Switchboard

The visible part of an AI feature is small. Around it sits the same machinery every time — holding a conversation, remembering what was said, finding the right documents, letting the assistant use a tool, keeping the bill under control.

Building that again per project meant each one was a bit different, none of them were good enough to trust with something that mattered, and the interesting work always started six weeks late.

An agent is a record rather than a codebase. Its instructions, which documents it can consult, which tools it can use, which model it runs on, whether it speaks — all of it is configuration, changed and running in minutes without anyone deploying anything.

Giving one a new ability is the same kind of change. Point it at a system, describe what the ability does and when to use it, and the assistant can use it. Notion, spreadsheets, email, and a client's own systems are all just abilities it was granted.

A conversation can also be handed between specialists rather than served by one assistant pretending to know everything. The one that answers is the one that should.

Point it at what you already have — contracts, manuals, past quotes, the shared drive nobody has tidied since 2019 — and it becomes the thing the assistant reads before answering.

Documents are broken into passages, and each passage is filed alongside a short note about where it came from and what surrounds it. That note is what stops a paragraph lifted out of a contract from losing the section it belonged to, which is the usual reason these systems confidently answer with half a rule.

Finding a passage runs two searches at once — one on meaning, one on the actual words — and merges them. Meaning alone misses a part number; words alone miss the customer who asked about "cancelling early" when the document says "termination for convenience". Together they catch both.

Each agent has its own limits: how many passages it may pull, which material it's allowed to see, and how relevant something must be before it's allowed to use it at all. That last one is the difference between an assistant that says it doesn't know and one that reaches for the closest thing it found.

Not every question deserves the expensive model, but almost every system sends all of them there anyway. Ours sorts the incoming question and sends it to the smallest model that can actually answer it, stepping up only when it needs to. The quality is the same and the bill isn't.

The other half of that is knowing where the money went. Every answer, every document search, every tool the assistant reaches for is counted and priced as it happens. So when a client asks what this feature costs to run per month, there's a real number, broken down, rather than a shrug and a cloud invoice.

That accounting is also what makes it safe to be honest. We can look at a feature, see it costs more than it returns, and say so with the figures in front of us.

The same engine drives our own back office — a portal of scheduled jobs that chase the numbers, flag the client nobody has spoken to, and prepare what's due to go out. Eleven of them, running whether or not anyone remembered.

It's here because it's the honest answer to the obvious question. If you say you build AI systems that businesses can rely on, the first business to rely on one should be yours.