AI integrations into the business.

What we mean by “AI integrations”#

We mean putting AI into the parts of the business where it does real work — first-touch intake triage, content drafting, document classification, ops research, support tier-1 — and out of the parts where it just generates noise. We don’t ship a chatbot on the homepage and call it a strategy. We integrate the model into the function it should already be doing better.

Where AI earns its keep#

Intake and triage#

A tuned model that does the first 90 seconds of intake — captures the prospect’s situation, classifies it against the firm’s actual practice areas (or service tiers), flags urgency, and hands a clean record to a human within minutes. Bilingual where the buyer market needs it. Bar-compliant or sector-compliant disclaimers wired in.

Drafting that doesn’t sound like AI#

Internal writing tools tuned on the firm’s voice and previous work, with retrieval over the firm’s own documents, so first drafts don’t read like the generic LLM mean. The savings aren’t in the drafting time — they’re in the editing time, which collapses by roughly half once the system has the firm’s voice and the firm’s facts in it.

Classification and routing#

Inbound email, support tickets, document review, contract review — wherever an analyst is reading something to figure out what bucket it goes in. Models do this competently, faster, and with auditable explanations attached to every decision.

Internal research#

Retrieval-augmented systems over the firm’s own corpus (case files, prior engagements, internal documentation) so the senior people stop being the firm’s search engine.

How we approach it#

We start from the function, not the model. Every engagement begins with: what is the team doing right now that an LLM could do at 80% the quality and 10% the time? Then we build the minimum system that does that thing well, measure it for two to four weeks against the baseline, and either expand the surface area or kill it. We don’t roll out company-wide LLM access as the deliverable — we ship one function at a time.

We use the API directly. We don’t sell our clients on hosted “AI platforms” that charge a markup on a thin wrapper around a model the client could call directly. Where a third-party tool genuinely earns its keep — call transcription, structured extraction — we’ll integrate it and tell the client what it costs.

What we don’t do#

We don’t deploy customer-facing chatbots that answer substantive legal, medical, or financial questions. The model produces statements the client is responsible for, the disclaimers protect less than people think they do, and the failure modes are public. We deploy customer-facing AI only for surfaces where the failure mode is “ask a human” — appointment booking, FAQ surfacing, intake routing — never substantive advice.

Frequently asked questions#

Which model do you use?

Whichever one is best for the task on the day we’re building it. We default to Anthropic’s Claude for drafting and long-context tasks, OpenAI’s models for tool-use-heavy workflows, and open-weights models (Llama, Mistral) where the data has to stay on premises. The choice is per-function, not a religious commitment.

How do you handle data privacy?

We use the API tier with zero-retention enabled where the provider offers it, route through a private endpoint where the data is sensitive enough to require it, and host on-premises with open-weights models where the client’s policy or regulator requires it. We document the data flow in every engagement.

Will this replace people on the team?

It replaces tasks, not people. The tasks it replaces best are the ones the team is least excited to be doing — the form-filling, the bucket-routing, the first-draft generation. The freed time goes into the work that compounds. We have not, in any engagement, used “headcount reduction” as a benefit. It’s not the value.

Can you build us a custom model?

Probably not, and probably not what you actually need. Fine-tuning a frontier model with a few thousand internal examples produces a marginal improvement at significant cost. Retrieval-augmented generation over your existing corpus produces a much larger improvement at a fraction of the cost. We default to RAG; we only fine-tune when the task is structured and high-volume.