements. That comparison costs you nothing you were not already spending and it settles the argument with evidence rather than opinion.

One thing an agent working in the background will not do is answer the phone while it works. If your counter and will-call lines are going to voicemail during the hours your people are heads-down on bids, month-end or a physical count, CallSphere builds voice and chat agents that pick up, take the year, make and model, capture the account, and book the callback. The bid packet gets finished; the shop calling about a water pump still gets a human answer on the other end.