AI enablement in the field
As implementation tutor on Elvante's Level 4 AI & Automation apprenticeship, I'm dropped into client organisations with almost no handover — a funded-learning compliance operation, a software agency, more to come — and expected to find the workflow, set the data boundary, deploy AI usefully, and leave auditable evidence. Two hours a month per client. Forward-deployed work in miniature.
The engagement
Elvante delivers a 14-month Level 4 apprenticeship in AI and automation. I own the implementation lane: taking each learner's real workplace — their actual systems, policies, and constraints — and getting AI deployed usefully inside it, with portfolio evidence an assessor can understand.
The structural challenge is the interesting part. The client passes over almost no context in advance: no employer briefing, no systems inventory, no statement of what's off-limits. The first email and first session are the discovery phase. Every learner is a different organisation with different data sensitivities, and the time budget is roughly two hours a month each.
In the field: the compliance operation
First deployment: a compliance officer at a funded-learning provider, working with sensitive learner records in a Microsoft/MIS environment. No handover beyond a name.
- Discovery: an intake email followed by live mapping of her report-checking workflow — where the hours actually went.
- Constraint: GDPR-sensitive learner data that could not touch public AI tools. The boundary was designed first; the automation had to fit inside it.
- Deployed: scheduled-task and email-triage patterns, plus a recurring regulatory-change monitoring task — value from AI without any sensitive record leaving her environment.
- Field lesson: mid-session, my own paid AI account hit usage limits. Since then every build runs on the client's own account, with a backup open. Never depend on your tooling in someone else's environment.
In the field: the software agency
Second deployment: a developer at a software agency, no AI in the company workflow, copy-pasting fragments into ChatGPT under a hard confidentiality boundary. The planned session was a redacted prompt-pattern workflow.
Ten minutes in, it was clear the plan was wrong. For a developer, the step-change is integrated tooling, not better copy-paste — so the session pivoted live: an AI coding agent connected to his IDE, GitHub, and deployment, then to his real ticket queue. Watching the agent read his actual tickets was the moment the entire programme became concrete for him.
Touching a company system mid-demo is also exactly where governance lives, so it became the lesson: supervised demo only, connection revoked afterwards, and a written employer-approval requirement set as the gate before any real use — with the specific questions his manager needed to answer. The target workflow (ticket → AI-assisted fix → pull request, human review at every step) is designed and waiting on that approval.
The demo that touches their system is ten times the demo that touches yours — but it drags governance forward with it. Stage the approval conversation before the wow moment, so it lands inside cleared boundaries.
What this practices
- Walking into unknown organisations and finding the real workflow fast, with no handover
- Treating data boundaries and tool policy as design inputs, not afterthoughts
- Reading the room and re-planning live when the prepared session is the wrong session
- Turning governance from a blocker into deliverable structure — approval gates, evidence trails, written boundaries
- Leaving auditable artefacts behind: session recaps, attendance records, portfolio evidence, reflection prompts
Status
Deployments
Two client organisations live, two more onboarding
Cadence
~2 hours per client per month, 14-month programme
Evidence
Delivery log, session recaps, and assessor-ready learner portfolios
Why this project
Causeway shows I can build AI systems inside one product over a long engagement. This shows the other half of forward-deployed work: parachuting into organisations I've never seen, under real confidentiality and governance constraints, and getting AI adopted — usefully, safely, and with evidence — on a tiny time budget. Learner and employer details are anonymised; the delivery structure and lessons are real.