The problem with "chatbot" as a category
The word "chatbot" covers an enormous range of capabilities - from a simple FAQ widget matching keywords to pre-written answers, to an agentic AI system executing complex multi-step workflows end-to-end. Treating these as the same category makes evaluation impossible and sets the wrong expectations.
The distinction worth making: reactive language models (what most people mean by "chatbot") vs agentic workflow executors (what Eshal means by "AI concierge").
| Dimension | Chatbot | AI Concierge |
|---|---|---|
| Primary function | Answer questions from KB | Execute customer workflows end-to-end |
| System integration | Reads from FAQ / knowledge base | Reads and writes to live CRM, OMS, TMS, ERP |
| Action capability | Provides information; human acts | Creates orders, processes returns, books appointments autonomously |
| Resolution definition | Customer received an answer | Customer's request was completed |
| Typical resolution rate | 40–60% | 80–90% |
Execution vs answering: a concrete example
A customer contacts support to return an item.
Chatbot: Explains the returns policy → provides the returns portal URL → customer escalates because they can't find their order number → agent spends 8 minutes resolving it.
AI Concierge: Identifies the customer, pulls their order, checks return eligibility, initiates the return in the OMS, generates a return label, sends it to their email - in 90 seconds, no human involved.
The chatbot answered every question correctly. The concierge completed the customer's objective.
Dynamic Action Gating - the key safety mechanism
The natural concern with an AI that can execute actions: what happens when it does something it shouldn't? Dynamic Action Gating is the answer.
- Fully Automated - tracking queries, return labels (under threshold), appointment booking in available slots, invoice resending. Low-risk, high-volume, easily reversible.
- Human Approval Required - refunds above threshold, account closures, price overrides, KYC completion in banking. AI prepares and queues; human authorises.
- Never Permitted - pricing rule changes, access to other customers' data, any action requiring human judgment under compliance rules.
This three-tier architecture is what allows regulated businesses in banking, healthcare, and government to deploy powerful agentic AI while maintaining control and compliance.
When does a simpler chatbot make sense?
To be fair: if your customer service need is genuinely limited to answering questions from a static knowledge base - product docs, basic FAQs, opening hours - a simpler chatbot is cheaper and adequate.
The test: look at your support queue. If more than 30% of contacts require an agent to take an action in a system (not just send information), you need execution capability - not just answering. An AI concierge will deliver significantly higher ROI.
See execution in action - not just answeringWe'll demonstrate the concierge completing a real workflow for your use case.