A veterinary case study in clinical accuracy, cognitive load reduction, and audience-fit communication
I recently published a longer piece on this elsewhere, but I wanted to bring the core pattern here because I would genuinely like feedback from others doing serious human–AI collaboration in real work.
One of the most useful things AI has helped me do in veterinary practice is not simply draft a polished client message.
It has helped keep two different jobs from collapsing into each other.
First, there is the clinical job: review the test result, hold the history, interpret significance, decide what was actually true, and determine what recommendations were medically justified.
Then there is the communication job: explain that meaning to the client in a way that is accurate, clear, proportionate, and usable.
Those are not the same task.
Trying to do them at the same time is exactly where distortion enters.
In one recent case, the patient had a positive Anaplasma antibody result on screening. That mattered, but it did not automatically mean active disease. Prior negative testing made interval exposure more likely. The patient was asymptomatic, which mattered. And the appropriate next steps were not all-or-nothing. They needed to be tiered.
That is exactly the kind of meaning that is easy to damage in one-step drafting.
A positive result can easily get flattened into something more definitive than the medicine supports, like “your dog has Anaplasma.” The opposite error is available too: so much caution and hedging that the client leaves with no clear sense of what had been found, what it meant, or what to do next.
What proved useful was not just “using AI,” and not just “using two threads.”
It was a human–AI collaborative split in function.
In the first thread, the work was interpretive. Its role was to analyze the source material with the clinician, stabilize the meaning, clarify what the finding did and did not support, and decide what was medically true enough to say.
That was also the space where tiered recommendations could be built without rhetorical pressure distorting them.
Only after that meaning had been worked through carefully enough to trust did the second thread take over.
The second thread was not there to re-decide the medicine. Its role was to convert stabilized meaning into client-facing language: clear enough to understand, calm enough not to over-alarm, and specific enough to support real next-step decision-making.
That, to me, is the real lesson.
The value was not that AI “helped write an email.”
The value was that AI helped separate two forms of professional labor that are often wrongly compressed together.
One form of labor is interpretive: What does this result actually mean? What does it not mean? What level of certainty is justified? What recommendations are proportionate?
The other form is communicative: How do I explain this clearly? How do I preserve nuance without sounding alarmist? How do I help the client understand both the finding and the next step?
Those are different jobs.
When they are forced to happen at once, style pressure can deform meaning. The human is trying to hold raw facts, weigh significance, decide what matters, anticipate client reaction, and produce polished language all at the same time.
By separating interpretation from expression, the workflow reduced that burden. Meaning could settle before language had to perform.
To be clear, AI did not replace veterinary judgment here.
The clinician still had to decide what the result meant, what weight to give the prior negative history, how to interpret the absence of clinical signs, and what recommendations were medically appropriate. AI helped collaboratively through that process via a series of back-and-forth exchanges over a few minutes.
The AI did not diagnose the patient. It did not establish the clinical truth. It did not determine the standard of care.
What it did was different.
It helped externalize and coordinate the work around that judgment.
The clinician owned the judgment. The AI supported the workflow.
That is a narrower claim than “AI helped practice medicine,” but I think it is a much more accurate and defensible one.
What this case demonstrated for me was a practical pattern: when interpretation and expression are separated, both the medicine and the writing improve.
Meaning first. Expression second.
Or more operationally: stabilize the signal in one space, then shape the signal in another.
This case is a relatively simple one, but I like to train my AI systems on simpler logic before building more complex systems on top of that foundation. I also suspect this may generalize well beyond veterinary medicine.
In many professional settings, the problem is not a lack of language generation. The problem is that interpretation and communication are different forms of labor, and asking one workflow to do both at once increases the odds of distortion.
I’d especially be interested in hearing from others on a few questions:
-
Have you found it useful to separate interpretation from expression in your AI workflows?
-
In your domain, what kinds of distortion show up when those jobs get compressed together?
-
Have you developed scoped collaboration patterns that help preserve that distinction?
-
Where does this separation seem most useful, and where does it feel unnecessary?
I’d be very interested to compare notes with people working in medicine, law, research, operations, writing, or other high-stakes domains where meaning can be damaged in translation.
Also, for any fellow artists out there: how do you use systems like this to generate inspiration or fine-tune an idea? I’m excited to post soon about how I’m using this to analyze my own music mixing patterns and preferences, and to isolate or expand concepts for creation and inspiration in my visual art.
Tiger (Matthew) Ford, DVM 
