What we are measuring as recruiters start working with MCP agents
We have begun tracking how recruiters use AI agents against live hiring data: the baseline metrics, where MCP fits the workflow, and the controls behind it.
Aman Kumar Tiwari
We have started tracking how recruiters actually use AI agents against live hiring data. Here is the baseline, what the external evidence says, and how we govern it.
Xobin MCP shipped as a product. The more interesting question is behavioural: when a recruiter can ask an agent anything about their pipeline, what do they ask? We have begun instrumenting that, anonymised and aggregated, so the recruiting community has data rather than anecdotes.
This post sets out the baseline we are measuring against, the external research that frames it, where MCP realistically fits in a recruiter's week, and the controls that make agent access to candidate data defensible.
The problem the numbers describe
Microsoft says 62% of respondents struggle with spending too much time searching for information during the workday.
Microsoft Work Trend IndexSHRM's 2026 recruiting benchmark reports a median 39 calendar days to fill nonexecutive positions.
SHRM Recruiting BenchmarkingLinkedIn reports that 61% of TA professionals believe AI can help improve measurement of quality of hire.
LinkedIn Future of RecruitingSHRM's 2026 State of AI in HR reports that 39% of surveyed HR professionals currently have AI adopted in their HR functions.
SHRM State of AI in HR 2026These studies describe the conditions around hiring teams. They are not evidence of Xobin MCP's impact. They are the reason we think recruiter-agent behaviour is worth measuring properly.
What we are tracking
Our instrumentation is aggregate and anonymised: no candidate records, no prompt contents, and no customer-identifying data leave the account boundary. We are tracking the shape of usage, not its subject matter.
Four families of signal
Intent mix
What proportion of agent requests are search and discovery, comparison, creation, or workflow actions.
Read-to-write ratio
How much agent usage is read-only evidence retrieval versus actions that change a record or contact a candidate.
Session depth
How many tool calls a single recruiting question turns into, and how often a session chains search → compare → create.
Approval behaviour
How often prepared write actions are confirmed, edited, or abandoned at the review step.
The baseline period opened with the launch of the Xobin MCP server. We will publish the first aggregate read-out once the sample is large enough to be meaningful rather than anecdotal.
The hypothesis we expect to be wrong
The intuitive assumption is that recruiters will reach for agents to do things, such as create the assessment or send the invite. Our early expectation is the opposite: the dominant behaviour will be search and retrieval of existing evidence, because that is the work that is currently manual and unbounded.
We will report the number either way.
That distinction matters. The goal is not to prove that agents make recruiting faster. It is to understand which recruiting tasks agents actually take on, which actions remain human-led, and where agent access creates measurable value.
Where MCP fits in the recruiter workflow
Work that starts and ends in hiring
- Running the pipeline
Stage management, bulk actions and day-to-day pipeline hygiene are faster in the product.
- Structured evaluation
Designing and calibrating assessments deserves the full interface.
- Reporting to a cadence
Recurring dashboards do not need a conversation.
Work that spans systems
- Ad-hoc evidence questions
One-off questions whose answer lives across résumés, scores and transcripts.
- Preparing for a conversation
Assembling context before a hiring-manager or panel discussion.
- Cross-referencing outside data
Comparing pipeline reality against a headcount plan or a document that lives elsewhere.
- Turning a discussion into a next step
Converting what was just agreed into an assessment or interview, ready for approval.
Rule of thumb: use Xobin for focused recruiting work; use Xobin MCP when hiring context has to work alongside something outside Xobin.
How we keep agent access defensible
Candidate data is among the most sensitive data an employer holds, and hiring is a consequential decision under emerging AI frameworks. Our controls are built for that standard.
Identity, not API keys
Access is granted through an OAuth sign-in to Xobin. No password or API key is handed to the assistant, and any connection can be revoked.
Permissions carry over
Every request is bound to the authenticated company, user and role. An agent can never see more than the person driving it.
Reads separated from writes
Actions that change records or contact candidates are declared separately, so compatible clients can require explicit confirmation before a write action executes.
Auditable by design
Agent-initiated activity is attributable to the authenticated user, so it can be reviewed like any other action in the account.
Human decision authority
Agents assemble and explain evidence. Selection, rejection and offer decisions stay with the hiring team.
Compliance posture
Xobin MCP inherits the security and data-protection commitments of the Xobin platform, including our existing certification and audit programme.
The intended workflow is deliberately simple:
Evidence is retrieved by the agent. Authority stays with the recruiter.
Two frameworks shape this directly. NIST's AI Risk Management Framework asks organizations to establish clear accountability and human oversight when AI is used in consequential contexts. The EU AI Act classifies certain AI systems used for recruitment and candidate selection as high-risk, with requirements that include human oversight, logging and transparency.
An agent that retrieves evidence for a human reviewer, under that human's own permissions, is a deliberately conservative design against those requirements.
What comes next
We will publish the first aggregate read-out of the metrics above once the sample supports it, alongside the categories of questions recruiters ask most.
The first release will focus on behaviour rather than performance claims: what recruiters search for, how often sessions move from retrieval to action, how frequently humans intervene, and which workflows remain inside the product.
If your team wants to be part of that cohort, or wants the raw governance detail for a security review, talk to us.
Want the security and compliance detail, or early access to the research read-out?
Talk to the Xobin teamSources
- Microsoft, "Will AI Fix Work?" — Work Trend Index.
- SHRM, Recruiting Benchmarking research.
- LinkedIn, The Future of Recruiting.
- NIST, AI Risk Management Framework (AI RMF 1.0).
- EU Artificial Intelligence Act.
- Model Context Protocol, official documentation.
Aman Kumar Tiwari
Aman writes about hiring practice, HR regulation and the assessment market. He tracks vendor releases and compliance changes so talent teams do not have to.
More from the Xobin hub
- NewsFive Things to Try First with Xobin MCPFive worked examples in a live assistant window, from evidence search to a context-built assessment, showing exactly how Xobin MCP works.
- NewsWhy XobinA perspective on building trusted measurement infrastructure for talent in the AI era.
- NewsIntroducing Xobin MCP ServerXobin now supports MCP, letting Claude, ChatGPT and other AI assistants securely search hiring evidence and act on recruiting workflows.