Building AI Agents for Professional Services: What Works Today

Every time I talk about AI agents, I get asked some version of the same thing: "Can we build an agent that just handles all client intake?" Or: "Can an agent do all our research?" The answer is more nuanced than yes or no. Some types of agents work today. Others don't, and probably won't for a while.

The line between "working agent" and "interesting research" is where you expect it to be: simple, well-defined workflows work. Complex, open-ended tasks don't.

What's an Agent, Really

An AI agent is a system that:

• Takes a goal (e.g., "research the competitive space for this company")
• Autonomously decides what steps to take
• Executes those steps (search, read documents, call APIs, send messages, etc.)
• Evaluates the results
• Iterates until the goal is achieved

This is more complex than "call Claude with a prompt." The agent has to reason about what's needed, not just execute one task.

What Works Today

Structured Task Agents: If your goal is well-defined and the steps are known, agents work well. Example: "Extract the following fields from this contract: parties, terms, payment schedule, termination clause." The agent knows what to do. It reads the contract. It extracts. Done.

Workflow Routing Agents: "Intake a new client. Classify them by industry and complexity. Assign to the right team." The workflow is defined. The agent executes. It works.

Fact-Finding Agents: "Look up the current board of directors for Acme Corp." The agent knows how to search, cross-reference, verify. For factual queries with known answers, agents work well.

Iterative Research with Constraints: "Research three competitors in the healthcare software space. Focus on pricing, features, and customer reviews." The scope is limited. The agent can do targeted searches instead of unlimited browsing. It works.

What Doesn't Work Yet

Open-Ended Research: "Understand the competitive space." The agent doesn't know what "understand" means. Is that pricing? Features? Market share? It could spend 100 API calls and never reach a coherent answer.

Strategic Judgment: "Should we enter this market?" Agents can gather information, but they can't make strategic judgments. That requires context, experience, and risk tolerance that agents don't have.

Creative Work: "Write a proposal." Agents might be able to draft it, but the quality will be mediocre. Human creativity and domain expertise still win here.

Long Chains of Reasoning: If your task requires 10+ steps of reasoning and decision-making, agents often lose the thread. They drift. They hallucinate. They get stuck.

The Implementation Reality

Building a working agent for professional services today looks like this:

1. Define the task narrowly. Not "do research." "Research competitor pricing for these three specific companies."
2. Define the tools available to the agent. It can search, read PDFs, call your internal APIs, but not browse the internet at random.
3. Set success criteria. "Agent should return a document with pricing for at least two of the three competitors."
4. Build in human oversight. The agent provides a draft. A human reviews and verifies before it's used.
5. Monitor and refine. When the agent fails, understand why and tighten the constraints.

This is not autonomous AI. This is AI-assisted workflows where the AI handles repetitive steps and a human handles judgment.

The Cost-Benefit Calculation

Building an agent that handles one task (say, client classification) costs roughly 3-6 weeks of engineering time. That's $15-30K.

The task needs to be high-volume enough to justify that cost. If you only do it once a week, an agent doesn't make sense. If you do it 50 times a week, it does.

The breakeven is roughly: 20+ occurrences per week, where each instance takes 10+ minutes of human time. At that threshold, building an agent usually pays for itself in 6-12 months.

What to Build First

If you're thinking about agents, start with this checklist:

• Is the task well-defined? (Yes / No)
• Does it happen frequently? (20+ times per week / Less frequently)
• Does it take significant time? (10+ minutes per instance / Quicker)
• Is the required workflow stable? (Mostly the same each time / Varies widely)

If you answered "yes" to all four, an agent might make sense. If you answered "no" to more than one, you're better off with simpler automation or staying with manual processes.

The Honest Take

AI agents are real and they work, but not in the way hype suggests. They don't replace knowledge workers. They augment them by handling the repetitive, well-defined parts of their job. The bottleneck is usually defining what "well-defined" means for your specific workflow.

Most firms would get 10x more value from well-executed AI-assisted workflows than from trying to build autonomous agents that don't quite work.