Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

Moving From a Single Chatbot to a Governed AI Agent Layer

Step-by-step path to move a company from chatbot-only use to a governed AI agent layer that keeps conversation history, business rules and approval limits.

Be Human typographic cover with the title: Moving From a Single Chatbot to a Governed AI Agent Layer
Conceptual illustration created with AI · Be Human

Moving a company from chatbot-only use to AI agents means turning a conversation window into a governed layer: one defined role at a time, with its own memory, tools, approval limits and a success test you can observe in the business. Start by inventorying what the chatbot already holds, write down the business rules it has been applying by habit, then hand one narrow task to an agent under human review before widening access. The sequence below is the one a leader can follow, with the checks that keep history and rules intact along the way.

Key takeaways

  • A chatbot holds conversation history, but an agent needs that history sorted into records, rules and permissions.
  • Write down the business rules the chatbot has been applying by habit before any agent touches customer or operational data.
  • Start with one role, one set of tools and one success criterion that someone in the business can observe.
  • Keep approval boundaries explicit, and raise them only after evidence from real use.
  • Run the agent beside the chatbot until it passes its own review, then retire the chatbot task by task.

Why a chatbot is the wrong place to keep company knowledge

A chatbot is good at answering a question in the moment. It is weak at holding what the company has decided, because those decisions tend to live in other places: old conversation threads, prompts that staff copied from a shared document, and the memory of the two people who always know how a refund is handled. When the chatbot is the only layer, nobody can say which answer reflects current policy and which reflects a conversation from last year.

Moving to agents forces a clean separation. Conversation history records what was said. Business rules record what the company has decided. Permissions record what a system is allowed to do. Each needs its own owner, its own review and its own place in the architecture. Treat them as one pile and an agent will repeat a superseded rule with the same confidence as a current one, which is the failure you most want to avoid.

The migration path, step by step

  • List every place the chatbot is used, who uses it, and what it has seen, including uploaded files, saved threads and copied prompts.
  • Sort each item into one of four groups: conversation history, reusable business rule, customer or staff record, or disposable draft. Only the first three need retention decisions.
  • Rewrite each rule as a short statement that someone can test, with a named owner who can say when it last changed.
  • Choose one role for the first agent, such as answering internal policy questions for the operations team or preparing draft replies for human sign-off.
  • Define the tools it may call, the memory it may read, and the actions that always need a human yes.
  • Set one success criterion the business can observe, such as drafts accepted without rewriting in a weekly review, rather than a judgment about how fluent the output sounds.
  • Run the agent beside the chatbot with a small group of users, record every disagreement between the two, and widen access one step at a time.

Keep history, memory and business rules in separate places

Conversation history is evidence. It shows how people asked questions and where the old setup failed them. It should be kept under your own retention policy and privacy obligations, and it should not be treated as authority on what the company believes. Use it to find gaps, not as a source the agent quotes as policy.

Memory is curated. It is the set of facts an agent may retrieve when they are relevant, usually drawn from approved documents at the moment of the question rather than fixed into the model. Retrieval-augmented generation is the common pattern for this, and the AWS guidance comparing retrieval with fine-tuning explains why changing a document is usually easier to correct than retraining a model.

Business rules are the standing decisions: discount authority, escalation paths, what counts as a complete order. Store them as versioned documents with an owner. When a rule changes, the agent's memory changes with it, and the old version is archived rather than left buried in a thread. A useful test: if you cannot point to the document an answer came from, the answer is not yet governed.

Set permissions and approval boundaries before the first run

Start with read-only access. An agent that can read the order system and draft a reply carries very different risk from one that can send the reply, change the order or issue a credit. Grant write and send permissions only after the read-only version has produced reliable drafts for a sustained period, and tie each permission to a named tool with its own log.

Treat any text the agent reads from email, web pages or uploaded files as data, not as instructions. Documents can contain hidden or hostile instructions. The OWASP prompt injection prevention guidance describes the basic controls: separate untrusted content from instructions, limit what each tool can do, and require confirmation for consequential actions. The NIST AI Risk Management Framework offers a broader vocabulary for mapping these risks to owners and controls.

Least privilege applies to people as well. The person who approves an agent's action should have authority over that action in the business, not merely access to the screen where the approval appears. For the recurring review of who and what has access, see how to run a permission review for your AI agents, step by step.

A decision guide for the first agent (illustrative)

Illustrative scenario: a small equipment distributor, described here only as an example, has a chatbot that its sales team uses to draft replies to dealers. Replies often repeat warranty terms that have changed twice, and the sales lead has been pasting the current wording into each new thread. The team wants an agent, but the first question is which task is safe to hand over.

Use four questions to choose the first task:

  • Does the task repeat often enough that a review habit will form around it?
  • Does the answer depend on documents that change, so that memory and rules must be maintained?
  • Is a wrong output reversible before it reaches a customer or a system?
  • Can the result be checked against a record the business already keeps?

A task that answers yes to all four is a strong first candidate. In the distributor example, drafting warranty replies with a person sending them passes all four. Issuing credits automatically does not, because it fails the reversibility question, so it belongs in a later stage with stricter approval rules.

How Be Human's managed approach handles the migration

Be Human's first engagement defines one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion, so the migration starts from a written scope rather than from whatever the chatbot happened to be doing. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, all carried out remotely in English or Portuguese. Changes to memory, rules or permissions follow a written request and approval process, as described in how to request and approve changes to a managed AI agent, so that the agent's operating context is never edited informally.

Watch the agent so problems surface early

Without observation, an agent failure looks like a vague sense that answers have gotten worse. Record each run as a trace: which memory documents were retrieved, which tool was called and with what arguments, which approval was requested and who gave it, and what the user did with the output. The OpenTelemetry observability concepts describe the traces, metrics and logs that make this kind of record useful across systems.

Evaluations are the second tool. Keep a fixed set of realistic test cases drawn from real requests, include several where the correct answer is to escalate to a person, and rerun the set after every change to memory, rules or the underlying model. Log the results so a regression appears as a measured change rather than a complaint. Decide in advance what pauses the agent, for example a rule that any incident stops write actions until a reviewer has looked at the trace. Teams that use the agent daily also need a plan for trust, which is covered in how to get your team to actually use an AI agent.

If you are not sure which task should move first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Do we have to shut down the chatbot before starting with agents? No. Run the agent beside the chatbot for a defined period, log where their answers differ, and retire the chatbot only for the tasks the agent has passed. Keeping both avoids losing the thread history people rely on during the transition.

How do we keep old conversations without letting the agent treat them as policy? Store conversation history under your retention rules, and give the agent access to approved business documents rather than to past threads. Reviewers can search history to find gaps, but it should not be a retrieval source for answers about what the company has decided.

Who should approve an agent's actions? The approval should sit with the person who owns the business outcome, such as a customer service lead approving a credit, not with the person who built the system. Log each approval with the request it answered, so that the record shows who authorized what and when.

How do we know when the first agent is ready to do more? Widen its access only when the logs and the fixed evaluation set show that it behaves as expected on the current task. Review the evidence in a regular meeting, and treat a new permission as a change that goes through the same written approval as any other.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →