Freehand raises $75M to run AI agents on supply-chain spend

Good morning 👋 A $75M round landed yesterday with a Fortune 500 client list most startups don't get until Series C, and neither the company nor the outlets covering it will agree on what stage it actually is.

In today's issue:

• 🔭 Freehand raises $75M to run AI agents on supply-chain spend
• 🔬 Research: whether agents can actually do research, not just execute it
• 🛠️ Try this: wire a spend-anomaly agent from systems you already have
• 🔗 Quick links: ARC-AGI-3 triples on two API settings, JWT auth for agents, faster support triage

🔭 THE ONE THING

📦 Freehand raises $75M to run AI agents on supply-chain spend

Freehand closed a $75M round co-led by Battery Ventures and NewRoad Capital Partners, with Penny Pritzker's PSP Growth and Nexus Venture Partners also in. The pitch: AI agents sitting inside procurement and spend workflows for Meta, Unilever, J&J, Pfizer, Dunkin, and Cardinal Health, an enterprise roster most startups don't get until Series C. And that's the part worth noticing: outlets can't agree on whether this is a seed or a Series B, with Axios pricing it as a $300M post-money seed. Founders Nitin Jayakrishnan and Abhijeet Manohar already built and sold one supply-chain company, Pando, which is probably why investors let the stage label stay fuzzy. Fortune 500 logos plus a nine-figure valuation on a round nobody will name cleanly: less a red flag than a sign enterprise AI-agent money is still getting shoved into whatever wrapper closes fastest.

🔬 RESEARCH HIGHLIGHTS

• Testing whether AI agents can actually do research, not just tasks. Most agent evals either hand agents narrow, easily-verified problems or slip AI-written papers into blind peer review, neither of which touches the actual open-ended part: picking the question worth asking, not just executing a known one. Two case studies here try to close that gap directly. The forecast that agents will soon automate AI research rests on a capability nobody had rigorously tested until now. Paper

• An AI teammate changes how humans talk to each other. Using Group Communication Analysis, this study ran small teams through decision-making tasks with a conversational AI seated at the table, then measured the human-to-human chatter, not the bot's output. Adding the AI teammate reshaped team communication dynamics on its own, a cost that shows up nowhere in a benchmark scoring the AI's answers. Paper

• Everyone drafting through the same handful of models looks like a monoculture risk. The concern: as writers lean on shared LLM assistants to draft and polish, population-level language variety may shrink even as any single piece of writing reads cleaner. Worth watching if you're worried about what "human" text even means once the next training set gets scraped. Paper

🛠️ TRY THIS

Wire a spend-anomaly agent from systems you already have, before you buy one

1. Pick two systems that hold your spend data (ERP export, PO tool, expense reports) and connect them with prebuilt MCP connectors instead of custom integration code, the approach AWS just laid out for Bedrock AgentCore.

2. Define a scoring rule in plain language: vendor concentration, price drift, duplicate POs. Same "score and flag" shape Amazon Quick uses for its no-code retention pipeline, just pointed at spend instead of churn.

3. Ask the agent a cross-system question your finance team currently answers by email.

4. Check its citations. If it can't name which system a number came from, don't trust the flag yet.

Prompt: Query our ERP and PO systems for vendors whose spend grew more than 20% quarter over quarter with no matching PO increase. List each vendor, the dollar delta, and which system each number came from.

Worth a look

• Amazon Bedrock AgentCore ships prebuilt MCP connectors, access control, and persistent memory, so cross-system queries become configuration, not an integration project. AWS
• Amazon Quick runs the same no-code score-and-act pattern on customer retention, at-risk accounts flagged and save letters drafted automatically. Worth stealing the pattern even if churn isn't your problem. AWS

🔗 QUICK LINKS

• OpenAI Two API settings, retained reasoning and compaction, tripled GPT-5.6's ARC-AGI-3 score. Worth checking if your own eval harness is leaving that on the table. OpenAI
• AWS Bedrock AgentCore Identity now supports Private Key JWT auth, so agents can authenticate with a signed KMS key instead of a shared secret sitting in an env file. AWS
• Deepgram Cut SageMaker AI support ticket triage time by adopting AWS IAM temporary delegation, letting AWS engineers get scoped, time-boxed access instead of full account handoffs. AWS

Enterprise AI money moves faster than the paperwork describing it. Worth remembering next time a stage label looks too clean.

Pradeep Perugu