--- canonical: https://tanveerriaz.me/blog/local-ai format: text/markdown --- Someone in the group asked a simple question: is there a way to run AI without paying every month and without my data leaving my machine? Short answer - yes. For a big chunk of everyday work, you can. That question turned into a full session, and this is the write-up of what we actually demoed live. No theory. Every tool here is running on my machine right now. Let's start with the thing everyone feels: cost. ChatGPT Plus is about $20 a month. Do the math over a year and that's roughly $240 - per person, per tool. Multiply that across a team and it adds up fast. The frontier models are genuinely brilliant, but here's what most people don't realise: you don't need the frontier for 40–50% of what you do daily. Drafting, summarising, reformatting, analysing a document - a free model running on your own laptop handles all of it. The unlock isn't a cheaper AI. It's knowing which work needs the expensive model and which doesn't. The on-ramp ## Ollama - local AI in minutes [Ollama](https://ollama.com/) is the simplest way in. It's a free app for Mac, Windows, and Linux that lets you download AI models and run them directly on your own computer. Once a model is on your machine, it works with no internet at all - nothing you type ever leaves your device. For anyone handling client documents or financial data, that privacy line is the whole point. The flow is boring in the best way: install the app, pick a model, start chatting. I asked it to write a 1,000-word story live in the room - it started generating instantly, fully offline. That "wait, this is running on your laptop?" moment is what makes it click for people. The trick worth remembering You don't need to be technical to pick a model. Open the free version of ChatGPT and ask: "I have a laptop with [your specs]. Which local model should I run in Ollama for document analysis and writing?" It'll tell you exactly what fits. You don't need to memorise any of this - you just need to know how to ask. Setting expectations ## What local models can and can't do Be honest about the trade-off, because it builds trust: local models run on your hardware, so speed depends on your machine. I demoed on a Mac Mini M4 Pro with 64GB - strong, always-on, basically a little home server. A lighter laptop will run smaller models slower. That's the ceiling. But within that ceiling, the range is wide: text generation, document analysis, image understanding, all local, all free. The one thing that genuinely needs heavy hardware - a proper Nvidia GPU - is image generation . Everything else we showed runs without spending a cent. One step beyond chat ## AnythingLLM - a workspace, agents, and jobs [AnythingLLM](https://anythingllm.com/) gives you a workspace around the model. It's also free and open-source, and it's where local AI stops being a novelty and starts being useful. You point it at your own documents, images, web pages - even a YouTube video - and it works with that content directly, on your machine. ### Running local agents This is where it gets interesting. AnythingLLM lets you run agents locally - the model can browse the web, pull from a document set, and chain a few steps together to actually complete a task, not just answer a question. Same privacy guarantee, more capability: trigger, gather, act, output - all on your side of the wall. ### Scheduling jobs And you can schedule those agents to run on their own. Set a job to run on a cadence - summarise, monitor, prepare something - and it executes without you sitting there. This is the shift from "AI I talk to" to "AI that quietly does work for me in the background." For a small team, that's leverage you own outright. The privacy demo ## Analysing your own documents, offline The demo that lands hardest with a finance and property crowd: drop in a private document and analyse it without a single byte going to the cloud. A contract, a statement, a travel report - the model reads it and answers your questions locally. No upload, no third party, no subscription. If data sensitivity is a real constraint in your work, this is the answer to it. For this region ## English to Arabic, on a local model We also ran translation entirely offline - English to Arabic through a local Ollama model. For anyone producing content across both languages, this matters: you can translate and draft in Arabic on your own machine, no cloud tool in the loop. A small demo with a big implication - language work, done privately, at zero marginal cost. In the real business ## Faizan's flow - AI that makes money Then Faizan took over to show what this looks like when it's making money, not just running in a demo. He runs Forward Digital Consultancy - 15 years in digital marketing - and over the last 6–8 months he's rebuilt how the agency works around AI, across real estate and B2B. His headline use case: turning a client meeting into a finished proposal deck. The old way took hours. His flow takes about 30 minutes after the call ends. 01 Record the meeting. The client call is captured and transcribed with Read.ai. 02 Parse the transcript. The conversation becomes structured input: what the client needs, their industry, their persona. 03 Automate with [Make.com](https://www.make.com/en). The scenario takes that structured content and drives the next step automatically. 04 Generate the deck in Google Slides. Out comes a formatted proposal with industry insights, audience personas, and recommended solutions. Trigger, process, output - with a human check at the end. That last part matters. Faizan was clear it's not fully hands-off: generic or messy meetings still need a person to shape the result. The automation removes the grind, not the judgement. Also worth knowing Use a mix of tools for their strengths - Claude, ChatGPT, Gemini, and [GenSpark](https://www.genspark.ai/) (a Singapore-built platform that bundles several models under one subscription). And for property specifically: DXB Interact for live market data, and the DLD public APIs if you want to build your own analysis on top. The toolkit ## Everything we referenced, in one place [Tanveer Riaz Personal site and projects open →](/) [Session deck Slides from the live demo open →](/deck/local-ai) [Ollama Run AI models locally, free open →](https://ollama.com/) [AnythingLLM Local workspaces, agents & jobs open →](https://anythingllm.com/) [Make.com No-code automation between tools open →](https://www.make.com/en) [GenSpark Several models, one subscription open →](https://www.genspark.ai/) ## What to take away - A free local model handles 40–50% of your daily AI work. Save the paid frontier models for the hard 50%. - Ollama gets you running in minutes. AnythingLLM turns it into agents and scheduled jobs that work while you don't. - Local means private. Your documents never leave your machine - the real edge for regulated or sensitive work. - The winning move isn't one tool. It's a workflow: record, parse, automate, generate - with a human at the end. - You don't need to be technical. You need to know how to ask the right question. That was the session. The goal was never to make anyone an engineer in 90 minutes - it was to show that the door is open and the tools are free. If you want the exact prompts, or want to go deeper on any of these, [reach out](/#contact) or [leave session feedback](/feedback). Thanks to Wali for co-hosting, and to everyone who showed up with sharp questions.