Skip to content
Dylan Taylor
All writing
Essay4 min read

A workflow beats another chat box

Most "AI for X" products are a chat box pointed at X's documents. But professional work is not answering a question — it is completing a repeatable unit of work with a decision and a paper trail.

Here is the shape of most "AI for [profession]" products: a chat box, pointed at that profession's documents, that answers questions in the profession's vocabulary. It's a reasonable first move, and by now it's a commodity. Retrieval-augmented question answering is table stakes — a capable team can stand one up in a week, and so can their competitor.

The problem is that answering a question is rarely the actual job. Professional work is made of repeatable units of work that end in a decision and leave a record: this contract was reviewed against our standard, these three clauses were flagged, this one was escalated to a partner, here's the redline, here's who signed off. A chat box helps you think. It doesn't do that.

Firms don't want a smarter search box. They want the review done — the same way, to the same standard, every time, with a trail they can defend.

Chat is stateless; work is stateful

A conversation is individual and ephemeral. You ask, you get an answer, the state evaporates. A workflow is organizational and durable. It encodes the firm's standard, the steps, who reviews what, the conditions that trigger an escalation, and what gets written down at each stage. It turns judgment that currently lives in one senior person's head into a process the whole firm can run — and improve.

That reframing changes what you build. Instead of optimizing for a fluent answer, you optimize for a decision with a known provenance: approve, flag, or escalate, each carrying the rule and the evidence behind it. You can watch a version of exactly that in the decision-trace sandbox — a clause moves through a rules engine and comes out the other side as an auditable decision, not a paragraph of advice.

The hard part isn't the model

The difficult, defensible work is upstream and downstream of the model. Upstream: capturing the firm's real playbook — the standards and exceptions and house style that make their review theirs — and turning it into something executable. Downstream: the review tables, the permissions, the escalation paths, the audit trail. None of that is a prompt. It's product and systems work, and it's where the moat is, because a competitor can copy your chat box but not a firm's encoded judgment.

This is the bet behind Littman: that the valuable thing isn't a legal chatbot but a system that turns a firm's own standards into structured, repeatable decisions. Chat is the least interesting thing you can do with retrieval. It's the front door, not the building.

LegalAIDecision systems