Your organization runs on workflows, now so does EMET

Join us as we introduce EMET Workflows: October 8th 12-1 pm EST
Sign up here

Jensen Huang said something on stage at GTC this year that I keep coming back to:

"The era of AI as a conversation partner is giving way to the era of AI as a digital workforce. Every company will have AI employees — agents that reason, plan, use tools, and deliver outcomes. This is not a feature update. This is a platform shift."

In the same stretch, he gave the clearest mechanical definition of an agent I've heard from anyone: "an agent, which consists of a large language model, or many, sitting inside a harness, and that harness helps it, orchestrates it, to do productive work."

We share that belief. AI is not here to be a chatbot — it's here to deliver the work.

Codifying what that work actually is turns out to be the hard part, and today we're taking a real step at it. We're introducing workflows in EMET: predefined, repeatable units of work that bring together tens of thousands of parts — data, models, bioinformatics, connectors, dedicated endpoints, encoded judgment, and human guardrails — to deliver the units of work that move commercial research forward.

It has been done before. Three industries have already proved it.

Nobody reading this has to take a leap of faith. In three professions, the useful version of AI stopped being a conversation and became a piece of work someone signs their name to. It's worth looking at what those actually are, because the shape is the same every time.

Software engineering. An engineer files a task — upgrade this library across the repository and fix what breaks. The agent reads the codebase, makes the change, runs the test suite, works through the failures, and opens a pull request. What lands in the engineer's queue is a diff: here is exactly what changed, here are the tests passing. They review it and merge it, or send it back.

That's a different object than an answer about how to do the upgrade. And the people doing the work voted with their hands: in March 2025, two and a half times as many Cursor users relied on tab completion as on agents. That ratio has now completely inverted. Nobody moved because the models got chattier. They moved because the thing they received changed from a suggestion to a finished change they could review.

Legal. This is the profession where being wrong is expensive, where the work has to survive review by someone looking for a reason to reject it, and where nobody gets to say "the AI said so."

A deal team drops several thousand contracts into a workspace. The workflow classifies the documents, runs the firm's own diligence protocol across the entire set, pulls the key provisions, flags terms that deviate from the firm's playbook, identifies missing clauses, and produces a diligence report with citations back to specific pages. An associate reviews it. A partner signs it. One firm's lease summary workflow went from four hours to four minutes.

The part I'd underline: what makes it useful is the firm's playbook, not the vendor's. Harvey shipped the workflows, but their customers have built more than 25,000 of their own — because a diligence standard is something a firm spent decades arguing its way to, and no outsider can hand it to them.

Healthcare. A physician sees a patient and talks to them rather than typing. Afterward, a structured clinical note is waiting — the history, the assessment, the plan, the diagnostic coding, the orders, the after-visit summary for the patient. The doctor reads it, corrects what's wrong, signs it into the chart. One health system measured an 86% drop in the effort required to finish notes.

The product isn't a conversation about the visit. It's a document you can sign.

Three professions, one shape. In each case, the deliverable is an object, a resolved case, a cited report, a signed note — produced by a repeatable path, with a human at the end who takes responsibility for it.

And in each case, the version that was actually used encoded the organization's own standards: the team's conventions, the company's refund policy, the firm's playbook. A generic approach was never good enough.

Figuring out disease biology and discovering therapeutics are next

Drug discovery has fragmented past what any individual can hold in their head. Thousands of data sources. Thousands of software packages. 700+ models. 500+ SaaS tools. Internal findings are trapped across teams in incompatible formats. Roughly 60% of a scientist's time goes to finding, cleaning, and organizing data, and experimental costs have risen two to fivefold over the past decade. It keeps fragmenting faster than any scientist or any company can manage.

Science does not need another model. It needs a conductor — something that fuses a fragmented stack into one system where the whole is worth more than the parts, that is Emet.

And it isn't one workflow. Walk into any preclinical organization, and you'll find hundreds of distinct, repeatable units of work: a target safety assessment, a developability profile, a dose-selection rationale, a DDI risk review, a biomarker strategy, an IND evidence package, a portfolio prioritization score. Count the variants by therapeutic area and modality and, in a large pharma, the number runs into the thousands.

None of these are administrative. They are the moments where a program advances, stalls, or dies. A target gets dropped. A molecule moves to IND-enabling studies. A program takes one of five slots at a portfolio review.

And today most of them exist nowhere you can point to. They live in a senior scientist's head, in a slide template someone inherited from a colleague who left, in an SOP written six years ago that no longer matches what the team actually does. The expertise is real and hard-won. It just isn't encoded anywhere, which means it can't be reused, can't be audited, and walks out the door when the person does.

That's the true scale of the problem. Not "can AI answer a scientific question." Can a company take the several hundred decisions it makes over and over again, and make every one of them reproducible?

But there's a second problem, specific to enterprise science, and it's the one that convinced us workflows weren't optional.

Nine programs come to a portfolio review. Five slots. Four get killed. Two of the teams did their target safety assessment with AI. One asked a chat product directly and got a thorough, cited, entirely defensible answer. The other ran a certified workflow that pulled an internal dataset the first team didn't know existed, and scored it against the company's own prioritization thresholds.

Both answers are good. Neither is hallucinated. And they aren't comparable, because they weren't produced the same way. So the committee may kill the wrong program, and nobody in the room can tell.

Everyone worries about the model being wrong. Inside a large organization, the bigger problem is a hundred answers that are all right and none of them comparable.

Free form is how a scientist explores. A workflow is how a company decides.

That's the sentence our team keeps coming back to.

Free form is a scientist asking Emet a question and Emet choosing their own steps and sources. It's genuinely valuable — it's how exploration should feel. It's also a different path each run, so it isn't reproducible.

A workflow is a predefined sequence of skills, methods, models, data sources, and decision-making. Same question in, same path every time. Fifteen-plus evidence sources queried in parallel — genetics, expression, pathways, safety — against your assay data and your thresholds. Weighted, scored synthesis. A quality gate that flags a missing piece of evidence and cites the gap instead of quietly routing around it. Human sign-off before any verdict. A cited output. Run, date and version sealed, so the next execution repeats identically.

Repeatable, grounded, gated, human-verified, audit-locked. That's the design.

The vocabulary matters because the industry has turned it to mush. Tools answer. Skills encode one analysis — which tools, in what order, what counts as enough evidence. Workflows decide. The agent is just the runtime: it plans, calls tools, handles failure, and stops to ask for a signature. The model is swappable, and it knows language and reasoning — not your science, your data, or your standards.

What we're shipping isn't intelligence. It's an encoded judgment — one company's specific definition of what counts as evidence and what the threshold means.

We took the model out of the decision on purpose

This is the design choice I expect to get argued with the most.

Once a workflow reaches the tiers that synthesize weighted evidence and issue a verdict, it runs as a deterministic expert system. Rules defined by the organization, executed by Emet. Not a model reasoning its way to a judgment call fresh each time.

A company's decision rules are the most valuable thing it owns — years of accumulated institutional scar tissue about what a signal is actually worth. Handing that to a model to re-derive on the fly, slightly differently every Tuesday, isn't sophistication. It's throwing the asset away.

Let the model read, retrieve, reason, and draft. Let the company's own rules decide. Then a verdict is explainable by construction: you don't trust the word, you open what produced it.

And having no sanctioned version isn't a neutral choice. It's uncontrolled growth — every team quietly inventing its own standard.

What we've actually built

Live to start workflows will span eleven areas of preclinical R&D.

Discovery biology — target ID, target validation, disease landscape.

Small molecule discovery — med chem, CADD, analytics and purification, bioassay, binding, material management.

Large molecule and biologics — binding, developability, candidate screening.

Instrumental analytics — biophysical, structural biology, proteomics and mass spec.

Quantitative biology and pharmacology — biochemical and cell assays, high-content screening, mechanistic pharmacology.

Translational and biomarker — biomarker discovery, functional genomics, systems biology, patient stratification.

DMPK — ADME and metabolite ID, PK/PD, dose selection, bioanalytical.

Comparative biology and safety — in silico, animal models, functional assays, DDI and risk, pathology.

Digital tech and data science — computational biology and chemistry, bioinformatics, data analysis, lab automation.

Portfolio management and research ops — prioritization, asset tracking, milestones, reviews, decision tracking.

Regulatory — IND support, strategy, submission planning, CTD assembly, evidence packages, health authority responses.

They're layered in four tiers that map to who does the work. Property-level characterization for bench scientists. Domain-level evidence synthesis for scientist-managers and function reps. Multidomain synthesis for therapeutic project leaders. Decision-tier verdicts — a target safety assessment resolving to low, moderate or high — for portfolio reviewers and R&D leadership.

Deliberately bottom-heavy, because that's how adoption grows: one scientist at a time, not one CIO contract at a time.

Depth is uneven and we say so. ADME and dose, genetics and omics, and safety risk go deep. Genotoxicity and bioanalytical method development are a single workflow each today. Coverage without depth would be a marketing claim, not a product.

How we know they're any good

This is the part that separates a workflow library from a prompt library, and it's where most of the actual effort has gone.

Workflows are validated and review-ready before they ship, and the validation is done by scientists, not by engineers approximating scientists. We're a science-first company, PhD scientists doing quality assurance at scale, curating the underlying data before anything gets processed, running evaluations and optimization as a standing function rather than a launch checklist.

But scientific rigor alone doesn't make a workflow useful in a pharma company. A protocol that would satisfy a journal reviewer can still be useless at a governance checkpoint, because the deliverable is different — the question isn't is this interesting, it's does this survive the committee.

So the workflows are designed with people who ran these functions inside large pharma. Our scientific advisory board and our own team include people who built discovery, safety and translational processes at the companies we sell to, and who know what an internal target safety assessment actually has to contain to move a program forward. One of the more useful things they've given us is the unit-operations lens — the way pharma already decomposes work into discrete wet-lab and in-silico steps — which maps almost one to one onto what we call skills. We didn't invent that taxonomy. We adopted theirs.

Then we go further: every deployment reconciles our workflows with the client's existing process, with forward-deployed PhD scientists working alongside their teams. Traceability, alignment to their standards, and their sign-off. Confidence isn't asserted in a demo. It's built by showing your work.

The goal is for you to make it your own

We keep building. This is where the library starts, not where it stops. We're adding workflows continuously — going deeper where coverage is still thin, opening new areas, and letting what customers actually ask for set the order. This is a standing effort by our scientific team, not a launch batch.

Some of those workflows will be run exactly as shipped. Where a company doesn't have a strong internal opinion — a target essentiality read, a standard developability profile, an antibody selector — the generic version is the right version. It runs on day one, no setup, and it's already in production at customers today.

Most, though, are a starting point. And they should be.

Take one workflow — cardiotoxicity risk synthesis — and consider what it has to become. A version for small molecules and a version for antibodies. A version tuned to oncology and one tuned to neuro. A version that reads the internal tox dataset that never leaves the building, scored against thresholds this company argued its way to over fifteen years and would never publish.

Multiply that across eleven areas and every program in a portfolio, and you're back at hundreds of variants — each one carrying a piece of judgment that belongs to that company and to nobody else. A workflow only becomes valuable when a company can make it its own.

Which is why we don't ship a library and walk away.

Our forward-deployed scientists are not a support function we bolted on. They are the mechanism. One of our scientists sits with your team, takes a shipped workflow apart, and rebuilds it around your question, your data, and your thresholds — then hands it back as something your organization can validate and sanction. That is a scientific conversation.

What this is actually for

Every company in this industry is running AI pilots right now. Most of them are conversations, and most of them will quietly end — not because the models weren't good, but because nothing came out the other side that anyone was willing to sign.

The test isn't whether your AI is impressive. It's whether two of your teams, asking the same question next quarter, would hand a committee two answers you could actually put side by side.

We built workflows in Emet so the answer to that is yes — on your data, against your thresholds, traceable to the evidence, with a scientist's name at the bottom. And we send our own scientists to sit with yours and make each one genuinely theirs.

Drug programs die on decisions made with incomplete evidence, and every one of those years is a year a patient doesn't have. That's the whole reason to encode this properly.

Huang called it a platform shift. He's right, but I'd put the emphasis somewhere else. The shift isn't that AI finally got smart enough to do our work. It's that we finally stopped asking it to talk about it.

Chat was the demo. The work is the product.

Join us as we introduce EMET Workflows: October 8th 12-1 pm EST
Sign up here