Framework
Published thesisEVOLVE
A public thesis for moving from isolated experimentation to enterprise capability: context, control, intelligence in the workflow, adoption, an operating model and a learning loop.
What capability did the organisation actually build?
EVOLVE is a maturity journey that gives organisations a shared discipline for embedding AI in owned workflows, governing agent autonomy, and learning from every run. This page sets out the public thesis.
The Challenge
AI activity is accelerating. Capability is not keeping pace.
Adoption stays uneven, so value remains locked in individuals. Outcomes stay anecdotal, so activity is mistaken for progress. Governance fragments, so trust and compliance do not travel with the work.
The Risk
The risk is not that AI makes mistakes. The risk is that every team makes different mistakes.
The Destination
A governed intelligence layer across the enterprise
The endpoint is not ubiquitous AI, but a deterministic system where humans remain at consequential decision points.
People + Clients
Outcomes, judgement, and trust remain human-owned.
Workflows + Agents
Assistance, action, and routing sit inside designed work.
Control Plane
Rules, permissions, and evaluation bound what AI may do.
Data + Knowledge
Context stays current, connected, and governed.
Continuous feedback connects outcomes to the next decision.
A Real-World Example
What is a Client Review Pack?
Imagine you're meeting an important client to discuss their investments. Before the meeting, you need to prepare a comprehensive briefing document that shows:
- • Portfolio performance: How their investments have performed over time
- • Market context: What's happening in the markets relevant to their holdings
- • Changes and alerts: Significant movements or risks in their portfolio
- • Recommendations: Suggested actions or adjustments based on their goals
- • Talking points: Key discussion topics and explanations for the meeting
This same process applies to many other business functions: client reports, performance reviews, compliance summaries, or any situation where you need to synthesize information and prepare a human judgement call.
Stage 1: The Manual Work
Why does it take 4 hours to prepare manually?
At the experimental stage, a relationship manager must do this work alone, using AI only for scattered help:
Gather data (45m)
Pull client data from multiple systems, retrieve recent portfolio statements, research market news relevant to their holdings.
Analyse and synthesise (90m)
Manually review performance data, calculate returns, identify trends, spot risks, connect portfolio changes to market events. Maybe use AI to help summarize, but results are inconsistent.
Draft the briefing (75m)
Write the summary, create talking points, format everything attractively. Different RMs use different approaches, templates, and language.
Review and refine (30m)
Read through everything, check for errors, ensure compliance, ensure the tone is appropriate. This step happens in the individual's head, with no documented standard.
The problem: Each RM does this slightly differently. If they leave, that approach leaves with them. Quality depends on who prepares the pack, not on a clear standard.
The Journey
The maturity journey: How AI changes the work
The same Client Review Pack task, through five stages of AI integration, shows how organisations move from isolated experimentation to governed intelligence.
Experimental
An RM uses AI individually to summarise notes.
4h baseline
Enabled
RMs use an approved method and shared template.
2h target
Embedded
AI is built into the end-to-end preparation workflow.
45m target
Empowered
Agents prepare proactively; humans handle exceptions.
15m RM effort
Evolved
Every correction and outcome improves the next pack.
RM reviews exceptions only
*These timelines are illustrative. Confirm your baseline during discovery and agree targets with the process owner.
What changes at each stage?
Experimental
Each RM reinvents the process. No standard. Quality depends on the person.
4 hours
Enabled
Everyone follows a standard template. Consistent approach. Still manual work.
2 hours
Embedded
Automation handles data gathering and drafting. Humans review and approve only.
45 min
Empowered
Agents prepare everything proactively. Humans handle only exceptions and edge cases.
15 min
Evolved
Systems learn from every outcome. Future packs improve automatically. Humans only review exceptions.
Exception reviews only
Stage 2: Enabled
Building a shared standard: From vague to reviewable
When RMs work alone, their AI requests are vague and unpredictable. Stage 2 introduces a structured approach.
Stage 1: Vague Prompts
❌ Unclear purpose · No audience · No constraints · No review standard
Stage 2: Structured Prompt
✓ Clear, reviewable, repeatable · Time cut from 4h to 2h
Stage 3: Embedded
Moving into workflows: From manual to automated
Stage 3 embeds AI into the actual business process. No more manual pulling of data or scattered AI help.
The automated workflow
Retrieve data
System pulls authorised client data automatically
Summarise changes
AI analyses performance and flags changes
Flag uncertainty
AI marks anything requiring human judgement
Draft briefing
AI generates a complete first draft
RM reviews
Human makes final judgement and approves
Publish
Document is routed to the client
Time reduction: 4h → 45 minutes (repeatable, owned, governed, measurable)
The Accountability Rail
AI capability expands; human accountability never transfers
Across all five stages runs a constant: humans remain accountable for consequences. This is not a limitation. It is the design.
AI can scale decisions. Humans remain accountable for consequences.
- 01
Judgement
When there is ambiguity.
- 02
Context
When nuance matters.
- 03
Ethics
When trade-offs affect people.
- 04
Responsibility
When consequences must be owned.
- 05
Critical thinking
When outputs must be challenged.
- 06
Trust
When confidence must be earned.
Responsible AI
The MAS FEAT Framework
To promote public trust, financial institutions must use AI responsibly, ethically, and accountably.
Fairness
No systemic bias
Ensure the AI doesn't disadvantage groups based on race, gender, age, or other protected characteristics.
Ethics
Value alignment
Confirm the AI operates within the firm's values and its intended purpose, with clear ethical boundaries.
Accountability
Human responsibility
Humans (not just code) are ultimately responsible for AI-driven outcomes and their consequences.
Transparency
Explainability
Be able to explain how and why an AI decision was made to a customer or regulator.
Stage 4: Empowered
Governing agent autonomy with explicit boundaries
Agents become trustworthy only when boundaries are defined before deployment. The three-tier system clearly marks what agents may always do, what requires human approval, and what is permanently out of scope.
Always Do
- • Retrieve authorised data
- • Summarise portfolio changes
- • Draft internal briefing
Ask First
- • Recommend products
- • Create client-facing content
- • Apply judgement thresholds
Never Do
- • Execute transactions
- • Change permissions
- • Send client communications
Evidence Gates
Expanding autonomy only when quality improves
Agent autonomy should expand only when evidence crosses predefined gates. This ensures capability grows with demonstrated reliability.
Autonomy Levels
- 1. Draft: Agents draft the review pack for human review.
- 2. Reversible: Agent prepares the review pack automatically.
- 3. Defined categories: Agent handles standard reviews independently.
- 4. Orchestrate: Agent coordinates the entire RM review workflow.
Quality Gates
- Quality ≥ 95%: Outputs meet the agreed standard.
- Exception rate < 5%: Exceptions are handled appropriately.
- Fully traceable: Every action and decision is explainable and auditable.
Stage 4-5: Institution
The Knowledge Lead: Human owner of standards and outcomes
Every AI-powered workflow needs a named Knowledge Lead who ensures human expertise becomes institutional capability rather than individual memory.
Defines
The quality standard and success metrics
Approves
Consequential outputs before deployment
Escalates
Ambiguity and exceptions requiring judgement
Improves
Standards, controls, and performance
Stage 4-5: Learning
Closing the loop: From execution to continuous improvement
In Stages 4 and 5, something fundamental changes. The system doesn't just execute. It learns.
Agent prepares the review pack proactively
Track what the RM did and how they changed the output
Improve the standard based on real outcomes
The shift: A traditional organisation learns through projects (quarterly reviews, annual audits). An AI-native organisation learns through operations (every single execution teaches the system).
Stage 5: Evolved
Building the company brain
Agents are only as reliable as the context and controls that shape them. The AI-native enterprise connects trusted knowledge to action through a deterministic control plane.
Three Layers
- Layer 3: Intelligence. Agents and models
- Layer 2: Control Plane. Definitions, policies, permissions, evaluation, governance
- Layer 1: Enterprise Context. Data, documents, workflows, institutional knowledge
The Feedback Loop
Signals → Portfolio changes, market movements, client interactions
Consolidate → Update context and aggregate new information
Validate → Apply current policies and rules
Act → Prepare reviews, recommendations, communications
Learn → Capture corrections and outcomes for continuous improvement
Know. Do. Prove.
Information governance is becoming AI's context layer. Imagine the smartest analyst in the world, with a bad filing cabinet: three versions of the same policy, some outdated, some never approved, some they shouldn't even be able to open. The AI may be brilliant. Its information may not be. Before AI can give us the right answer, the organisation must know which information is right.
Know
Trusted context
- More data is not better context. AI needs the right document, for the right purpose, at the right time.
- One authoritative source, effective dates and version status, access and purpose restrictions, and an owner, classification and lifecycle status for every item.
- Without an effective date and a “supersedes” link, AI will confidently cite an old policy.
- Fast-moving facts should be looked up live, never stored for AI.
Do
Permissions and authority
- Permissions before retrieval. Authority before action.
- Filter before ranking, and if nothing is configured, the answer is no. If AI reveals that an inaccessible document exists, the design is wrong.
- Being allowed to read information is not the same as being authorised to act on it.
- Sending to a client or placing an order stays human-initiated: the AI prepares, a person sends.
Prove
Evidence
- An AI answer receipt, for approvals and denials alike: request, source, version, access check, decision and the rule that made it, output and review.
- Log “the model declined” separately from “the policy blocked it”, or you can only show the model usually behaves, not that your controls work.
- The receipt shows what the AI was given and what the system did, not the model's internal reasoning.
- Six months later, can you reconstruct the answer?
Three questions before AI touches your information
If you can't trust what it knows, control what it does or prove what happened, you're not ready to scale that use case. And if it goes wrong, can you stop it and say who was affected?
Start with one repository, one use case and one measurable improvement. Any “no” is where your AI readiness work begins. Before you give AI more intelligence, give it better-governed information.
The regulator is moving the same way. MAS issued its Guidelines on Artificial Intelligence Risk Management on 7 October 2026. They apply to all financial institutions and all forms of AI technologies, and take effect on 7 October 2027: Sections 3 to 4 from 7 October 2027, and Sections 5 and 6 by 7 October 2028.
From the talk “Your AI Is Only as Good as Your Filing Cabinet”, AIIM ASEAN Forum 2026, Singapore, 22 October 2026.
Try the interactive Know, Do, Prove demo (illustrative example data)
Implementation
12–24 month implementation roadmap
Build foundations first; expand autonomy only as evidence and governance mature.
Enable: Prompting Literacy
Months 1–6 · Build AI literacy across the organisation
Diagnose current maturity
Assess where teams are on the EVOLVE journey
Establish acceptable use policy
Define what AI can and cannot do in your organisation
Deliver prompting literacy training
Train teams on structured prompting techniques
Establish baseline metrics
Document cycle time, effort, and quality at Stage 1
✓ Outcome: AI value identified and measured across the organisation
Embed: Workflow Redesign
Months 6–12 · Move AI into owned business processes
Document AI-driven processes
Map where AI can add value in key workflows
Assign process owners
Designate accountability for each AI-enabled workflow
Design review and approval points
Define where humans make final decisions
Prove cycle time and quality improvements
Demonstrate measurable progress towards Stage 3
✓ Outcome: Cycle time reduced from 4h to 45 minutes; processes are repeatable, owned, and governed
Empower: Bounded Agents
Months 12–18 · Give agents autonomy within clear boundaries
Define the three-tier boundary system
Always Do · Ask First · Never Do
Assign Knowledge Leads
Name humans responsible for standards and outcomes
Set evidence gates for autonomy expansion
Quality ≥95% · Exception rate <5% · Fully traceable
Evaluate and audit agent outputs
Measure against quality standards; expand carefully
✓ Outcome: Humans focus on exceptions and edge cases, not routine preparation
Evolve: Institutional Intelligence
Months 18–24 · Build the company brain and continuous learning
Build semantic layer
Create a single source of truth for definitions and policies
Connect knowledge systems
Link data, workflows, and intelligence layers
Enforce temporal governance
Ensure policies stay current; flag expired rules
Scale closed-loop learning
Every execution teaches the system; continuous improvement
✓ Outcome: Every review improves future reviews; organisation learns through operations, not projects
Key metrics to measure throughout
Reliability
Quality of outputs; error rates
Cycle Time
Time to complete key processes
Decision Quality
Outcomes and correctness of decisions
Knowledge Retention
What remains when people leave
Status
Published thesis
EVOLVE is a public teaching thesis for enterprise AI capability, not a product. It is released under the same Creative Commons licence as PINPOINT and BLUEPRINT.
© 2026 Ahmed Muzammil · CC BY-SA 4.0