People often ask me which AI tool they should use. I wish there were one clean answer, but my own daily routine suggests otherwise. I move between several tools, sometimes within the same hour, because each one is useful for a different kind of job.

I think of AI as a toolbox. The first step is figuring out what you are trying to do. Then you can decide which tool fits, whether you are allowed to use it for that purpose, and how closely you need to check the result.

There is a classic idea in biomedical informatics that captures this well. In 2009, Charles Friedman wrote that “a person working in partnership with an information resource is ‘better’ than that same person unassisted.” It is often written like this:

Person + information resource > person alone

The person is a central part of that equation. We define the problem, give the tool context, notice when something seems off, and decide what to do with the answer. The same AI system can be extremely helpful in one person’s hands, and much less useful in someone else’s (and when used wrongly, it can be flat out dangerous in a third person's). 

The tools I use every day

For quick, everyday questions, I usually start with the free version of Gemini. This is where I go when I need a fast answer, a quick thought, or a starting point that I probably will not revisit. I also use it for preliminary image generation. It helps me get an initial concept on the screen before I spend much time refining it.

When I know I am going to work through something over several rounds, I usually open ChatGPT. I pay for ChatGPT Pro because I use it frequently, although a paid plan is certainly not necessary to start experimenting. I use it to draft emails, work on documents, think through abstract ideas, and figure out how to communicate something clearly. I also use it when I want more control over an image.

The back-and-forth is a big part of the value for me. I may begin with a rough thought, react to what it gives me, explain what feels wrong, and keep going until the result sounds like me. I rarely take the first response and send it unchanged. In fact, this exact article you're reading was mapped out for me by ChatGPT!

At work, I use my institution’s version of Microsoft 365 Copilot. This use is less flashy, but it may save me the most time. Copilot can work with the emails and files I already have access to. Instead of searching through months of email, I can ask, “Who was the person who brought up this project?” or “What did we decide the last time we discussed this?” I can also ask questions about my Word documents, PowerPoint presentations, and Excel files.

A surprising amount of my day used to involve looking for things I knew I had seen somewhere. Copilot helps me find that information without remembering the exact sender, file name, or date.

How I use AI clinically

Clinical use requires a different level of caution. I start by checking whether my organization has approved the tool for the specific use and what types of information I am allowed to enter.

One of my favorite clinical tools right now is Doximity. Weill Cornell Medicine recently made Doximity’s clinical AI tools available to us in approved workflows that allow the use of protected health information.

I use Doximity Scribe for telemedicine visits. I built my own template so that the note comes out in the format I actually want. That customization makes a major difference. I spend much less time rearranging a generic note after the visit.

I also use Doximity Ask. I created a response template for patient-specific questions, including questions about laboratory results. I can provide the relevant clinical context and the results, then use the response as a starting point. I read it carefully, make changes, and remain responsible for the final interpretation that reaches the patient.

Before using the scribe, I tell the patient what I am doing and obtain consent. This is a good habit in any setting. It deserves particular attention in California, where consent and notification requirements may apply depending on how the technology records a conversation or generates patient communications. Doximity’s own terms also require users of its transcription and summarization features to obtain the necessary consents. Every clinician should understand the rules that apply to the tool, the organization, and the state where care is delivered.

For planning visits, I often use OpenEvidence. I may ask what parts of the history I should clarify, what diagnoses should be on my radar, what I cannot afford to miss, or what recent evidence is relevant. This gives me a useful structure before the visit and can point me toward sources worth reviewing.

These are examples from my current routine. They are not endorsements, and I expect my routine to change. AI products evolve quickly. Their features, pricing, and institutional approvals change too.

What the comparison studies can tell us

There is growing research comparing general-purpose models with clinical AI products. A 2026 study inNature Medicine found that frontier models such as GPT, Gemini, and Claude outperformed OpenEvidence and UpToDate Expert AI on the questions and benchmarks included in that study. A later preprint used a different set of real clinical questions and specialty-matched physician reviewers. In that study, OpenEvidence performed better. 

I find the disagreement useful. The studies asked different questions, used different model versions, and evaluated the answers in different ways. Their results show how much performance depends on the task. A leaderboard can be interesting, but it cannot fully predict which tool will fit your day-to-day work.

Experience helps fill that gap. The more you use these systems, the more you notice that one is better at a certain kind of writing, another is better at retrieving your own information, and another gives you a more useful clinical answer. You also start to recognize their recurring weaknesses.

A practical way to start

If you are unsure where to begin, choose one small task that regularly wastes your time. It could be finding an old email, rewriting a patient handout, creating an agenda, planning a teaching session, or getting an initial answer to a clinical question.

Then work through a few basic steps:

1. Check what your organization permits. Know which tools are approved and what information can be entered. Keep patient information out of any system that has not been specifically approved for it.

2. Ask whether AI is useful for this problem. Sometimes a template, a better search, or a simpler workflow will solve the problem more reliably.

3. Start with something low risk. Learn how the tool behaves before using it for work with meaningful clinical consequences.

4. Try the same task in more than one tool. Compare the quality of the answer, the sources, the tone, the omissions, and the amount of editing required.

5. Save what works. If a prompt, template, or sequence consistently helps, make it part of your workflow.

6. Keep checking the output. The amount of verification should rise with the potential harm of an error.

The goal is to get your hands dirty in a safe way. Reading about AI is helpful, but regular practice builds a different kind of understanding. You learn how to phrase a request, when to add more context, when to ask for sources, and when an answer simply does not feel reliable.

That experience will become more valuable as new tools appear. The specific products will keep changing. The ability to test a tool, understand its limitations, and decide whether it belongs in your workflow will carry over.

What comes next

I expect more of this technology to move directly into the electronic health record. Epic and other EHR vendors are already developing tools that summarize charts, draft hospital courses and handoffs, assist with documentation, and suggest billing or coding information. Integration should reduce some of the copying and pasting that happens when clinicians move between separate systems. It can also place AI within the privacy, security, and governance structures that healthcare organizations already use.

Those tools will still need thoughtful users. A summary can miss an important detail. A suggested billing code can be wrong. A polished draft still needs someone who understands the patient and the purpose of the work.

My goal with all of this is fairly ordinary. I want to take better care of patients, communicate more clearly, and spend less time digging through emails, reformatting documents, or reconstructing notes after a visit. If technology can give me some of that time back, I can put it toward work that requires a human being.

Start with one problem. Find an approved tool. Try it, check it, and compare it with something else. Keep the uses that genuinely improve your work. With enough practice, you will build your own AI toolbox, and it will probably look different from mine.

At the annual meeting, we will work through specific examples of where AI helps, where it creates new risks, and how PALTC clinicians can begin using it responsibly in their own work.

Jared Silberlust, MD, MPH
Assistant Medical Director
Physician Organization Information Services
Weill Cornell Medicine