Training Isn't Prompting

It's Behaviour Change Under Regulation.

Most AI training programmes teach people how to write a better prompt. That is not what regulated financial services firms need. What they need is training that changes three specific behaviours: how people handle data when AI is available; how they treat AI outputs before acting on them; and how they escalate uncertainty rather than improvise around it.

In a regulated environment, those three behaviours are the operational difference between AI as an asset and AI as a liability. The distinction also matters commercially: Gartner's research consistently finds that productivity gains only convert to financial impact when workflows are redesigned and behaviour changes to match the new model.

The Skills Gap in Jersey Is Real, Specific, and Consequential

The Jersey AI Playbook (March 2026) identifies skills development as the most consistently cited priority across every major stakeholder group—Digital Jersey, Jersey Finance, the JFSC, the IoD, and the Government of Jersey itself. All of them highlight the same gap: AI capability is advancing faster than the workforce behaviours needed to deploy it safely and effectively.

In a small island jurisdiction with natural talent constraints, the skills gap has a direct economic consequence. The AI Playbook is explicit: the cost of delayed investment in skills development exceeds the cost of the investment itself. Every month of inaction widens the gap relative to competitor jurisdictions moving in a more coordinated way.

What makes the skills gap in regulated financial services distinctive is that the competencies most urgently needed are not primarily technical. They are behavioural and judgement-based:

• Knowing precisely what is permitted, what is prohibited, and when to escalate before proceeding
• Treating every AI output as a draft until the required validation has been completed
• Understanding which records must be retained and how that retention is evidenced
• Recognising when a workflow requires stronger human oversight and acting on that recognition

Why Prompting Training Alone Fails

The most common AI training failure mode is the 'prompt engineering' programme: sessions that teach staff how to get better outputs by writing clearer inputs. Most failures in AI-enabled regulated workflows are not caused by poor prompt construction. They are caused by:

• Data handling errors: using the wrong data in the wrong environment because the boundary was not made explicit
• Validation failures: acting on AI output without the required human review
• Record-keeping gaps: failing to document human review in a way that produces usable evidence
• Workaround behaviour: creating routes around the approved boundary because it feels slower

Fixing these requires governance-first training—curriculum that begins with boundary discipline and works outward to tool capability. Leading with prompting and hoping governance follows is the wrong sequence.

A Minimum Viable Training Curriculum

Base module — all staff, mandatory, with attestation

• Acceptable use: what is permitted, what is prohibited, what requires escalation
• Data handling: what must never leave the controlled environment, and why
• Validation: what 'AI output as draft' means for different workflow risk classes
• Escalation: who to contact, how to document the interaction, and why asking is the safe behaviour

Role modules — targeted and workflow-specific

• Client-facing and relationship roles: confidentiality boundaries, safe drafting, mandatory review before any client-facing output is used
• Operations teams: AI within controlled workflow steps, exception documentation, escalation triggers, record-keeping requirements
• Risk, compliance, and audit functions: what monitoring signals matter, how oversight is evidenced, governance cadence

Manager module — the most critical investment

• How to reinforce boundary discipline in daily team interactions, not just formal reviews
• How to respond to a bypass request or exception submission—what to ask and what to document
• How to review monthly KPI signals for drift before it becomes a pattern

The manager module is where most AI training budgets are under-invested. Culture is transmitted through the middle layer. If managers treat AI controls as optional, the rest of the organisation will treat them as optional.

Validation as a Trained and Measured Skill

Gartner's AI accuracy research makes a point that should shape every training programme: AI errors are inevitable. The management question is not how to eliminate errors—it is how to detect and manage them proportionately.

• Low-stakes internal drafting: lightweight review before use, periodic spot-check sampling by management
• Client-facing communications: mandatory human review before sending, documented in the workflow record
• Regulated outputs (suitability, compliance, financial crime): explicit human-in-the-loop gates with documented rationale forming part of the audit trail

A simple rule that works: AI output is a draft. It becomes a decision input only after the required checks have been completed for that workflow's risk classification.

The Four-Week Deployment Plan

Week 1: Publish the permitted/prohibited/escalate guidance in plain language. Brief all managers on their role as the primary reinforcement layer. Define escalation routes and record-keeping expectations before any tools go live.

Week 2: Complete mandatory base module rollout with attestation recorded. Deploy role-specific scenario tables giving concrete examples for daily situations each function will encounter.

Week 3: Run focused sessions for operations, risk, and compliance functions. Collect high-friction points from early users and feed them into the controls backlog. Identify where guidance needs refinement.

Week 4: Review early signals: bypass attempts, exceptions raised, questions via the escalation route. Refine guidance where needed. Confirm manager cascade completion—not assumed, but confirmed.

Measuring Training Effectiveness: Behaviour, Not Attendance

Training completion rates measure whether people were present. The indicators that a training programme is having a real effect include:

• Bypass attempts trending down in the weeks following rollout
• Exceptions attributed to misunderstanding—rather than genuine risk appetite edge cases—declining
• Quality outcomes in workflow QA sampling improving over the first two cycles
• Manager cascade completions confirmed by name, not estimated by assumption
• Escalation route usage increasing in the first month (an indicator that people are asking rather than improvising)

What Only Executives Can Do

Training fails when it is treated as a delegation item. Executives control four things that no training programme can substitute:

• Tone: Articulating clearly, and repeatedly, why boundary discipline matters—as operational foundation, not compliance checkbox
• Incentives: Making explicit what behaviours are rewarded and what are not
• Resourcing: Protecting the time required for training, workflow redesign, and management reinforcement
• Reinforcement: Asking for evidence of behaviour change rather than assurances of compliance

If the C-suite treats AI training as an IT or HR delivery item, the organisation will treat AI controls as optional.

Closing question: Does your AI training programme change how people handle data, validate outputs, and escalate uncertainty—or does it primarily teach them how to write a better prompt?

Next in the series — Part 5: Audit-ready AI is an operating model, not a policy document. What regulators and auditors will actually ask, and what evidence you need ready.

Sources: Jersey AI Playbook, March 2026; Gartner AI accuracy and workforce research 2025–2026; PwC Channel Islands Wealth Management Insights 2025.

Professional disclaimer: The views expressed are my own and are shared for general informational purposes. This content is intentionally non-confidential and does not include proprietary, client-specific, or employer-sensitive information. It is not legal, regulatory, investment, or professional advice; readers should seek appropriate advice for their circumstances.

Financial services leader (20+ years) | Lean Six Sigma Green Belt | London School of Economics (LSE) AI Leadership Accelerator

A version of this article was first published on LinkedIn as part of the Sequence body of work. This is the canonical version.

The views expressed are the author's own and shared for general information. This content is intentionally non-confidential and does not include proprietary, client-specific or employer-sensitive information. It is not legal, regulatory, investment or professional advice. © 2026 Paul Soares. Sequence is the trading name of Sequence Advisory Limited (in formation) · hello@sequenceadvisory.com