Automation & Integration
Which manual steps a language model can take on, and which we never let it.
In the manual steps we already automate — moving data, filling a queue, drafting a reply — a language model sometimes does part of the work now. This page is about where that helps, where it stops, and the one step a model here never takes on its own: writing anything back without a person approving it first.
evidence status · capability and method; no named AI engagements published
Human-in-the-loop, by design
An assistant proposes; a person disposes. That is the whole design, and it holds regardless of how good the draft is: nothing the assistant produces reaches a system of record until a named person has looked at it and said yes.
The workflow, step by step. A source system sits inside a boundary that marks what the assistant may read. Nothing outside that boundary is available to it.
- The assistant reads the source system, and only the source system — nothing outside the boundary shown.
- It drafts a proposed action from what it read: a reply, an extraction, a routing decision. Nothing is sent or written at this point.
- The proposed action goes to a named person at an approval gate, with the source shown alongside it, so the reviewer can check the draft against what the assistant actually read.
- Only after that person approves does the action write to the system of record. A rejection sends the proposed action back to the assistant for revision; it never reaches the system of record on its own.
The one step in this sequence that never happens without a person is the fourth — see "What we will not build" below.
company delivery history · founder-asserted ·
What we will not build
the boundary · the decision stays with a person · Scoring and decisioning
Fully autonomous decisions are out of scope — the same limit the rest of this site holds to, not a weaker version of it for this silo. An assistant here drafts, extracts, routes and proposes; it does not execute, and it does not decide.
We will not wire an assistant to send data anywhere a client has not explicitly agreed to send it — a document handed to a script that calls out on its own, a request routed through a service nobody signed off on. What a workflow can read and where it can write is agreed before anything runs, not discovered afterward.
And we will not ship a flow nobody can audit. If a person cannot see what the assistant read, what it proposed, and why a request was declined, the flow is not finished — the logging is not optional.
What we build
Three things, all inside workflows our automation and integration work already touches:
- Assistants inside a workflow — LLM-assisted development applied to steps we already automate: a model drafts the reply, the extraction or the routing, and the source it read is shown alongside the draft, not hidden behind it.
- Human-in-the-loop approval designs — the assistant proposes, a named person disposes, and nothing writes back until that person says so.
- Build, buy, or hybrid evaluations — run before any platform commitment, so a client is not sold a subscription when a scoped flow would do.
company delivery history · founder-asserted ·
Build, buy, or hybrid
Before any platform commitment, we work out which answer fits the workflow in front of us. The candidates are a scoped flow built specifically for it, an existing platform configured to do the same job, or some of each. A subscription is not the default answer. It is one option we cost against the others, the same way we cost any other automation work before we build it.
company delivery history · founder-asserted ·
How it is engineered
The same production discipline as the rest of our automation work applies here, without exception:
- Fail-closed gates — when a validation check on the assistant's output fails, or the input does not match what the flow expects, the flow stops and waits for a person instead of guessing.
- Logged actions — what the assistant read, what it proposed, and what a person decided, all kept.
- Telemetry — the same monitoring discipline we run on the rest of our production automation.
- A defined rollback — agreed before the flow goes live, not improvised after it breaks.
company delivery history · founder-asserted ·
Continue
Start with a call about the process someone still re-keys by hand.
30 minutes · with the engineer who would scope the work
one-page recap within 2 business days of a call that progresses