Every professional services firm that has tried to deploy AI in the last two years has hit the same wall. The tools work in the demo. They fall apart in production. Outputs are inconsistent. Partners distrust the results. The rollout stalls, and someone declares that AI “isn’t ready for this kind of work.”
The tools are ready. The operations are not.
The pattern we keep seeing
When we do a diagnostic engagement, we ask firms to walk us through how a piece of work actually gets done — not how the playbook says it gets done, but how it really gets done. What we find, almost without exception, is that the real process is held together by experienced humans filling gaps.
A senior associate knows that the client brief is always incomplete and emails the right person before starting. A partner knows that the standard template doesn’t apply to this jurisdiction and adjusts accordingly. A coordinator knows that the system of record is three days behind and pulls numbers from a spreadsheet instead.
None of this tacit knowledge is written down. None of it is structured. And none of it can be handed to an agent.
When firms wire AI tools into these workflows, the agents encounter the same gaps the humans do — but they have no intuition to fill them. So they hallucinate an answer, skip the step, or stall. The output looks wrong, because it is wrong. Not because AI is inadequate, but because the process it was given was never designed to be executed by anything other than a trained human who already knew what to do.
What operations actually means
Operations, in this context, means the explicit, structured, transferable specification of how work gets done. Not a policy document. Not a general description. A specification precise enough that something with no prior context — human or agent — could follow it and produce the expected output.
Most professional services firms have never built this. They haven’t needed to. Experienced professionals internalize firm knowledge over years. The knowledge lives in people, not in the system.
That works fine when the firm scales by adding senior people. It breaks entirely when you try to scale by adding agents.
The seams that block agent execution
The most common failure points we find are not in the hard parts of the work. They’re in the connective tissue — the small decisions, handoffs, and judgment calls that sit between steps.
Who decides when a deliverable is ready to review? What happens when a data source returns an error? How does a task get escalated when a threshold is crossed? When two pieces of guidance conflict, which takes precedence?
For humans, these questions are answered by experience and hierarchy. For agents, they need to be answered by the system design. If they’re not, the agent makes a choice — and that choice is frequently wrong.
The work of making a firm agent-ready is largely the work of finding these seams and closing them. Not by adding AI. By redesigning the process so that the implicit is made explicit, the tacit is made transferable, and the gaps that humans filled by intuition are filled instead by structure.
Build the foundation first
The temptation is to start with the AI. It is more exciting to build than to document. It is easier to demo a tool than to redesign a process. And vendors are eager to sell you a solution before you’ve diagnosed the problem.
But firms that start with the AI spend most of their time managing failures. Firms that start with operations — mapping how work actually flows, identifying the seams, redesigning around them — find that the AI integration is the easy part.
The order matters. Operations first. Agents after.
That’s not a slower path to value. It’s the only path that actually gets there.