AI data market
Human-agent handoff data: the opportunity sits in the outcome gap
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
The human agent handoff data opportunity is narrower than it looks. Agent stacks now standardize the pause and the person's reply: MCP's elicitation, A2A's input-required state, and approval and interrupt tools in OpenAI's Agents SDK and LangGraph. None of the four files SourceX read defines the business outcome after work resumes. The scarce asset is past handoffs joined to verified outcomes, which live in business records.
Agent run
Human-agent handoff data
Key takeaways
- A handoff has three moments that need data, the escalation, the person's correction and the agent's resumption, and each is useful only with the outcome attached.
- MCP (specification 2025-06-18), A2A 1.0.0, OpenAI's Agents SDK and LangGraph all record a pause and a human reply; none of the files reviewed defines the eventual business outcome.
- Logs from new runs will be plentiful; historical handoffs joined to verified outcomes will not, and that join is where a supplier could earn a premium.
- Demand signals are dated forecasts and surveys, not purchases: Gartner's 2029 autonomy forecast for common service issues and its August 2026 finding that 87% of customers want a route to a person.
- The thesis fails if episodes cannot be replayed to a checkable outcome, if platforms keep the logs, or if simulated handoffs prove good enough.
Is human-agent handoff data an opportunity separate from agent data in general?
Yes, as a thesis to test rather than a fact: a handoff is the point where an agent stops, a person decides and the agent carries on, and training or scoring that behavior needs records of all three steps plus what happened afterwards. Ordinary trajectories show actions. Handoff data shows judgment at a boundary, the moment an agent should not continue alone.
Three moments define the need. Escalation: the state when the agent stopped and the reason it stopped. Correction: what the person changed, rejected or added, and on whose authority. Resumption: the state the agent picked up from, and whether it carried the correction through. Each is useful only with the outcome attached, because a handoff that looked tidy can still end in the wrong refund or a reopened ticket.
An illustrative case shows the pieces. A support agent drafting a refund stops because the order is past its return window; a team lead approves an exception for a long-standing account and halves the amount; the agent resumes, issues the partial refund and closes the ticket; two weeks later the customer disputes the charge anyway. Only the last fact says whether the handoff worked, and it lives in the payments system, not in the agent's log.
The investor question is whether that need creates an asset someone can own. The test used here has three steps: list the inputs a handoff episode requires, check which ones agent tooling now captures on its own, and find the inputs only business records hold. Whatever lands in that last group is the candidate scarce supply; everything else is likely to become a byproduct of running agents.
What four agent stacks record when an agent stops for a person
The pause itself is no longer scarce. SourceX read the current specification or documentation files of two agent protocols and two agent frameworks on October 27, 2026, with the comparison criteria fixed before reading: how work stops, what form the person's answer takes, how work resumes, and whether anything records the outcome after the handoff.
The differences matter for anyone hoping to sell or buy correction data. MCP's three responses separate an explicit no from a dismissal, a distinction a training label needs. OpenAI's SDK reduces the person's answer to approve or reject plus an optional message, so an edit arrives as text that someone must parse. LangGraph accepts any value unless the developer supplies an optional response_schema, which types the reply and validates it when the schema is a Pydantic model, TypedDict or dataclass, so the structure of a correction still depends on the developer who wrote the graph.
Each stack defines a way for work to stop for a human and continue, and each turns the person's answer into structured data a developer can log. None of the four files defines a field or state for whether the business result of the handed-off work turned out right:
| Stack and file read | How work stops | What the person's answer becomes | How work resumes | Outcome after the handoff |
|---|---|---|---|---|
| MCP specification 2025-06-18, elicitation (new in that version) | A server requests information from the user through the client | Accept with schema-checked content, decline or cancel | Inside the same server interaction | None in the file read |
| A2A specification, latest released version 1.0.0 | The task enters the input-required state (auth-required when authorization is needed) | A new message sent on the same task | Client continues with the same taskId and contextId | The agent's own terminal state, such as completed or failed |
| OpenAI Agents SDK, human-in-the-loop guide | A tool marked needs_approval raises an interruption | Approve or reject, with an optional rejection message | RunState is serialized, the decision recorded, the run resumed | None in the file read |
| LangGraph, interrupt in langgraph.types | interrupt() raises a resumable GraphInterrupt and surfaces a value | The value the client passes back with Command, optionally typed by a developer-supplied response_schema and validated against it when that schema is a Pydantic model, TypedDict or dataclass | The node re-runs from its start; a checkpointer is required | None in the file read |
The handoff input map: what tooling captures and what only records hold
Set the four stacks against what an episode needs and the gap is specific. Tooling captures the pause, the reply and the resume state for new runs. The reason a case needed a person, the authority behind a correction and the verified outcome sit in business systems such as ticket histories, approval logs and refund ledgers, and those systems also hold the years of handoffs that happened before any agent was deployed.
The map is SourceX's framework. Capture ratings come from the four files above; the last column names where such history typically lives, not any company's records:
| Input | Moment | Captured by agent tooling? | Where history usually lives |
|---|---|---|---|
| State when the agent stopped | Escalation | Yes for new runs: checkpoints and saved run state | Ticket fields at transfer, often partial |
| Reason for stopping | Escalation | Only if the developer writes it into the request | Escalation codes and transfer notes |
| The person's decision and authority | Correction | Decision yes (approve, reject, accept); authority no | Approval logs and role or permission records |
| What the person changed | Correction | As a free-form reply or rejection message | Edited orders, overridden fields, comments |
| State at resumption | Resumption | Yes: saved state or the same task identifiers | Rarely recorded; the case simply continues |
| Verified outcome | After the handoff | No | Refunds, reopen events, chargebacks, audits |
Where value would collect along the handoff data chain
Value tends to collect at the hardest link. The chain runs from companies whose operations already pass cases between automated steps and people, through rights review and preparation, to the assembly of linked episodes, to builders of evaluation sets and training environments, and finally to agent developers. Protocol and framework owners supply the plumbing for new runs; they do not own the history of past handoffs or the systems that confirm how those handoffs ended.
That points to two places where a supplier could earn more than a commodity rate: assembling past handoffs into episodes with outcomes attached, and keeping a live link to the systems that confirm outcomes as new runs accumulate. Raw transcripts without outcomes sit nearer the commodity end, because a simulator can imitate how people write far more easily than it can imitate what happened to the account a week later.
Three kinds of holder could own that join: companies that operate the systems where outcomes are confirmed, outsourced service desks that run escalations for many clients, and data suppliers that license records across many companies. Each faces a different limit. The first holds depth for one business, the second holds breadth under client contracts that decide what may be reused, and the third depends on persuading each company to license what it holds.
The second point is a judgment, not a measured market fact. It is the claim an investor should put to suppliers by asking for episode samples and counting how many carry a verified outcome.
Three dated signals that handoffs are not going away
The demand evidence is forecasts and surveys, so read it as direction, not market size. Gartner predicted on March 5, 2025 that agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029. By its own wording the forecast concerns common issues, which leaves the uncommon cases, and the transitions into them, with people.
Customers expect the route to a person to stay open. In a Gartner survey of 3,566 customers published August 4, 2026, 87% said it is essential for companies using generative AI in customer service to offer a way to reach a human agent. In January 2026, Gartner also said full automation would be prohibitively expensive for most organizations.
Agents are also taking on longer tasks, and every added step is another point where one may need a person. METR measured the length of software tasks that frontier models complete with 50% success and found it doubled about every seven months since 2019 (arXiv 2503.14499), while cautioning that the tasks are software work. In the EU, Article 14 of the AI Act requires high-risk AI systems to be designed so they can be effectively overseen by natural persons while in use; the duty covers high-risk systems, not every agent.
Strongest alternative: a pile of transcripts is not an environment
The most serious objection is that handoff records may lack what an agent needs in order to learn from them. A transfer note says a case went to a person; it rarely captures the tool state the agent saw, the options it had, or an end state a grader can check. Without those, a record collection supports the study of handoffs, not the training or scoring of an agent's handoff decisions.
The test is replayability. Take a sample of episodes and ask whether each can be rebuilt into a starting state, a set of permitted actions, the correction as it was made and an outcome a grader can verify. If most cannot, value sits with whoever builds environments, and records are an input priced accordingly. If most can, the records are the scarce part, and the supplier holding them has something the plumbing cannot produce.
What could make the handoff data thesis wrong
Six developments would weaken or break the thesis, and investors should watch for each in supplier conversations and in the market:
- Logs going forward crowd out history: if the stacks above become standard, developers may gather enough handoff events from their own deployments and need fewer licensed past records.
- Platforms hold the data: escalation histories often live in software whose terms can limit export or training use, and access is not a license.
- Privacy cuts context: handoff transcripts carry names, accounts and complaints, and de-identification can strip details an episode needs.
- Simulation becomes good enough: simulated users and generated escalations may cover common handoff patterns cheaply, leaving real records valuable only for rare ones.
- The trend evidence is contested: a 2026 statistical reanalysis of METR's time horizons (arXiv 2610.12466) recomputes them under looser assumptions about how task difficulty scales, and Gartner's figures are forecasts and survey answers, not observed spending.
- Outcome labels are messy: a refund reversed later or a ticket reopened weeks afterwards makes a verified outcome a judgment call with its own error rate.
Diligence questions for a handoff data supplier, SourceX included
The same questions apply to SourceX. Its work sits between a company's records and an AI developer's license: finding and qualifying sources, checking rights, contracting, preparing and delivering the data, and paying the company. The company approves what is licensed and on what terms, and signs, before SourceX receives any records.
SourceX publishes no episode counts, completeness rates or commercial figures for handoff data or anything else; those numbers go to investors who request them. The company is raising capital under Rule 506(c), which limits purchasers to accredited investors, and a reader weighing it as an investment can ask for materials through the investors page.
Put these questions to any supplier, and ask for the answers measured on a sample rather than asserted in a deck:
- What share of sampled episodes include the state at the stop, the person's change and a verified outcome, and over what denominator?
- Which system confirms each outcome, and how long after the handoff is it observed?
- How is the authority behind a correction recorded: a role, an approval record, or an inference?
- Can an episode be replayed to a gradable end state, and who built the replay?
- What rights and privacy review did each source pass, and does the license cover training as well as evaluation?
- How many distinct source companies stand behind the supply?
Methods note: files read, versions and limits
Files read on October 27, 2026: docs/specification/2025-06-18/client/elicitation.mdx and the 2025-06-18 changelog in github.com/modelcontextprotocol/modelcontextprotocol at commit 54b91cd, with the 2026-07-28 version checked for the response actions; docs/specification.md in github.com/a2aproject/A2A at commit 98279d6; docs/human_in_the_loop.md in github.com/openai/openai-agents-python at commit b040da5; libs/langgraph/langgraph/types.py in github.com/langchain-ai/langgraph at commit 858891a.
Criteria: the four comparison columns were fixed before reading. Outcome after the handoff is marked absent when the file read defines no field or state for whether the business result was correct; any stack can carry such a field in developer-defined data. Not measured: adoption of any stack, volumes of handoff events, prices for handoff data, or the completeness of any company's records. The input map and value chain are SourceX frameworks built from these files and the cited forecasts, not market measurements.
Frequently asked questions
What is human-agent handoff data?
Records of the moments when an AI agent stops and a person takes over or decides, and of what happens when the agent resumes: the state at the stop, the reason, the person's decision and authority, what changed, and the verified outcome. Without the outcome, the records describe handoffs but cannot score them.
Don't agent frameworks already capture handoffs?
For new runs they capture the pause, the person's reply and the resume state; MCP elicitation, A2A's input-required state, and approval and interrupt features in OpenAI's Agents SDK and LangGraph all do. None of the files reviewed defines the business outcome, and none reaches back to handoffs that happened before an agent was deployed.
Why would verified outcomes be the scarce part of handoff data?
Outcomes are confirmed in systems the agent never touches, such as refund ledgers, reopen events and audits, often weeks later. Joining them to a handoff takes access to those systems, rights to license them and work to link the records, which is harder to copy than collecting transcripts.
How should an investor test a handoff data supplier?
Ask for a sample and measure it: the share of episodes with state, correction and outcome present, how outcomes are confirmed, whether episodes replay to a gradable end state, and how many source companies contribute. Claims without denominators should carry no weight.
Does SourceX publish metrics on handoff datasets?
No. Episode counts, completeness rates and commercial figures are not public; SourceX provides company metrics to investors who ask for them through sourcex.si/investors. Any offering is open only to accredited investors whose status has been verified.
Sources
- MCP's 2025-06-18 specification introduced elicitation, letting servers request structured information from users, with accept, decline and cancel responses. Source
- The A2A specification (latest released version 1.0.0) lets an agent move a task to the input-required state; the client continues with the same taskId and contextId. Source
- OpenAI's Agents SDK pauses runs for approval of tools marked needs_approval, records approve or reject decisions, and resumes from a serialized RunState. Source
- LangGraph's interrupt raises a resumable GraphInterrupt, resumes via Command by re-executing the node, and requires a checkpointer. Source
- Gartner predicted (March 5, 2025) that agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029. Source
- A Gartner survey of 3,566 customers (August 4, 2026) found 87% say it is essential for companies using GenAI in customer service to provide an option to reach a human agent. Source
- Gartner (January 26, 2026) said full automation of customer service will be prohibitively expensive for most organizations. Source
- METR found the length of software tasks frontier AI completes with 50% success doubled about every seven months since 2019, and cautions about extrapolating beyond software tasks. Source
- A 2026 statistical reanalysis recomputes METR time horizons with methods that relax the assumption that difficulty scales linearly with log human time. Source
- EU AI Act Article 14 requires high-risk AI systems to be designed so they can be effectively overseen by natural persons while in use. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.