Prebuilt AI Agents for Law Firms and CPAs

Most AI advice for professional services is generic. "Use ChatGPT for intake." "Ask Claude to summarize this document." It works for a few minutes and then falls apart the first time the output is wrong in a way that matters. When we started building SUDO, our private AI platform for law firms and CPAs, we stopped writing prompts and started shipping prebuilt agents. Here is what we learned.

The Problem With "Write Your Own Prompt"

A partner at a law firm can write a fine prompt. They can ask ChatGPT to summarize a deposition or draft an engagement letter. The output looks great the first time. The second time, it looks different. The tenth time, someone on the team has rewritten it for their own style, and now the firm has ten versions of the same tool doing ten slightly different things.

That inconsistency is a problem anywhere. In professional services, it is a liability.

When the output is going to a client, touching privileged information, or feeding into a regulated workflow, "it mostly works" is not the bar. The bar is: the output is the same shape every time, the constraints are honored every time, and when something is off, you can look at the agent definition and see why.

That is what a prebuilt agent gives you that a blank prompt does not.

What A Prebuilt Agent Actually Is

Strip it down and a prebuilt agent for a professional services firm is four things:

• A role: narrow, specific, and scoped. Not "legal assistant" but "conflicts-check intake agent for a five-attorney commercial litigation practice."
• A structured input: the fields the agent expects. Client name, matter type, opposing party, referral source, jurisdiction. No free-form paragraphs.
• A deterministic output format: a defined schema. Engagement memo. Conflicts grid. Drafted intake summary. Every run produces the same shape.
• Guardrails: the things the agent will refuse to do. No legal advice. No privileged output from non-privileged input. No client-specific financial guidance unless authorized. Explicit, written, checked before output.

Everything else, the prompt wording, the model choice, the UI, is implementation detail. If you nail those four, the agent survives daily use. If you skip them, you are back to "it worked on Tuesday."

The Agents That Actually Get Used

We have shipped dozens. The ones firms return to look remarkably similar across law and accounting.

For Law Firms

• Conflicts Intake Agent. Takes a structured intake form, cross-references prior representation data, flags potential conflicts, and produces a memo for the partner to review.
• Engagement Letter Drafter. Takes matter type, scope, fee structure, and client name. Produces a draft engagement letter in the firm's standard format. Will not fill in unknown fields with guesses.
• Deposition Summary Agent. Takes a transcript. Produces a summary, a witness outline, and a list of open questions. Does not produce legal conclusions.
• Discovery Response Drafter. Takes a production request and a document set. Produces a first-pass response with citations. Flags documents it is uncertain about for attorney review.
• Client Update Agent. Takes matter status notes. Produces a plain-language update for the client. Never invents progress that is not in the notes.

For CPAs

• Section 7216-Safe Client Prep Agent. Takes prior-year return data that the client has consented to share internally. Produces a preparation checklist. Will not produce anything that would require Section 7216 consent to disclose externally.
• Engagement Letter Drafter. Same pattern as legal. Scope, services, fees, and boundaries.
• Client Ledger Reconciliation Helper. Takes uploaded bank data and general ledger extracts. Produces a reconciliation worksheet with flagged variances. Does not post entries.
• Memo to File Agent. Takes a decision or judgment call. Produces a memo documenting the reasoning. Read-only, append-only.
• 1099 / W-9 Intake Agent. Takes vendor submissions. Produces a compliance checklist and flags missing or inconsistent data.

The catalog keeps growing, but the pattern does not change. Narrow role. Structured input. Deterministic output. Guardrails. Everything else is detail.

Outside regulated practice the same shapes hold, minus the compliance work. We ranked them for owners in AI agents for small business, which covers the five workloads worth funding first and what each tier costs.

What Holds Up Across Verticals

Four patterns show up in every agent that has survived real use, regardless of whether it is for a law firm or an accounting practice.

Structured intake beats free text. Every time. Forcing the user to fill specific fields removes 80% of the variability in the output. If a partner pastes an entire email chain into a chatbot, the chatbot does its best. If the agent asks for client, matter, jurisdiction, and scope, it knows exactly what to produce.

Output schemas prevent drift. A deposition summary that has the same sections in the same order every time is reviewable in minutes. One that is shaped differently every time takes twice as long to read and generates its own set of questions about what is missing.

Guardrails are the hardest part and the most valuable. Deciding what the agent will refuse to do is a firm-level conversation, not a technical one. "We will not generate legal advice. We will not fabricate citations. We will not produce output about a party with an active conflict." Those rules are encoded in the agent and they do not move.

Refusal is a feature. When an agent says "I don't have enough information to produce that" instead of guessing, the firm trusts it more. The agents that confidently produce plausible-looking wrong answers are the ones that get turned off.

The Failure Modes We Stopped Repeating

A short list of things we learned not to do, usually by doing them first.

• Starting with the model and working backward. It does not matter which model you use if the role is vague and the output is unstructured. Pick the use case, define the agent, then worry about the model.
• Writing a single "mega-prompt" that tries to do everything. One agent, one job. An agent that drafts engagement letters is not the same agent that summarizes depositions. Combining them produces worse output for both.
• Letting the model "help" with compliance. The model will happily produce plausible reasoning about Section 7216 or attorney-client privilege. That reasoning is not authoritative. Guardrails have to be encoded by a human who understands the rule. The agent enforces them.
• Assuming the cloud is fine. For professional services, data residency and confidentiality are not settled by a generic SOC 2 line. That question deserves its own post, and we wrote one.

How to Evaluate a Prebuilt Agent

If a vendor sells you "prebuilt AI agents for law firms," here is what to ask before signing anything.

1. Show me the role definition in writing. If the agent's role is fuzzy, the output will be too.
2. Show me the input schema. Structured inputs or free text? If it is free text, the agent is a prompt template, not a prebuilt agent.
3. Show me the output format. Is every run the same shape? Can I write a checklist against the output?
4. Show me the guardrails. What will the agent refuse to do? Who decided? When was the list last reviewed?
5. Show me where the data goes. Does the prompt, the input, or the output leave the firm? If yes, the conversation is about cloud AI, not private AI, and the risk calculus is different.

If the vendor cannot answer those five questions clearly, the agent is not prebuilt. It is a prompt with a nice wrapper.

The Bottom Line

Prompt engineering is not the hard part of AI for professional services. The hard part is shipping agents that do the same thing well, every time, with the firm's constraints baked in. That is not a prompt problem. It is a product problem.

A firm that adopts prebuilt agents with structured inputs, deterministic outputs, and written guardrails gets the productivity lift of AI without the drift, the liability, and the "it worked on Tuesday" problem. A firm that hands its staff a generic chatbot and hopes for the best gets the opposite.

We built SUDO because professional services firms need the first option, not the second. If you are thinking through what AI should look like inside your firm, let's talk.