Questions answered from the source
Operators, analysts and study teams get the answer from your own procedures, methods and protocol, with the document, the version and the step cited.
Life sciences and pharma · AI agents inside the business
Life sciences agents. A person decides. Data stays.
Katonic runs AI agents inside your own data centre or private cloud. Agents read your procedures, your batch records and your safety reports, prepare the work, and stop for a named person.
No agent signs a record, releases a batch, closes a deviation, submits to a regulator or decides what caused an adverse event. It prepares, and a person decides.
Batch, patient and safety data stay on infrastructure you control.
Batch, patient, safety and unpublished product data stay on your infrastructure
A model you serve
Each person's permissions kept
No agent signs a record or releases a batch
Manufacturing, quality, safety and regulatory work run on reading, checking and drafting, and each piece ends with a named person's review. That is where an agent helps most, and where it has to be held most firmly.
One workflow template we opened takes in adverse event reports and triggers a regulatory submission for serious cases, with no step at which a person approves.
Running on the premises is not enough. What matters is which tool waits for a person, whose identity the agent uses, and what is recorded.
Batch records, deviations and out-of-specification investigations had no ready-made agent in the template libraries we searched.
Tools must be chosen one by one: of 2,719 operations in 31 public life sciences interface descriptions, 1,214 are sent as changes, and the method alone cannot tell which of them only read.
Batch records, patients' health data, safety cases and unpublished product data are confidential or personal. They have to stay inside the company.
What the rules ask
A named person reviews and clears what an agent produces. A signature belongs to one individual. A qualified person certifies the batch. The company validates what it uses.
Agents read and draft. People keep every decision about the batch, the medicine, the patient and the member of staff.
Questions answered from the source
Operators, analysts and study teams get the answer from your own procedures, methods and protocol, with the document, the version and the step cited.
Work prepared for a decision
Batch records, deviation reports, adverse event reports and submission documents are checked and set out for the person who reviews them.
Open items in one list
Corrective actions, protocol deviations, gaps in a trial file and calibrations due are listed and counted, with the records behind each.
A record of every step
Who asked, which tool ran and what was decided.
From the production floor and the laboratory to quality, regulatory affairs, safety, clinical operations, medical information and supply.
Manufacturing and quality
Batch records, procedures, deviations, changes
Laboratory and validation
Results, methods, stability, test documents
Regulatory and safety
Submissions, labelling, cases, literature
Clinical, medical and supply
Trial files, visit reports, replies, suppliers
Each is a use case a drug maker, a biotech, a contract research organisation or a laboratory can build in Katonic Studio. A name with a link has its own page.
Every agent follows one rule: it prepares, and a named person decides.
Go deeper
Guides and answers
A life sciences company that wants AI agents on its own infrastructure has more than one kind of product to choose from. The products sold for this fall into four types: agent platforms you run yourself, infrastructure platforms, automation platforms with agent features, and developer frameworks. Our comparison describes fifteen of them with the same fields, read on each vendor’s own pages: where each runs, who operates it, governance, price and limits. For validated work, look for a record of each agent version, the test results behind its release and who approved it.
On-premises AI agent platforms compared: 15 options by type →An agent reaches your documents and systems in one of four ways, all through the same tool gateway and the same approval rules.
Connection · 01
A batch record, a deviation report, a source document for an adverse event, a site visit report.
Connection · 02
Standard operating procedures, specifications, validated methods, the approved product information, the protocol and regulatory guidance, indexed with each reader's permissions kept.
Connection · 03
↗Give Katonic the description of an interface you already run and choose which operations become tools. This is the route to quality events, documents, laboratory results, submissions, safety cases and trial documents.
Connection · 04
↗Servers from the catalogue, deployed inside your own cluster, for the data warehouse, reporting databases and web search of the published literature.
System names are left out on purpose. You supply the address and the sign-in. Batch, laboratory, quality, safety and trial records reach an agent read only, or through an imported API whose change operations are left out or held for a person: no agent writes a regulated record on its own.
It runs on the company's own infrastructure or private cloud, with a model the company serves. Batch, patient, safety and unpublished product data stay there.
Every tool call passes the company's policy. A policy can allow a tool, record its use, hold it for a person's approval, or block it, for the whole organisation or for one team.
A tool that changes something waits for approval unless the company has decided otherwise for that tool. That is how no record, case, reply or filing changes unapproved.
No agent signs. Review, approval and release signatures are made by the person, in the system of record, under the company's own procedure.
The agent acts as the person using it. Where the company's API checks the caller, Katonic passes the signed-in user's identity, and the agent cannot claim another.
Every call is recorded, with who asked, which tool, and what was decided, and an agent, a tool or a whole server can be switched off at once.
From visual workflows to full code: pick the path that fits your team and project.
Visual Builder
Follow guided steps to build agents. Configure each stage through the UI. No code required.
Best for
Quality and regulatory specialists
Conversational Builder
Reviewers describe how they check a batch record or prepare a deviation report. AI generates the agent. Developers refine together.
Best for
Quality specialists + Developers together
Full IDEs
JupyterLab, VS Code, RStudio with GPU access, pre-installed frameworks, and direct deployment.
Best for
Data Scientists & ML Engineers
The rules on manufacturing practice, electronic records, drug safety and AI differ by region and agree on four things. Each is a matter of how the system is built, and of what the company itself decides.
A person answers for the output
The European Commission's draft Annex 22 to its rules on good manufacturing practice, still a draft in November 2026, says large language models "should not be used in critical GMP applications", and that in other use a trained person stays responsible for what the model produces. In a warning letter of 2 April 2026 the United States Food and Drug Administration wrote that the output of an AI agent "must be reviewed and cleared by an authorized human representative" of the firm's quality unit. Here every agent supports a named reviewer. None accepts or rejects anything, and a tool that changes something waits for that person.
A person signs and releases
United States rules on electronic records, 21 CFR Part 11, say each electronic signature "shall be unique to one individual". European rules on computerised systems, EU GMP Annex 11, say a system "should allow only Qualified Persons to certify the release of the batches". Here no agent signs, releases or certifies: review, approval and release are made by the person, in the system of record, and the agent acts only as the signed-in user.
A person decides the safety case
European guidance on pharmacovigilance, GVP Module VI, asks four things of a valid case: an identifiable reporter, an identifiable patient, a suspect reaction and a suspect medicine. A serious case is reported "in no case later than 15 calendar days after initial receipt". Here an agent sets out the four elements and flags what is missing. A safety officer decides validity, seriousness, expectedness and causality, and the safety team keeps the clock and makes every submission.
The company validates what it uses
European AI law applies its high-risk rules from 2 December 2027 for the uses in its Annex III and from 2 August 2028 for AI in the products in its Annex I, under Regulation (EU) 2026/1744, and asks that such AI can be overseen by people. Whether and how an agent is validated for its intended use is the company's own decision under its quality system. Katonic runs on infrastructure you control, with a model you serve, including with no connection to the internet, and records every call for that decision.
Katonic AI is certified to ISO 27001 and holds no other certificate. Validating an agent for its intended use is the company's own work under its quality system. Draft Annex 22 may change before it is final. This is a summary of published rules, not legal advice.
The company is made up. On the production floor, an operator asks what the procedure says about a step, and the agent answers with the document, the version and the step cited. After a batch is made, an agent reads the batch record against the master record and lists every missing entry and every value outside its limit for the quality reviewer. In the safety team, an agent screens the results of the literature search and sets aside the articles that may hold a case, with the passage for each, for a safety reviewer to read. Each one reads what the company already has, changes nothing by itself, and leaves a record.
Seven questions to ask any supplier before an agent touches a regulated record, with our answers.
On infrastructure you control: your data centre or your private cloud. The model is one you serve there.
What the person using it can read. Knowledge sources keep each reader's permissions, and where your API checks the caller, Katonic passes the signed-in user's identity.
Only what you allow. A tool that changes a record, a case, a reply or a filing waits for approval unless you have decided otherwise for that tool.
A named person. No agent releases a batch, closes a deviation, submits to a regulator, decides whether an adverse event is valid, serious, expected or caused by the product, or decides about a patient, a trial participant or a member of staff.
No. Review, approval and release signatures are made by the person, in the system of record, under your own procedure. Regulated records reach an agent read only, or through an imported API whose change operations are left out or held for a person.
Who asked, which tool ran and what was decided, for every call. Whether and how an agent is validated for its intended use is your decision under your own quality system, and these records are what the platform supplies to it.
At once. An agent, a tool or a whole server can be switched off by your administrators.
A GMP procedures assistant can run on the first day, because no system of record is touched. Talk to our team about building it on your infrastructure.