The fastest way to spend $200,000 on an AI project that nobody uses is to start with the agent and look for the problem.
We see it most weeks. A founder reads a launch post about Anthropic’s Computer Use or Microsoft’s latest Copilot variant, decides their business needs one of those, and books a strategy session. We listen, we ask three questions, and we usually end up recommending against building anything for at least 30 days.
That’s an awkward way to start a sales conversation. It’s also the only honest one.
A good conductor knows that adding more instruments doesn’t make a piece sound bigger. It usually makes it sound noisier. The same is true of AI agents. The question isn’t can you put an LLM in this workflow, it’s whether the workflow will be measurably better when you do, and whether you’d be willing to put your name on the result.
There are three places we routinely tell teams not to build.
The workflow that doesn’t exist yet
If a process is run by one person, in their head, in a way they can’t fully describe, the dissonance isn’t the absence of automation. It’s the absence of a process. Building an agent on top of a non-process gives you a faster, scarier version of the same chaos. The Score for these clients is almost always the same: write the SOP first. Then we’ll talk.
The workflow that already works
Sometimes a team has built a beautiful little Zapier flow, or a Power Automate routine, or a junior staff member who turns 200 emails into 12 actionable summaries every morning. It works. It’s cheap. It doesn’t need a multi-agent architecture and a vector store. Replacing what works with what’s fashionable is the fastest path to a regretful invoice.
The workflow where the cost of being wrong is high and the cost of being slow is low
Tax classifications. Legal contract terms. Anything that goes to a regulator. Anything that touches a customer’s money. These are not places to deploy an agent that’s right 94% of the time. They’re places where a human checks the work, slowly and on purpose, and the agent’s role is to make the human’s checking faster, not to replace it. We build a lot of these. They look less impressive in a demo. They tend to last.
The pattern under all three is the same. The Score isn’t a sales pitch. It’s a triage. We’re trying to find the workflows where the cost of an AI mistake is low, the cost of doing nothing is meaningful, and the upside is a multiplier rather than a marginal saving.
Most businesses we look at have between two and five workflows that pass that test. Not fifty. Not one. Two to five. The honest version of an AI roadmap names them, sequences them, and says nothing about the rest until something changes.
This is uncomfortable for vendors. Their incentive is to find more places to build. It’s uncomfortable for consultancies whose business model is the next deliverable. It’s uncomfortable for executives who’ve been told AI is everywhere and are wondering why their roadmap is suddenly only three items long.
It’s not uncomfortable for the operations teams who’d otherwise be the ones cleaning up after a half-deployed agent that nobody asked for and nobody knows how to switch off.
If you’re sitting with a vendor pitch in front of you and a vague sense that something’s missing, that something is probably this question: what does this look like when it’s wrong, and who deals with it? If neither of you has a clean answer, the agent isn’t the answer.
The conductor’s first job is to listen. Sometimes the room is asking for silence.