Moving an 80 Person Singapore Tier 1 Bank Product Team to an AI-Native Agile Operating Model

The situation

An enterprise of this profile is usually one of the Singapore tier 1 commercial or wholesale banks, regulated by the Monetary Authority of Singapore and subject to the Banking Act, the Technology Risk Management Guidelines, the FEAT principles for AI use, and the Risk-Based Capital regime that informs how product profitability is measured. The product team that triggers the engagement is rarely the digital channel alone. It is more often a horizontal product line such as cards, mortgages, SME lending, or transaction banking, where a P and L owner has been told to deliver new AI-augmented features without expanding headcount and without breaching the model risk perimeter.

The participant pool sits around seventy to ninety people. Product owners, senior business analysts, engineering leads, design, the embedded risk and compliance partners, and a handful of operations leads who own the customer journey end to end. Some are ICAgile certified from earlier waves. Some have a Scrum Master certificate from elsewhere. Most have used a generative AI assistant in a personal capacity. Very few have written a single requirement that contemplates an AI-augmented control or has thought about what changes in the definition of done.

The business pressure that triggers the work is usually one of three. A regulator request that lands as a model risk review and exposes a gap. A competitor launch that compresses a planned roadmap. A senior leader returning from a peer bank visit overseas who has decided the current operating cadence is too slow. Engagements of this archetype begin with the recognition that the gap is not tooling. It is the way the team writes work, validates work, and signs off work.

What we observed

Across a Singapore tier 1 bank product team of this scale, four capability gaps recur. None are theoretical. Each was raised inside the discovery conversation, before the founder proposed a single course.

1. The backlog still treats AI as a feature, not as a way of working

User stories continue to be written as if a human will execute every step. When the team experiments with an AI assistant, it is bolted on as an additional acceptance criterion rather than redesigned into the user journey. The first effect is that the backlog inflates without delivery cadence improving.

2. Risk and compliance partners are pulled in too late

The embedded risk partner is consulted at the end of refinement, not at the start of discovery. By that point, the team has invested in a design path that the model risk function has to reject. The reject does not happen explicitly. It happens as a delay that nobody wants to attribute to risk.

3. Product owners cannot brief an AI workflow well

The single highest leverage skill in an AI-Native Agile product team is the quality of the brief that the product owner writes for an AI-augmented capability. In most banks we walk into, that brief reads like a prompt rather than a product specification. There is no statement of what is in scope for the model, what evidence the model should consume, what the human signs off on, or what FEAT obligation the workflow satisfies.

4. Engineering leads cannot evaluate model output quality

The engineering lead can review code. They cannot, yet, review the output of a retrieval-augmented workflow against a representative sample. That single missing skill is what turns AI-augmented features into a slow trickle through the change advisory board, instead of a confident release.

The engagement approach

Engagements of this archetype run as a structured cohort with founder-led discovery and senior-trainer cohort delivery. The shape below is representative of what a Singapore tier 1 bank product team commits to.

Discovery, week zero

The founder, Prashant Shinde, runs a two day discovery with the sponsor group. Output is a one page operating model gap statement, a participant segmentation, and a target outcome map that names the three product line metrics the engagement will move. This is signed off by the head of product, the head of risk and the head of learning before the cohort opens.

Foundation, weeks one and two

The cohort splits into two tracks. The product, design and BA track runs AI Agile for Singapore Banking and the AI Product Pathway. The engineering and tech lead track runs the AI Agile Coach Foundations module focused on AI-Native Agile system design. Both tracks meet in joint workshops on FEAT, model risk and the Technology Risk Management Guidelines reading of an AI-augmented release.

Practitioner, weeks three to six

This is where the work product changes. Each squad inside the cohort brings a live backlog item. The PO rewrites it as an AI-Native Agile work item with explicit model risk, explainability and human-in-the-loop statements. The engineering lead designs the evaluation harness. The risk partner reviews. The founder runs a fortnightly senior reviewer panel where a sample of work items is critiqued in front of the cohort. By week six, a working pattern library is emerging.

Certification and embed, weeks seven and eight

Participants who choose the certification path complete ICP-ACC or ICP-APO with an Agile Visa licensed trainer. The cohort closes with a sponsor showcase where each squad presents two pattern library entries and one set of metrics they have moved. The founder hands the sponsor group an embed plan, including the next cohort wave and the senior coach who will continue the cadence.

The total commitment for a participant is between forty and sixty hours over eight to ten weeks. The sponsor commitment is a fortnightly checkpoint with the founder. The pattern, our research and our learner record across 75,000+ professionals across 140+ countries, is that this is the minimum viable shape that moves the operating model rather than just the slide deck.

What changed

Engagements of this archetype move four categories of outcome. We deliberately do not publish dollar figures or single-number percentages here. Each engagement starts from a different baseline and the bank owns its own scorecard.

1. Time from regulatory ask to AI-augmented control implementation

The cycle from a model risk question landing on a product line to a documented, defensible AI-augmented control compresses. The change is not the AI doing the work. The change is the product owner and the engineering lead writing the brief and the evidence pack in a form the model risk function can review the first time.

2. Proportion of AI features that survive the first model risk review

Before the engagement, the first-pass survival rate is low enough that teams self-censor and stop proposing AI work. After the engagement, the first-pass review is treated as a normal gate, not as an existential threat to the roadmap.

3. Backlog quality and refinement throughput

The pattern library begins to act as a force multiplier. New work items are drafted faster because the team is no longer inventing the structure each time. Refinement sessions get shorter while the depth of risk thinking gets deeper.

4. Senior engineer engagement with AI tooling

The category we are most interested in is whether the most experienced engineers, the ones who hold the architectural memory of the bank, are now comfortable evaluating model output. When this lifts, the bank stops being dependent on a small AI-tooling vanguard and starts running AI as a normal engineering discipline.

Lessons that transfer to other Singapore tier 1 banks

Four lessons recur across this archetype. They cost nothing to apply, and we share them because they reduce the risk of the next bank running a worse version of the same programme.

• Treat the brief as the product. The single highest leverage skill is how the product owner writes the AI brief. Invest in that one skill before any tooling rollout.
• Bring risk into discovery, not into refinement. The model risk function is a designer, not a gate. Train them to sit in discovery and the rework cost falls dramatically.
• Build a pattern library, not a slide deck. The cohort output should be reusable artefacts the next squad copies, not a deck nobody opens again.
• Sponsor cadence is non-negotiable. If the sponsor cannot give a fortnightly hour, the cohort will revert. The engagement does not survive a sponsor who delegates entirely.