adragan - stock.adobe.com
LinkedIn touts agentic AI to slash recruitment time
LinkedIn’s head of engineering for talent solutions explains how fine-tuned LLMs and agentic AI architectures are replacing traditional search methods to save recruiters four hours per role
LinkedIn has built an artificial intelligence (AI) infrastructure stack around agentic workflows, moving beyond standard generative text features to autonomous AI agents capable of managing complex recruitment tasks.
The technology is now being used to support the global rollout of LinkedIn Hiring Assistant, a product the Microsoft-owned company claims is already saving recruiters an average of four hours per role and reducing candidate profile reviews by 62%.
Speaking to Computer Weekly in a recent interview, Prashanthi Padmabhan, vice-president for engineering at LinkedIn Talent Solutions, said the company’s infrastructure stack has evolved to support an agentic era where AI doesn’t just summarise text, but executes multi-step workflows.
“We took all those manual, labour-intensive parts of the recruitment process and used a combination of agentic flows and a multi-modal agent-based architecture,” Padmabhan said.
Padmabhan said LinkedIn Hiring Assistant is built on the LangGraph agent orchestration framework, allowing AI agents to perform the hard work of scouring LinkedIn’s billion-member database to identify job candidates for hiring teams.
The move comes as the recruitment market faces increasing friction. According to new data released by LinkedIn, about three in four recruiters in key Asia-Pacific markets say finding qualified talent has become significantly harder.
“We are using the power of LLMs [large language models] to process a large set of data and then come back with candidates for the given role,” Padmabhan said. “We’re not just coming back with the results; we’re also giving explanations and evidence for why we think these top candidates are best fit for the role.”
However, standard off-the-shelf LLMs are not effective in catering to the nuances of specialised enterprise hiring. Instead, LinkedIn uses an ensemble of models fine-tuned on its massive dataset of skills, job changes, and professional relationships.
This allows AI agents to interpret natural language requests, such as a recruiter describing a role conversationally, rather than forcing them to construct complex Boolean search strings.
“You need a lot of domain-specific intelligence. We use our own secret sauce – our data and insights – to make sure we are able to bring that into context,” Padmabhan said, adding that the engineering team employs retrieval-augmented generation, reinforcement learning and other techniques to continuously improve model performance.
For enterprise CIOs and HR leaders, the integration of AI agents into existing HR management applications is necessary so companies can manage the entire employee lifecycle, from recruitment and onboarding to engagement and training.
Padmabhan said LinkedIn’s recruitment capabilities can integrate with popular HR platforms and applicant tracking systems (ATS) such as Workday and SuccessFactors, blending LinkedIn profile data with an enterprise’s candidate records.
Inside LinkedIn’s agentic architecture
For many engineering teams, the move to agents requires a departure from traditional application architecture. LinkedIn uses LangGraph’s hierarchical approach to task management, with supervisor agents and sub-agents constantly communicating with each other.
The supervisor agent acts as an orchestrator, interpreting a recruiter’s intent, such as a conversational request to find a Java engineer in Bangalore. The supervisor then invokes sub-agents to perform specific tasks, which could be querying an ATS, searching LinkedIn for candidates, or writing outreach messages. Sub-agents can be decommissioned when needed.
LinkedIn has also invested in platform-level capabilities to manage memory, context, and versioning for AI agents. With these capabilities, product teams can, for example, track changes to prompts and roll them back if necessary. “Just like how code is versioned and launched, prompts are now the new code, so you want to have a good versioning system for that,” Padmabhan said.
Memory management is also key. For Hiring Assistant to be effective, it must retain context across different sub-agents and sessions. If a recruiter prefers candidates with a certain level of experience, that preference should persist in future searches and interactions, regardless of which sub-agent is executing the task.
However, with automation comes the risk of AI hiring biases. Padmabhan stressed that LinkedIn adopts a human-in-the-loop approach, meaning the AI agent provides evidence and reasoning for its selections, but the recruiter makes the final call.
“We are not making the Hiring Assistant make autonomous choices. Its job is to do the hard work and tell you why it chose those candidates with clear evidence,” she said.
She added that before any agentic AI product reaches production, it must pass a battery of tests from a responsible AI team to check for gender and demographic biases, as well as security vulnerabilities like prompt injection attacks.
Early adopters of LinkedIn Hiring Assistant, such as United Overseas Bank (UOB) and blockchain technology company OKX, have benefited from the technology.
Speaking at the LinkedIn Talent Connect event in Singapore, Jay Chan, executive director and head of talent acquisition at UOB, whose team was a charter customer for LinkedIn Hiring Assistant, used the tool to identify the specific candidate he ultimately hired, helping him justify the return-on-investment (ROI) to business leadership.
Tracy Mao, director and head of HR for Singapore and Malaysia at OKX, reported that the tool’s candidate matching capabilities were outperforming manual outreach, saving about six to eight hours of recruiter time. However, Mao noted that the tool currently faces challenges when “pairing up with multi-language candidates”.
Read more about AI in APAC
- Japanese banking giant MUFG aims to transform into an AI-native company by using agentic AI, changing how it handles data, and inking key partnerships with OpenAI and Sakana AI.
- Malaysia’s Ryt Bank is using its own LLM and agentic AI framework to allow customers to perform banking transactions in natural language, replacing traditional menus and buttons.
- Singapore researchers show how adapting pre-trained AI models can solve data scarcity issues in countries with limited resources.
- Lenovo’s CIO Playbook 2026 reveals that 96% of APAC organisations are planning to invest more in AI, with a growing reliance on hybrid infrastructure to manage rising inference costs.
