What 10,000 Scientific Workflows Taught Us About AI in Biopharma

Consider a literature review. Finding relevant papers is one step. A complete workflow also applies the team's inclusion criteria, records why a paper was excluded, compares published evidence with internal results, keeps conflicting findings visible, and produces a brief or evidence table that another scientist can review. The search matters, but the handoffs around it determine whether the result is usable.

The same distinction appears in every function. Generating code is not a biometrics workflow. Drafting a paragraph is not a regulatory workflow. Summarizing a monitoring report is not clinical oversight. Each output has upstream inputs and downstream consequences. A useful agent has to carry context between those stages and know when the next step requires a person.

This changes how we evaluate AI. Fluency matters less once the work leaves a demo. Teams care whether the right source version was used, whether a required check ran, whether an exception remained visible, and whether the result arrived in the form the next reviewer expects.

A drug program may use an electronic lab notebook, a LIMS, a clinical data platform, a controlled document vault, a statistical environment, file storage, and internal software built for one company. These systems hold different parts of the same decision. None has a complete view of the work.

People bridge the gaps manually. They download a file, check a tracker, ask which version is current, copy a result into a document, and send a message to confirm what changed. This effort is easy to dismiss as administration. In practice, it consumes the people who best understand the science and the program.

Enjamb gives an agent access only to the systems and actions required for the assigned workflow. The agent carries the requester's permissions into those systems, keeps returned artifacts distinct, and associates tool activity with the run. If two sources disagree, the workflow can surface the conflict instead of blending them into a clean answer.

That is why connectors are only the beginning. Retrieval without continuity creates a larger pile of context for someone to sort through. The workflow has to reconcile definitions, preserve versions, run the required checks, and deliver the result where the work continues.

A locked database does not produce a submission package on its own. The protocol and statistical analysis plan have to become specifications. SDTM and ADaM datasets have to be built against those specifications. The datasets and define.xml need validation. Tables, listings, and figures have to come from the approved analysis data, and independent quality control has to resolve the differences that matter.

Enjamb can run that sequence as connected work. Estimands, derivations, and open decisions from the protocol and SAP stay visible in the specifications. Dataset variables remain associated with their derivations, input versions, programs, and logs. Pinnacle 21 findings can be read against the data that produced them, then cleared or justified in the same working context.

For output QC, exact matches should not consume a statistician's attention. Enjamb can compare primary and validation outputs, clear what agrees, and route the remaining differences with the relevant programs, logs, and inputs attached. The statistician spends time on the discrepancy, not on rebuilding the path to it.

This is a useful boundary for automation. Statistical judgment stays with statisticians. The surrounding work arrives organized enough for that judgment to be applied quickly and recorded beside the result.

Clinical teams already have dashboards and trackers. The recurring problem is that the status in one place does not always agree with the evidence in another. A CRO report may say a site is on plan while the CTMS, EDC, or monitoring record shows an unresolved dependency. By the time someone reconciles the difference, the milestone may already be at risk.

A clinical operations workflow can follow study startup across contracts, IRB approvals, and initiation visits. It can read monitoring reports beside the EDC records they describe, surface findings that never became follow-up, and keep interim or database lock dates tied to the enrollment, cleaning, and verification work that must happen first.

Protocol deviations show why ownership belongs inside the workflow. Logging the deviation is not completion. The record needs a source, an owner, a clock, and a route for sponsor-reportable cases. The follow-up remains part of the work until the responsible person closes it.

Vendor oversight follows the same pattern. Enjamb can compare a CRO's status report with the study's own records and raise the differences with supporting context. The clinical lead sees the issue that requires action instead of another summary of the report they already received.

A submission draft is only useful if its claims come from the right evidence. Regulatory teams need to know which table supports a statement, whether terminology matches across modules, how the current position differs from a prior filing, and which points still need functional review.

Enjamb can draft IND and NDA sections from approved evidence, analyses, tables, and prior filings while keeping the source behind each statement. It can compare claims and terminology across CTD modules, the clinical study report, and earlier submissions so differences appear before the review cycle hardens them into the package.

A health authority response has another chain behind it. The team must locate the original commitment, gather the relevant evidence, assemble a response, and identify the questions that cannot be settled from the record alone. Enjamb keeps those pieces together and routes the remaining points to the appropriate reviewer.

Commitments continue after the document is sent. Owners, due dates, supporting work, and the correspondence that created the obligation need to remain connected. Treating authoring as an isolated task misses most of what makes regulatory work difficult.

A deviation investigation may draw on executed batch records, laboratory results, equipment history, procedures, and communications. The useful output is not a generic root-cause summary. Investigators need a timeline they can inspect, the source behind each event, and a clear account of what remains uncertain.

If the investigation leads to corrective and preventive actions, those actions need owners and evidence that the work happened. A proposed change must be assessed against the procedures, systems, products, and open quality work it affects. Batch record review must compare executed records with approved instructions and surface missing, inconsistent, or out-of-sequence entries.

Enjamb can prepare this work and preserve the record around it. When a proposed action would update a governed system, the run can stop for a named reviewer. The reviewer sees the requested action and its supporting context before deciding whether it should proceed.

Inspection readiness is therefore a property of daily work, not a package assembled at the last minute. Current procedures, decisions, actions, approvals, and source evidence stay connected as the work progresses.

Research workflows often start with a broad question and a mixed evidence set. Publications may disagree with each other. An assay result may challenge the mechanism described in the literature. An internal analysis may rely on a definition that changed between campaigns.

Enjamb can screen publications beside internal campaigns and ELN records, keep inclusion and exclusion reasons with the review, and surface papers that contradict a program assumption. For assay work, the workflow can take instrument output through analysis and QC, then prepare the ELN entry with the method and figures attached.

A candidate progression package may combine potency, selectivity, and liability evidence. A program brief may need the current protocol, internal data, prior decisions, and recent external research. In each case, the output should let the committee interrogate the claim and open the supporting material without starting the review again from scratch.

Agents become less useful when they smooth away uncertainty. A conflicting result, an untested assumption, or a missing acceptance threshold belongs in the deliverable. The reviewer should know where the evidence ends and where judgment begins.

Telling users to check an AI output is not a review system. The workflow has to identify what needs review, carry the relevant evidence to the reviewer, and record what the reviewer decided. Otherwise the person repeats the work to feel confident in the result.

Different decisions need different gates. Reading a permitted source may proceed automatically. Writing to a system of record may need approval. A statistical discrepancy may require a statistician, while a submission claim may require a regulatory owner. The agent should route each exception to the person with the authority and context to resolve it.

This approach also changes what the activity history must contain. A useful record includes the requester, instruction, sources, system actions, intermediate results, approvals, and final output. A chat transcript captures only a fraction of that history.

Across 10,000 workflows, this has been one of the clearest dividing lines between an interesting result and dependable work. Teams delegate more when they can see where the agent acted, where it stopped, and what a person approved.

Two companies may use the same name for a workflow and mean different things. Their templates differ. Their procedures define different acceptance criteria. Their systems divide responsibility in different ways. The approval that one company assigns to a study lead may belong to quality or regulatory at another.

A generic agent can know the domain and still miss the organization's method. Enjamb lets teams encode the sources, instructions, templates, review points, and delivery format that make a workflow theirs. Once the method is approved, the organization can run it again without asking an expert to reconstruct the process every time.

This is where institutional knowledge becomes operational. The useful knowledge is not limited to documents. It includes which evidence the team trusts, how it handles exceptions, what a reviewer expects to see, and why a prior decision was made.

Repeated workflows also expose where the method needs work. If the same exception appears in every run, the problem may be an unclear procedure, a missing source, or an approval boundary that no longer matches how the team operates. The workflow makes that friction visible.

We started Enjamb to give the people developing medicines more capacity. Ten thousand completed workflows have made that idea more precise. Capacity does not come from producing more text. It comes from carrying work through the systems, checks, and handoffs that stand between a question and a decision.

The strongest workflows have a defined finish. They preserve the source trail. They clear routine work without hiding exceptions. They place human review at the point where judgment or authority is required. They also produce something the next person can use: a validated dataset, a reconciled status, a sourced draft, an investigation timeline, or a decision package.

That is what we mean by an AI workforce for biopharma. Enjamb does not sit beside the work and offer suggestions. It takes on the parts a team delegates, works within the organization's permissions and methods, and returns the result with enough context for an expert to decide what happens next.

More than 130 biopharma companies have now used Enjamb to run this kind of work. The next 10,000 workflows will cover different programs, systems, and operating methods. The standard remains the same: the work must be useful, reviewable, and connected to what produced it.