Loading page
AI capabilities
Capability lists are easy to write and hard to trust. This one is organised around the work: what goes in, what comes out, and which part a human still owns.
Seven things it does
Intake is a structured questionnaire, or a conversational interview that fills in the same answers. Either way it covers your tools, your team and the work that repeats, and the answers are turned into a structured picture of the process.
In the interview, anything inferred rather than stated in your own words is labelled as inferred.
A plan is not prose. It is a dependency graph of tasks with acceptance criteria, owners, risk levels and the approval gates between them — which is what makes progress measurable rather than reported.
Each task states what it needs before it can start and what proves it is done.
Agents write and edit code, configure automation platforms, set up data pipelines and prepare content, working inside an isolated per-client workspace with a filesystem, a shell and version control.
One client, one container, one workspace. Never shared.
Agents never hold your credentials and never call your systems directly. Every write is carried out by the platform, and the endpoint, the exact payload and the idempotency key are all inside the signature a named person gives it.
A plan that changed by one character after it was signed is refused, not quietly re-approved.
A separate, deliberately stronger model checks the output against the acceptance criteria agreed before the work started, and writes the verification report a human engineer reads before signing.
The reviewing model is never the model that produced the work.
Deployed automations and connected systems report health, failures, drift and cost. A high- or critical-severity signal opens an incident automatically, tied to the event that raised it, and our staff are alerted.
Repeats of the same problem fold into the one open incident instead of piling up beside it, and every failed run is recorded against its automation.
Run history and monitoring signals are analysed for steps that fail often, run out of time, or retry without ever succeeding. Your engineer runs that analysis, reviews what it finds, and presents the ones worth doing.
Proposals enter the same approval loop as everything else.
The workforce
Work is routed to the role that owns it. Narrow roles can be held to an acceptance criterion; a single general assistant cannot.
Maps how work actually moves through your business today, including the steps nobody documented.
Reads your tools, your data and your market so recommendations are grounded in your situation, not a template.
Ranks opportunities by payoff, effort and risk — and says plainly which ones are not worth doing.
Designs the automation: triggers, branches, failure paths and the human checkpoints in between.
Writes and refactors the code, in a sandboxed workspace, against acceptance criteria agreed up front.
Connects the systems you already pay for. Credentials are scoped to a single run and revoked at the end of it.
Moves, cleans and models the data behind reporting and knowledge systems, with lineage kept intact.
Tests the build against the acceptance criteria and produces the verification report an engineer signs.
Screens what an outbound write says before it leaves, and refuses one whose content breaks a rule you set.
Writes the runbook, the SOP and the handover notes so your team can operate what was built without us.
Watches the signals live automations report, and opens an incident on a high- or critical-severity one for our engineers to work.
Reads run history and proposes cheaper, faster or more reliable versions of what is already deployed.
Sequences the plan, chases the blockers and keeps the delivery dates you were given honest.
Under the hood
Each kind of work is bound to a model chosen for that work: careful analysis, fast classification, long-form documentation, high-volume agent turns. The binding is configuration, not code, so a better model is a settings change rather than a rebuild.
Which model produced an artifact, and under which settings, is recorded with the artifact. If a claim in a report is ever challenged, we can say exactly what generated it.
When the platform runs without a live model connection — in a demo environment, for instance — everything it produces is stamped as simulated. We would rather show you an obvious placeholder than a convincing one.
Limits
A capability page that only lists capabilities is a sales document. These constraints are enforced by the platform, not left to good intentions.
The assessment uses the same intake and analysis described on this page, and the report tells you which findings were confirmed and which were inferred.
No payment details. No sales call required.