Why do some vendors call themselves AI agent platforms when they're just chatbots?

Some vendors call chatbots "agent platforms" because agents are what buyers now budget for, the word has no agreed definition, and a language model that calls one API looks agentic in a demo. The difference shows when you ask to see one action proposed, approved by a person, run on a real system and found in an audit record, under a budget cap. A real platform shows that live in minutes. A chatbot wrapper changes the subject.
What exactly is a chatbot wrapper?
A chatbot wrapper is a thin layer over a language model. It takes a message, returns text and stops. It may have a nice interface, a knowledge base and even one or two integrations, but it does not:
change records in your systems as part of a defined process;
remember what happened last time in a way other agents can use;
run under its own permissions and budget;
hand work to another agent without losing context;
leave a record of what it did and who approved it.
An agent platform does all five. The gap is architectural, not cosmetic. For what the five look like in one product, see Nagent's platform, covered at the end.
Why do vendors relabel chatbots as agent platforms?
Six causes, roughly in order of how often you will meet them:
The budget moved. Buyers now have line items for "AI agents". A product that does not use the word does not make the shortlist. The practice is sometimes called agent washing.
There is no fixed definition. "Agent" can mean anything from a prompt with a tool call to a governed fleet acting on live systems. Vendors pick the definition that fits their product.
Demos flatter thin products. A model that calls one API to look up an order looks like an agent for the length of a demo. The missing pieces (permissions, approvals, audit, failure handling) only show up in production.
The roadmap is sold as the product. Approval flows, audit export and budget caps are "coming next quarter". A control plane that does not exist yet is not a control plane.
Chatbots genuinely got better. Language models made old chatbots far more capable at conversation. Relabelling them feels honest to the vendor, even when nothing can act.
Funding and analyst pressure. Category language follows money. When "agentic" is the category investors and analysts track, product pages follow.
None of this makes the vendor dishonest by intent. It does make your evaluation harder.
What does it cost you to buy a chatbot by mistake?
No time saved. People still copy answers into the CRM, the ad account and the CMS. You paid for an agent and got a drafting tool.
Hidden risk. If the product can act but has no approval model, the first mistake happens in public.
No audit trail. When a client, regulator or board asks what the AI did, there is no record to show.
Lock-in to a rebuild. When you need real agents later, nothing carries over: no permissions model, no history, no learning.
What are the red flags in a demo?
The sandbox demo. The agent "acts" only in a demo environment. Ask to see it on a connected test account of yours.
The roadmap answer. "Approvals are coming." "Audit export is on the roadmap." Note every one.
The single API call. It looks up one thing and replies. Ask what happens when the task needs three steps across two systems.
The autonomy boast. "It runs everything without anyone checking." Full autonomy with no controls is a liability, not a feature.
No answer on permissions. Ask whose permissions the agent uses. If nobody knows, nobody designed it.
Pricing per seat only. Not wrong on its own, but agent platforms usually meter work: actions, credits, conversations or outcomes.
How do you test a vendor in 30 minutes?
Ask for these six things live, in order. A real platform shows all six. A wrapper deflects from the second or third.
Show an action being proposed. Not a chat reply: a concrete change, such as a budget update, a CRM record change or an email to a contact.
Show the approval step. A person approves, edits or rejects it before it runs, and the action actually waits.
Show it running on a real system. The connected account changes.
Show the audit record. What was proposed, who approved it, what ran and when, readable without the vendor's help.
Show the limits. The agent's budget cap, its tool grants and what happens when it hits the cap.
Show a handoff. One agent passes work to another. Whose permissions does the second agent use?
Score what you saw. Six out of six is a platform. Four or five is a platform with gaps: get dates for the missing pieces in writing before you sign. Three or fewer is a chatbot or a workflow tool, which may still be useful, but should be priced and governed as one.
Which vendors publish their controls openly?
A useful signal is whether a vendor documents its controls in public. Some examples, checked on 24 October 2026:
Relevance AI lists role-based access control, human-in-the-loop approvals, version control, audit logs and full agent tracing on its own site (Relevance AI).
WRITER AI Studio describes policies set once and enforced at runtime, activity logs exportable to a SIEM, and permissioned access to connected systems (WRITER).
Salesforce Agentforce documents a low-code Agent Builder for defining subagents, instructions and actions, and agents that act within set guardrails (Salesforce).
Sierra describes agents that connect to systems of record to complete tasks such as returns and claims, with testing, monitoring and experiments (Sierra).
Public documentation is not proof. It does tell you what to ask to see.
How do you fix a bad purchase already made?
If you suspect you bought a chatbot labelled as an agent:
Run the six-step test on the product you have. Write down what it cannot show.
Narrow its job to what it does well, usually drafting and answering, and stop expecting actions.
Move live actions back to people until a platform with approvals is in place.
Check the contract for audit export, data use and exit terms.
Pilot a real platform on one workflow with live actions, and compare the record after two weeks.
How does Nagent answer the test?
We make Nagent, so judge this section as a vendor's own claims, then ask us to show them live. Nagent is multiplayer AI for growth: people and AI agents in one shared context, delivering the number. See the platform.
Proposed action: agents propose concrete actions on growth systems. NIA, for example, proposes changes to live Google Ads campaigns.
Approval: anything above an agent's level waits on a card for a person, in the team room or the Inbox. Every agent arrives at L2 Assist on the Earned Autonomy Ladder; below L5 Lead, every public post and reply waits for a named person; deleting and governance changes always wait.
Running on a real system: NIA's live changes run on the connected ad account after approval; campaigns arrive paused.
Audit record: every agent decision and action is written to an append-only record the customer can read and export.
Limits: each agent has its own budget in credits and its own tool grants; a breach stops the agent, and so can a person.
Handoff: when MIRA hands work to another agent, it runs under the receiving agent's own level and budget, and that agent cannot hand it on again.
What Nagent does not do: run customer support conversations in its core plans (customer experience is Convexa, a separate product), post to social media (MOXA is coming soon) or act as a general-purpose assistant. Some seats, such as HOOK and Echo, are in Beta.
For more on the category, read what an AI agent platform is and how it differs from AI tools.
Frequently Asked Questions
What is the core difference between an AI agent platform and a chatbot?
A chatbot returns text. An agent platform runs agents that act on real systems under permissions, budgets and approvals, share memory and leave an audit trail.
What is agent washing?
Relabelling a chatbot or simple automation as an AI agent or agent platform without the capabilities that make agents useful and safe, such as acting on systems, approvals and audit.
Is Sierra a chatbot wrapper?
No. By its own description, Sierra builds customer agents that connect to systems of record and complete tasks such as returns and claims, with testing and monitoring. It is an agent platform focused on customer experience.
Do AI agents need human approval before they act?
On a well-designed platform, it depends on the agent's level and the action's risk. New agents should wait for approval on anything that spends, publishes or contacts customers.
How long should an agent platform pilot run?
Two to four weeks on one real workflow with live actions is usually enough to see whether approvals, audit and limits hold up.
Run the six-step test on Nagent with your own workflow: book a demo.