Automations.

What we mean by automations#

Custom-built workflows, agents, and internal tools that take real recurring work off the team’s plate. We’re not selling Zapier templates with a markup. We build software that does the specific job the business needs — read this inbox, classify these documents, draft this email, watch this report, escalate this exception — and we ship it, document it, and operate it for as long as the function compounds.

How the work actually runs#

We start from a process the business is already doing manually. A team member walks us through it; we map the steps, the decisions, and the data the steps depend on. Within two to three weeks we have a working v1 the team is using alongside the manual process. Within six to eight we’ve replaced the manual process and freed the time. We then operate the system — the alerts, the retries, the audit log, the edge cases.

Where it earns its keep#

• Inbound routing and classification. Email, support tickets, intake forms, contract review queues.
• Drafting at the start of a task. First-draft proposals, follow-up emails, internal memos — written from the firm’s own corpus.
• Reporting that’s actually read. Daily/weekly digests that surface the three things that changed and what to do about them.
• Background watchers. Monitoring services where the business cares about state changes — pricing, rankings, news, regulatory feeds, competitor sites.

Human-in-the-loop, on purpose#

Every automation we ship has an explicit human-in-the-loop step at any point the failure mode is “we sent something we shouldn’t have sent.” The model drafts; a human approves. The model classifies; a human spot-checks the audit log weekly. The model proposes the action; a human signs off before it executes. We don’t build “autonomous agents” that act on external systems without supervision. The promised cost savings of removing the human are almost always outweighed by the cost of cleaning up a hallucination after the fact.

What we don’t sell#

We don’t sell access to a vendor automation platform. We don’t sell “AI agents” that we white-label from someone else. We don’t bill for “implementation” on tools the client could deploy in a week themselves. The deliverable is the running system, the documentation, and our team operating it — not a setup fee on a SaaS subscription.

Frequently asked questions#

Which model do you build on?

Whichever fits the task. We default to Anthropic’s Claude for drafting and long-context work, OpenAI’s models for tool-use-heavy workflows, and open-weights models on-premises where the data is sensitive enough to require it. The choice is per-function.

Do these run on our infrastructure or yours?

Whichever the client prefers. Default is our infrastructure during build (so we can iterate quickly), then we hand off or co-operate against the client’s infrastructure once the system is stable. Sensitive workloads run on the client’s infrastructure from day one.

How long until we see results?

A v1 you can use against a single workflow within two to three weeks. Real time savings within six to eight. The compounding work — the documented audit log, the edge-case handling, the routing to the right human — accumulates over months and is what separates a useful automation from a brittle one.

Can you replace our existing automation stack?

Sometimes. We audit what’s there before recommending anything. Sometimes the right answer is to keep Zapier or n8n for the simple triggers and build new only for the parts that are actually breaking. We don’t reflexively recommend replacement.