The work before the work
Why mapping and diagnosis — not the model — is where every AI implementation has to start.
The seductive lie of this moment is that an AI transformation is an AI problem. Pick the model, write the prompts, wire up the tools, and the work reorganizes itself around the machine. But making a process AI-native is a long sequence of steps, and only one of them is the model. The others — finding the right process to AI-ify, understanding how it truly runs, earning the right to change it, proving the result — are the unglamorous 90% that decide whether any change lands. Almost all of this happens before a single agent is built.
We’ve spent the last several months sitting with the people who own this work — transformation leads, COOs, heads of operations, and forward-deployed teams pushing agents into real businesses. There’s a really obvious pattern emerging: the teams that stall, or don’t build anything beyond toy use-cases, are the ones who skip to agent-building. The teams that move start somewhere way less exciting: a map.
Automating the ideal, not the reality
Every process looks simple from the outside. The golden path of a procurement workflow is twelve obvious steps. Then the exceptions, the workarounds, and the “what ifs” start to multiply, and what looked like twelve steps is really seventy-two. The quick win becomes a multi-month slog, and the math that justified the project quietly falls apart.
(We think this is a big part of the recent executive burnout around agent pilots.)
A head of transformation at an enterprise payroll processor put it sharply: her people can perform a process flawlessly and still not tell you why each step exists. “You do the thing for so long that you lose the critical thinking around why a process needs to run the way it does.” The documented process and the practiced one have quietly diverged over the years. And when the operating team talks to the AI team, they talk past each other. Both sides are making assumptions neither has said out loud.
When you ship an agent without closing this gap, you ship garbage.
A map is how buy-in gets built
The hard part of implementation is rarely technical — and as building agents gets easier and easier, this becomes even more true.
A transformation lead at a global investment bank told us they spent 12 months gearing up to try some automations. They brought in consultants for four months. The consultants declared that there was opportunity to save 80% OpEx across certain functions.
And then reality set in. Getting alignment on the opportunity was near-impossible. There were dozens of stakeholders and massive changes proposed. But no objective truth about what’s happening today, what the post-AI work actually looks like, or the business cases for the changes. So the team spent another eight months validating the consultants’ work to build political momentum internally.
This is what actually moved the org: a patient, bottom-up calculus of where the value really was.
We heard the same thing from an applied-AI team that embeds forward-deployed engineers inside large enterprises. Their scarcest resource wasn’t engineering — it was agreement. A single workflow might exist in a hundred variants, and deciding which one to start with took weeks of back-and-forth with executives who felt every choice as risk. What they wanted most was an inventory of workflows to target against: something concrete enough to point at and say, start here.
The diagnostic IS the real work, not a tax paid before the real work begins. It’s the artifact that aligns the org and provides the certainty needed to make massive changes.
The diagnostic was always the work
The instinct to start with a diagnostic is old, unglamorous, and 100% correct. It just used to be unaffordable to do properly. A global investment bank can afford to spend millions on consultants and internal diagnostic teams, but most orgs can’t spend that much or wait that long.
But we think the economics of diagnostic can change. Mapping that took a year before can take weeks today. There are new tools available today: user session recordings paired with video language reasoning models; limitless document synthesis; AI-assisted interviews. These enable live, granular, and objective mappings of how orgs actually work.
Come map it with us
This is what Cobalt builds: the map that comes first, made from how the work actually happens, current enough to trust and concrete enough to act on. We’re in private beta with a small set of AI implementation teams. If you’re running a transformation or FDE team, we’d like to help. Drop us a line at contact@cobalt-research.com.