← back to the talk

AI Fluency Kit

A working toolkit, not a set of tips. Tool-agnostic — it applies to any coding assistant. Three tiers of increasing stakes: individual fluency to build the judgment, team governance to run it at scale, and regulated deployment — the controls a bank needs to put AI behind a real process. Leading a team or carrying the risk? Tiers 2–3 are written for you.

The scarce skill is no longer can you build it? — it's can you judge it, cost it, and govern it? Judgment is what turns AI spend into value; governance is what lets a regulated organisation act on it at all.

Tier 1

Individual fluency

build the judgment

A standing file the AI reads every session — your rules, context and invariants, so it stops re-making the same mistakes.

One line every time the AI burns you. It remembers nothing between sessions — so you become the memory. Seeded with real finance-domain failures.

How to write a brief the AI can't misread — and risk-stratify it by what the work actually touches.

writing-specs.md

Verify-with-data checklist

The green check isn't proof. How to verify with data — sized to the stakes, because the AI optimises the number you gave it.

verify-with-data.md

Model-routing cheat sheet

There's no "an AI" — one test ran 143× apart in price for the same job, and price didn't predict quality. Which to reach for, and when to pay up.

model-routing.md

Tier 2

Team & governance

run it at scale

Cost controlshero

Turn AI spend into a managed line item: estimate tokens before you call, budget per team, alert on the trend, right-size by task, mind the cache and channel.

cost-controls.md

Model-tier policy

A standing data-classification → model-tier policy: which model is allowed on which class of data. Sensitivity is a gate; blast radius is a dial.

Log every AI call, put humans on the gates that matter, and verify probabilistic output — the record that lets a regulated team use AI at all.

Adoption playbook

From one person to the org: a 30-day on-ramp that earns a measured pilot, and an honest pilot that earns the rollout.

adoption-playbook.md

Tier 3

Regulated deployment

put it behind a real process

How a model earns the right to be used: a standing register, named owners, time-boxed approval by data class. "It was already there" is not a governance answer.

model-approval.md

Vendor evaluation

Buying a model is buying a dependency that sees your data and sets your costs. The benchmark is the easy 20%; data terms, residency and exit are the other 80%.

Incident response

AI fails silently — a confident wrong answer, not an exception. Detect → contain → correct → learn, with a kill switch you've actually tested.

incident-response.md

Regulatory mapping

What these practices satisfy: a generic mapping to APRA CPS 230/234, the Privacy Act, and responsible-lending obligations — the bridge to your compliance team's framework.

regulatory-mapping.md

The evidence behind the cost case — one brief, eight models, full telemetry: the 8-model experiment →