Behruz Kholov on AI in Pharma: Where Machines Take Over the Routine, and Where Human Intelligence Remains Essential

In 2024, investments in AI solutions for healthcare exceeded $11 billion, and the pharmaceutical industry emerged as one of the most active adopters of these technologies. Most public discussions focus on drug discovery, the application of artificial intelligence to identify new molecules. But there is a less discussed yet equally significant area where AI is reshaping processes right now: visual communications in pharmaceutical and medical fields. This represents tens of billions of dollars in annual budgets traditionally spent on slow, fragmented content production processes, and it is precisely here that automation demonstrates the fastest return on investment.

However, opportunities come paired with risks. Generative models produce molecular images with distorted ring structures, invent nonexistent receptors, and confuse competitive with noncompetitive binding. In an industry where visual inaccuracy directly affects a physician's trust in a drug, the cost of a neural network hallucination is significantly higher than in any other sector. Behruz Kholov, a pharmacist and multidisciplinary designer, has spent the past year and a half building workflows in which AI and scientific expertise operate together without replacing one another.

The Real Maturity Level: Where AI Stands Today

The industry tends to discuss AI in general terms, as an abstract powerful technology. In practice, it is more important to understand exactly what stage of maturity it has reached and what tasks it can actually perform. The conventional AI progress scale distinguishes five stages: large language models, AI agents, multi-agent systems, AGI, and artificial intelligence surpassing human capabilities. Today's mass applications of AI are limited to the first two stages. Everything else remains in the realm of research and theoretical development, which is why any ambition for fully replacing experts with AI is currently premature.

Today's AI is essentially LLMs, large language models, and the initial level of agents. We have not yet reached multi-agent systems, AGI, or human-level AI. My work is built on the use of LLMs and custom agents, as well as narrowly specialized models such as LoRA, for generating visual content adapted to pharmaceutical and medical brands. I am not trying to replace the human. I am building a platform called AI-Integrated Digital Design & Marketing Ecosystem for Pharmaceutical and Medical Communications, an ecosystem that operates autonomously. Every client, whether a pharmaceutical company, a clinic, a pharmacy chain, a medical representative, or a laboratory, receives their personal Neuron within the platform: a fully customized ecosystem tailored to the brand, with its own visual language, scientific context, and understanding of the target audience. This fundamentally changes the operating model. Instead of a team of five to seven specialists and losses at every handoff, a single specialist working with the platform achieves results that previously required weeks of coordination. One specialist plus one platform, and the task is closed. This is a matter of principle: in pharma and medicine, the cost of error is too high to be entrusted to a machine that does not understand the meaning of what it generates. The platform amplifies the specialist, while the specialist remains accountable for every scientific decision.

Where AI Delivers Manyfold Acceleration: Concrete Coefficients

When AI is deployed within the right zone of responsibility, the effect is measured not in percentages but in multipliers. This applies to tasks whose outcomes can be formally described, visually verified, and validated through automated criteria: brand style compliance, geometric parameters, format characteristics. In this zone, AI and related automation tools reduce execution time by an order of magnitude or more, freeing up expert capacity for tasks that genuinely require human involvement.

Kholov identifies three specific tasks where automation has produced the most measurable effect in his practice. The first is the generation of brand-styled iconography through trained LoRA models. The second is the automated postprocessing of generated images via Photoshop scripts. The third is the assembly of 3D packaging models in Blender. Each of these tasks belongs to the technical category, requiring no scientific judgment, which is precisely why delegating them to machines is both safe and productive.

First, brand-style icon generation through LoRA. Previously, drawing 20 icons manually took two to three days. Now, model training takes 8 to 20 hours, performed once per brand, plus generation of 20 icons in 20 to 30 minutes. Acceleration is by an order of magnitude. Second, postprocessing of generated icons in Photoshop through scripts. Previously, centering, cropping, and exporting 20 icons required two to three hours of manual work. Now the script handles it in two to three minutes. Acceleration is 40 to 60 times. Third, assembly of 3D packaging models in Blender through scripts. Previously, creating a 3D package model with UV mapping and textures required 30 to 40 minutes per unit. Now the script does it in one to two minutes. Acceleration is 20 to 30 times. These numbers matter because they demonstrate that automation in the right zone is not a marketing claim, but a measurable release of resources for tasks where the expert is truly needed.

Where AI Remains Dangerous: Three Red Lines

The same models that deliver manyfold acceleration on certain tasks become a source of serious risk on others. The boundary lies where a task requires understanding of scientific context rather than mere reproduction of visual form. AI operates on statistical patterns and does not engage with meaning, so in any task where meaning is primary, its involvement must be limited and expert-verified. In pharmaceuticals, several such zones exist, and they touch the very core of what communication materials are created for.

The first red line is the scientific accuracy of MoA visualizations. A neural network can produce an impressive animation, but it does not understand the logic of chemistry: for instance, it may draw a six-membered ring instead of a five-membered one. For the pharmaceutical industry, such a 'hallucination' in molecular structure renders the content professionally unfit. The second is the generation of medical texts and context overall. AI is excellent at condensing and structuring, but it is entirely blind to regulatory requirements. It does not know which wordings are mandatory in a particular country and which are inadmissible. Without an expert filter, generated text is merely a set of words that may fail legal or ethical review. The third line is the visualization of molecular interactions. AI often simplifies processes to the point of losing scientific meaning. It may depict a drug simply 'blocking' a receptor while ignoring the type of binding, whether competitive or noncompetitive. For a physician, these are not details but critical nuances that determine the entire clinical picture. Without an understanding of molecular pharmacology, the neural network turns science into decoration.

AI Agents in Pharmaceutical Practice: From Chatbots to Assistants

Most users today perceive AI as an advanced chatbot that answers questions. But the real transformation of workflows occurs at the next level, where AI becomes an agent capable of not only responding but executing actions within real systems. An AI agent works with email, calendars, databases, performs tasks, and remembers conversational context. For a specialist managing several pharmaceutical projects simultaneously, this means hours saved per day and a fundamentally different mode of operation.

Kholov has developed and deployed his own AI agent accessible via Telegram. It understands voice messages and text, retains previous conversations, searches emails, schedules calendar meetings, and manages tasks. The architecture is built on OpenAI's Whisper for speech recognition, LLMs for query processing, and instrumental APIs for action execution. This is not a theoretical concept but a working tool that is already accelerating operational processes in his practice.

The first level of AI is the chatbot: ask a question, get an answer. The second level is AI agents that actually perform actions. They reach into your inbox, calendar, and tasks; they work on your behalf and save hours of your life. You simply say, 'Schedule a meeting for tomorrow at three p.m.' The agent schedules it, confirms it, reminds you. You never even opened the calendar. In my case, this means I can focus on scientific work and design, while the agent handles the entire operational routine. This is not the replacement of the human, this is the removal of actions that do not require human participation.

The Pharmacist's Assistant: AI That Speaks Three Languages

A distinct class of AI agents in the medical field consists of specialized assistants that work with pharmaceutical information and adapt it to the audience. Kholov developed such an agent through his studio, Cr8.Dsgn. The assistant provides information exclusively from verified sources, including PubMed, FDA, and official drug labeling, and does not draw from Wikipedia or forums. It automatically adjusts the complexity of its response depending on who is asking: a patient, a pharmacist, or a physician.

Using the example of a single drug, amlodipine, the assistant provides three fundamentally different responses. To a patient, it explains that the drug lowers blood pressure by relaxing the walls of blood vessels, notes basic intake rules, and lists red flags for consulting a doctor. To a pharmacist, it describes in detail the mechanism of blocking calcium channels, dosage ranges, contraindications, and CYP3A4-related interactions. To a physician, it provides the molecular mechanism with binding to the α1-subunit of voltage-dependent calcium channels, pharmacokinetics with specific half-life values, indications with levels of evidence, and references to key studies such as the ALLHAT Trial. Every fact is accompanied by a source citation.

The main thing to understand about this assistant is that it does not diagnose and does not prescribe treatment. It does not draw on questionable sources. It adapts to the interlocutor, and every fact it presents is backed by a verifiable reference. This is an applied example of how AI can operate safely in pharmaceuticals: with strict boundaries, traceable sources, and no claim to medical accountability. The assistant amplifies the specialist but does not substitute for them.

Custom LoRA Models: Why Standard AI Is Not Suitable for Pharma

Standard generative models such as Midjourney, DALL-E, or base Stable Diffusion are trained on enormous datasets of internet imagery. This makes them versatile but creates serious problems in specialized tasks. They produce molecular structures with distortions, medical illustrations with incorrect anatomy, icons in arbitrary styles, and fail to maintain visual consistency across different materials of a single brand. For pharma, where trust is built partly on subtle visual brand details and on the scientific accuracy of every element, such models are unacceptable.

The solution Kholov uses in his practice is the training of custom LoRA models for a specific pharmaceutical brand. The process works as follows: a set of reference visual elements is assembled, each is described in detail in a language understood by Flux during training, and the model is then trained on servers within one to two hours or on a personal computer within 8 to 20 hours. After training, the model is able to generate new elements that precisely match the brand's visual system, without drifting toward arbitrary "creativity".

created 20 icons in the visual style of a pharmaceutical brand. I exported each in the format required for training. For every icon, I wrote a description: what the icon depicts, what the primary colors are, what the secondary colors are, what exactly is shown, what elements are included, what the background is, what the style is, what the material of the style is. I paired each icon with its description and configured the training parameters. After training, the result is a LoRA model that can generate through a system prompt. Standard models are trained on the internet, so they generate anything. A custom model knows only what it has been trained on, in the style of a specific brand. This is the technical implementation of the zero-distortion principle: the neural network simply has no freedom to generate anything beyond the defined brand.

The Boundary of Responsibility: What AI Does and What the Human Decides

The central question that any specialist integrating AI into pharmaceutical practice must answer is where the line lies between tasks delegated to the machine and tasks reserved for the expert. There is no universal answer, but Kholov has formulated a working rule developed through years of practice: AI is responsible for the "how", the human is responsible for the "what". AI can determine how to generate an icon, how to crop an image, how to assemble a 3D model, how to structure text. But decisions about what exactly to display, which detail to retain, where to place the emphasis, and where the boundary of acceptable simplification lies remain with the specialist who carries scientific expertise.

In his own practice, this boundary is operationalized through a verification system embedded into every stage of the workflow. Generated icons are checked for brand style compliance both visually and through scripts. Molecular structures are cross-referenced against the PubChem database. All scientific assertions, from mechanisms of action to indications and clinical data, are verified against the evidence base: clinical studies, meta-analyses, and systematic reviews in the Cochrane Library, PubMed, and other specialized sources. This is a matter of principle: verification rests on the same evidence-based medicine standards that all medical and pharmaceutical specialists rely on. 3D packaging models are controlled through scripts for geometry and proportions. The final material, in any case, passes through an expert who bears scientific responsibility for it.

AI can be responsible for the 'how', how to generate, how to crop, how to export, how to structure. The expert is responsible for the 'what', what exactly to show, which detail to keep, where to place the emphasis, where the line of acceptable simplification runs. We have not reached AGI. AI cannot grasp the idea inside your head. It predicts the next pixel, the next token, the next word. A pharmacist, by contrast, understands what is happening inside the patient's body at the molecular level. The gap between these two levels of understanding is precisely the boundary I do not cross. And I will not cross it until AI learns to think like a scientist. But even when AGI arrives, the question of responsibility will remain. And responsibility for the life and health of patients I cannot transfer to a machine.

A Systemic Industry Response: Platform Instead of Tool Chaos

All the principles Kholov articulates, the boundary between AI and expert tasks, evidence-based verification at every stage, training models for specific brands, the rejection of standard generative solutions in favor of customized ones, are today consolidated in his own practice into a working methodology. But the methodology of a single specialist, however refined, does not scale to the industry. Pharmaceutical companies, clinics, laboratories, and medical agencies around the world face the same questions daily: how to accelerate content production without losing scientific accuracy, how to maintain visual brand consistency, how to shorten approval cycles, how to integrate AI safely. Every player has its own context, regulatory requirements, visual language, and audience. A universal solution does not exist, but a universal architecture is possible, one within which each client receives a customized environment. This is the task Kholov is addressing at the platform level.

This is precisely the problem I am designing SynaptIQ AI to solve. It is the AI-Integrated Digital Design & Marketing Ecosystem for Pharmaceutical and Medical Communications, in which every client receives their personal Neuron. A pharmaceutical company, a clinic, a pharmacy network, a laboratory, a medical representative: each has its own operating context, its own brand, its own audience, its own regulatory requirements. And each needs an ecosystem that understands precisely its specifics. SynaptIQ builds these ecosystems automatically: a single scientific input passes through the entire chain, from training the LoRA model for the brand to generating the full set of materials, with evidence-based verification embedded at every stage. What in today's industry requires a team of five to seven specialists, weeks of coordination, and substantial budgets is closed in SynaptIQ by a single specialist working with the platform. My goal is not to build yet another tool. My goal is to ensure that pharmaceutical and medical communications cease to be a slow, fragmented, and costly process. That scientific accuracy no longer conflicts with speed, and that speed no longer undermines scientific accuracy. And that the cost of error in communication is no longer paid by the patient, neither in time, nor in money, nor in trust.

The position Kholov articulates unexpectedly aligns with the direction in which global regulators are moving. In January 2025, the FDA published its first official draft guidance titled "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products", a document based on the analysis of more than five hundred AI-involving submissions received by the agency between 2016 and 2023. The key innovation of this guidance is a risk-based credibility assessment of AI models, requiring demonstrated control and verification at every stage of their application. In parallel, the market for AI solutions in clinical research grew from $7.73 billion in 2024 to $9.17 billion in 2025, and is projected to reach $21.79 billion by 2030, with a compound annual growth rate of nearly 19%. The industry is simultaneously accelerating AI adoption and formalizing requirements for human oversight. It is precisely at this intersection, rapid scaling of capabilities alongside tightening verification requirements, that specialists who from the outset build their workflows on the principle of "AI as instrument, expert as accountable party" hold the advantage. The future of pharmaceutical communications, judging by all indicators, lies not with fully autonomous AI but with hybrid systems in which the machine amplifies the human, and the human guarantees the scientific and ethical integrity of the result.

Alexander Brooks
February 15, 2025