---
title: "What are the different types of AI in the market, and which is suited to my business?"
description: "Types of AI for business, explained: what each kind does, open-weight vs proprietary, cloud vs your own hardware, and why a smaller model is often enough."
url: "https://www.lmtek.com/source-code/types-of-ai-for-business"
locale: "en"
updated: "2026-10-15"
---

# What are the different types of AI in the market, and which is suited to my business?

A plain-language map of the AI on the market, for business readers deciding what their organisation actually needs. It isn't a model ranking and it isn't a supplier recommendation: LM TEK makes the infrastructure AI runs on, not the AI itself. Read it before you choose a tool, a model or a supplier.

By LM TEK · Published 15 Oct 2026

## Key points

- Business AI comes down to three separate questions: what it does, how you get it, and how big it has to be. Answer them separately.
- What it does: language, vision, forecasting, search over your own documents, or agents. How you get it: proprietary or open-weight, and, separately, a provider's cloud or your own hardware. Open-weight is not the same as local.
- The assistants most people know, such as Claude and ChatGPT, are general-purpose frontier models. You can rent them but not own them, and open-weight models of similar scale need data-centre-class hardware.[^dsv3]
- For many business jobs, a smaller single-purpose model is big enough, and that's where the efficiency is.[^loraland][^nvidia] For a general assistant for every employee, renting a frontier model may well be the right call.

You ask for a short overview of the AI options for your business. What comes back is a slide with a dozen terms on it: generative AI, LLMs, open-source, cloud AI, copilots, agents, RAG. They're laid out like a menu, as if you pick one.

You don't. Some of those words describe what the AI does. Some describe how you get hold of it. None of them tells you how big it needs to be, which is the question that decides most of the bill.

Picture a vehicle catalogue that lists "refrigerated", "diesel", "leased" and "van" as four alternatives. Every word is accurate. The list is still useless, because each word answers a different question. Most AI guides are that catalogue.

This one keeps them apart.

## Three questions most AI guides mix up

**What does it do?** Does it read and write text, look at images, predict a number, find an answer in your files, or carry out a task across several systems? This is the question a business owner asks first, and rightly, because it starts from the work.

**How do you get it?** Who owns the model, what the licence lets you do with it, and where it runs. This is the question that decides who controls the price, where your data goes, and whether you can keep a version that works.

**How big does it have to be?** Size drives what a model can do. It also drives the hardware, the energy and the running cost. It's the question most guides leave out, and the one this piece ends on.

The three are independent. A language model can be proprietary or open-weight. It can run in a provider's cloud or on your own hardware. It can be enormous or small. Knowing you need a language model tells you nothing yet about how to get one, or what size to get.

## What it does: five types of business AI

**Type · 01 · Language models**
Read, write, summarise, translate and draft. Typical uses: customer replies, contract summaries, internal knowledge questions, first drafts of reports. They need text in, and they work best with clear instructions and examples of good output.

**Type · 02 · Vision**
Inspect images and video. Typical uses: visual quality control on a production line, reading scanned forms and invoices, checking goods against a label. They need images and, for a narrow inspection task, labelled examples of good and bad.

**Type · 03 · Forecasting and prediction**
Predict a number or a category from past data: next month's demand, which customers are likely to leave, which machine is likely to fail. They need clean historical data from your own systems. *Generative vs predictive:* a generative model produces new content, such as a paragraph or an image; a predictive model produces an estimate, such as a figure or a yes-or-no. Plenty of organisations have run predictive models for years without ever calling them AI.

**Type · 04 · Search over your own documents**
Find and answer from the organisation's own files: policies, contracts, manuals, case notes. The system first retrieves the passages relevant to a question, then a language model writes the answer from those passages, so the answer can point back to its source. It needs your documents, organised well enough to search.

**Type · 05 · Agents**
Carry out multi-step tasks by using other tools: look something up, fill in a form, update a record, send a message. Typical uses: back-office workflows that today pass between several systems and a person. Of the five, agents are the least mature. They turn a model's mistakes into actions rather than text, so they need tighter limits on what they may do and closer human oversight than the other four.

Most real deployments combine types. A support assistant that answers from your knowledge base is type 04 built on type 01. An agent usually has a language model at its core.

## How you get it: two choices, not one

Getting a model involves two separate choices. Mixing them up is the most common mistake in conversations about AI and data, and it leads to wrong conclusions about where your data goes.

**Choice one: proprietary or open-weight.** A *proprietary* model is owned by its maker, and you use it only on the maker's terms, usually through a paid service. You can't download it, inspect it or keep a copy. An *open-weight* model has its trained parameters, the "weights", published for anyone to download and run under a licence. You can keep a version that works for as long as you like, and test and inspect it yourself.[^licences]

Open-weight is not the same as open-source. Under the Open Source Initiative's definition, an open-source model also comes with the complete code used to train and run it, and enough detail about its training data for a skilled person to build an equivalent, and it can be used for any purpose without asking permission. Many open-weight models publish the weights only, and some licences restrict how they can be used.[^osi][^licences]

**Choice two: a provider's cloud or your own hardware.** This decides where the model runs, and so where your prompts, documents and outputs go while it works.

|  | **Runs in a provider's cloud** | **Runs on your own hardware** |
|---|---|---|
| **Proprietary** | The usual way to use the big assistants: rented through an API or a subscription | Rare; only where a maker licenses an on-premises deployment[^onprem] |
| **Open-weight** | Rented from a cloud or hosting provider that runs the model for you | Downloaded and run on infrastructure you control |

> **Open-weight is not the same as local**
> An open-weight model can run on a cloud provider's servers, and then your data goes to that provider. A proprietary model is available only where its maker offers it. Openness decides what you can inspect and keep. Where the model runs decides where your data goes. Two decisions, made separately.

## What you know: the big general-purpose assistants

The AI most people have used is a general-purpose assistant such as Claude or ChatGPT. These are *frontier models*: the largest and most capable models available at a given time, built to do almost anything reasonably well. Writing, coding, analysis, translation and conversation, all from one model. That breadth is exactly why they are so large.

One way the largest models are built is called **Mixture of Experts**. Instead of one network that does all the work for every word, the model is split into many sub-networks, called experts, and a small routing layer sends each word to only a few of them. The rest sit idle for that word. The idea dates back to 1991; a 2017 paper showed how to scale it to very large networks, and several of the largest open-weight models now use it.[^moe] The benefit: each answer takes less computation than a model of the same total size that uses all of itself every time.

The name invites a misreading. The experts are not specialists in law, finance or medicine that the model consults by topic. The model learns its own way of dividing the work during training, and that division doesn't follow business subjects. A Mixture of Experts model is still one general model, not a panel of specialists.

Which of the assistants you know are built this way isn't public. Neither Anthropic nor OpenAI has published the internal architecture of its flagship models, such as their size or whether they use Mixture of Experts.[^closed] Open-weight models show the technique in published numbers:

**Mixture of Experts in three numbers: DeepSeek-V3**[^dsv3]
- **Total parameters:** 671 billion
- **Used per word:** about 37 billion
- **Held in memory to run it:** all 671 billion, roughly 700 GB of GPU memory for the weights alone, at the precision it was released in

The third number is the one that matters if you want to own the model. Think of a workshop with many benches. Each job visits only a few of them, so each job is quick. But you still need floor space for every bench, because you don't know which ones the next job will need. A Mixture of Experts model uses only part of itself for each word, but the whole model has to sit in memory, ready for whichever experts the router picks next. Fewer calculations per word don't mean less hardware to hold the model.

## What you need: a model big enough for the job

That's why you can't put Claude or ChatGPT in your own office. Their weights aren't released, so they are available only to rent. And open-weight models of comparable scale need multi-GPU, data-centre-class servers just to hold them: DeepSeek-V3, at roughly 700 GB, takes a server with several of the largest data-centre GPUs.[^dsv3] How to work out the memory any model needs, and what that means for the hardware, is the subject of [*How much GPU does a model actually need? The VRAM math*](/source-code/how-much-gpu-does-a-model-actually-need).

What you can own is a **single-purpose model**: a smaller model, chosen or fine-tuned for one job, sized to fit hardware an organisation can run itself. Fine-tuning means training an existing model further on examples of the specific task, so it gets better at that task in particular.

Nobody buys a 40-tonne truck to deliver sandwiches across town. A van does the job, costs a fraction to run, and fits in your own yard. The evidence that the same holds for AI is growing. In the *LoRA Land* study, researchers fine-tuned 310 models, built on ten small open-weight models of 2.5 to 8.5 billion parameters, for 31 specific tasks. On average the fine-tuned models scored 10 points higher than GPT-4, and GPT-4 still did better on only 6 of the 31 tasks.[^loraland] That result holds for those tasks and those test conditions. It is not a general claim that small models beat large ones. Researchers at NVIDIA have gone further and argued, in a position paper, that small language models are powerful enough, better suited and more economical for many of the repetitive tasks inside agent systems.[^nvidia]

The model makers are moving the same way. OpenAI's open-weight gpt-oss-120b is a Mixture of Experts model that its maker states runs on a single 80 GB GPU.[^gptoss] Right-sizing isn't a compromise forced on smaller buyers any more. It's a design goal.

> **What you know vs what you need**
> What most businesses know is the biggest assistant on the market. What most business jobs need is a model big enough for that job, and no bigger.

None of this makes the big assistants the wrong choice. For broad, open-ended work, such as a general assistant for every employee, a frontier model rented through an API may well be the right tool. The point is that it shouldn't be the default for every job just because it's the one you've heard of. Why it matters who owns the model at all is argued in [*Renting intelligence is the new outsourcing mistake*](/source-code/renting-intelligence-is-the-new-outsourcing-mistake).

## Match the job to the type, the size and the delivery

The table puts the three questions together for some common business jobs. It describes fit, not products: it names no model and no supplier, and your own data, volumes and constraints can move any row.

| **Job** | **Type of AI** | **Size it usually needs** | **Delivery that usually fits** |
|---|---|---|---|
| Drafting customer replies in your house style | Language (01) | Single-purpose, fine-tuned on your past replies | Open-weight, on your own hardware or in a cloud |
| Searching and answering from contracts | Search over your documents (04) | Single-purpose language model plus retrieval | Open-weight on your own hardware, where the documents are sensitive |
| Visual inspection on a production line | Vision (02) | Small, task-specific | On your own hardware, close to the line |
| Demand forecasting | Prediction (03) | Small; often not a language model at all | Wherever your data already lives |
| Multi-step back-office tasks | Agents (05) | Single-purpose models for the routine steps; a larger model only where judgement is needed | Mixed; with strict limits on what the agent may do |
| A general assistant for every employee | Language (01) | General-purpose frontier model | Rented through an API or subscription |

Read down the size column and the pattern shows. Only the last row really needs a frontier model. The other five need a model that does one job well.

That's where the efficiency in business AI comes from: a tool big enough for the job, and no bigger. A single-purpose model needs less hardware and less energy, it can run inside your own operational boundaries, and it's a model your organisation can own rather than rent.

So start from the job, not from the model you've heard of. Pick the type the job needs, then the smallest size that does it well, then decide where it runs. In that order.

## Frequently asked questions

### What are the different types of AI used in business?

By what it does, business AI falls into five broad types: language models that read and write text, vision models that inspect images and video, predictive models that forecast numbers or categories, search systems that answer from an organisation's own documents, and agents that carry out multi-step tasks using other tools. Separately, each can be proprietary or open-weight, and can run in a provider's cloud or on your own hardware.

### What is a Mixture of Experts model?

A Mixture of Experts model is split into many sub-networks, called experts, and a routing layer sends each word to only a few of them, so each answer takes less computation. The experts aren't topic specialists; the model learns its own division of work. The whole model still has to be held in memory to run, so it needs as much GPU memory as its total size suggests.

### Is an open-weight model the same as running AI locally?

No. Open-weight means the model's weights are published and can be downloaded under a licence. Local means the model runs on hardware you control. An open-weight model can run in a cloud provider's data centre or on your own servers, and only the second keeps your data inside your own boundaries.

### Can I run a model like ChatGPT on my own hardware?

Not ChatGPT itself: its models are proprietary. You can use them through OpenAI's own services or through Microsoft's Azure cloud, but you can't download them or run them on your own hardware.[^azure] You can run open-weight models, but those of comparable scale need multi-GPU, data-centre-class servers. For most business jobs, a smaller single-purpose open-weight model is enough and runs on far less hardware.

### What is the difference between generative AI and predictive AI?

Generative AI produces new content, such as text, images or code, from a prompt. Predictive AI produces an estimate from past data, such as next month's demand or the likelihood that a customer leaves. Many businesses have used predictive models for years; generative models are the newer type behind today's assistants.

## Read next

If a single-purpose model sounds like your answer, the next question is how much hardware it needs.

- [How much GPU does a model actually need? The VRAM math](https://www.lmtek.com/source-code/how-much-gpu-does-a-model-actually-need): How model size, compression and the number of users decide the GPU memory a model needs, and what that does to the budget.
- [Renting intelligence is the new outsourcing mistake](https://www.lmtek.com/source-code/renting-intelligence-is-the-new-outsourcing-mistake): Why core, steady AI work shouldn't stay rented by default.

## Sources

Model cards and makers' pages were accessed on 14 October 2026.

[^osi]: Open Source Initiative, [The Open Source AI Definition 1.0](https://opensource.org/ai/open-source-ai-definition), 28 October 2024: an open-source AI system can be used "for any purpose and without having to ask for permission", and comes with "the complete source code used to train and run the system" and "sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system". See also OSI, [Open Weights: not quite what you've been told](https://opensource.org/ai/open-weights), January 2025.
[^licences]: Two examples. Meta, [Llama 3.3 Community License Agreement](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/LICENSE), 6 December 2024: use is subject to Meta's Acceptable Use Policy, and a licensee with more than 700 million monthly active users "must request a license from Meta". OpenAI, [gpt-oss-120b model card](https://huggingface.co/openai/gpt-oss-120b): released under the "permissive Apache 2.0 license".
[^onprem]: For example, Google, [Run Gemini anywhere, including on-premises, with Google Distributed Cloud](https://cloud.google.com/blog/topics/hybrid-cloud/gemini-is-now-available-anywhere), 28 August 2025.
[^moe]: Jacobs, R. A., Jordan, M. I., Nowlan, S. J. and Hinton, G. E. (1991), "Adaptive Mixtures of Local Experts", *Neural Computation* 3(1), 79 to 87, [doi:10.1162/neco.1991.3.1.79](https://doi.org/10.1162/neco.1991.3.1.79); Shazeer, N. et al. (2017), "Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer", [arXiv:1701.06538](https://arxiv.org/abs/1701.06538).
[^closed]: Anthropic's [Transparency Hub](https://www.anthropic.com/transparency) and [models overview](https://platform.claude.com/docs/en/about-claude/models/overview), and OpenAI's [GPT-5 System Card](https://openai.com/index/gpt-5-system-card/) (August 2025), describe the models without giving their size or internal architecture.
[^dsv3]: DeepSeek-AI, "DeepSeek-V3 Technical Report", [arXiv:2412.19437](https://arxiv.org/abs/2412.19437): "671B total parameters with 37B activated for each token". [DeepSeek-V3 model card](https://huggingface.co/deepseek-ai/DeepSeek-V3): the published model totals 685 billion parameters, including a 14-billion-parameter prediction module, and "we only provide FP8 weights". At FP8, about one byte per parameter, the weights alone take roughly 685 GB.
[^loraland]: Zhao, J. et al. (2024), "LoRA Land: 310 Fine-tuned LLMs that Rival GPT-4, A Technical Report", [arXiv:2405.00732](https://arxiv.org/abs/2405.00732): "4-bit LoRA fine-tuned models outperform base models by 34 points and GPT-4 by 10 points on average"; "GPT-4 continues to outperform fine-tuned models in 6 out of 31 tasks". The GPT-4 version tested was gpt-4-0613 (June 2023).
[^nvidia]: Belcak, P. et al. (2025), "Small Language Models are the Future of Agentic AI", NVIDIA Research, [arXiv:2506.02153](https://arxiv.org/abs/2506.02153), a position paper: small language models "are sufficiently powerful, inherently more suitable, and necessarily more economical for many invocations in agentic systems".
[^gptoss]: OpenAI, [Introducing gpt-oss](https://openai.com/blog/introducing-gpt-oss), 5 August 2025: gpt-oss-120b runs "efficiently on a single 80 GB GPU" and "activates 5.1B parameters per token"; the [model card](https://huggingface.co/openai/gpt-oss-120b) gives 117 billion parameters in total.
[^azure]: Microsoft Learn, [Foundry Models sold by Azure](https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure): "They include all Azure OpenAI models".

Free to cite with attribution: LM TEK (2026). What are the different types of AI in the market, and which is suited to my business? LM TEK Source Code. https://www.lmtek.com/source-code/types-of-ai-for-business