In Agents: Where GenAI Value Actually Lives, Ruark Chick and Floor Blindenbach-Driessen sort agent ideas into three categories — personal, workflow, and enterprise — because each carries different rules for ownership, testing, and what "deployed" actually means. They also flag a distinction that cuts across all three: AI agents that surface, summarize, or structure tend to hold up well; agents expected to validate, approve, or decide tend to struggle. A policy finder is useful on day one. A policy checker has to be right every time, or it's worse than useless.
In this blog, I stay inside the personal AI agent category and go deeper on two things: why an imperfect personal agent is often more valuable than it looks, and why the moment a firm decides to help people build them is more consequential than it seems.
Building can be how we understand the work
It's tempting to say we should fully understand a process before building an agent around it. In practice, we often build the first version because we don't yet understand everything. Trying to build it forces questions ordinary process discussion leaves unresolved: what information does it actually need, which instructions are obvious to an experienced attorney but invisible to the system, where does judgment enter, which results can be checked quickly. A personal agent makes those questions visible because it produces something you can use, test, and challenge. That's more than a productivity tool — it's process discovery, a rough but functioning representation of what someone is trying to accomplish, without having to debate it in the abstract.
The prototype is not the product
A personal agent works because the person using it is quietly holding the whole system together — they know which documents to select, which cases are the exception, which prompt needs to run twice. None of that means the agent is defective. It means it's still a prototype.
The risk appears when a personal experiment gets shared without anyone reconsidering what changed. The ILTA blog names this well: unplanned sharing is the central risk of the personal-agent category — a tool that's never been stress-tested can quietly become informal firm-wide infrastructure, carrying an access and data-leakage risk nobody consciously took on. And the transition is rarely a clean handoff. It's a slow leak: someone builds something useful, a colleague asks to borrow it, six months later three paralegals depend on it, and no one ever made an ownership decision.
This is where ILTA's sorting questions actually earn their keep — not as a checklist you run before you start building (most people don't, and probably shouldn't; you build the first version because you don't yet understand the task), but as a signal once "just for me" stops being true. Is this still one person's workflow? Does it still rely on judgment, or is someone now leaning on it to be right every time? Could the colleague asking to use it actually configure something like it themselves — or are they asking precisely because they can't, meaning they also can't tell when it's wrong? If the answers have shifted since you built it, that is a warning sign that it may no longer be a personal agent.
Not everyone wants to build their own agent
The framework above assumes an owner willing to build their own Agent. Plenty of good lawyers have no interest in that, which raises a reasonable-sounding question: could a firm just hire someone to build personal agents for people?
I'd push back. The value of a personal agent isn't the tool. It's the discovery that happens while building it — being forced to answer what you actually check before trusting the output, where your judgment kicks in, which cases are exceptions. Hand that off to a specialist and the owner gets a working tool but skips the part where they had to articulate any of it. That produces the exact failure mode ILTA's timesheet example describes: the agents that stick are used by people already fluent enough to know when the output is wrong. An agent built by someone else and handed over is a black box with one user — not a personal agent in the sense that matters. So I'm skeptical "build-for-me" pays off as a service; it quietly reintroduces the unowned, unstress-tested risk the personal-agent category was supposed to avoid. Whether the economics of a dedicated done-for-you function ever work out is also a question.
There's a middle path, though, that does preserve the mechanism: build-with-me. A specialist interviews the owner into their own answers rather than supplying them — same discovery process, just facilitated. The owner still handles the judgment work; the specialist makes it faster than fumbling through the agent-building process alone. That also changes the economics: a firm can't easily justify headcount for "one person's productivity gain," but it can justify headcount for a structured discovery process that produces personal agents as a byproduct — because a specialist running dozens of these sessions starts to see the same needs recur across a practice group. That pattern-matching is exactly the raw material the workflow-agent category needs: cheap evidence of which repeatable workflows are worth the heavier investment of shared data, failure handling, and defined ownership.
Recognizing when it has stopped being an experiment
None of this argues for stopping people from building until every requirement is known — that would remove the most useful way to discover and build agents in the first place. It argues for being honest about which activities preserve that discovery process and which only look like they do. A personal agent, built or co-built by its owner, is a useful tool and could be the wireframe for something bigger. An agent handed over by someone else is just software with one user and no one accountable for it, regardless of who built it. The first is worth encouraging broadly. The second is worth being skeptical of.
Building with people, not for them, should be the default format for individual agents. And as a general principle, no personal agent should scale without a look from IT or development first — a quick check against the basics of responsible software: who owns it, what happens when it's wrong, who's accountable once someone else starts relying on it. That's a different skill from building the agent itself, and worth treating with the same respect.
The original ILTA article, Agents: Where GenAI Value Actually Lives, lays out the full personal/workflow/enterprise framework, the sorting checklist, and proportional governance approaches for each category — read that first if you haven't.
Rise Above AI Chaos: A Business Fable About Leading Organizations Through the AI Revolution