Welcome! This training is designed to get you up and running with Compass. It’s meant to be used while you’re actively working inside Compass, so if you haven’t already, make sure you have access to a Compass environment before diving in — otherwise, this will be a lot of reading about buttons you can’t press yet.
1. What Compass Is and Isn't
Before we start clicking around though, let’s take a minute to understand what Compass actually is — and maybe more importantly, what it isn’t.
At this point, you’ve probably heard more about artificial intelligence (AI) than you ever wanted to — in social media, the news, professional conferences, and possibly at dinner. AI can be broadly divided into two categories. The first is chatbots, which is what most people picture when someone says “AI.” You open up Claude or ChatGPT, ask it something, and it answers. Simple enough. But the chatbot doesn’t have access to your data or your environment; it only knows what you explicitly tell it.
The second category is AI agents. Unlike chatbots, agents are furnished with tools that let them operate directly within your environment. Compass is a custom-built AI agent equipped with tools to query the Ursa database and look up metadata in the Ursa data model — which is what makes it useful for healthcare data analytics specifically. In practical terms, this means you can give Compass a set of instructions, and it will go off and work autonomously for a while, deciding which tools to use and when to come back to you with results or questions.
Compass is currently built on Anthropic’s Opus models — the same underlying models that power the Claude AI products. That said, Compass is specifically designed for healthcare data, so it’s a distinctly different beast from general-purpose Claude in terms of what it can do and how it’s configured.
Alright, enough theory. Let’s actually use it, and the similarities and differences will become clear quickly.
2. Landing Page and Menu
The landing page of Compass is the Explorer New Session tab. That’s the most advanced tool in Compass and we’ll save it for a later session — don’t worry, you’ll get there. For now, notice that the Explorer tool is represented by a compass icon, highlighted in turquoise on the left menu bar.
Clicking the arrow at the top of the menu bar will expand it so you can see what each icon represents. In this session we’re focusing on Opportunity Scan; the other tools will get their own time in the spotlight in future sessions. (The links in the Applications section take you out of Compass entirely to other views in the broader Ursa product — helpful to know, but not our focus today.)
3. Opportunity Scan
Let’s get started with Opportunity Scan — go ahead and click its icon in the left menu bar.
Opportunity Scan is the most hands-off tool in Compass — in a good way. You give the agent specific instructions up front, and after that, your main job is to kick things off and review the results. The user selects one or more areas to investigate, clicks one button, and then the agent takes it from there: running the full analysis, curating the results, and generating an annotated presentation ready for review.
The Opportunity Scan landing page lists any opportunity scans that you've run or have been shared with you (so this page may be blank right now). Try clicking the Graphical button at the top of the screen. This link brings you to a graphical representation of an Opportunity Scan, likely the Cost of Care Opportunity Analysis. You’ll see a set of boxes representing different subject area domains which make up the Opportunity Scan — click on any one of them to expand it and see the specific analyses inside. Each of these analyses is called a Diagnostic.
Let’s run a simple Opportunity Scan to see how this all fits together. Click the turquoise New Scan button at the top of the screen. In the first dropdown (labeled Diagnostic Suite), scroll down and select User Training, then click the turquoise Create Scan button.
The screen will update to show only the domain(s) for the scan you selected. In this case, there’s just one domain — called Training — and expanding it reveals the one diagnostic inside, called User Training. You can watch the agent’s progress right here as it runs through the diagnostic and starts building the Presentation. You can also click into the Diagnostic to see the next layer of the agent’s work — go ahead and click on User Training.
4. Diagnostics and Sessions
On this screen you can see the three Sessions in this Diagnostic. Sessions are the workhorse that power much of the functionality in Compass. Sessions represent a single focused interaction with an agent.
Click into the first session and scroll all the way to the top. You’ll see a large block of text in a teal box listing all of the tools available to the agent in this session, along with instructions on how to use them. Don’t worry about reading every word — the important thing to know is that this block is included in every diagnostic as the agent’s operating manual.
Near the bottom of that first box, you’ll see the line “The starting request from the user is this:”. What follows is the Session Prompt which gives the agent specific instructions on how to perform the analysis for this session. It typically begins with the specific Ursa object that the agent should use (URSA-PHF-R01 Patients with Active Membership in this case), any calculation instructions that the agent may need, and the key results that are being requested.
Now the ball is in the agent’s court. It will pick from its available tools using its own judgment — that autonomous decision-making is exactly what makes it an agent rather than just a fancy search box.
The agent will probably call GET_MEASURE_METADATA on the relevant table first, or it might skip straight to QUERY_DATABASE — either way, it’s making its own judgment call.
Each teal chat bubble that follows shows the response from a tool call, generated by Compass code. You can expand any bubble to see the full response. When the agent uses QUERY_DATABASE, expanding the bubble lets you see the raw database result — unfiltered by the agent. The agent also shows its thought process and SQL queries as it goes, so you can follow along if you’re curious.
After the agent has run its analysis and found the requested results, it will use the PROMOTE_FINDING tool to lock in any QUERY_DATABASE response from the session as a worthwhile artifact or Finding.
These Findings will appear on the right-hand side of the screen, and you can expand them by clicking on it. The table of data for a particular Finding and the SQL (which you can see by clicking the SQL button) are historically accurate and hallucination-free. The title and the annotation are created by the agent, but they can be changed via the edit button.
Click the Collapse Findings button in the top right corner to return to the main session window, then scroll to the bottom. The agent will usually wrap up with a text summary and may ask if you want any follow-ups. Keep in mind: this summary is the agent’s interpretation — useful context, but should be grounded in the actual SQL results from the session.
5. Back to the Opportunity Scan
While you’ve been exploring that one session, the agent has quietly been running the rest of the diagnostic, reviewing findings, and putting together a presentation. Multitasker. Let’s go see what it built — click the Opportunity Scan icon in the left menu bar.
Click the List button at the top of the screen to see your completed and in-progress scans, then click the one for today’s date.
The scan expands to show two sections. The Training section at the bottom is everything we just walked through — the User Training diagnostic and its three sessions. The top section is Presentations. Click on the Full Presentation to take a look. (Note: Each of these elements will get green checkmarks as the agent completes them. If the checkmark by the Presentation is not yet green, give the agent a few minutes to complete it before clicking on it.)
6. Presentations
You are now on the Presentation screen where the agent has created an exportable set of slides based on a prescribed presentation prompt. (This prompt is editable for your specific needs, but we’ll talk about how to change it in a future session.)
The default prompt told the agent to identify the top 3 recommendations for the selected population, starting with a high-level summary slide, followed by a more detailed slide and supporting data for each recommendation, and wrapping up with a context and trending section. Not bad for something that ran on its own.
Use the Next and Prev arrows to page through the slides, or click a slide number directly to jump there.
Per the prompt instructions, the presentation ends with a slide listing the other drivers that were reviewed but didn’t make the top 3 cut — useful context for anyone who wants to dig further.
Just as you could peek under the hood in the diagnostic sessions, you can do the same for the presentation. Click the turquoise Build button at the top of the screen to see how the agent put it all together.
In the Build screen, scroll to the top to see the large prompt text block. Same idea as the session prompt — it lists all the tools available for building the presentation and how the agent should use them.
Scrolling further, you’ll again hit the line “The starting request from the user is this:” — which is where the prebuilt instructions for constructing this particular presentation live.
Keep scrolling to watch the agent’s thought process unfold — reviewing the findings, picking its top recommendations, and assembling the slides.
At the bottom, the agent wraps up with two useful lists (as it was instructed to do). The first flags any data that wasn’t included in the diagnostic but could have strengthened its recommendations. The second notes any visualizations requested for the Context & Trending section that it couldn’t produce from the available findings.
On the right side of the screen you’ll see thumbnails of each slide — handy for reference while you’re editing. Try clicking Slide 3 and a small preview will pop out. Click it again to close.
The chat window at the bottom lets you ask the agent to revise the presentation in plain language. Try typing
Delete Slide 8
and clicking Send. The agent will take care of it and let you know when it’s done — and probably ask if there’s anything else.
You can also ask the agent to reformat existing slides. Try:
Change Slide 9 to just be a data table instead of a chart
You can also add new slides — either pulling in findings the agent set aside, or from deeper analyses in existing diagnostics. We’ll cover that workflow in our next session.
For now, let’s assume you’re happy with the presentation. Click the presentation name (in turquoise, top left) to return to the slide view. There are a few more things you can do from here.
7. Editing and Exporting Presentations
Click the Edit Presentation button in the top right. This screen is for editing the presentation’s metadata — not the slide content itself. You can rename it, add a description, apply tags, or share it with other users. (We’ll cover Entity and Vintage in the next session.)
Go ahead and type Training in the Tags window to create a tag. To share with someone, you can grant them Editor or Viewer access — click the dropdown beside the Viewer box to see the available users. Click Save Presentation when you’re done.
You can also export the presentation to PowerPoint for further editing or to apply your organization’s template. The exported file includes the full underlying data — not just screenshots — so you have complete edit capability. Click the Export to PowerPoint button in the top right corner. (Note: If you are unable to export, your permissions may need to be updated, so please reach out to your Ursa contact for assistance.)
Open the exported file in PowerPoint and click Enable Editing at the top. Navigate to slide 4 and select the chart — from here you can edit the underlying data, show or hide rows and columns, add data labels, change the visualization type, and so on.
That’s a wrap on Session One! You’ve now learned how to run an Opportunity Scan, dig into the Diagnostics and Sessions behind it, and view, edit, and export a Presentation.
In the next session, we’ll go deeper: next-level analysis from a scan, customizing diagnostics and prompts, focusing a scan on a specific population segment (like a payor or provider group), and saving segment-specific context for the agent to use in its work. See you there!