How is generative AI different from your machine learning service?
Generative AI works with language and content: assistants, retrieval over documents, drafting, summarizing, and extraction from unstructured text. Machine learning as covered on our other page works with structured data to predict numeric or categorical outcomes such as demand, risk, or churn. They use different techniques, need different inputs, and fail in different ways. If your problem is a number, that page applies. If it is language, this one does.
Will the system make things up?
Sometimes, yes. Producing fluent and confident text that is incorrect is a property of how these models work, not a defect that can be fully engineered away. Grounding responses in your own documents and citing sources reduces it substantially and makes remaining errors checkable, but does not eliminate them. Any supplier telling you their system does not hallucinate is either mistaken or misleading you.
What is retrieval augmented generation and do we need it?
It means the system retrieves relevant passages from your own content and answers using them, citing what it used, instead of relying on what the model absorbed during training. If the system needs to answer questions about your specific policies, products, or documents, you need it. If it is only drafting generic text, you may not.
How is this different from RPA?
Robotic process automation executes deterministic, rule-based steps and does exactly the same thing every time, which is what you want for a defined process. Generative AI handles steps requiring language judgment, where the input varies and no rule can cover it, and it does not guarantee identical output for identical input. Many real workflows want both: RPA for the deterministic path, a language model for the step that needs interpretation. Our RPA development page covers the first.
What does it cost to run?
Cost is driven by interaction volume, the amount of content sent with each request, and the model chosen, and it can rise faster than expected once usage grows. We measure cost per interaction during development and project it at your expected volume rather than quoting a figure here, because the range across use cases is very wide.
Is our data used to train the provider models?
That depends on the provider and the contract tier. Business and enterprise agreements from the major providers generally exclude customer data from training, but the terms differ and change. We establish which provider processes your data, under which terms, and whether it satisfies your obligations, before anything is built. Where the answer is unacceptable, self-hosted open models are an alternative we can assess.
Can it work in languages other than English?
Yes, though performance varies considerably by language and by model, and it is generally weaker for languages with less representation in training data. Where multilingual performance matters we test it explicitly against your own content rather than relying on the provider claim.
How do you stop the system revealing information a user should not see?
Access controls are applied at retrieval, so the system can only draw on content the requesting user is entitled to see. Applying permissions after generation is unsafe, because the content has already influenced the answer. This is a design decision made at the start and is one of the first things worth asking any supplier about.
What happens when the model provider updates their model?
Behavior can change without warning, including on cases that previously worked. This is why we build an evaluation set: it can be re-run against a new model version to detect regression before your users encounter it. Systems without one discover the change through user complaints.
How do we get started?
Describe a task in your organization that involves reading or writing text at volume, and tell us what a wrong answer would cost. We will assess whether it is a suitable use case, propose the smallest version worth building, and tell you plainly if we think the risk cannot be contained well enough to proceed.
