HFS Research

Suplari Launches Data Assistant for Procurement Data

Suplari announced Data Assistant, an AI agent that automates the full lifecycle of procurement data ingestion and transformation. The agent accepts multiple file formats (Excel, CSV, Word, PDF, zip), identifies content and destination, corrects data faults without human escalation, monitors connectors for data drift, and repairs connectors when source systems change. Suplari positions the product to address persistent pipeline fragility that blocks procurement AI programs, citing research that poor data quality is a major barrier to scaling procurement AI. The company emphasizes traceability by recording every run, transformation and automatic fix.

Opendoor Shuts India Operations, Sparks AI-Offshoring Debate

Opendoor is closing its India operations less than two years after expanding into Chennai and Bengaluru, a move CEO Kaz Nejatian said aims to bring operational work closer to U.S. customers and shift toward smaller, AI-native teams. The company had roughly 250 India employees when it opened offices in 2024; broader headcount reductions make it unclear how many were affected. Opendoor reported 1,042 global employees at the end of last year (down from 1,470) and a non-U.S. workforce decline to 184 from 342. The exit has reignited debate over whether generative AI is changing the economics of offshore back-office work and Global Capability Centers in India, a market with more than 2,100 centers employing about 2.36 million people and generating nearly $100 billion annually.

akm 0.7.0 Released with Proposal Queue and Bench

akm 0.7.0 is a technical release that introduces a durable proposal queue for agent-generated changes, three new non-mutating CLI commands (reflect, propose, distill), a first-class "lesson" asset type for synthesized knowledge, opt-in LLM feature gates, and a paired-run benchmarking tool (akm-bench v1). The proposal queue stores drafts outside the asset tree until explicitly accepted, and proposed assets are excluded from default search unless included. Seven bounded in-tree LLM call sites are gated behind llm.features.* flags (all false by default) with a tryLlmFeature() wrapper providing fallbacks and timeouts. The release also includes multiple security, UX, and hygiene hardenings and is opt-in for users upgrading from 0.6.x.