We use machine learning at Semperr. We have used it quietly for years, in the corners of our products where it measurably improves the work and where the cost of being wrong is borne by us, not by a client. Where those conditions do not hold, we do not use it. The distinction is worth spelling out, because our clients increasingly ask.
The short version: models assist; people decide. No model writes a client memo without a human name attached. No model files on our behalf. No model is permitted to be the last pair of eyes on anything that leaves the firm.
Where the models earn their keep
Summarization, search, classification, and first-pass document review. These are the tasks at which the current generation of models is genuinely good — not because the models are perfect, but because the downstream human can verify the work in less time than doing it from scratch. The math works out. The client gets the same deliverable faster, and the human who signs it still signs it.
We also use models to surface anomalies. A Trace pipeline runs a statistical check against a client's historical data and flags rows that look unusual. A human analyst decides whether the flag is real. The model narrows the haystack. The human finds the needle.
Where we do not use them
We do not use models to generate the final text of a client deliverable. A memo, a brief, an investment note, a legal summary — these leave the firm under a human name, and they were written by that human. If a model drafted a version that helped the human think, the draft is not the deliverable.
We do not use models to price a deal, render a figure for a report, or decide a citation. The cost of being wrong in those settings is high, the ambiguity is real, and the appropriate accountability structure is a person, not a probability distribution.
And we do not use models we cannot explain. A black-box recommender embedded in a client-facing tool is a liability our clients did not agree to take on. If we cannot, in a meeting, describe how the thing works and why we trust it, we do not ship it.
Disclosure, in plain terms
We tell clients where we use automation. Not in a terms-of-service footnote — in the deliverable itself, where it matters. If a Clad report used a model to pre-classify documents before a human reviewed them, the report says so. If a Trace flag came from a statistical anomaly detector rather than a human analyst, the flag says so.
We keep a copy of every model output we shipped or nearly shipped. If the tooling changes in a way that matters, we tell the clients it affects, and we tell them before the next deliverable, not after.
A note on the next few years
Some of what we do today, we will automate tomorrow. Some of what we think of as requiring a human will, on closer inspection, turn out not to. We will move the line. We intend to move it slowly, publicly, and with the clients it affects in the room.
What will not move is the rule that every output leaving the firm carries a human signature. That is not a technical claim about what models can do. It is a claim about accountability, and about what a firm owes the people who rely on it. At Semperr, both claims resolve the same way: to a person, with a name, who picked up the phone.
— Marco