b12 Labs Wants to Give Chemistry an Autopilot

There is a familiar number that gets thrown around whenever anyone talks about making a new drug: ten-plus years, roughly two billion dollars, and a hit rate so brutal that only about two of every ten thousand candidate compounds ever make it to a patient. b12 Labs, a three-person startup out of Y Combinator's Summer 2025 batch, looked at that math and picked a narrow, stubborn part of it to attack - not the biology, not the clinical trials, but the plain chemistry of actually making the molecules.

The company builds AI agents for the chemistry lab. Feed it a target molecule, and its software proposes how to synthesize it, figures out which reaction conditions have the best odds, and then generates the instructions a robotic lab needs to run the experiment. The tagline on its website is unfussy about the goal: "AI that designs, optimizes, and runs chemical reactions. Accelerating pharma R&D from years to months."

The problemThe slow, unglamorous middle

In drug discovery, a lot of attention goes to the flashy ends of the pipeline - the target you're chasing, the trial that proves it works. The middle, where a chemist has to actually build the molecule, is quieter and often slower than outsiders expect. Knowing what you want is not the same as being able to make it. Synthesis runs on cycles of experiment design and lab trial-and-error, and each cycle takes time.

b12 Labs argues the tooling around this work is fragmented - separate software for planning routes, for predicting conditions, for driving instruments - with poor handoffs between them. A chemist might sketch a route in one program, reason about conditions in a spreadsheet, and then hand a protocol to whoever knows how to run the instruments. Every seam between those steps is a place where time leaks out. And there is a second, almost ironic bottleneck: many labs already own capable robots, but the machines sit underused because operating them takes a specialized automation engineer who can write the low-level code. The hardware is expensive and idle, and the reason is human, not mechanical.

Compounded across a discovery program with hundreds of molecules to make, those small frictions add up to real calendar time. That is the wedge b12 is prying at - not the science of what to build, but the operational drag of building it. It is an unfashionable place to start a company, which is part of why it is interesting: the bottleneck is real, measurable, and mostly ignored by the parts of the field chasing splashier problems.

The robots are already in the building. b12's bet is that the missing piece is the software that knows how to talk to them.

The productThree agents, named for the periodic table

b12 Labs organizes its platform into three tools, each named after an element - a small piece of chemist-humor branding that also maps cleanly onto the workflow.

Put together, the three describe a single arc: a chemist types a request in plain language, hands over any prior run data and relevant documents, and the agents move from a route on paper to a job a machine can execute. No coding required is the recurring promise.

The differenceFull-stack, and deliberately vendor-agnostic

Plenty of groups work on pieces of this - retrosynthesis predictors, reaction-optimization models, self-driving-lab research. b12 Labs' distinguishing bet is to be full-stack and vendor-agnostic at the same time: one agent system that spans planning, optimization, and execution, and that speaks to several robotics platforms rather than locking a lab into one. The company lists Chemspeed, Opentrons, and Unchained among the systems it integrates with, and says it can drive manual labs too.

That neutrality is the strategic point. A lab rarely rips out its instruments to adopt a software vendor. By sitting on top of whatever hardware is already on the bench, b12 is trying to be the connective tissue for a stack nobody else wants to own end to end. It is a classic wedge: start where the buyer has already spent the money, and sell them the layer that makes the spend pay off.

There is a subtler advantage to being full-stack, too. When planning, optimization, and execution live in one system, the results of a run can flow back into the next plan without a human retyping anything. b12 describes this as closing the experimental feedback loop automatically - the sort of thing that sounds mundane in a slide but is exactly where most AI-for-science demos quietly stall. Getting a model to propose a route is comparatively easy. Getting the bench result back into the model, reliably, day after day, is the hard part, and it is the part b12 says it is building for.

Fig. 1 - Platforms b12 Labs states it works with. Illustrative; not a benchmark.

The peopleAn olympiad champion and a Nature author

The two founders met at EPFL, in the Lab of Artificial Chemical Intelligence, where Zlatko Joncev was doing a postdoc and Andres Bran was a PhD student. Joncev, now Co-Founder and CEO, trained as a medicinal chemist at Roche; he finished near the top of Serbia's national chemistry competition three years running and went on to the International Chemistry Olympiad, with published work on atroposelective synthesis and, later, on retrosynthesis using large language models.

Bran comes from the AI side of the same field. He built an early AI agent for autonomous molecular planning in robotic labs - work published in Nature Machine Intelligence that has drawn hundreds of citations and picked up a best-paper nod at a NeurIPS AI-for-Science workshop. The pairing is the pitch in miniature: someone who has spent years actually making hard molecules, next to someone who has spent years teaching machines to plan them. Neither is a generalist software founder who wandered into chemistry; both come out of the bench, which matters when your customers are chemists deciding whether to trust a machine with a reaction.

That credibility is not incidental to the business. Selling automation into a wet lab means persuading people who have watched plenty of promising tools fail to survive real chemistry. A founder who can speak fluently about atroposelective synthesis and about robot code is easier to believe than one who can only speak about one. The team rounds out its resume with stints at Roche, Chemspeed, and the Max Planck Institute - a spread that covers big-pharma medicinal chemistry, lab-automation hardware, and fundamental research, which is roughly the exact triangle the product has to operate inside.

"We believe the future of synthetic chemistry is agentic."

The modelSoftware for the wet lab

b12 Labs sells business-to-business software into pharmaceutical and biotech R&D - the medicinal and process chemists and discovery teams who spend their days on exactly the synthesis cycles the platform targets. As an early-stage company it has run pilot demonstrations rather than published a roster of marquee customers, and it has been hiring founding engineers to build out the product. The company is backed by Y Combinator, with a reported first raise in the low six figures tied to the S25 batch.

The name itself is a small in-joke worth pausing on. Vitamin B12 is one of the most notoriously difficult molecules ever synthesized - a decades-long saga in organic chemistry. Naming your company after it, then setting out to make hard molecules easier to plan, is either a dare or a thesis statement. Probably both.

The marketWhere it fits

b12 Labs sits at a busy intersection: AI-for-science, lab automation, and drug-discovery tooling, each of which has its own incumbents and its own hype cycle. Retrosynthesis predictors have existed for years. Reaction-optimization models are an active research area. Self-driving labs get written up regularly. What is scarcer is a product that ties those threads together and points them at a paying customer's actual instruments.

The harder, less-demoed part of that world is closing the loop between what a model proposes and what a robot actually does on the bench. That is the seam b12 is aiming at. Whether the agents prove reliable enough for chemists to trust across real programs is the open question - a synthesis plan that looks elegant on screen still has to survive contact with a stubborn reaction. But the wedge is clear, and it is a specific, unglamorous one: make the making faster. For a company named after one of chemistry's hardest molecules, that is a fitting place to plant a flag.