AI automation, past the demo stage
Retrieval that cites its sources, classification at scale, and agent systems that run unattended behind review gates and audit-grade logs. Built to survive production, which is a different problem from building something that demos well.
What this practice covers
Applied AI past the demo stage. Retrieval over private corpora with citations back to source, classification at scale, and multi-agent infrastructure with persistent memory, review gates, and audit-grade logging. The unglamorous parts are the ones that decide whether AI survives contact with production.
Retrieval is the most common starting point. An organization already holds the answer somewhere across contracts, email, case files, or documentation, and the cost is not that the answer is missing but that finding it takes a person an hour. Retrieval over a private corpus turns that into a query, and citations back to the source thread are what make the result usable rather than merely plausible.
Classification at scale is the second. Sorting, routing, and tiering a volume of inbound that a person cannot keep up with, where the cost of a miss is a real deadline. Third is agent infrastructure: multi-step systems with persistent memory that carry work forward across runs instead of starting cold every time.
What makes it survive production
The unglamorous parts decide whether AI survives contact with production. A demo has to succeed once on an input somebody picked. A system running unattended has to behave on inputs nobody anticipated, and has to be answerable afterwards for what it did and why.
Three things carry that weight. Verification gates built as first-class agents rather than as a check bolted on at the end, so the step that decides and the step that acts are separate and a person can sit between them. Audit-grade logging of every action, because a system whose behavior cannot be reconstructed cannot be trusted with anything consequential. And citations, so a retrieved answer can be opened and checked against the source rather than believed.
This is also the honest limit of the practice. Where an answer has to be right and cannot be verified, the correct design keeps a person in the loop rather than claiming a confidence the system does not have.
Systems already running
The problem and the build for each of these sit on the work page, alongside the rest of the practices.
- 01
Mid-size law firm
Case-relevant correspondence resolves by search rather than by manual review. Results return with thread citations attorneys can open and verify.
- 02
Internal AI agent platform (HDS)
One principal runs several workstreams in parallel, because review and verification are automated rather than queued behind a person. The agents that built this site are the same ones used in client work.
Common questions
How is this different from wiring up a chatbot?
A demo has to work once, in front of you, on an example you chose. A production system has to work unattended, on inputs nobody anticipated, and be answerable afterwards for what it did. The difference is almost entirely in the parts nobody demos: retrieval that cites its source, review gates before anything acts, and logging good enough to audit.
Can it work over our own documents and email rather than public data?
That is the common case. Retrieval runs over a private corpus you control, and results come back with citations pointing at the source document or thread, so a person can open the original and check it. An answer you cannot trace is not usable for work that carries consequences.
What stops it from acting on something it got wrong?
Review gates. Verification is built as a first-class part of the system rather than a wrapper added at the end, so the steps that change something are separated from the steps that decide, and a human sits on the ones that matter. Every action an agent takes is logged at audit grade.
Which models do you use?
Claude and OpenAI models, routed through a gateway rather than called directly, with pgvector for retrieval and LangGraph and n8n for orchestration. Routing through a gateway is deliberate: it keeps model choice a configuration decision rather than a rewrite, which matters in a field where the best option changes every few months.
Do you use this yourselves or only sell it?
HDS runs its own multi-agent platform for delivery, with persistent memory, code review and verification gates as first-class agents, and audit-grade logging of every action. The agents that built this site are the same ones used in client work, which means the failure modes described here were found the hard way rather than read about.
What would you point it at first?
Start with a free scoping conversation with Mike Hyams, the person who builds and supports the work.