I used to run my own machine shop. Now I’m shop floor supervisor at Ravnsgaard Metal in Randers. Manufacturing still spends real money on machines, tooling, CAM, ERP seats, and power. The office side is simpler — Word, Excel, PowerPoint, a shared drive, backups. Yet a lot of shops still pay per-seat for Microsoft 365 Business because that’s what everyone defaults to.
I’m not against paying for software when you use it. I’m against paying for a bundle where most of the bundle sits unused.
What a manufacturing office actually needs
Strip it down and the list is short:
- Word, Excel, PowerPoint — or something that opens the same files without breaking them
- A network folder everyone can reach (
\\server\shareor equivalent) - Backups that get tested, not just configured once
- Email — often already hosted somewhere else anyway
What most small shops don’t need as the center of their IT:
- Teams as the company nervous system
- SharePoint workflows and Planner boards
- OneDrive sync for every laptop
- The rest of the M365 collaboration stack that gets bundled into every seat
Sales might use Outlook heavily. Production might never log into Teams. Admin lives in Excel. Engineering keeps drawings on the server. Paying per-seat for software half the company ignores is a quiet leak.
A real bill: 30 M365 seats and 20 ChatGPT accounts
Picture a mid-size manufacturer — 30 people with Microsoft 365 Business Basic, 20 of them with ChatGPT Plus because sales, engineering, and admin all found a use for it. Not unusual. Not extravagant.
What they pay today (EUR, annual billing):
| Line item | Seats | Per seat | Annual |
|---|---|---|---|
| Microsoft 365 Business Basic | 30 | €5.60/mo | €2,016 |
| ChatGPT Plus | 20 | €23/mo (EU) | €5,520 |
| Total | €7,536/year |
That’s €628 every month, recurring, before anyone buys Copilot, extra storage, or a consultant to wire up SharePoint. Microsoft is raising Business Basic to €6.53/mo from July 2026 — same shop becomes €7,768/year on office alone.
ChatGPT Business (the team plan) runs about €20/user/mo on annual billing in the US; in the EU many shops just buy Plus individually because it’s simpler. Either way you’re in the same ballpark: twenty AI seats is four to five thousand euros a year.
What self-hosted costs for the same jobs:
Not zero. You’re trading subscription fees for hardware and someone’s time. But the hardware is yours, and the subscriptions stop.
| Item | Role | Typical cost |
|---|---|---|
| Refurb server + disks | Nextcloud, Samba, backups (restic), OnlyOffice | €1,400 |
| UPS + cold backup disk | Power fail + offsite copy | €350 |
| GPU workstation (used, 24 GB VRAM) | Ollama + Open WebUI for ~20 users | €2,200 |
| LibreOffice on PCs | Word/Excel/PowerPoint replacement | €0 |
| Proxmox, Nextcloud, Ollama, Open WebUI | Software | €0 |
| Year-one hardware | ~€4,050 |
Ongoing (after year one):
| Item | Annual |
|---|---|
| Electricity (two boxes, business hours) | ~€200 |
| Hardware refresh (spread over five years) | ~€810 |
| Software subscriptions | €0 |
| Equivalent annual run rate | ~€1,010 |
You still need a person who updates the server and tests backups — maybe a few hours a month if someone inside the company can do it, or a local IT contractor if not. That’s real cost, but it’s not €7,500 locked in forever.
Five-year comparison:
| SaaS stack | Self-hosted | |
|---|---|---|
| Year 1 | €7,536 | €4,050 (hardware) + ~€200 (power) |
| Years 2–5 | €7,536 × 4 = €30,144 | ~€1,010 × 4 = €4,040 |
| Five-year total | ~€37,680 | ~€8,290 |
Roughly €29,000 over five years — mostly for cloud AI and office bundles a lot of staff barely touch.
The trade-off is honest: local LLMs are not GPT-5 on every hard question. LibreOffice occasionally chokes on a complex Excel macro workbook. You own the problem when the server fan dies at 2 a.m. But for email drafts, summaries, formula help, shared files, and backups — which is what most of those 50 subscriptions are actually doing — you’re paying a premium for convenience and best-in-class AI, not because the cheaper path doesn’t exist.
A stack that covers the basics
This isn’t a single prescription — shops differ. But the menu is well known and mostly free:
| Need | Practical options |
|---|---|
| Office docs | LibreOffice on each PC (free, local). OnlyOffice or Collabora if you need better fidelity with .xlsx and .docx — both can run on a server you control. |
| File share | Samba on a NAS or a small Linux box. A lot of shops already have this and don’t think of it as “open source infrastructure” — it’s just the server in the back room. |
| Sync + optional extranet | Nextcloud — files, calendar, basic office integration if you want browser access without VPN gymnastics. |
| Backups | restic or BorgBackup to another disk, another site, or both. The tool matters less than testing a restore once a quarter. |
| One box, several services | Proxmox on a used workstation — run Nextcloud, OnlyOffice, and your backup target as VMs without buying three physical servers. |
This isn’t exotic. A used PC, a UPS, someone who checks the backups monthly, and you’re most of the way there. The scenario above maps directly onto this menu — same boxes, same software, real numbers.
When Office 365 still makes sense
I’m not selling an ideology. Microsoft 365 is the right call when:
- The whole company actually lives in Teams and the Outlook calendar
- You co-edit documents with customers on SharePoint all day
- Nobody inside the company will touch updates or backups
- A customer or auditor mandates the Microsoft ecosystem
- Sales and travel staff need seamless mobile sync without VPN
If that’s your shop, pay for the bundle and use it. If it isn’t, you’re probably overpaying for defaults.
What manufacturing is actually using ChatGPT for
I’ve watched this spread through offices and shop floors — at Ravnsgaard and elsewhere. It’s rarely “replace the engineer.” It’s mostly text work people were already doing slowly.
Office and admin
- Polish emails — especially English for export customers
- Summarize long PDFs: quote packages, standard snippets, customer specs
- Excel formula help, quick scripting when someone remembers VBA exists
- Draft job postings, SOP outlines, safety bulletins
Engineering and production (light touch)
- “Explain this tolerance stack” at a basic level
- Compare materials in plain language before you dig into the handbook
- Turn bullet notes into a first draft of work instructions
- Translate drawing notes or customer correspondence
Where it goes wrong
- Anything that needs shop-specific judgment — same as any AI tool
- Uploading customer drawings, part numbers, pricing, RFQs — that data leaves the building
- Treating the output as gospel for tolerances, material certs, or compliance
Most of this is text transformation. That’s exactly what a smaller local model can handle if you accept it’ll be worse than GPT-4 on hard reasoning.
Self-hosted LLM — doable, with eyes open
You don’t need OpenAI’s servers to draft an internal email. You need a spare machine, a model runner, and a web UI your staff can actually open in a browser.
A practical path on a shop LAN:
- Ollama — simplest way to run models locally
- Open WebUI — browser chat, multi-user, history stays on your server
- Your existing SMB share or Nextcloud for the files people might want to summarize
Hardware reality:
- 16 GB RAM — small models (7B–8B) for drafting and summaries. Fine for email polish, not magic.
- 32 GB RAM or a GPU — noticeably better responses, still not ChatGPT on the hardest questions
- A ten-year-old office PC — probably not worth it. A used workstation with a decent CPU and RAM is.
What local LLM is good enough for: email drafts, internal summaries, formula hints, rough SOP text, “explain this paragraph.”
What to expect: worse than the best cloud models on complex reasoning. Fine for “don’t send customer data to OpenAI.”
Security note: local doesn’t mean air-gapped by default. Lock down the WebUI. Don’t expose it to the internet without auth and VPN. Same rules as any internal tool — your ERP, your file server, your QC database.
There’s also a difference between me using Cursor on my code repo — where I control what’s in context and I read every diff — and the whole office pasting customer RFQs into a public chat window. Different risk profile. Different tool.
Bottom line
Manufacturing offices are a bad fit for default SaaS bundles. Match the tool to the job: documents, shared drive, backups. Pay for cloud collaboration when you actually use it.
Local LLM follows the same logic. Keep sensitive data in-house when the task is simple enough that a local model can handle it. Save the cloud for when you need it and the data isn’t customer-owned.
The office stack is the easy place to stop overpaying. The machines were never going to be free. The spreadsheets don’t have to be expensive either.