This Is Your Practice Management System On AI

AI is set to drastically improve the ways our practice management systems support us. From generative email replies, to agents we assign tasks to that will do the work.

It's an opportunity for PMs, but also a tremendous risk. If we're slow to move, mainstream solutions like Microsoft 365 will pull users out of the space altogether. It's perhaps the greatest existential risk there has been for the accounting firm software category.

The step-change in productivity enabled by AI will be introduced to the world not by the small companies in our space, but by Microsoft & Google, and it necessitates a response to retain users.

So here's a roadmap for how to get your organization working toward an AI-enabled version of practice management. I'll touch on some of the more technical aspects of how to implement these changes here. For a less developer-ey, accountant-frendly version I'll have a full accountant-facing rundown on my daily show tomorrow. Pod | Video

17 ways AI will make your practice management system MUCH more helpful than it is today:

Search

1. One search to rule them all

A practice management system is the cornerstone of a firm. It's home to more context than any other system a firm runs on. In the age of AI whoever has access to the greatest amount of context wins out.

The first step is to simply make it searchable. A single, unified search across: Email history, files, projects, meeting transcripts, team chat, SOPs. Something akin to your Slack quick switcher, or Raycast.

Not just text-matching search - vector search. The ability to surface relevant context in a more meaningful way across your entire firm, deep search into files, into all comms, will be transformational.

The secret sauce is embedding all sources of context. Store the embeddings in a vector database like Pinecone, Weaviate or Chroma, then make it accessible to the user via orchestrators like GPT and LangChain.

2. Extended search

Certain users will search in very contextual ways. For example: "What W-2s did John Smith have in 2021?"

Once all documents are embedded, classification of docs, and classification of pages within docs enables extended search. This once required your own costly classification model - now vanilla GPT excels at basic document classification to get you to something that'd good enough.

Start small here. What are the 5-10 most common things users search for, and identify those types of documents. Tax forms are an easy example since the text on those pages follow a common structure.

Ultimately the goal is to move from "here's some possibly helpful source material" to "here's the answer, and the source material if you want to see it". The general process here is (1) embed the query (2) fetch the highest scoring context (3) use GPT + possibly LangChain map reduce to synthesize the context into a concise answer.

3. External search

Since context wins in the age of AI, connections to third-party apps are valuable. If your app connects to each client's QuickBooks file for example, you have a tremendous advantage.

Extending search into third-party apps likely means embedding third-party data sources as well. A unified search experience across all internal context + the ledger would be phenomenal, especially once you're leaning into semantic question & answering, not just retrieval.

Other non-obvious sources of context:

1. Tax returns - embedded returns aren't particularly valuable without human-provided semantic context. GPT can get you box 7, but pairing box 7 with an explanation of what box 7 represents, or the form instructions for box 7 is ultimately more helpful.
2. Transcripts - look to pull in transcripts & recordings from the services accountants already use. This is an overlooked treasure trove of context today.
3. Engagement letters - the terms of engagement are a helpful framing, particularly when it comes to client comms.
4. The firm's accounting file - payment & invoicing history for a client is often overlooked and impacts much of the work that's done for a client.

Client Communication

4. Generative replies

Arguably the holy grail of AI applications - the steps above are a prerequisite to doing this well - suggested replies that take into account all of the other context of the firm, as well as the tone of past comms with this client.

Your reply should take into account that a colleague emailed the client yesterday. That you're generally snarky with this client. That they have a project that is near completion. That their engagement letter stipulated certain expectations of the client. The details they provided in a Zoom meeting last week. Practice management systems should do this better than anyone else, because they hold the most context.

Among the most significant ripple effects of AI in the category will be the importance of email communications. Where email could've been an optional aspect of a PM in the past, a fantastic email experience now ought to sit at the center of every PM.

I have a hard time seeing any future alternative. Email must live where the context lives. If a user is writing emails from Outlook, yet their business context lives in the PM, they'll either move to a PM that supports email, or move their context to Microsoft365. The clock is ticking on contextless email.

Due to the volume of context necessary to create a rich response, the solution here is likely a LangChain map reduce of several sources of context, synthesized into a single output with GPT. In some applications where a huge amount of context is necessary, some researchers are exploring summarizations of summarizations to pull in even more context.

Chunk size, top-k values, prompt engineering, email classification, few shot training, chain of thought prompting, conditioning of the input prompts and conditioning of the embedding results will wildly impact the quality of what you get here.

5. Suggested documents

The answer to a client question will often either live in a document, or be the document itself. If they're looking for a copy of the document, a suggested attachment makes sense here.

If they're asking a question for which a document can provide the answer, a suggested document for review, or even a summarized excerpt from the document is helpful. For example an engagement letter, or past research a colleague has documented in the file.

From a technical standpoint this is simply a matter of embedding the email to find the most similar documents, or storing document summaries, and finding the most similar summary.

6. Generative attachments

Lastly the on-the-fly generation of new documents becomes an exciting opportunity. Need a letter for a banker? Our organizational history contains 200 other such letters. Select three similar suggested docs, and generate a starting point from which to draft a new doc. GPT-4 does this nicely.

Documents

7. Renaming

A couple approaches to file naming could serve as both a time-saver, and a helpful source of standardization across a firm:

1.) This could be done simply via GPT + similar documents. Find similar other documents in the firm, and have GPT generate a concise file name using similar files for reference.

2.) Or this could be done according to a specific nomenclature. Most firms haven't leaned into file-naming conventions - AI can help. The naming process here would be a combination of GPT-based document classification and GPT summarization.

8. Summarization

A one-or two line summary of what the document contains saves the user from opening & closing just to figure out the file contents. A LangChain map reduce is usually the solution here to avoid GPT context limits.

9. Document Similarity

Surfacing similar documents is often helpful when the user is authoring or modifying a doc. This a way to better bring colleagues into alignment, and leverage the reusability of past work. Vector similarity will do the trick here.

Requests

10. Do We Already Have It?

Nothing makes an accountant look worse than asking for something you already have. AI can surface relevant context wherever it lives. This should simply be a matter of embedding the request, and returning a list of the most similar sources of context across the rest of the firm. You'll come out the hero and genuinely improve the client experience.

11. Is A Request Still Valid?

When information is provided via a different channel, the PM can check whether it resolves an outstanding request. This could be as obvious as an unlabeled document upload, or as non-obvious as a simple statement by the client in a Zoom call that answers a request a colleague makes three months from now.

Client Portal

12. File Splitting + Classification

Oftentimes batch files are submitted in a way that isn't particularly helpful. The portal here can suggest splitting & renaming uploads in a more helpful manner. Because time is of the essence when the user uploads the file, GPT-3.5 is your best bet to get a quick classification & suggested file splits. A map reduce may be necessary for large documents.

13. Request Association

If documents are uploaded without context, look to associate them based on their contents with outstanding requests. This prevents future auto-reminders going out to the client, and should kick the project status back to active once outstanding requests are resolved. If the client can confirm this association, it removes any AI guesswork after submission.

CRM

14. Sentiment Analysis

Now that context is centralized across the firm, sentiment analysis becomes more meaningful when weighted across email, chat and meeting transcripts. This can be benchmarked at the client level, but also at the staff-client level to identify friction with specific staff members.

15. External Updates

Fetch external sources of context for clients from public sources like LinkedIn or the company website. This context may be something the accountant needs to be alerted to, but can also serve to enrich suggested client interactions. You'll look like a real smartypants if a suggested reply includes reference & link to a LinkedIn post they made a couple days ago.

Agents

16. Administrative Agents

The ultimate long-term moat for practice management systems beyond mainstream services will be agents that do the work. A point of entry here is an agent that provides administrative, organizational support.

Agents could function from a set of pre-built tasks, but the real opportunity is the ability to assign tasks to an agent. The same way a task is assigned to a human user, it can be reassigned to the agent to complete.

While in the past this would've seemed wildly optimistic, self-prompting agents now make this feasible. AutoGPT is the best point of reference today, but this reality is closer than you may realize. Ultimately the agent is self-guided, and would report back to its supervisor or the assigner when taking action.

The sections before this one kept us competitive with Microsoft & Google. This section 2x's the fees accountants will happily pay their practice management provider.

17. Domain-Specific Agents

If you integrate with the client's QuickBooks for example, how can the agent begin performing tasks in those files? Agents could extend to third-party services, particularly those you already authenticate with, completing recurring tasks as a human would.

There are also opportunities here around tax as well. Fetching and filling to and from government forms, performing research, and drafting tax resolution documentation are all valuable tasks that in the past a human needed to complete.

Agents represent a massive revenue opportunity. The ability for a PM to frame an incremental feature through the lens of the alternate cost of the human worker. It's also the ultimate source of differentiation, not only from other solutions in the space, but perhaps more importantly, from solutions outside our space.

What do you think? What did I miss? This is the summation of conversations I've had in the past few months with 50+ cos, mostly folks in our space, and with the CEOs & product leaders of most practice management systems.

While I'm nervous of the impact things like Microsoft Copilot could ultimately have on industry-specific tooling, I'm confident the long-term, high-context applications of AI & agents will ultimately win out.

If this is something your team is exploring, how to rebuild a culture around investment in AI, shoot me a DM. Needless to say things are heating up ??