Knowledge base, RAG, agent, and workflow are not four alternatives but four layers: first decide whether each step is “look up”, “compute”, or “act”, then choose the layer. The most common mistake is skipping the layering and going straight to an agent — asking an autonomous system to do deterministic work is both expensive and uncontrollable.

Questions to ask before procurement

  • Which questions can be answered by retrieval alone, and which need an action? Mixing the two is where most agent projects start losing control.
  • What are the writes — changing data, sending mail, placing orders, calling systems? Who approves each write, and can it be reversed?
  • What interfaces do the existing systems expose? Where is knowledge scattered (documents, wikis, tickets, CRM), and how is permission managed in each place?
  • How fast does knowledge change? What updates weekly, what stays fixed for a year? How is outdated knowledge found and removed?

How to judge technically

Layer by determinism, from simple to complex:

  • Knowledge base: structured facts first. Product parameters, prices, policies, procedures — structure whatever can be structured. Answers from structured knowledge are verifiable and do not depend on model guesswork.
  • RAG: unstructured retrieval. Documents, manuals, and ticket history go through retrieval-augmented generation: retrieve the relevant passages first, then let the model answer from them, with citable sources. Microsoft's RAG technical overview states the pattern clearly: retrieval quality decides answer quality, and traceable citation is a baseline requirement.
  • Workflow: deterministic multi-step work. When the steps can be drawn in advance, run in fixed order, and branch only a little — use a workflow, not an agent. Every workflow step is predictable, testable, and auditable.
  • Agent: non-deterministic multi-step work. Only when the steps cannot be fixed in advance and the next step depends on judging intermediate results is an agent worth it. Anthropic's practical guide gives the test: the more a task needs autonomous decisions, the more an agent is worth; otherwise a workflow is cheaper and more reliable.

One hard rule spans all four layers: reads and writes run in separate lanes. Retrieval and answering can be open; writes and execution must pass permission checks and human approval — that is gated writing. An agent may propose, but it may not execute a write on its own.

Where this applies

This combination fits knowledge-intensive work: customer service, internal support, compliance queries, sales support — questions with checkable sources, answers that need citations, and actions that need approval. It does not fit control scenarios with extreme real-time demands, businesses with no accumulated knowledge to retrieve, or pure computation tasks (which do not need a language model at all).

Risks

  • Retrieval hallucination and distorted citation. A RAG answer may cite a passage that does not exist, or misstate what a passage says. The evaluation set must cover citation accuracy, not just answer fluency.
  • Stale knowledge. A retrieval system does not automatically know a document is outdated. The knowledge base needs a named update owner and an expiry cleanup mechanism.
  • Permission leakage. RAG can retrieve document content the user is not allowed to see and write it into an answer. Retrieval must be bound to the caller's permissions, enforced at the retrieval layer — not at the display layer.
  • Uncontrolled agent execution. Giving an agent write permission without an approval gate hands the system keys to a probabilistic system. Writes must be gated; this is a non-negotiable floor.
  • Over-orchestration. All four layers plus multi-agent coordination sounds advanced; in practice it complicates a simple problem. Get one layer running before deciding whether to add the next.

Deliverables checklist

  • Knowledge source inventory: sources, formats, update owners, confidentiality levels, permission mapping
  • Retrieval evaluation set: real question samples, expected citations, pass thresholds
  • Workflow diagrams: steps, branches, and exception handling for deterministic processes
  • Agent permission matrix: the read / write / needs-approval operation list
  • Approval rules: which writes need human approval, approval time limits, fallback process
  • Audit plan: how retrieval and execution records are stored and queried

Sources

Next step: if you are planning a knowledge base and agent combination, start with Negency enterprise AI agent development, consulting & systems integration — enterprise knowledge base & RAG, agent workflows, and MCP/API tool integration are part of its eleven capabilities; or take the AI scenario screening to confirm which service line fits your scenario.