From Chatbot to Agent
AI is moving beyond chat. The next wave is practical systems that retrieve trusted knowledge, use tools, follow real workflows, remember what matters, escalate to humans, and prove their value. This series explains how to build those systems responsibly — in plain English, with the architecture shown.
15 articles
Start here
What an AI system actually is, and how a model gets access to your own documents.
- Part 1Why a chatbot is not an AI systemMost AI projects start with 'can we add a chatbot?'. It is a reasonable place to start and a poor place to stop — because the chat window is the one part of an AI system that does the least work.8 min read
- Part 2RAG explained simply: how AI answers using your own documentsRetrieval-augmented generation is the pattern behind almost every 'AI that knows our business' project. It is not complicated — and understanding it tells you exactly why these projects succeed or fail.9 min read
- Part 3Building a simple RAG backend: the minimum useful architectureThe first version of a document AI system should be boring, traceable and easy to test. Here is what actually belongs in it — and the long list of things that do not belong in it yet.9 min read
- Part 4From RAG to agents: when AI needs to take actionRetrieval gets you an answer. An agent gets you an outcome. Knowing which one your problem needs — and resisting the upgrade when it does not — is one of the more expensive judgement calls in AI delivery.8 min read
Build the agent
The loop an agent runs, and the four things around it that make it safe to deploy.
- Part 5The AI agent loop: perceive, reason, plan, act, observeEvery agent, from a simple document classifier to a multi-step research assistant, runs the same loop. Understanding it is what lets you set limits on it — which is the whole job.9 min read
- Part 6Memory for AI agents: what should be remembered, and what should notAn agent that remembers nothing is a stranger every morning. An agent that remembers everything is a POPIA problem with a chat interface. The design question is which facts earn a place in between.9 min read
- Part 7Tools: how AI agents actually do workThe model reasons. Tools are what let an agent do anything at all — and the moment you grant one, you have given a non-deterministic system production access. That deserves the same scrutiny as any other integration.8 min read
- Part 8Evaluation: how do we know an AI agent is working?If there are no tests, you are guessing. Evaluation is the difference between an AI system you can defend in a governance meeting and one you can only describe enthusiastically.9 min read
Run it in production
Visibility, cost control, and the delivery discipline that gets a pilot over the line.
- Part 9Observability and cost: why production AI must be visibleAI workloads bill by usage and behave non-deterministically — a combination that turns an unmonitored system into an unpredictable invoice and an unanswerable audit question.8 min read
- Part 10From local prototype to production-ready AI agentThe demo took a weekend. Getting it into production has taken seven months and it is still not there. That gap is not a technology problem — it is everything a prototype is allowed to leave out.9 min read
- Part 15The CloudNala Agent Development LifecycleFourteen steps, eight of which happen before anyone writes a prompt. That ratio is the method — because the expensive AI failures are design failures, and design failures are cheapest to fix on a whiteboard.9 min read
Practical use cases
What this looks like in tendering, public service delivery, internal operations and proposals.
- Part 11Building a tender AI agent: from document to bid/no-bid recommendationReading tenders faster is not the prize. Making better bid decisions, with the evidence attached, is — and that changes what you build and where the human stays.9 min read
- Part 12Public-sector AI workflows: from citizen request to service actionPublic-sector AI should not start with a chatbot on the website. It should start with the service workflow — because a citizen who gets a fast answer and no action is worse off than before.9 min read
- Part 13Building an internal operations agent before you sell oneThe most honest way to learn what agentic AI is actually like to run is to point one at your own operations, where the consequences of getting it wrong land on you.8 min read
- Part 14AI agents for proposals and RFPs: where automation helps and where humans still matterAI can take days out of proposal production. It cannot take on any of the accountability — and confusing those two things is how organisations end up committing to work they cannot deliver.8 min read
Exploring AI agents in your organisation?
CloudNala can help you start with the workflow, the data, the governance and the architecture — before jumping into tools. Most engagements begin with a short design phase, because that is where the expensive decisions get made.
Book an AI Readiness Workshop