Our thinking

ML, LLMs and agents: how the pieces fit

On this page
  1. 01Machine learning: patterns in your data
  2. 02Deep learning: the engine underneath
  3. 03Large language models: language, at scale
  4. 04Retrieval: answers grounded in your data
  5. 05Agents: from answers to work
  6. 06How they work together
  7. 07Choosing what to build

Most AI products aren't built from one technique. Machine learning finds patterns, large language models handle language, and agents turn both into work that gets done. Knowing which does what makes it much easier to decide what to build.

Machine learning: patterns in your data

Machine learning is the broadest layer: models that learn from examples instead of rules. It forecasts demand, flags fraud, scores churn risk and powers recommendations like Netflix's. It works best on structured data you already have, like transactions, behavior and sensor readings.

Deep learning: the engine underneath

Deep learning uses layered neural networks to handle messier inputs: images, audio and free text. It's how medical scans get checked for anomalies and how photos get tagged. Most of today's language and vision models are built on it.

Machine learning, deep learning, and NLP each handle a different part of the same problem. Photo: Iván Antonio on Unsplash
Machine learning, deep learning, and NLP each handle a different part of the same problem. Photo: Iván Antonio on Unsplash

Large language models: language, at scale

Natural language processing used to mean separate tools for sentiment, translation or picking out names. Large language models now do most of that in one, and they can read, summarize and write. On their own, though, they only produce text. They don't know your business unless you connect them to it.

Retrieval: answers grounded in your data

Retrieval connects a language model to your own documents and records, so answers come with a source. It's the difference between a confident guess and an answer you can check.

Agents: from answers to work

Agents combine all of the above with the ability to act: look something up, update a record, send a message, then check the result. They follow rules about what they may do alone and when to hand over to a person. This is where AI stops being a feature and starts owning a piece of work.

How they work together

A real product usually uses several layers at once. One Sogody-built example, TweetPeek, combines sentiment analysis with recommendation models to read a profile and suggest relevant content in seconds. Today, the same idea might add an agent that acts on those suggestions, within rules a team sets.

Choosing what to build

  • Predicting a number? Start with machine learning.
  • Reading or writing language? Use a language model, grounded in your own data.
  • Doing a task start to finish? Build an agent, with clear limits and a person to hand over to.

Tell us what you want to build.

A 30-minute call with our CTO to explore what we could build or automate for your business.