What is Generative AI?
Generative AI is a class of models, mainly large language and multimodal models, that produce text, code, images or structured outputs from natural-language and document inputs. In an enterprise it is used to summarise, retrieve, interpret, draft and assist.
The model is trained on broad data and does not know your business. Enterprise Generative AI is therefore mostly an engineering problem: supplying the right context, constraining the output, integrating it into a workflow and measuring whether it is correct.
What is Agentic AI?
Agentic AI is software in which a model plans and executes multi-step work by calling tools and systems, keeping state across steps, and escalating to a person when it should not act alone. It differs from a chatbot in that it does things, not only says things.
Production agents run inside a defined identity, an explicit permission set and a tool registry, with traces of every step. An agent without those is a prototype, whatever the demo looked like.
Generative AI vs Agentic AI: what's the difference?
Generative AI produces an output for a person to use. Agentic AI takes actions across systems to complete a task. The first is a capability inside a workflow; the second is a workflow participant with permissions, state and accountability.
Most enterprise systems start generative (a knowledge assistant, a drafting copilot) and become agentic only for steps where the action is well defined, reversible or approved by a person. Moving a system from generative to agentic raises the bar for evaluation, observability and security, which is why we treat it as a separate engineering decision.
RAG vs fine-tuning: which should you use?
Use retrieval-augmented generation (RAG) for knowledge the model does not have and for content that changes. Use fine-tuning when evaluation shows the base model consistently fails on behaviour, format or domain style that prompting and retrieval cannot fix. They solve different problems and are often combined.
RAG keeps facts current and traceable to a source. Fine-tuning teaches a model how to respond, not what is true this week. Choosing fine-tuning to fix a knowledge gap is the most common expensive mistake we see in GenAI programmes.
When does an AI agent need memory?
An agent needs memory when a task spans more than one interaction or one session, when earlier decisions change later ones, or when personalisation is part of the value. It does not need memory for a single-turn retrieval or a stateless classification.
Memory is a data-governance decision as much as a design one. What is remembered, for whom, for how long and who can inspect it has to be decided before the feature exists, because it is difficult to retrofit.
What is MCP and when is it useful?
MCP, the Model Context Protocol, is an open standard for connecting AI applications to tools and data sources through a common client-server interface. It is useful when several AI applications need the same tools, or when you want tool integrations that outlive a particular model provider or agent framework.
It is not useful as a reason to expose every internal system to an agent. Every MCP server is a permission boundary and needs the same identity, scoping and audit treatment as any other integration.
Single-agent vs multi-agent architecture?
Start with a single agent and a well-designed tool set. Move to multiple agents only when the workflow has clearly separable roles, different permission boundaries or different models per step, and the added coordination is justified by the outcome.
Multi-agent architectures multiply the surfaces to evaluate, trace and secure. They should be justified by the workflow, not chosen because they are fashionable.
How do you evaluate an AI agent?
With a set of realistic tasks that have known correct outcomes, run against the agent before every change, measuring task success, tool-call correctness, steps to completion, unsafe-action rate and escalation behaviour. Production traces are then sampled and scored the same way.
Agent evaluation is the subject of its own page, because it is the part most teams skip and the part that decides whether an agent can be trusted with real work.
How we evaluate agents and GenAI systems
How do you prevent unsafe agent actions?
By making unsafe actions impossible rather than unlikely: scoped permissions, a registry of approved tools with side-effect classification, output validation, prompt-injection defences, and human approval for any action that is irreversible or consequential.
Guardrails on the model's text are the last line, not the first. The first line is that the agent simply cannot call a tool it was not granted.
How do you control GenAI cost?
By measuring cost per completed task, routing each request to the smallest model that meets the quality bar, caching what repeats, trimming context to what retrieval proves is relevant, and designing the workflow so people handle the cases the model should not.
Cost is a design input from the first architecture review, not an optimisation pass at the end. A system that is correct and uneconomic will still be switched off.