Loading page
How it works
This is the whole engagement, start to finish. The point of writing it out in this much detail is that you should be able to tell, before you spend anything, exactly where a human is accountable and where you have a veto.
The core loop
Phases five through nine run continuously once you are live. The two lanes below are the reason the loop is trustworthy: work never crosses into production without passing through the human lane.
Diagram of the seven-stage operating loop. Stages three and four sit in a human lane; the rest are performed by AI, and the sequence repeats.
Specialist agents write the automation, the integration or the software, working inside an isolated workspace against your real requirements.
A second, deliberately stronger model checks the output against the acceptance criteria. The model that reviews is never the model that built.
A human engineer reads the diff, the verification report and the risk level, then signs. Nothing reaches your systems without this signature.
Anything that touches money, customers or production waits for your explicit approval. High-risk changes need more than one approver.
Release runs through sandbox and staging first. Every deployment is reversible and recorded with who approved it and what exactly shipped.
Live automations are watched for failures, drift and cost. Incidents open themselves and are routed to an engineer.
Run telemetry feeds opportunity discovery, so the next improvement is proposed from evidence rather than from a sales calendar.
Phase by phase
We build a picture of how work moves through your business: the tools, the handovers, the steps that only exist in somebody's head. This comes from your answers, not from system access.
AI, reviewed by an engineerA written map of your current process
The map is analysed for waste, risk and fragility. Findings are separated into what was confirmed, what was inferred and what was projected — and labelled accordingly.
AI, reviewed by an engineerAn assessment report you keep
Opportunities are ranked on payoff against effort and risk, with a recommended first project scoped tightly enough to quote at a fixed price. We also tell you what not to automate.
AI, priced by a humanA ranked opportunity list and a fixed-price proposal
Nothing starts until you approve the scope, the price and the acceptance criteria. High-risk changes require more than one approver, and the approval is recorded against that exact version.
YouA signed scope with acceptance criteria
Specialist agents build inside an isolated workspace. Credentials for your systems are issued for a single run and revoked when it ends; a workspace can only reach endpoints on your allowlist.
AIA working build in a sandbox
An independent model tests the build against the acceptance criteria and produces a verification report. Then a human engineer reads the change, the report and the risk level, and signs.
AI, then an engineerA verification report and an engineer's signature
Release moves through sandbox and staging before production, and every stage is reversible. Anything touching money, customers or production data waits for your explicit release approval.
AI, released by youA recorded, reversible deployment
Live automations are watched for failure, drift and cost. Incidents open themselves and are routed to an engineer, so a break is our problem before it is your customer's.
AI, escalating to a humanLive health, incident history and monthly reporting
Run telemetry feeds opportunity discovery. The next thing we suggest comes from evidence in your own data, and it enters the loop at Recommend like everything else.
AI, proposed to youNew opportunities, evidenced and ranked
What you control
These are structural, not policy. They are enforced by the platform, and they apply to every project regardless of size.
Not a checkbox and not an agent reporting success. A person reads the change and the verification report, and their signature is recorded against that specific version.
An agent saying it finished is an input, not a conclusion. A task only moves to verified when an independent check against the acceptance criteria passes.
System credentials are issued when a run starts and revoked when it ends. They are never persisted into an agent's workspace and never shared between clients.
A workspace can only reach the systems you approved. Anything else is refused at the boundary rather than logged after the fact.
Releases are staged and recorded with what shipped and who approved it. Rolling back is a normal operation, not an incident.
Honest expectations
We will not give you a delivery date before we know the scope, and we will not put a payback figure on a proposal without showing the arithmetic behind it and labelling it an estimate.
Integration work is where timelines slip, and it slips for reasons outside the build: a vendor’s API access takes two weeks to approve, a legacy system has no export, a data set turns out to be in worse condition than anyone believed. When that happens we tell you the same week, with the revised date and what it depends on.
We will also tell you when the answer is that you should not automate something — because the volume is too low, because the process should be fixed before it is automated, or because the judgement involved should stay with a person. That is part of what the assessment is for, and it costs you nothing.
The assessment takes about fifteen minutes of your time and returns a ranked, costed opportunity list reviewed by a human engineer. You keep the report either way.
No payment details. No sales call required.