I don’t know much about veterinary medicine, so I decided to pretend I was pitching an AI company that sells software to vets.
I wanted to see what happened if I tried to develop an idea in a field where I couldn’t lean on subject-matter expertise: AI in pet care.
I also wanted to test whether a framework I’ve been building to make better content decisions would change anything.
Given the same research and the same vague topic, how would a chat using my framework differ from one without it?
I gave ChatGPT a real company and vague territory
I chose VetRec, an AI company that sells software to veterinary practices, because there was enough public information to work with without making the exercise a foregone conclusion.
The product itself gave me several places to look.
VetRec’s AI scribe turns consultations into clinical notes and summaries. Its receptionist answers calls and messages, books appointments, collects intake information, and handles other front-desk work.
The company had also published customer stories that offered some evidence about how veterinarians were using the software in actual practice.
I assembled a small research packet from product and blog pages, customer stories, industry research, and competing products. I deliberately avoided telling either chat what I thought the interesting story was.
Then I ran the same first prompt twice.
I’m considering pitching VetRec. I’ve attached some research about the company and its market. The broad territory I’m thinking about is AI in pet care.
What content opportunity do you see here?
One prompt went into a temporary, non-personalized chat with only the research packet. The other went into a project running the content framework.
The plain chat delivered a polished idea
Within one response, the plain ChatGPT conversation had moved well beyond “AI in pet care.”
It noticed that VetRec’s products largely automate work surrounding veterinary care rather than clinical judgment. Then it proposed broader editorial territory: the future of veterinary work.
The angle didn’t wow me. “The future of X” is an obvious choice when you’re trying to turn new technology into a bigger industry story.
From there, it fleshed the angle out fast:
- a POV about reducing work without removing clinical judgment
- supporting themes around economics and adoption
- an annual research report
- competitive positioning
And finally, a line you might see in a pitch deck:
VetRec can become the company that explains what veterinary medicine looks like when clinicians get their time back.
I wasn’t itching to make the series. I was too overwhelmed by how finished the answer felt. A vague topic had turned into something with the shape of a strategy before I’d interrogated anything.
The framework wouldn’t settle yet
The framework chat found almost the same opportunity. It saw more potential in the work surrounding clinical care (documentation, intake, scheduling, client communication) than in speculation about AI replacing veterinary judgment.
Still, this chat treated the idea as a hypothesis. Before building on it, the framework asked who the idea was for, what VetRec could credibly claim, and what kind of evidence the argument would need.
Then it noticed two details that pointed somewhere else.
One VetRec customer, a large specialty and emergency veterinary hospital, had only partially adopted VetRec across its departments until leadership mandated use of the phone integration.
Another VetRec customer, Veterinary Emergency Group, only expanded its rollout after a six-month pilot involving frontline staff, operations, and clinical leadership.
Getting the software into a veterinary hospital was one thing. Getting it embedded in the way people actually worked was another.
Together, these anecdotes opened up a competing story: what makes AI actually stick in a veterinary hospital?
Instead of choosing between ideas, the framework left both possibilities open.
To decide between them, it wanted evidence I couldn’t get from vendor case studies alone—namely, direct input from veterinarian teams about:
- what they were comfortable delegating
- where they drew hard boundaries
- what had to change inside a practice before the technology became routine
That was the first meaningful difference. One response had already packaged its interpretation. The other was still trying to determine which interpretation the evidence could carry.
Then I confessed I can’t interview anyone
The first turn was a clean comparison. Same research packet, same prompt.
After that, I let the conversations develop naturally, so I don’t treat the second turn as an A/B test.
I asked the plain chat:
Of these directions, which one would you actually pitch to VetRec first, and why?
In the framework chat, I asked:
I’m doing this speculatively from outside the company, so I don’t have access to VetRec customers or private interviews yet. What would you do next with public sources?
The plain chat committed to The Future of Veterinary Work and promoted the idea to a “content platform” complete with four pillars—an impressive amount of infrastructure for an idea I’d met five minutes earlier.
The framework changed direction. With public sources alone, the adoption story was easier to support than the broader future-of-work thesis.
There were deployment histories I could reconstruct, practitioner objections to investigate, product changes to trace, and category evidence I could keep separate from VetRec’s own claims.
The future-of-work angle hadn’t suddenly become a bad idea. I just had less reason to believe in it from the outside.
A good answer can arrive before the decision does
A 2025 paper on LLM decision-making called DecisionFlow approaches a very different problem from a surprisingly relevant direction.
Researchers argue that LLMs can generate decisions and plausible justifications without an explicit representation of the decision space connecting the two.
Their framework makes the model reason over structured actions, attributes, constraints, and tradeoffs before producing a decision. On two high-stakes decision benchmarks, they reported accuracy gains of up to 30 percent over strong prompting baselines.
DecisionFlow tests a structured approach to LLM decision-making in high-stakes domains like healthcare and finance. It has nothing to do with content strategy, and it doesn’t validate my framework.
But when DecisionFlow made the model identify options, constraints, and tradeoffs before choosing, accuracy improved. When the researchers collapsed those steps into one long prompt, performance got worse.
That’s close to what I’m doing with my own framework.
I’m not trying to stuff more context into ChatGPT in hopes that a better answer fell out. I want the model to move through important decisions separately.
Could VetRec credibly make this argument? Was I seeing a pattern across veterinary practices or being generous to a few vendor case studies? What happened when a common evidence constraint changed what I could learn?
ChatGPT had no trouble finding a plausible direction. The framework’s value showed up in what happened when the idea had to survive the evidence, VetRec’s authority, and the limits of what I could learn from the outside.
I still don’t have a BS detector for this
I would still want input from people who actually understand veterinary medicine before I get anywhere near a serious piece about the field.
But by the end, I knew what the public evidence could support, where I was making an inference, and what I’d need access to before pushing the argument further.
The experiment also gave me a better sense of where my own knowledge stopped.
Subject expertise gives you context, pattern recognition, and a much better BS detector. Editorial judgment tells you how far the evidence can carry the argument.
AI can blur that boundary. A model can hand you the finished shape of an idea—positioning, pillars, formats—while many of the decisions behind it remain murky.
The plain chat was perfectly capable of finding a strong direction. The framework kept changing how much confidence that direction deserved as the evidence and constraints changed.
Of course, asking the right questions doesn’t require a framework. But embedding those questions into the process gives the model a decision structure built around your actual constraints, rather than letting it supply the decision logic on its own.
AI can come up with something to say. Before I build around it, I want to know what the idea’s done to earn my confidence.
