How to Actually Use AI in Probation

Probation serves two goals:

• Protect public safety
• Re-integrate justice involved individuals into society as positive contributing members.

We've spent decades figuring out how to do that; You already know the "Evidence Based Practices".

But there's a problem.

Today, most of the work officers do is documentation. We're talking about 5 hours a day just filling out forms, writing documents, or evaluating reports.

That's over 1,000 hours a year typing/staring into your computer! Think about that for a second. You're spending 3 days a week (7 months a year!!!) documenting the work you did.

Don't get me wrong - this stuff is important. Disseminating information about your clients is important. It would be irresponsible to not record your work given the implications of your job.

But the reality is the urgent nature of documentation is getting in the way of the important nature of helping people.

Most of the hours you spend at work are on reporting the work you did, not actually doing the work...

But, It's not your fault!

Case management systems are designed for officers to document the work they do. They're not designed to actually help officers do the work.

Now your software is THE job.

That's backwards.

The good news is, there's a set of concrete tasks you probably do on a daily basis that AI can help accelerate and improve.

We're going to talk about what those are and how you can use AI to save yourself 3-4 hours a week.

What Probation Tasks Are The Best Fit for AI?

AI is a statistics engine. It excels at taking in a bunch of data, synthesizing it, and predicting what should happen next.

The primary example of this is analyzing a large corpus of data and spitting out a summary or condensed version of that information.

And guess what? Most of the documentation and report evaluation work you do as a probation officer fall into this category!

At CaseBase we've studied how officers spend their day. From that study we've found 5 administrative/documentation tasks that most commonly take up officer time.

• Chronological notes
• Evaluating reports (interlocks, UAs, etc...)
• Writing reports (court reports, PSIs, violations, etc...)
• Building case plans
• Conducting assessments

Again, these are all important. But by and large officers spend far more time on the documentation part of each of these than the judgement, intervention, and engagement side.

I believe the opposite should be true.

How Can You Accelerate and Improve These Tasks With AI?

As an officer, you can use AI to accelerate the documentation part of all 5 of these tasks. As a result, you'll free up 5+ hours a week that you can spend engaging with your clients and thinking critically about how to better support them.

At CaseBase, we're building an AI assistant built specifically to help probation officers do their job better and more efficiently. It's easy to use, feels intuitive, is CJIS compliant, and officers love it.

If you want to learn more about it visit casebase.co or shoot me an email at tyler@casebase.co

That being said I acknowledge that not everyone is in a position to buy new software. Budgets are tight and AI is new.

So, I'll walk you through how to accelerate these tasks in your consumer AI tool of choice for free!

Keep in mind standard consumer AI like ChatGPT, Claude, Gemini, and Copilot don't have the same security/compliance protocols baked in that CaseBase does so you'll have to take some extra steps to keep your data secure.

How to Build Your Own AI Assistant

Using ChatGPT, Gemini, Claude, or Copilot to accelerate your documentation tasks is simple. In fact, every single task we're going to accelerate comes down to the same 3 principles.

• Give the AI a set of instructions for each task.
• Give the AI the relevant policies, procedures, templates, and guidelines (context) it needs for each task.
• Give the AI the redacted information it needs to review to complete each task.

Step 1: Create an Account

The first step you MUST follow before using any AI tool to help you as a probation officer is to create an account. If you do not create an account you are unable to control how your data is used by the AI providers.

Keep in mind that creating an account does not make you automatically CJIS compliant or automatically protect your data.

What it does is give you the opportunity to limit how the AI providers use the data you give them. It's still crucial that you redact PII/CJI before giving any data to the AI.

Note Don't forget to ask your chief/director for permission to do this first!

Step 2: Configure Your Privacy Settings

After you've created an account all AI providers give you the ability to restrict what data they use for training that you give them.

Typically you can do this by going to "settings" and then selecting something like "privacy settings", "security settings", or "data settings".

Often AI providers will use terminology like, "Help improve the model for everyone". Turn this off! This is code for, "we will use your chats and the data you upload for model training".

Search for any other settings about data sharing and turn them off. We want to limit data sharing as much as possible.

A quick note. AI providers won't share your data with the general public. And model training does not mean anyone can ask about your client and get their information. But it's best practice to limit the data surface area you expose to AI.

Step 3: Create a Project

For every workflow we're going to accelerate with AI (we'll demonstrate 3 here) you'll want to create a "project". A project is a contained space to put your instructions, context, and prompts for a specific task you will conduct repetitively.

ChatGPT and Claude call these projects. Gemini calls them "Notebooks" and Copilot might call them something different. But the idea is the same!

Give the project a clear name that describes the task you want to complete.

For example; Case notes, Risk Assessments, Interlock Report Analysis, etc...

Inside the project the next most important thing to do is give the model "context".

Context is what the model references to understand how you operate. For example, for case notes, you could upload your case note template.

Typically there is an area for you to drag and drop your files. In ChatGPT is says, "Soruces". Find that area and add all of the files that contain information relevant to the task you are working on in the project.

The last and most important step is to add your instructions. Your instructions will tell the AI how to go about creating the work product you want.

This can include steps to go through, what information you need, and any non-obvious things about what you expect the AI to do for you.

Now that you know how to create a project, we're going to demonstrate how to accelerate all of these different tasks with a project in ChatGPT for each!

Accelerating Documentation Tasks With AI

The way to use AI to accelerate all 5 of these tasks is nearly identical. Instead of going through all of them we'll focus on 3. But know that the 4 principles above apply to every task.

Chronological Notes

The task officers spend most of their time on is writing chronological notes. On average we see officers spend about 8-10 hours a week on this alone.

You probably do this after every office visit, phone call, field visit, or check in with a client.

And what exactly are you doing? You're summarizing a contact with a client so you can retain the information you discussed for later contacts and as evidence if ever required.

This is just the kind of thing AI is amazing at!

If you use CaseBase, there's a built in case note generation feature that does all of this for you. But if you're not ready for a purpose built system, let's build a project in ChatGPT to help us do the same thing!

Project Setup

In ChatGPT (or your AI of choice) create a new project and title it something like "Chronological Notes".

Inside of the project start adding your context. Some examples of files that are good to upload would be:

• A blank chronological/case note template
• A completed note template with personal information redacted
• Some rules about what you are expected to put into the case note
• Any information about policies, procedures, or guidelines depending on the clients risk or supervision level, case type, etc..

Finally, add your instructions for creating case notes. In ChatGPT you can add instructions by clicking the 3 dots at the top right of the page and clicking "Project Settings".

Inside of this menu add your "instructions". The instructions should include things about what to put into the note, what sources to pull from, how long the note should be, and any other non-obvious things that make your notes unique.

Generate a Note

Now that we have our project, it's going to be super easy to create a new case note. Instead of writing down everything, we're going to let the AI do what it's best at - summarize.

Obviously, it needs some information to summarize, right? Instead of typing everything in, we're going to use voice-to-text to save even more time.

If you have Microsoft teams I recommend recording your contact with a client to generate a transcript. Once we have this, instead of having to type everything into ChatGPT, we can just copy/paste the transcript in.

Now go back to your ChatGPT project and plug in the transcript. If you've already created your instructions, ChatGPT will pump out a new chronologic note for you on the spot!

Remember, always redact PII like names/identifiers/addresses before giving the information to ChatGPT and make sure to check the information is correct.

This gives you the ability to fully engage in your sessions with clients while still getting the quality of note you want. This alone will save you 10+ minutes if you were to write your notes after a contact.

Evaluating Reports

The second most common and time consuming task we see officers do is evaluate reports. We're talking interlocks, UAs, breathalyzers, RemoteCOM, etc...

You know the drill. You receive a weekly/monthly report for a client. Your job is to assess it for any abnormalities, violations, etc.. If you find no issues, you're good to go! If you find a problem, it's time to figure out what to do next. Could be writing a violation, could be a quick call to your client, could be running it up the chain.

All of these report evaluation tasks come down to exactly what AI is great at. Look at a large corpus of data, summarize it, and predict what should be done next!

Project Setup

We're not going to beat this one with a dead horse. Setting up your project for evaluating reports is just like creating a project for writing chronological notes.

Instead I'll focus on a specific example and talk about how I would set up my projects for this. We're going to set up an interlock report evaluation assistant

Instead of setting up a single project to do the generic process of "evaluating reports", I would make a unique project for every type of report you evaluate.

For example, an interlock project, a RemoteCOM project, etc...

For each report I would simply add a few pieces of context into the "sources" section and a very simple set of instructions.

For the interlock report assistant, I would first upload an example of a marked up report with anything concerning.

For example, circle some violations or things that would flag concern to you.

For additional color, tell the model what specific things constitute a violation or a concern.

If you evaluate reports from different vendors I'd do this for every different vendor report.

Then for the instructions you create for the project tell it exactly how you evaluate reports and what to look for.

For example, "Check every reporting period for the client to make sure a test was conducted within the window. If a test was not conducted in the window, flag it for me and find the nearest test to that window if it exists. If you see anything that is not necessarily a violation but seems odd (like potential tampering or trying to trick the system) please flag that as well."

It may take a few iterations of instructions to get exactly what you want.

Evaluating Reports

Now that you have your first project set up, let's have our AI assistant actually evaluate a report for you!

Before you put anything in, make sure to redact PII like name, birthdate, home address, etc...

Then create a new chat in your project, attach the file, and let ChatGPT do its thing!

It may not be perfect the first time around but what you can do is give it feedback in the chat thread. ChatGPT, Claude, and Gemini will use this to learn about its mistakes and improve next time around.

For bonus points, ask it to write up a report and export it to a word doc for you!

Build a Case Plan

The final example we'll share with you today is how to build a case plan. Just like the other projects, the workflow is almost exactly the same. Give some context, add instructions, and let AI work its magic.

Project Setup

Creating good case plans is unique in that they should be SMART. You want to combine the clients needs with actual services and programming your agency offers for it to be effective. We don't want generic and non-measureable goals.

This means we need a few pieces of important context.

In the sources section of your new project try to add your programming, classes, and service providers.

We want the case plan to have REAL examples of what clients should do.

This will let the AI actually add tools and techniques to your case plans that your agency actually uses

Add an example of a case plan you're proud of. Not just any old case plan, but one you'd be excited to share with the world. One that is your gold standard.

If you have it, add your agency guidelines for case plans. This should include things like:

• How many domains/goals/objectives should a case plan
• How long do you typically set objectives out into the future
• Do you use any special vocabulary or styling techniques

Finally add your instructions. Generally this will be something like, "Review the client's risk assessment, suggest some domains and goals to the officer, and then let the officer confirm what they want the focus of the case plan to be. When done, export a word document with the case plan."

This is your time to shine. Customize this how you see fit and iterate on it as you continue to build more case plans.

Generating a Case Plan

Now it's time to go! Generally when having AI create a case plan we recommend you give it a risk assessment in addition to some thoughts following your conversation with the client about what they want to focus on. Give the AI some guidance so it can meet the client where they are.

Attach the risk assessment in a new project chat.

Give it some background on what the client wants to focus on.

Let it rip!

Just like your other projects, it will take a few attempts to get exactly what you want. I encourage you to experiment and give the AI feedback to improve it's outputs.

Now What?

You've created some projects in ChatGPT, Claude, Gemini, or Copilot. Now what?

AI requires an investment. Unlike other software AI improves with time. BUT, it needs your feedback to improve. Don't just expect AI to do all of the work for you - it won't.

It's an investment, but one that pays off in spades if you put the effort in up front to build your AI assistant in a way that works how you want it to.

If you don't want to make that investment, you need something more secure, or you want something purpose built for your agency reach out to us at casebase.co.

We're building the world's first AI assistant for probation officers and we'd love to show you what it can do. Today, 23 agencies across the country trust CaseBase to help with things like writing case notes, building case plans, evaluating reports, and writing documents.

If you just want some help on understanding AI and how you can implement it with your agency don't hesitate to reach out to me at tyler@casebase.co

The worst thing you can do is nothing. Start using AI and figure out how to make it help you!