Phone Icon1-919-645-1070
Phone Icon1-919-645-1070
SCHEDULE DEMO

Agentic AI and InstantGMP: What Works Now

Home Blogs Agentic AI and InstantGMP: What Works Now

Agentic AI and InstantGMP: What Works Now

Posted on by Rick Soltero
July 1, 2026

Before I started really using AI in depth, I wasn’t sure what it could do besides answer the same questions I put to Google, only with more detail. Now that I have used Claude in connection with an MCP server connected to my InstantGMP database, I have a whole new perspective, and I have watched it change the way our quality operations actually run.

So what does it look like to use AI inside a quality system, and more importantly, is it compliant with the regulations that govern our operations? To understand the shift that has happened over the last twelve months, it helps to start with the difference between AI that answers questions and AI that takes action, which is what people mean when they talk about agentic AI.

From Chatbot to Agent: The Shift That Matters

For the first few years of the AI era in GMP software, most chatbot tools offered the same experience: you typed a question and the system returned an answer. That was helpful, but it was not transformative. A quality team managing open deviations, a backlog of CAPAs, and an upcoming FDA inspection does not need a smarter search bar.

Agentic AI is different because an agent does not just answer one question. It executes a sequence of steps on your behalf, pulling from multiple data sources, applying logic, and returning a synthesized result, so you can ask one thing and the system does work that would otherwise have taken your team hours or days.

In a GMP facility you can ask something like this: “Show me all deviations from the last twelve months grouped by product, equipment, operator, material, and root cause.” Instead of a list of records that you then export to a spreadsheet and sort yourself, you get a structured analysis across every relevant dimension. Or you can ask for a chart of all batches produced by month, and the system queries the entire batch record system and returns an interactive chart that can show batches for a specific year or across all years, grouped by month, product, or status.

Those are not hypothetical examples. Those are queries running against live InstantGMP data today, and they are the basis of the AI Insights features built into the platform.

Where Agentic AI Earns Its Place

The manual processes that currently consume the most analytical time in a GMP operation are exactly the ones where agentic AI delivers the most value.

Quality log analysis is a good example. The traditional workflow is to pull the deviation, CAPA, complaint, and training log reports, export them to Excel, build pivot tables by product and equipment, look for the pattern, and then write the summary, and that process can take a day or two. With agentic AI connected to your quality management system, the same analysis takes minutes, and the analyst’s job shifts from assembling data to interpreting it and deciding what to do about it, which is where their expertise belongs.

Inspection readiness is another area where the difference is immediate. We are adapting AI as a tool to improve our electronic software solutions and our quality management systems, and inspection readiness is where I have seen the clearest gains. I have been involved in enough regulatory inspections to know what the weeks beforehand look like, with teams manually compiling status reports, someone hunting down training records, someone else checking calibration logs, and a third person verifying the status of every open CAPA. It is a form of regulatory archaeology. An agentic AI connected to all modules of your quality system, from electronic batch records and inventory through document management and learning management, can assemble a complete readiness summary on demand, covering open deviations with age and status, CAPAs past their target date, batch records sitting in QA review longer than thirty days, and personnel with overdue training, all from one query.

Batch record status meetings are another casualty of the old approach. Production and quality teams routinely hold status meetings to answer a single question, which is where we stand with the open batches. Agentic AI makes that instant, because asking which batches are currently under QA review and which have been open more than thirty days returns a precise answer that includes the specific approval steps blocking advancement and how overdue each one is. That is information your team can act on immediately, without a meeting.

The Compliance Question

Here is where most quality professionals pump the brakes, and reasonably so. How do we validate this, and what does FDA think?

Those are the right questions, and I want to be careful about how I answer them, because the honest answer in the middle of 2026 is that the regulatory direction is much clearer than it was a year ago while the rules themselves are not yet final.

In Europe, the European Commission published a draft Annex 22 on Artificial Intelligence in July 2025, alongside a revised Annex 11 and an updated Chapter 4 on documentation. Public consultation closed in October 2025, EMA convened a multistakeholder workshop at the end of June 2026 to gather expert input on guardrails and mitigation measures, and a final text is expected later in 2026. Until that text is adopted, Annex 22 is a strong signal of regulatory direction rather than an enforceable requirement, but it is a signal worth reading closely, because it is considerably more restrictive than much of the AI marketing in our industry would lead you to believe.

The draft applies to AI used in critical applications, meaning applications with direct impact on patient safety, product quality, or data integrity, and within that scope it permits only static, deterministic models: models that are locked after training and that return the same output when given the same input. Dynamic models that continue to learn during use are excluded from critical applications. So are probabilistic models. And so, explicitly, are generative AI and large language models. The reasoning is a change control argument at heart, because a model that keeps changing its own behavior after deployment introduces uncontrolled change, and uncontrolled change is incompatible with the validated state that GMP is built on.

What the draft does allow is the use of generative AI and large language models in non-critical applications, provided a qualified person is responsible for evaluating whether the output is suitable for its intended use. That is the human-in-the-loop principle, and it is carrying a great deal of weight in the current draft.

In the United States the picture is less codified. There is no AI carve-out from 21 CFR Part 211 or Part 11, so if an AI tool generates, modifies, or archives electronic records in a GMP context, the existing audit trail, electronic signature, and data integrity requirements apply exactly as they always have. FDA issued a draft guidance in January 2025 on the use of artificial intelligence to support regulatory decision-making for drug and biological products, setting out a risk-based credibility assessment framework, and FDA and EMA jointly published Guiding Principles of Good AI Practice in Drug Development in January 2026. On the industry side, ISPE released its GAMP Guide: Artificial Intelligence in July 2025. None of these is a GMP manufacturing rule in the way Annex 22 will be, but taken together they point in the same direction: risk-tiered expectations, documented human oversight, and validation effort proportional to how consequential the output is. We have pulled the whole picture together on our AI Systems Validation page for anyone who wants the detail in one place.

FDA has also started enforcing. In April 2026 the agency issued a warning letter to Purolea Cosmetics Lab citing, among numerous other CGMP violations, the use of AI agents to create specifications, procedures, and master production records without adequate review before use, a failure the agency tied to 21 CFR 211.22(c) and the responsibilities of the quality control unit. What makes that letter worth reading closely is not the technology failure but the human one, because when investigators asked about the compliance gaps, personnel said they had not known about certain requirements since the AI agent had not raised them. AI answers the question you ask it. It does not tell you what you forgot to ask.

Where a Query Layer Fits

So where does an agentic AI that reads your GMP records land within that framework? The honest answer is that it most likely sits on the non-critical side, and that the boundary has not yet been drawn precisely enough for anyone to claim certainty.

Here is the argument for the non-critical reading. In the InstantGMP MCP Server, the AI does not create records and it does not make release decisions. The underlying data, meaning the deviation records, the CAPA statuses, the batch approval steps, was entered and approved through validated workflows with electronic signatures and audit trails. The AI queries that data through our validated API layer and presents it in a synthesized view. It is a lens, not a data entry point. Human review remains the control, and any action taken in response to what the AI surfaces goes through the same documented workflows and approval chains it always did.

That argument is sound, but it is not a free pass, and I want to be direct about that. The moment a synthesized view starts driving a quality decision, you are considerably closer to the critical side of the line than the architecture diagram suggests. Hybrid systems that combine deterministic retrieval with a probabilistic language layer are precisely the boundary case that industry commenters have asked EMA to clarify, and that clarification has not arrived. Anyone telling you their agentic AI is definitively outside the scope of Annex 22 is telling you more than the current draft supports.

What I would say instead is this. Grounding matters enormously, because a language model constrained to cite records from your validated system behaves very differently from one generating text freely, and that distinction is exactly what FDA’s own Chief AI Officer pointed to when defending the agency’s internal AI tool against reports of hallucinated citations. Traceability matters just as much, because if the system tells you a vendor carries quality risk, you need to be able to drill down to the specific deviations and inventory records that produced that conclusion. Black box summaries are not audit ready, and explainability is something regulators are increasingly going to ask for rather than something you get credit for volunteering.

And you still have to validate it. We built a protocol for our AI features rather than simply asserting they are validated, comparing AI-generated output against reference data pulled directly from the application, defining acceptance criteria before testing rather than after, running each query repeatedly to confirm the factual content stays consistent, including negative test cases to confirm the system reports that no data was found instead of fabricating something plausible, and documenting in advance which changes trigger a revalidation run. The full protocol, including the change control triggers for model version and MCP server updates, is written up on our AI Systems Validation page, and it sits on top of the GAMP 5 validation foundation the platform already ships with. A good architecture reduces the scope of what you have to validate. It does not remove the obligation.

One more point worth stating plainly, because it is easy to lose in a vendor conversation. Under the draft Annex, accountability for AI used in your GMP operation stays with the manufacturer. Vendor documentation supports your validation, but it does not transfer the responsibility. That applies to us as much as to anyone else.

What to Look for in a Vendor

Not all AI-powered GMP software is the same, and here is what actually separates productive AI from marketing language.

Ask whether the AI has access to your full data or only one module, because a deviation analysis that cannot see inventory data cannot tell you which deviations are linked to specific incoming lots. Look for AI that operates across electronic batch records, quality management, inventory, and document management simultaneously, since that cross-module view is the whole point.

Ask whether the AI is querying your validated data directly or running a separate model that might fabricate details, because structured queries against your actual records are both more accurate and more defensible in an inspection, and the AI should function as a query layer rather than a second database.

Ask whether you can trace the records behind any AI-generated insight, and then actually drill down until you reach the source records that produced it.

Ask whether AI-assisted action passes through your existing approval workflows, because if a vendor’s AI can modify records without going through validated change control, that is a validation problem rather than a productivity gain.

And ask to see the validation protocol and the results, not the attestation. “Our AI is validated” is the beginning of a conversation, not the end of one.

The Point

Agentic AI for GMP operations is not a replacement for your quality professionals. It is a replacement for the data assembly work that currently occupies a significant fraction of their time, whether that is deviation trend analysis, inspection readiness summaries, batch status reviews, or vendor quality risk assessments. Those are the weekly tasks every quality department runs, and they currently require skilled people to do work that software can now do in seconds.

InstantGMP is a global manufacturing platform built for regulated industries, serving pharmaceutical, biotech, dietary supplement, medical device, and food and beverage manufacturers who need GMP-compliant electronic batch records, inventory management, LIMS, and quality management systems. The agentic AI built into InstantGMP PRO reflects the same principle that has guided everything else we have built, which is that compliance should not require heroics, and that the right information assembled from across your system and available in a single query lets your team spend their time on the work only they can do.

The technology works. The regulatory framework is taking shape but is not finished, and the manufacturers who come out ahead will be the ones who deploy carefully now and document what they did, rather than the ones who wait for final guidance or the ones who assume it will not apply to them.

InstantGMP provides cloud-based GMP software including electronic batch records, QMS, inventory management, and LIMS with agentic AI analytics. Schedule a demonstration or learn more at instantgmp.com.

Sources and Further Reading

share