Which problems suit AI
AI fits work that is repetitive, language-heavy, and tolerant of small errors. All three conditions have to hold.
Classifying inbound customer email and routing it to the right team — a good fit. It happens hundreds of times a day, it requires reading free-form text, and a misrouted message can be forwarded on.
Payroll calculation — a bad fit. It is repetitive, but it needs no language understanding and tolerates no error at all. That problem belongs to a clearly written function, not to a probabilistic model.
This is where many AI projects go wrong at the outset: they pick a problem demanding exact correctness, then are surprised the output is unusable.
RAG — letting AI read your data
An LLM — large language model — knows nothing about your internal documents. Ask it about your finance team’s approval process and it will invent a very plausible answer.
RAG (Retrieval-Augmented Generation) is the fix: before answering, the system searches your document store for passages relevant to the question and passes them to the model alongside it. The model answers from real documents instead of from a hazy memory of its training data.
Whether the result is good depends mostly on the retrieval step, not the generation step. Retrieve the wrong passage and even the strongest model answers wrongly. So most of the effort in a RAG project goes into chunking documents, indexing them, and checking that real queries surface the right passages.
Cloud or on-premise
Calling the OpenAI or Anthropic API is fast, cheap at low volume, and always gives you the newest model. In exchange, your data leaves your infrastructure.
For most workloads that is acceptable. For medical records, HR data, or documents covered by a confidentiality clause, it is not.
Running models on-premise solves that, at a cost: you need GPUs, people to operate them, and models that run on modest hardware are usually weaker than the best commercial ones. 9G Platform was built to manage both kinds in one system, so sensitive data can stay internal while everything else uses cloud models.
How to measure it
An AI project without a measure never ends — there is always another prompt to tune, another model to try.
Agree the measure before the first line of code, and make it a business number rather than a model number. Not “92% accuracy”, but “manual email triage drops from 400 a day to 60”. The second number tells you whether the system is worth paying for; the first does not.
Alongside it, build an evaluation set drawn from real work — enough to re-run whenever a prompt or model changes and know immediately whether things got better or worse.