80% Efficiency Gain in Benefit Extraction
Powered by AI.
AIRINC helps companies navigate employee relocation across cities and countries. At the centre of their service is benefit analysis: reading a company's relocation policy, identifying every benefit and sub-benefit it contains, and benchmarking that package against industry standards.
Skilled work.
Done manually.
AIRINC’s core service depends on benefit analysis—reading policy documents, identifying every benefit and sub-benefit, and benchmarking against industry standards. The process itself was well-defined. Analysts followed a structured approach, applying domain expertise to interpret complex policy language and extract the relevant details.
Each document often 40+ pages, had to be read end to end. Benefits were identified and catalogued manually, with accuracy dependent on the analyst’s attention and experience.
There was no automated way to track coverage, verify completeness, or standardise output across engagements. The work was precise, but it was time-intensive, and difficult to scale. AIRINC didn’t need to change the process. They needed a system to execute it.
Every single engagement.
Four problems.
One breaking point.
A system for extraction,
built for verification.
The Zig built an end-to-end AI extraction system designed around AIRINC's existing workflow.
Documents are uploaded and standardised before processing. Non-PDF formats are automatically converted before processing begins. A large language model then reads the entire document and identifies every benefit and sub-benefit it contains, categorising each one with precision, beyond simple keyword matching.
Critically, the system doesn't just extract. It traces. Each identified benefit is pinpointed to the exact location in the original document, with the relevant text highlighted so an analyst can verify the AI's work at a glance. Nothing is accepted blindly. Every extracted item passes through a human review layer before export.
The initial architecture used Azure Document Intelligence with vector embeddings and AI Search. After testing, The Zig pivoted to a full LLM approach, delivering meaningfully better accuracy on complex policy language. A custom recall function re-runs extraction if more than 10% of expected items are absent.
Not just faster.
Verifiable and
Consistent.
The system reduced processing time significantly, but the more important change was traceability.
Every extracted benefit now links directly to its source text, allowing analysts to verify results immediately without re-reading entire documents. Output is standardised across engagements, removing the need for downstream normalisation in the previous workflow and enabling consistent benchmarking.
Beyond the metrics, this project introduced AIRINC to AI as a practical capability.
AIRINC's analysts understand nuance in policy language that no system will catch perfectly. The right design puts AI in front of that expertise, not in place of it. Human-in-the-loop isn't a concession to imperfect technology. It's the architecture that makes automation trustworthy enough to actually use.
See what's possible.
Build what matters.
Tell us where manual work is holding your team back. We'll be direct about what AI can actually do for it (no slides, no theatre).