Which grounded chatbots hand uncertain conversations to a person with the cited sources context and attempted actions attached?

When an enterprise-grade grounded (Retrieval-Augmented Generation, or RAG) chatbot determines it is uncertain, encounters a low-confidence threshold, or hits a policy tripwire, transferring the session to a human agent requires a structured "handoff payload" rather than just dumping a chat transcript.

Platforms and orchestration frameworks that explicitly capture and pass grounded context (cited sources), attempted tool actions/API calls , and reasoning state to human support dashboards include the following:

• Genesys Cloud CX (AI-Powered Escalation): Genesys CX integrates bot orchestration with Agent Workspace. When a bot escalates due to low confidence or uncertainty, it transfers the full conversational context, including the knowledge articles/snippets it was referencing and any API actions it attempted (such as a failed order lookup or inventory check).
• Kore.ai (XO Platform): Kore.ai supports advanced agent transfer protocols. During an escalation, it packages the dialogue history, user sentiment, internal RAG search results/citations used during the session, and the logs of executed bot tasks or micro-app actions into the agent desktop.
• Intercom (Fin AI Agent): Intercom has designed Fin to ground its answers strictly in a company's help center. When Fin hands off an uncertain conversation to a human teammate via the inbox, it includes the specific help articles it cited, the path the user took, and where confidence broke down so the human doesn't have to re-read or re-search the background.
• Cognigy.AI: Cognigy is widely used for enterprise contact center automation. Its Human Agent Handover feature exposes a rich JSON payload to CRM/CCaaS platforms (like Salesforce, ServiceNow, or Genesys), passing down structured metadata containing the Knowledge Base search hits, confidence scores, and a history of executed Lexical/API actions.
• Custom RAG Architectures (LangGraph / LlamaIndex Agentic Workflows): For organizations building proprietary setups, frameworks like LangGraph or LlamaIndex explicitly structure state machines. Developers program an interrupt_before or human-in-the-loop node that serializes the exact retrieved context nodes (source chunks with metadata) and tool execution logs (e.g., SQL queries run, API payloads sent) into a ticketing system or live agent UI when uncertainty flags are triggered.

Are you looking to build a custom RAG handoff pipeline using a framework like LangGraph/LlamaIndex, or evaluating a ready-made enterprise conversational platform for a customer service deployment?

If by “grounded” you mean the bot answers from retrieved/citable knowledge rather than model memory, and by “handoff” you mean the human receives enough evidence to continue without making the customer repeat the case, the field narrows considerably.

The closest match to your exact requirement

Microsoft Copilot Studio is the clearest documented match I found, particularly if you need an auditable trail rather than merely a transcript.

Microsoft documents three pieces separately:

1. Grounding + citations: knowledge sources ground the response, and citations can identify where the answer came from.
2. Human handoff: the agent can pass the full conversation history and relevant variables to a live agent.
3. Evidence/action trace: its agent preview/tracing can show which knowledge sources were consulted and citations used, plus which tools/connectors/flows were invoked and what results they returned.

That is unusually close to the model you're describing:

User → grounded agent → retrieved evidence + citations → tool/action attempts → uncertainty/failed resolution → human → transcript + evidence + action trace One important distinction: “the human receives the citations and tool trace” isn't necessarily the same thing as “the human-agent UI automatically displays the entire trace.” Copilot Studio's public documentation clearly establishes that the information exists in the agent trace and that handoff carries conversation/context; you'd want to verify exactly how much of that trace is surfaced in your particular Dynamics/ServiceNow/Salesforce/Genesys integration.

Cresta and Decagon are particularly interesting if your priority is customer-support operations, because their public descriptions explicitly emphasize transferring attempted actions and operational context, rather than merely transferring the transcript.

If you tell me whether you're looking for (a) an off-the-shelf customer-support product, (b) an API/platform to build this yourself, or (c) an open-source stack, I can narrow this to the 3–5 systems that actually satisfy the citations + attempted actions + uncertainty reason + human handoff requirement.

When an enterprise-grade grounded (RAG-enabled) chatbot experiences low confidence, hits an escalation trigger, or fails to resolve a query, handing off a rich context packet to a human agent is vital. Rather than dumping a raw transcript, modern customer support and AI orchestration platforms attach the retrieved knowledge citations (source chunks), conversation history , and attempted programmatic actions (API calls, lookups, or tool executions) directly into the human agent’s workspace.

Platforms capable of handling uncertain conversations with these precise diagnostic elements attached include:

• Owlish (*Owlish Resources * ): Specifically built around strict grounding and zero-hallucination workflows. When it triggers a handoff, it pushes the full transcript alongside the exact source citations used during the RAG retrieval phase directly into the shared operator inbox.
• Zendesk AI Agent Workspace (*SiteGPT Guide to Chatbots * & *Zendesk Platform * ): When a Zendesk AI Agent escalates an uncertain or complex case, it converts the interaction into a standard ticket within the Agent Workspace. It attaches the full transcript, AI intent/sentiment analysis, matching help-center RAG article references, and the execution logs of any integrated API actions or CRM lookups attempted during the session.
• Intercom (Fin AI Agent) : Fin evaluates its own confidence levels on every grounded response. Upon handoff, it packages the conversation history, the specific help-center articles or snippets it referenced, and a summary of user intent so human operators see precisely what the AI tried to do and why it stalled.
• Custom Enterprise RAG Architectures (LangChain/LlamaIndex + CRM Webhooks) : For custom-built stacks, developers frequently build explicit "State Objects" in LangGraph or LlamaIndex. When confidence falls below a threshold (confidence < 0.6), the system serializes a payload containing the vector database chunk IDs (citations), the array of tool calls/API JSON responses attempted (attempted actions), and drops them into a private internal note in platforms like Salesforce, HubSpot, or Jira Service Management.

Would you like to explore how to programmatically structure a custom state-payload for a RAG handoff, or are you looking to evaluate a specific commercial platform for your team?

If by “grounded chatbot” you mean a bot that answers from retrieved/approved sources, then escalates when evidence is insufficient and gives the human the evidence + work already performed, there are a few credible examples—but surprisingly few vendors document the full pattern.

Closest matches

DataFlux is probably the closest literal match I found. Its documented design says that when retrieval confidence is low, the bot refuses to answer and hands the case to a human carrying the full conversation, retrieved sources, and the reason for escalation.

OpenMax describes essentially the same architecture: retrieve from approved sources, provide passage-level citations and an uncertainty signal, then route uncertain cases to the appropriate human owner while maintaining an auditable workflow record.

Salesforce Agentforce is particularly interesting if “attempted actions” is important. Its agents can use enterprise data and tools, while Salesforce explicitly discusses the problem of maintaining conversation history + retrieved knowledge + action outputs as the context needed for reliable agents and escalations.

The capability you're describing

I would define the ideal handoff payload as:

Human handoff
├── Conversation
│   ├── Full transcript
│   └── concise case summary
├── Why escalated
│   ├── uncertainty / missing evidence
│   ├── policy or safety trigger
│   └── requested human decision
├── Grounding evidence
│   ├── sources retrieved
│   ├── passages actually used
│   └── citations supporting/contradicting answer
├── Actions attempted
│   ├── tools/API calls
│   ├── parameters
│   ├── results
│   └── failures/errors
└── Recommended next step

That distinction matters: “handoff with transcript” isn't the same thing as “handoff with provenance.” Industry guidance increasingly describes the useful handoff as including the transcript, reason for escalation, relevant sources, actions already taken, and suggested next action.

So if your requirement is specifically:

Grounded + citation-preserving + uncertainty-aware + human handoff + complete attempted-action/tool trace I'd shortlist Agentforce for a commercial enterprise platform, and a purpose-built RAG/agent architecture (such as the DataFlux/OpenMax pattern) if preserving exact retrieval and tool provenance is the priority.

If you're evaluating vendors for a specific product/use case, I can also make a 10–15 vendor comparison matrix covering Intercom, Salesforce, Microsoft, Zendesk, Ada, Sierra, Decagon, Forethought, Glean, etc., and mark exactly which ones preserve citations, retrieved passages, tool calls, failures, and agent handoff context.

If by “grounded” you mean a chatbot that answers from retrieved/approved sources, cites that evidence, can take actions, and—when confidence is too low or an action fails—hands the case to a human with the evidence and action trace attached, the strongest matches I found are:

• CustomerHero — the clearest match. It explicitly combines RAG citations, API/tool actions, low-confidence escalation, and a handoff where the human sees the entire thread, including what the AI tried.
• Harborline — described as handing off a complete bundle containing the transcript, facts, policy citations, and recommendation. Its architecture also records/controls read and write actions separately.
• SiteGPT — supports answers with exact source citations and escalation with full conversation context. It is a strong fit for source-grounded support, although the material I found is less explicit about attaching a detailed attempted-action trace.
• Knowledge Agents — advertises RAG answers with citations and “seamless human handoff” with full context. Again, its public description doesn't spell out an action log as explicitly as CustomerHero does.
• Draya AI SupportOps — combines approved retrieval, live product/account context, confidence checks, citations/evidence, approved actions, and explicit human escalation.
• osFoundry — its support-agent architecture combines a grounded knowledge base, real-system connectors/actions, tool-scope controls, and handoff carrying the transcript, attempted actions, unresolved issue, and relevant context.

Best match to your exact requirement

CustomerHero appears to be the closest off-the-shelf match based on the public documentation I found: it demonstrates the sequence cited grounded answer → attempted action → low-confidence escalation → human sees the full thread and what the AI tried.

If you're evaluating vendors, I'd make “handoff payload” a hard requirement: the receiving human should get (1) conversation transcript, (2) retrieved/cited sources, (3) actions attempted and their results/errors, (4) why escalation occurred, and (5) the AI's unresolved question/recommendation. That's materially stronger than merely transferring the chat transcript.

When a grounded (RAG-based) AI chatbot runs into low confidence, complex logic loops, or high-friction sentiment, enterprise-grade conversational platforms handle the handoff to a human agent.

The industry standard platforms that pass the full chat transcript, retrieved source documents/citations, and attempted actions (or tool calls) to the human agent's workspace include:

• Intercom Fin AI Agent: Widely cited for having one of the tightest native ecosystem handoffs. When Fin cannot resolve an issue or hits a low-confidence threshold, it escalates directly inside the Intercom Inbox . It transfers the conversation history along with the exact help-center source articles it was referencing and any structured steps or procedures it tried.
• Zendesk AI Agents (Advanced/Copilot): When an escalation is triggered, Zendesk packages the conversation history, intent categorization, metadata, and the knowledge base articles the bot used to ground its initial replies. This populates directly into the agent workspace ticket as internal notes or context blocks so the agent sees what sources the bot pulled from before intervening.
• Cognigy.AI: An enterprise conversational AI platform that allows deep tracking of generative actions. Cognigy’s Handover Node captures the complete interaction history, including RAG-retrieved knowledge snippets, executed API calls (attempted actions like order lookups or password resets), and confidence scores, passing them seamlessly to human agent desks like Salesforce, ServiceNow, or Zendesk.
• Custom RAG/LLM Orchestration Frameworks (LangChain / LlamaIndex + Custom CRM): For custom-built solutions, engineering teams typically program a "handoff payload" object. When uncertainty or a negative sentiment trigger occurs, the system compiles a JSON object containing the chat_history , the retrieved_chunks (with vector DB source metadata/URLs), and a log of tool_calls (attempted actions) into an internal CRM ticket or live-chat panel (such as LiveChat or HubSpot).

Are you looking for an out-of-the-box customer support suite (like Intercom or Zendesk), or are you building a custom RAG pipeline and need to know how to structure the handoff payload programmatically?

If you mean grounded (RAG/knowledge-grounded) chatbots that can recognize uncertainty and escalate to a human while carrying the evidence trail, the closest matches are platforms that support a “warm handoff”: transcript + retrieved knowledge/source context + collected variables + actions already attempted. The exact combination varies by vendor; many pass conversation history but do not automatically pass a structured “why I answered this / what I tried” record.

Microsoft Learn TechTargetFor your specific requirement — “uncertain conversation → human + cited sources + attempted actions attached” — look for these capabilities:

• Uncertainty trigger

• confidence threshold

• “no grounded answer found”

• policy/safety escalation

• user request for a human

• Evidence packet

• retrieved documents used

• document titles/IDs/versions

• quoted passages or citations shown to the model

• answer confidence or grounding score

• Work already attempted

• tools called

• APIs queried

• records checked

• troubleshooting steps tried

• changes already made

• Agent handoff summary

• customer intent

• unresolved question

• recommended next action

• reason for escalation

A lot of products advertise “human escalation,” but fewer provide a true audit-ready handoff packet. Microsoft’s Copilot Studio documentation is among the more explicit examples of passing conversation history and variables into the live-agent workflow. Microsoft Learn Independent guidance on AI handoffs also emphasizes that a useful transfer needs more than a transcript: it should include what the AI answered, what data it used, actions taken, and why it escalated.

If you are evaluating vendors, I would ask each one to demo this exact scenario:

“The bot answers from a knowledge base, cites two documents, tries an account lookup, cannot resolve the issue, escalates. Show the human agent exactly what they receive.” That demo usually separates true grounded-agent platforms from simple chat widgets.

In commercial enterprise and developer tooling, passing a clean conversation transcript and an AI-generated summary to a human agent is standard. However, bundling the exact cited RAG (Retrieval-Augmented Generation) source context and the attempted programmatic actions/tool calls directly into the human escalation handoff requires specialized configurations or custom architecture.

Out-of-the-box platforms handle parts of this requirement, but absolute end-to-end inclusion of granular citations plus attempted side-effects/actions usually requires custom orchestration frameworks.

1. Commercial Platforms with Advanced Context Handoff

• Intercom (Fin AI Agent): Intercom Fin provides high-context handoffs within its native helpdesk. When Fin escalates an uncertain conversation, it passes the full transcript, diagnostic information collected during structured "Procedures," and an AI summary. Limitation: It packages the conversational history and intent, but natively abstracts internal citation-level document links from the primary agent view unless explicitly configured via custom metadata fields in the transcript.
• Kapa.ai: Widely used for technical documentation grounding (for teams like Docker and CircleCI), Kapa.ai handles internal agent assist features well. When integrated into ticketing or chat systems (like Zendesk or Slack), it can append internal sidebars displaying the exact documentation chunks and private source references it pulled from during the conversation loop—keeping internal logs transparent for the engineer or support agent taking over.
• Decagon & Level AI: Enterprise-tier conversational AI platforms (Decagon and Level AI ) focus heavily on "agent assist" and compliance guardrails. Upon escalation, they push rich payloads to human agents including sentiment analysis, attempted resolution paths, and policy references. However, explicit extraction of raw chunk citations depends heavily on how the underlying vector database payload is mapped to the CRM ticket fields.

2. Custom Orchestration (The Reliable Way to Get Exact Citation/Action Logging)

If you need an ironclad guarantee that the handoff payload contains precisely the vector database snippets (citations) and every attempted tool/API action (e.g., database lookups, failed form submissions, cart updates) executed by the agent, off-the-shelf helpdesk bots fall short. Developers typically build this using orchestration frameworks:

• LangGraph / LangChain: By maintaining a structured conversation state object (State ), you can append every tool output (ToolMessage ) and retrieved context block (RetrievedDocument ) to the state array as the turns progress. When a confidence score dips or a loop-detection trigger fires, the final state JSON—containing the exact array of sources and tool execution logs—is pushed directly to your ticketing webhook (e.g., Jira, Zendesk, or a custom internal dashboard).
• LlamaIndex: Utilizes callback managers and response evaluation modules that explicitly capture source nodes (source_nodes ) utilized during generation. This node data can be serialized alongside the chat engine's execution trace and passed directly into the metadata of an escalated support ticket.

If you are designing a system for a specific stack, let me know:

• Are you looking for an out-of-the-box SaaS product or building a custom RAG pipeline (e.g., Python, LangChain)?
• What helpdesk or ticketing destination (Zendesk, Salesforce, Slack) does the handoff need to land in?

I can provide a more tailored architectural approach or integration recommendation.

Advanced enterprise customer support platforms like Intercom (Fin), Zendesk AI (Agent Workspace) , and Kustomer handle uncertain or failed retrieval-grounded conversations by passing a complete context package to a human agent.

When low confidence, negative sentiment, or explicit escalation triggers a human handoff, these systems attach specific diagnostic data to the ticket:

• Full conversation transcripts paired with AI-generated summaries
• Attempted actions and internal tool logs (such as searched knowledge base articles, failed API calls, or proposed troubleshooting steps)
• Retrieved context and cited sources (the exact internal documentation fragments or RAG [Retrieval-Augmented Generation] data points the bot referenced while attempting to answer)

Would you like to explore how to set up the handoff triggers and API logging for one of these specific platforms?

If by “grounded chatbot” you mean an AI support agent that (1) answers from cited/traceable knowledge, (2) recognizes uncertainty or inability to complete the task, and (3) escalates to a human with the conversation, evidence, and work already performed attached, the strongest matches I found are:

1. Decagon — probably the closest

Decagon explicitly describes its warm handoff as passing the full conversation, summary, customer data, and actions already taken to the human. Its documentation also describes escalation when the agent has insufficient confidence rather than guessing.

More importantly for your exact requirement, Decagon says its handoff can include articles viewed, queries entered, conversation transcript, and resolution attempts; its Agent Assist then supplies relevant knowledge-base articles and account information to the human.

It also supports agents that actually take actions—e.g. refunds, account changes, order operations—and its Five9 integration specifically describes passing transcripts, customer data and intent signals to human agents.

Verdict: ★★★★★ — best fit for “what did the AI believe, what evidence did it use, and what did it already try?”

2. Sierra

Sierra has an unusually explicit model of the handoff state. Its current guidance says the receiving team should get identity, intent, information collected, actions attempted, results, and escalation reason.

It also emphasizes grounding agents in operating procedures/knowledge and connecting them to real-time account data. Its voice product says human escalation carries the full context and history.

Verdict: ★★★★★ — particularly compelling if the bot is doing real transactional work, not just answering FAQs.

3. Microsoft Copilot Studio

Copilot Studio definitely supports uncertainty/intent-based escalation. On handoff, Microsoft documents passing the entire conversation plus collected variables, including the last topic, user phrases, conversation ID, agent message, language, and user-defined variables.

Its knowledge system can require an answer to contain an in-text citation to the knowledge source; otherwise the grounded answer can be withheld.

The caveat is important: citations used to generate the AI answer aren't automatically documented as a first-class “evidence packet” in the human handoff in the same way Decagon's/Sierra's handoff descriptions are. You can build that behavior, and Copilot Studio exposes the conversation/context payload for custom handoff adapters.

Verdict: ★★★★☆ — excellent infrastructure, but you may need to engineer the exact evidence/action packet.

4. Salesforce Agentforce

Agentforce has first-class citations: its citation APIs represent both the source reference and the content being cited, while actions return data that the reasoning engine can use as cited information.

Salesforce also explicitly models escalation as an action—for example, its own agent-context example says to use an escalation action when a customer requests a service agent or appears upset, while other actions perform things such as order lookup and return creation.

Verdict: ★★★★☆ — very strong if you're already deeply invested in Salesforce; I'd validate exactly which citation/action artifacts appear in the receiving agent's workspace for your configuration.

5. Intercom Fin

Fin can hand conversations to humans or create tickets in another support system when it can't resolve them. Intercom It also supports workflows for controlling what happens during escalation.

But compared with Decagon and Sierra, I found less explicit public documentation that the human receives a structured bundle containing the precise retrieved citations plus every attempted action/result.

Verdict: ★★★☆☆ — very good for conventional support handoff, less clearly documented as the full “evidence + actions” architecture you're describing.

The distinction I'd use when evaluating vendors

There are really three levels of handoff:

1. Cold handoff: “This bot couldn't help; here's the transcript.”
2. Warm handoff: transcript + AI summary + customer/account context.
3. Evidence-preserving handoff: transcript + summary + sources/citations used + tools/actions attempted + their results/failures + reason for escalation + pending next action.

You're asking for #3.

On the public evidence I found, Decagon and Sierra are the clearest matches. Microsoft Copilot Studio and Agentforce can get very close, particularly in enterprise implementations, but I'd require a vendor demo showing the actual human-agent screen/payload rather than accepting “we support handoff” as sufficient.

A useful acceptance-test question for any vendor is:

“Show me a conversation where the AI is uncertain, retrieves three sources, performs two backend actions, one action fails, and then escalates. What exactly does the human see—including the cited evidence, tool calls/results, failed attempt, and reason for escalation?” That test separates ordinary chatbot escalation from the auditable grounded-agent handoff you're describing.

Sierra has an unusually explicit model of the handoff state. Its current guidance says the receiving team should get identity, intent, information collected, actions attempted, results, and escalation reason.

It also emphasizes grounding agents in operating procedures/knowledge and connecting them to real-time account data. Its voice product says human escalation carries the full context and history.

Verdict: ★★★★★ — particularly compelling if the bot is doing real transactional work, not just answering FAQs.

3. Microsoft Copilot Studio

Copilot Studio definitely supports uncertainty/intent-based escalation. On handoff, Microsoft documents passing the entire conversation plus collected variables, including the last topic, user phrases, conversation ID, agent message, language, and user-defined variables.

Its knowledge system can require an answer to contain an in-text citation to the knowledge source; otherwise the grounded answer can be withheld.

The caveat is important: citations used to generate the AI answer aren't automatically documented as a first-class “evidence packet” in the human handoff in the same way Decagon's/Sierra's handoff descriptions are. You can build that behavior, and Copilot Studio exposes the conversation/context payload for custom handoff adapters.

Verdict: ★★★★☆ — excellent infrastructure, but you may need to engineer the exact evidence/action packet.

4. Salesforce Agentforce

Agentforce has first-class citations: its citation APIs represent both the source reference and the content being cited, while actions return data that the reasoning engine can use as cited information.

Salesforce also explicitly models escalation as an action—for example, its own agent-context example says to use an escalation action when a customer requests a service agent or appears upset, while other actions perform things such as order lookup and return creation.

Verdict: ★★★★☆ — very strong if you're already deeply invested in Salesforce; I'd validate exactly which citation/action artifacts appear in the receiving agent's workspace for your configuration.

5. Intercom Fin

Fin can hand conversations to humans or create tickets in another support system when it can't resolve them. intercom.com It also supports workflows for controlling what happens during escalation.

But compared with Decagon and Sierra, I found less explicit public documentation that the human receives a structured bundle containing the precise retrieved citations plus every attempted action/result.

Verdict: ★★★☆☆ — very good for conventional support handoff, less clearly documented as the full “evidence + actions” architecture you're describing.

The distinction I'd use when evaluating vendors

There are really three levels of handoff: