Chatbot, AI Assistant or AI Agent? Choose the Smallest System That Solves the Job
2026-07-21 · Reviewed 2026-11-04

Chatbot, AI Assistant or AI Agent? Choose the Smallest System That Solves the Job

A chatbot answers through a conversational interface. An assistant helps a person prepare or interpret work. An agent can select steps and use authorised tools toward a goal. These are useful working definitions, not a universal certification scheme: products overlap and vendors use the terms differently. For a buyer, the more important question is what the system may actually read, decide and change.

Start from the job, not from the word agent

Write the desired outcome without naming a technology. For example: prepare a response using verified property facts, identify missing details and give a reviewer the source. That might need a template and a retrieval step rather than an autonomous agent. If input conditions and outputs are stable, a deterministic rule is often easier to inspect and test.

An assistant becomes useful when language varies, relevant information is scattered or a person needs a draft to review. It should identify uncertainty and show what supports its answer. An agent adds a different kind of responsibility when it chooses tools, performs successive actions or adapts its path. More autonomy increases the need to define boundaries; it does not remove operational ownership.

A decision table for one workflow

Working definitions and decision boundaries
RouteExampleImportant control
DECISIONUse the least autonomy that meets acceptanceA named owner and a measurable outcome
Rule or standard featureRoute a form with known fieldsValidate inputs and handle exceptions
Reviewed assistantDraft a sourced responseShow evidence and require the agreed review
Action-taking agentPlan authorised steps across toolsLimit tools, actions, budgets and recovery
Conceptual LEMAIT framework, not a measured performance comparison. A conversational interface does not reveal the permissions underneath it.

Consider the difference between suggesting a viewing slot and booking it. A booking needs current availability, a valid identity, the appropriate permission and a recorded confirmation. A confident sentence cannot substitute for those checks. Similarly, a summary is not authority to change a customer record, send a document or mark a payment as received.

Tool access is the business risk boundary

Inventory which systems the workflow reads and writes. Keep customer and tenant boundaries intact. Give the workflow only the permissions it needs; do not hand over a broad administrator account because setup is easier. Define which actions can run automatically, which require review and which are out of scope. Treat incoming documents and messages as untrusted data, not new operating instructions.

Retries also matter. A repeated request must not create duplicate appointments, records or messages. Record the action identifier, outcome and failure state, then decide when retry is safe. If the system cannot tell whether an action succeeded, it needs reconciliation rather than another blind attempt. A visible stop path and human escalation queue are features of a usable business system, not optional extras.

Evaluate the result rather than the conversation

Use examples that represent ordinary, incomplete, contradictory and adversarial inputs. Keep an evaluation sample separate from development. Measure accepted outcomes, severe errors, review minutes, cost and recovery. A fluent answer can be incorrect; a short chat can conceal an unresolved issue. Synthetic examples demonstrate specific controls, not production accuracy, customer satisfaction or revenue.

  • Inputs: authorised sources, an update owner and a clear response when information is missing or stale.
  • Actions: explicit permissions, approval gates, duplicate controls and an auditable recovery path.
  • Operations: a named reviewer, monitoring, change testing and a decision to pause when quality degrades.

Price delivery and operation separately

Do not buy an agent by counting how many virtual workers appear on a diagram. Ask what result is included, what evidence proves acceptance and what recurring operation costs. A funded milestone can make a bounded deliverable predictable; model usage, maintenance and new features still need explicit treatment. Saved time is released capacity until an economic benefit is demonstrated.

Use the free acceptance checklist to decide whether a rule, reviewed assistant or agent is justified. LEMAIT's diagnostic for one process is EUR 350–600; a bounded pilot starts at EUR 960 and broader implementation at EUR 3,000. Each agreed milestone is fully prepaid. Ongoing support is scoped after the actual workflow and responsibilities are known, without inventing an unlimited service promise.

Disclosure and context: these definitions and controls are LEMAIT's practical procurement framework, prepared with AI assistance and editorial review. France Num discusses starting from business tasks and connecting existing tools: https://www.francenum.gouv.fr/guides-et-conseils/intelligence-artificielle/comprendre-et-adopter-lia/lintelligence-artificielle-6 . This article does not claim that every business needs an agent, that the examples are deployed client systems or that a technology label establishes legal compliance.

Prepared with AI assistance and checked against the linked public sources by LEMAIT. Scenarios are illustrations, not customer results. No ranking, ROI or legal-compliance guarantee.

TRIQUETRA COMPANY W.L.L., Bahrain · LEMAIT · Privacy