What to Have in Place Before You Build Custom AI Agents
The 202-to-303 bridge. What firms need to have working before custom development makes sense.
There is an order to AI infrastructure that firms often try to skip. They hear about agents - AI systems that can take actions, automate workflows, deliver results without human intervention at each step - and want to build them immediately. The idea of automating due diligence, or running competitive monitoring on autopilot, or generating client reports without manual synthesis is genuinely compelling.
The firms that build agents without the foundation in place end up with fragile systems that require constant maintenance and produce inconsistent output. The ones that build agents on top of solid infrastructure get tools that actually work.
The foundation is what 202 builds. The custom development is what 303 builds on top of it.
What agents actually need to work
An agent is an AI system that takes a goal and acts autonomously to achieve it - searching, retrieving, processing, and producing output without requiring human direction at each step. What makes an agent work reliably is not primarily the AI model itself - modern models are capable enough for most professional services applications. What makes an agent work is the surrounding infrastructure it runs on.
Documented workflows
An agent executes a workflow. If the workflow is not documented clearly enough for a human to follow without asking questions, it is not documented clearly enough for an agent to follow either. The documentation work in 202 is the prerequisite for the agent development work in 303.
Working data connections
Agents that need to access your firm's data to do their work need those data connections already set up and tested. Building the connection while also building the agent doubles the complexity and doubles the points of failure. The connection work should be done, stable, and understood before the agent is built to rely on it.
Clear success criteria
An agent needs a definition of what good output looks like - specific enough to evaluate whether the agent succeeded. "Produce a good research summary" is not a success criterion. "Produce a structured summary that addresses all five questions in the brief, cites at least three primary sources, and is under 800 words" is. This specificity comes from having run the workflow manually enough times to know exactly what good looks like.
Skills that the agent can call
Complex agents are often compositions of simpler capabilities. An agent that handles a full due diligence workflow might call on a document review Skill, a risk flagging Skill, and an executive summary Skill as sub-tasks. Having those Skills already built, tested, and reliable means the agent is assembling proven components rather than building every capability from scratch.
The most common readiness gap
Firms that approach us about building agents are usually ready on the technology side - they understand what they want to build and why it would be valuable. The readiness gap is almost always the same: the workflow the agent should automate has never been fully documented.
The workflow exists, and people do it well. But when you ask them to write down exactly what they do - step by step, with every decision point and every edge case - they find that large portions of the workflow are implicit. They make judgment calls they have never articulated. They use information sources they access casually but have not documented. They handle exceptions in ways that they have never had to explain.
Documenting the workflow is where the real work of agent preparation happens. And it is work that benefits the firm regardless of whether an agent is built - because the documentation is useful to every human who does the same work.
The question that tells you whether you are ready
Could a smart new hire at your firm execute this workflow on their first week, using only the documentation you have written - without asking anyone for help?
If yes, you are probably ready. If no, the documentation work is what comes first.
This is not a high bar. A new hire has questions, makes mistakes, needs some guidance. But if the workflow documentation could get someone competent to a working result, it is specific enough to build on. If it could not, the gaps in the documentation are the gaps in the agent specification.
What being ready looks like
A firm that is ready for custom agent development has a few things clearly in place. The workflow is documented in detail. The data connections the agent will need are set up and tested. The success criteria are written down. And ideally, the core Skills the agent will use have been built and are already in the library.
At that point, the agent development conversation shifts from "can we build this?" to "how long will it take and what will it cost?" - which is a much more tractable conversation.
Building the infrastructure that makes agent development tractable is what Apparatus 202 is designed to do. If you are already there and want to talk about what custom development looks like, book a call and we can talk through what is possible.