Harini Gopalakrishnan, Founder of theHaze.ai & Industry GTM, Vespa,ai.gettyAI companies are entering the life science services fray. Anthropic’s recent launch of Claude Science is a good example of that shift, positioning an AI company not only as a model provider, but as part of the scientific workbench itself.For CIOs and CAIOs in life sciences, that changes the architecture conversation. The question is not simply where AI can add value. It is also a question of where to decouple, where to stay modular and where to resist the growing pull toward monolithic platforms. Here are five lessons I believe leaders should pay close attention to in light of these developments: 1. Treat AI Like SoftwareThe early excitement around generative AI often leads enterprises to use powerful models for almost everything. A simple question, a basic summarization task or a lightweight internal workflow may be routed to a heavyweight model because it is available and easy to use.In one customer discussion I had, the expensive model had effectively become the answer to everything—FAQs, summaries and simple lookups like asking for an update on the helpdesk ticket. Sometimes workers used it to synthesize documents or retrieve information.That approach works in pilots. It becomes expensive at scale. It becomes even more expensive if you're feeding it multipage documents all the time across many users. Token costs quietly add up. Many enterprise use cases are better served by retrieval, ranking, metadata, search and deterministic workflows. The best design approach is not to send everything to a large model. Rather, you need to retrieve the right context, use the smallest relevant chunk and reserve reasoning for the moments where reasoning is truly needed.AI needs the same discipline as good software engineering. Don’t overuse a model. Don’t over-engineer a workflow. Don’t call a frontier model when a lightweight model, retrieval layer or rules-based process will do.In real-world deployments, a well-tuned retrieval layer often does more for relevance, latency and cost control than defaulting every workflow to a higher-level reasoning model.2. Beware Monolithic Vendor LockYou must be careful about how much of your stack gets handed to one vendor. Leading models are compelling precisely because they bundle more of the workflow into one environment, including tools for research, data analysis and computing. That may accelerate adoption, but it also raises a real architectural question: Does an enterprise really want the same AI provider to be its platform, embedding layer, knowledge layer and science suite?This also has implications for the broader services ecosystem. Traditional consulting firms and boutique services companies will need to rethink their value proposition. If AI providers bring both the model and the services layer, enterprises will ask a fair question: Who is best positioned to build this workflow?For CIOs and CAIOs, the key is to be clear about what you are actually buying. Are you buying access to a model? Are you buying a workflow? Are you buying domain expertise? Or are you building a long-term dependency?Unless you are truly building a domain-specific model or a deeply differentiated capability, be cautious about large internal builds. The service arms of AI companies are still service arms. They can accelerate implementation, but they should not replace your enterprise architecture discipline.​The safer path is modularity. Keep the model layer, retrieval layer, semantic layer and integration layer loosely coupled wherever possible. That way, a change in model provider or an infrastructure decision will not trigger a rewrite across the rest of the estate.3. Be Pragmatic About SemanticsThe AI services excitement has renewed interest in terms involved in semantic modeling, ontologies and knowledge graphs, especially when it comes to medical vocabularies, including the Medical Dictionary for Regulated Activities, the Systematized Nomenclature of Medicine, etc. This pattern is not new. Synonym resolution, crosswalks across medical vocabularies and subscription-based public ontologies have existed for years. What is new is the scale of the aspiration. These systems aim to connect common vocabularies, graphs, data catalogs and enterprise metadata into one coherent semantic layer. The challenge is that a given business rarely has the time or incentive to curate metadata at that scale​. Pharma has learned a hard lesson from years of enterprise data cataloging: Global alignment is difficult. Instead, think in terms of smaller, business-owned semantic models. Use reference data where it exists. Build lightweight ontologies for specific use cases. Let business teams create the metadata that helps AI understand context. Over time, AI itself may help generate and maintain semantic models.​4. Move AI Closer To The BusinessBusiness teams are no longer passive recipients of technology. They understand AI, they experiment with models and, given the right access, they can quickly build prototypes and applications themselves.This means the traditional boundary between IT and business is changing again.The old idea of “shadow IT” may resurface, but resisting it is not the answer. The better answer is controlled enablement. Give business teams access to validated models, approved data environments and safe development patterns. Create governance councils that validate models, assess risk and guide adoption. But do not turn IT into a firewall that blocks experimentation.AI applications cannot be built in isolation from the business. There is no such thing as a useful generic enterprise RAG system if it does not answer the specific questions that medical affairs, commercial, clinical, regulatory or manufacturing teams actually ask.The closer AI teams sit to business workflows, the more useful they can become.​5. Keep Regulatory, Validation And Deterministic Workflows In FocusAI has enormous potential across pharma: productivity improvement, content authoring, call center optimization, CRM workflows, medical affairs intelligence, clinical trial acceleration and drug discovery. But not every part of pharma can move at the same speed.In clinical, manufacturing and any regulated environment, compliance remains non-negotiable. Validation, qualification, documentation, traceability and reproducibility still matter.AI is not fully deterministic. That matters in regulated workflows where repeatability is required. There will be interesting developments as newer models and agentic systems enter validation-heavy areas, but for now, CIOs should be careful. In many regulated environments, traditional data engineering, deterministic systems and plain old workflow controls remain essential.​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?