How to define the requirements for an AI agent?

by Ian Spektor

I don’t think I need to start off this blogpost by convincing you some patient-facing interactions — that you have highly skilled, and highly paid, people carry out today — can be improved, scaled, and fully automated by an AI agent by this point, do I?

If you’re here, you’re well past the point of evaluating whether it can be done, or even whether it makes sense for your business. It does.

You’ve probably even mapped out exactly which interactions you want to automate!

And whether you’re automating something as simple as an intake form or as complex as an end-to-end pre- and post-surgery suite of consultations, check-ins, and reminders, you’ll be haunted by one same question: “How do I define — precisely — what the AI should do?”

Agenda

At Puppeteer, we’re working with clients in each corner of the healthcare industry on automating the most diverse array of patient-facing conversational flows, both text- and voice-based. This guide provides a peek into the process we carry out when onboarding a new client.

It lays out each and every aspect you should be taking into account to properly define the requirements for an AI agent:

  • Identity, scope, tone and style
  • Patient journey & interaction points
  • Content guidelines and knowledge base
  • Example conversations

By thoughtfully defining these elements, you can be sure that everyone on your team (and your AI provider!) is on the same page on what the agent should, and shouldn’t, be.

Let’s jump right into them.

Identity, scope, tone and style

These are probably the simplest — both to define and to implement — but also the most important ones to get exactly right. Your patients will be communicating with your agent, and you want them to receive at least the same level of treatment they would if interacting with one of your employees.

Identity

Giving the agent a human name, like Jim or Martha, can go a long way in making it feel human and warm — although something along the lines of CareCompanion or Mr. Robot can still accomplish the same goal, without it seeming like you’re trying to pass your AI as an actual human being.

Let’s stick with Jim as our agent’s name for the remainder of this post.

Scope

Describe who your agent is, and what the main actions it should perform are.

For Jim, that could be “you’re an administrative assistant at Best Clinic”, and “you’re in charge of reaching out to patients to coordinate a first meeting, and periodically check on them once they’ve started treatment”.

Tone and style guidelines

  • Should your agent be professional, friendly, empathetic, or a mix?
  • Should it use formal language, or be more conversational?
  • Should it use emojis?
  • Should it use contractions and coloquial expressions?

All of these can work — it just depends on your company’s context and needs, and the specific workflow you’re trying to carry out.

Patient journey & interaction points

Alright, we’re done with the basic stuff.

Here’s the thing: anyone can deploy a “custom chatbot” as a widget on their website if all they need is to customize its tone and style. There’s a trillion companies that do just that. And if all you need is to have it answer some basic questions on your company’s data, they’ve got you covered.

What not anyone can do is have an agent carry out a complex, end-to-end patient journey, with a pre-defined structure of interactions, both proactive and reactive. And what use is an agent that should perform periodic check-ins after a surgery, if it will only show up when the patient themselves decide to talk to it?

List of interactions

Real value lies in automating proactive communications — check-ins, notifications, reminders, follow-ups, first-time reach outs, you name it.

All of these can be set up, and should therefore be properly defined.

  • First interaction: When, and how, does the patient begin interacting with the AI?
  • Scheduled interactions: Should the AI initiate conversations proactively? If so, when and how often?
  • Event-triggered interactions: Should the AI initiate conversations based on certain triggers? Think events firing off in your CRM, actions happening in your app, or the patient asking to be notified of something later on.
  • Patient-initiated interactions: How should the AI respond when a patient reaches out spontaneously? Should the AI assist with topics outside its predefined scope?
  • Last interaction: How should the AI handle end of support? How and then should it inform the patient that the service is ending? Should the patient still be able to initiate conversations with it after that?

Interaction detail

Additionally, for each point of interaction, all of these should be laid out in detail.

  • Timing and context: When does this interaction occur? Under what circumstances? Which conditions should be met for it to happen?
  • Purpose and goals: What is the AI supposed to achieve during this interaction?
  • Conversation content: What topics should the AI cover? Are there specific guidelines or materials the AI should base the conversation on?
  • Tone and style specifics: Any adjustments to the general tone and style for this interaction?
  • Frequency and duration: Is this a one-time interaction or recurring? How long should it last?

Content guidelines and knowledge base

Congrats! Jim now knows exactly when and how it should address the patient.

It still doesn’t have the slightest idea on what your company does, what services it does and doesn’t offer, or what guidelines or medical information it should use, though. Oops.

Source materials

You must define and provide any and all material you need the agent to have access to during the whole patient journey.

Providing documents that thoroughly describe your company and its services is recommended.

Additionally, most healthcare providers have specific medical knowledge that the AI should turn to before falling back to its broader knowledge base, which should also be provided — think the latest research on GLP-1 for a weight loss assistant, or specifics about the duration and costs of CPAP treatment for a sleep apnea clinic.

Speaking of its broader knowledge base - what exactly is that?

By this we mean the general knowledge it has gained during pre-training, i.e., the default knowledge that models come out-of-the-box with from your LLM provider of choice.

Allowing the AI to access this knowledge is generally OK — but can also be very much not OK, depending on your use case and how costly or harmful relaying incorrect information to the patient would be. That’s one more call for you to make!

Topics to address or avoid

We’ve defined the topics and goals the agent should touch on during each interaction, but it’s still a pretty good idea to tell it what topics it should definitely not touch on.

Some sane defaults are, for example, to tell the agent not to provide actual medical advice or claim it’s a doctor, and avoid political or religious discussions.

On top of these prohibited ones, some sensitive topics need to be handled with extra care — such as mental health concerns, complications, or self-harm. The AI should have a list of these to be able to act appropriately, and raise an alarm for a human professional to take over (or fire off some other escalation protocol) if needed.

Example conversations

You’re done! Each and every aspect of your agent is now precisely defined.

… which makes its behaviour about 10% less ambiguous than before we started.

Your AI provider and yourself can be looking at one same document containing all of these requirements, be sure to fully understand it, and be imagining two completely, and I mean absolutely, different conversations.

There’s no easier fix to this than providing a couple of example conversations for each interaction point! This will make sure you’re all on the same page before starting to set up the agent.

Make sure these conversations include patients taking on different roles — such as a fully compliant one, a dubitative one, and a completely erratic one that just won’t follow along with the agent’s goals.

Conclusion

OK! Now you’re done.

Well — done defining the requirements for the first version of your first AI agent.

So, just getting started. But definitely on the right path!

I hope this guide will prove useful in your journey towards making AI actually work for you and your business.

If you’d like to learn more on how a specification like this can be turned into a working, fully-managed, text- or voice-based agent in just a couple of days — we're just an email away.