Every clinical operations leader has heard the same directive twice this quarter. Adopt AI, pilot Claude or Copilot, find use cases. The message is consistent. The playbook for what a 30-person clinical ops team should actually have its people do with these tools on Monday morning is not.
MIT Technology Review ran a piece earlier this month on how small businesses are using AI. The case studies were a part-time tutor and a fabric shop. While these examples fit well with the training BC Consulting has been providing through the U.S. Chamber of Commerce Foundation's Small Business B(AI)sics program, they are not exactly rooted in clinical trial operations. The pattern they identified, though, is the same pattern that separates clinical teams getting value from AI from the ones still talking about it. Day-to-day use of AI earns its keep on the routine, time-consuming, low-stakes-on-mistakes work that eats team bandwidth. It does not earn its keep on the strategic, accuracy-critical, high-stakes-on-mistakes work that defines the actual study.
For a small clinical ops team, that distinction is the crux of the conversation.
Consider the work that fills a clinical ops director's Tuesday. There are sponsor meetings to prep for, meeting notes to write up afterward, action items to chase, vendor status reports to read and summarize, document templates to fill out, draft sections of SOP updates to prepare for review, and a backlog of email threads to triage. None of that is the work that wins or loses the study. All of it is necessary. All of it costs hours that could otherwise go to oversight, risk review, and the actual operational judgment the team is paid for.
That workload is exactly where AI is currently good. A trained meeting-notes pass takes a transcript and produces a structured action-item list in under a minute. Much of this is happening already with Copilot or other transcription tools built into Teams and Google Meet. A long vendor status report can be summarized into the three things that matter for the next steering call. A draft SOP redline can be produced from a markup of changes, leaving the clinical lead to focus on the parts that need real judgment. None of this is novel. The point is that small teams who name these use cases explicitly, train people on them, and standardize the workflow get the hours back. Teams that wait for a strategy document or an enterprise rollout do not.
The same MIT piece is equally clear about the other side of the line, and three of its cautions translate directly to clinical with higher stakes.
The first is accuracy-critical work. AI hallucinates. In a tutoring business this means a slightly wrong meeting summary. In a clinical operations team it means a misstated inclusion criterion, a wrong supply forecast, or a fabricated regulatory citation in a draft document that someone almost sends. The rule for clinical is the same as the rule for everyone else, with a sharper edge: anything that downstream readers will treat as authoritative needs a human verifier in the loop. Built right, AI accelerates the verifier. Built wrong, AI bypasses the verifier and creates risk that no efficiency gain pays for.
The second is platform lock-in. The MIT article frames this as a switching-cost concern. In clinical the cost has a different name. A team that builds its workflows around a vendor's chat model and then needs to migrate hits 21 CFR Part 11 validation rework, GxP procedure updates, and audit-trail reconstruction. The right posture for a small clinical team is to use AI tools that are easy to swap, to route prompts through abstraction layers when feasible, and to avoid building business-critical processes on top of features only one vendor offers. What we have stressed at BC Consulting in our AI training sessions is to learn patterns of prompting and designing agents and skills. Those patterns generally apply across any frontier model.
The third is data security. Pasting a protocol section or a patient-identifiable sentence into a public chat interface is a compliance event. Small teams without an enterprise AI subscription default to free-tier consumer tools, which is the riskiest possible posture. The fix is not to ban AI. The fix is to pick a tool with the right data-handling commitments, document what is and is not allowed inside it, and train the team on the line.
What separates teams that get sustained value from AI is not the tool they pick. It is having a written playbook that names the work AI does, the work it does not, the tools the team uses, and the rules for the boundary between them. It also takes a training pass so that every person on the team uses those tools the same way instead of inventing a private workflow.
That playbook does not write itself. The teams that have one usually got there by working through their own operations with someone who has done it before, identified the specific points in their study workflows where AI fits, and codified the standard. BC Consulting builds that playbook with clinical operations teams. Practical, study-specific, and oriented around the work that already fills the calendar.
The directive to use AI is going to keep coming. The teams that benefit will be the ones who picked their spots, trained their people, and resisted the urge to use AI everywhere at once.