London HR Startup
A voice agent and chatbot that handle 70% of candidate-ops admin, with human-in-the-loop gates on every sensitive action and a full action audit trail
We replaced spreadsheet-driven operations with a governed network of AI agents, an agentic voice system and chatbot backed by MCP servers, that now handle an estimated 70% of candidate-ops admin, with mandatory human-oversight gates on every candidate- and client-facing action and a complete audit trail of what each agent did and why.
The challenge
A fast-growing London HR startup was buckling under manual processes: candidate sourcing, interview scheduling, offer management and client reporting were all spreadsheet- and inbox-driven. They wanted to scale throughput without scaling headcount, but HR data is sensitive: candidate PII, consent and client confidentiality mean an off-the-shelf autonomous agent that acts unsupervised was never an option. Any automation had to be auditable and gated by human oversight on the decisions that matter.
What we built
Mapped every customer-facing and back-office workflow over a 2-week discovery, ranking each by automation feasibility, risk and ROI
Stood up internal MCP servers exposing Greenhouse, HiBob, Slack, Notion and the data warehouse as governed tools, each with scoped permissions and access logging
Built specialised voice and chat agents for sourcing, scheduling, candidate comms and weekly client reporting, each with explicit human-in-the-loop checkpoints before any sensitive or outbound action
Embedded assistants directly inside the team's existing tools (Slack, Notion) so adoption was friction-free, with every agent action captured in an audit trail
Defined an internal AI development lifecycle: prompts and agent definitions versioned in Git, evaluated against golden datasets, and deployed via CI
Architecture
Outcomes
The voice agent and chatbot now handle an estimated 70% of candidate-ops admin, with human-oversight gates on every sensitive action and zero consent or data-handling incidents to date
Roughly 3x pipeline throughput within 90 days with no added ops headcount
Client-reporting cycle cut from ~6 hours/week to ~20 minutes, with a human reviewing every AI-drafted report before it goes out
A fully auditable agent estate the founders can trust and extend, rather than an opaque black box
Results
68%
of candidate-ops admin handled
£120k
annual headcount avoided
0
consent / data incidents
3×
pipeline throughput
Other case studies
Ready to take AI from pilot to production?
Book a free discovery call with our team of ex-AWS engineers. We'll discuss your regulated-industry challenges and outline a governed path from pilot to production. No commitment required.
