Short answer: a capable AI model can now run on hardware in your own office, with nothing sent to an outside AI company. For most Indian businesses that's worth doing only for a narrow set of work: client files, financials and internal documents you'd never paste into a chatbot. For everything else, a good cloud service is still cheaper and smarter.
This became a practical choice quietly over the last year, and most owners we speak to haven't noticed yet.
What changed
Two things happened at once. Open-weight models (ones you can download and run yourself) got good, and the tools to run them got simple.
This week gave a good marker. On 15 October, Meta released Muse Spark Open, an 816-billion-parameter open-weight model, and showed it running fully on two of Apple's new M5 Ultra Mac Studios with no cloud servers involved. That's near the top end of what AI can do, sitting on a desk. The machines with enough memory for it cost far more than a normal office PC, so it's a signal of direction more than a purchase for most firms.
The more useful news for a 20- or 100-person company is lower down. Models in the 7 to 14 billion parameter range now run on a laptop or desktop with 16 to 24 GB of memory, according to hardware guides tested on everyday machines. Free tools like Ollama or LM Studio install them in minutes. They won't match the best cloud models on hard reasoning, but for summarising a contract, drafting a reply or answering questions from your own SOPs, they're good enough.
Which work belongs on a private model
Sort it by business process and by how sensitive the data is.
Operations is usually the best fit. Vendor contracts, HR files, internal SOPs, purchase records and staff questions like "what's our leave policy for contract workers?" are all text, all private, and all repetitive. A local model that reads your own documents and answers from them saves hours without anything leaving the building.
Leadership comes next. Board papers, monthly MIS, cash-flow sheets and salary data are things owners hesitate to upload anywhere. A private model can summarise the month's numbers or draft a note to the board from them.
Sales is mixed. Client proposals, pricing and negotiation notes are sensitive; standard follow-up emails aren't. Many teams end up splitting the work: private model for anything with a client's details or your margins, cloud model for the rest.
Marketing rarely needs it. Blog drafts, ad copy and social posts are meant to be public anyway, and the best cloud models write better. Paying for local hardware here is usually wasted money.
What a local model does and doesn't solve
It removes one specific risk: your data going to an outside AI provider, being stored on their servers, or being used under terms you didn't read closely.
It doesn't make you compliant with India's data protection law on its own. Under the DPDP Rules notified on 13 November 2025, most obligations for businesses, including consent, security safeguards and breach reporting, apply from 13 May 2027. Those apply whether the AI runs in your office or in someone else's data centre.
It also adds a risk many people miss: the machine itself. In January, researchers at SentinelOne and Censys found about 175,000 Ollama servers open to the public internet across 130 countries, India among them, most without any login. Ollama only listens on the local machine by default; these were set up carelessly. If you run a model for the whole office, someone has to own its access settings the way they'd own the accounting server.
What it takes to set one up
For a first try, the list is short:
- One machine with enough memory. A recent laptop or desktop with 16 GB handles a small model for one person. A shared office setup needs more, often a desktop with a dedicated GPU.
- A model with a licence you've read. Many open models use permissive licences like Apache 2.0, but terms differ by model and by size, and some larger releases add limits. Check the licence file on the official model page.
- One owner. Someone who keeps the software updated, decides who can reach it, and knows which documents it can read.
- A clear first job. "Answer staff questions from our 40 SOP documents" is a good start. "Use AI for everything" isn't.
Try it with one team for two weeks. If people keep using it, scale it; if they drift back to the cloud chatbot, find out why before buying more hardware.
When to skip it
If your data isn't sensitive, if nobody can look after a machine, or if the work needs the strongest reasoning available, stay with a cloud service on a business plan that doesn't train on your data. That's the right answer for a lot of companies, and there's nothing second-best about it.
The point is to choose deliberately. Most businesses today send everything to the cloud by default, or avoid AI for sensitive work altogether. There's now a reasonable middle option, and the work that suits it is easy to list. If you're mapping which of your processes fall where, our AI consulting and strategy page explains how we usually sort it.
Arthat AI
Working on something similar at your company?
Arthat AI builds these automations for Indian businesses every week. If you want to talk through whether this fits your situation, we'll give you a straight answer — even if the answer is "don't build it."
