Case Study · Enterprise AI & Service Design

AI for commercial lending

From an AI narrative writer to orchestrating an end-to-end credit application journey for Relationship Managers.

My roleLead Experience & Service Designer — shaped core agent workflows, validated prototypes with RMs, and aligned product & AI architecture roadmaps.
ClientLloyds Banking Group
TeamPO, CJM, AI Lead, Solution Architect
TimelineApril – July 2026
The Context

Three roles, one connected workflow

Client SME Business Owner

Provides financial statements and supporting documents to secure commercial lending.

Primary User Relationship Manager (RM)

Gathers data across 12+ systems, analyses risk, and writes the credit application for sanctioning.

Approver Credit Sanctioner

Reviews narrative, financials, and securities — returning incomplete cases for rework.

The Pivot

We started by designing a writer

Discovery proved writing is only the final articulation of a much larger preparation problem.

10InterviewsRMs, Credit Officers & Directors across Agriculture, Real Estate & Healthcare.
6RM SurveysValue-vs-effort scoring across the end-to-end credit preparation journey.
2–3 hrsPer NarrativeRaw manual effort spent compiling and writing a single credit application.
Key Insight

Discovery proved the friction wasn’t typing — it was fragmented data, manual spreading, and missing upstream context.

Original briefAI writes the credit narrative
Reframed scopeAI orchestrates evidence, analysis & rationale for a first-time-right application
The Problem

Manual ingestion slows preparation before writing begins

  1. 01 · NeedCustomer need
  2. 02 · ShapeShape lending request
  3. 03 · PrepareCredit application
  4. 04 · DecideCredit decision
  5. 05 · FulfilFulfilment
  6. 06 · ServiceServicing
5.5 daysIntent to first submission
26.6 days avgFirst submission to decision
01 · Fragmentation12+ disjointed systemsManual navigation and repeated rekeying create high cognitive load.
02 · ReworkMissing context drives loopsCredit Officers repeatedly chase RMs for unrecorded background.
03 · Late detectionGaps surface too lateMissing evidence is spotted at sanctioning instead of intake.
Design Process

From preparation friction to a validated North Star

Step 01 — Current-state discovery

Started by understanding the problem, not designing an ‘AI tool’

Method · Current-state journey mapping & research synthesis

Why is preparing a credit application difficult — and which parts genuinely require an RM’s human expertise?
As-Is Journey OverviewSME Commercial New Lending Journey (End-to-End)
Across 30+ days of deal shaping and 26.6 days to decision, RMs acted as “human middleware” between disconnected tools.
Research SynthesisResearch Theme & Problem Clustering Board
Step 02 — Reframing problems into opportunities

Translated research into 8 transformation opportunities

Method · How Might We synthesis mapped to root-cause problems (P1–P8)

How Might We Synthesis Board

Opportunity Theme How Might We Root-Cause Problem (P1–P8)
Theme 01First-time-right applications
Top Priority
1. How might we help RMs submit complete, consistent, decision-ready applications first time — so Credit don’t need repeated follow-ups?
P1 · Gaps & reworkUnclear narratives drive repeated Credit follow-ups.
P5 · Missing contextRM client knowledge isn’t captured systematically.
P6 · Unvalidated dataData progresses without quality or completeness checks.
Theme 02Structured flexibility
2. How might we create a consistent narrative structure while still letting RMs tailor what matters for the deal and sector?
P1 · InconsistencyVarying depth slows Credit interpretation.
P2 · Complex casesMulti-entity & sector-specific deals are error-prone.
Theme 03Complex case guidance
3. How might we guide RMs through complex cases (multi-entity, cross-guarantees, sector variants) with fewer errors and gaps?
P2 · High complexityManual assembly breaks down on multi-entity deals.
P8 · Stale copy-forwardBOLT copies old data, forcing full manual rework.
Theme 04Document variability
4. How might we reduce the cognitive load of reviewing client documents by standardising intake and flagging inconsistencies?
P3 · Variable formatsClient PDFs arrive in inconsistent structures.
P4 · Manual effortReviewing and comparing figures is mentally taxing.
Theme 05Serviceability & analysis
High Impact
5. How might we accelerate serviceability calculations and commentary while keeping outputs transparent for Credit to trust?
P4 · Slow spreadingManual calculations produce inconsistent depth.
P6 · AuditabilityCredit needs verifiable sources behind every figure.
Theme 06RM knowledge capture
Top Priority
6. How might we capture and reuse RM client knowledge in a structured way — so teams get context without RMs rewriting it?
P1 · Repeated chasingCredit chases context never recorded upstream.
P5 · Siloed notesClient context lives in emails and RM memory.
Theme 07Reduce admin burden
7. How might we make uploading, attaching, and managing documents effortless so colleagues focus on quality over admin?
P7 · Upload adminManual document linking creates heavy admin overhead.
Theme 08Safe data reuse
Top Priority
8. How might we enable safe reuse of existing data and documents — so RMs don’t pull forward stale info or restart after errors?BOLT currently pulls outdated info while omitting relevant historical documents.
P8 · Broken copy-forwardLegacy copy functions carry stale data and drop valid docs.
Step 03 — JTBD & Human–AI task allocation

Decomposed the RM’s job across Human, AI, and Automation

Method · Jobs-to-be-Done decomposition & opportunity mapping

Theme MappingAI Theme Opportunities Board
Opportunity TreeOutcome-to-Use-Case Opportunity Tree
Step 04 — Prioritise opportunities

Use-case & strategic-fit prioritisation

Method · Desirability, Value & Feasibility scoring

Criterion 01DesirabilitySolves the highest-effort RM pain points without sacrificing control.
Criterion 02ValueCuts preparation hours and downstream Credit rework loops.
Criterion 03FeasibilityMatches data readiness and foundational ingestion dependencies.
PrioritisationValue vs. Feasibility Matrix
Step 05 — AI value hypothesis

Converted opportunities into a testable AI value hypothesis

Method · Hypothesis-driven design

Core Value Hypothesis

If AI can orchestrate gathering and creation of application content, analyse financial, risk, and customer information, and provide contextual guidance, colleagues can spend less time compiling applications and more time assessing outcomes.

Evidence GatheringWould automated evidence gathering measurably reduce RM effort?
Analytical TrustWould AI-generated financial analysis be trusted by RMs and Credit?
Narrative AdoptionWould RMs adopt AI-drafted narrative content in live submissions?
Guidance ValueDoes conversational guidance help RMs navigate complex cases?
TransparencyWhat sources and rationale must be visible to build confidence?
Human ControlWhich decision points strictly require explicit RM sign-off?
Step 06 — End-to-end future state service model

Mapped an integrated future credit application around 7 states

Method · Future-state journey mapping & service blueprinting

Single Colleague Workflow

One connected flow from case setup to sanctioning
  1. 01Case set upIntake & deal framing
  2. 02Guidance & supportPolicy & doc checklist
  3. 03Data gatheringInternal & client docs
  4. 04Data capturingStructured ingestion
  5. 05Analysis & assessmentAI synthesis + RM review
  6. 06Quality checkFirst-time-right check
  7. 07SanctioningSubmit to Credit
To-Be JTBD & AI Capability Map
AI-Led Capability Human / Automated
Stage
1 · Need
2 · Shape Deal
3 · Prepare App
4 · Decision
5 · Fulfilment
6 · Servicing
As-Is
Ad-hoc capture
Manual prep
Manual spreading & writing
9–10 rework calls
Manual setup
Periodic review
To-Be
Pre-deal prepAuto-surfaces portfolio context
Early intakeFlags missing docs upfront
AI OrchestrationExtracts, spreads & drafts with source links
First-time-rightComplete, verifiable case
Auto-handoverTerms flow to fulfilment
Live alertsCovenant tracking
Capabilities
AIMeeting brief
AIDoc gap check
AIFinancial spreading AINarrative synthesis AUTOPolicy validation
RMRM sign-off
AUTOData sync
AIRisk alerts
Shift
Faster prep
Early clarity
Review vs. compile
Fewer loops
Zero rekeying
Proactive
Core Shift
Moving from fragmented manual assembly in Stage 03 to an orchestrating AI partner that prepares verified evidence for RM judgement.
Step 07 — North Star prototyping

Translated an abstract agentic-AI idea into something RMs could react to

Method · Scenario-based experience prototyping

The North Star prototype acted as a shared boundary object — aligning RMs, Product, Architecture, and AI around one concrete future state.

Boundary ObjectEarly Wireframe Flow & Cross-Functional Review Notes
Step 08 — Qualitative concept testing

Moderated concept testing with real RMs

Method · Moderated testing using future-state prototype

Source TraceabilityEvery AI figure links directly to the source page and document.
Visible AI RationaleShow how conclusions were derived — never a black box.
RM ControlAI suggests and drafts; the RM explicitly approves or edits.
Human OversightClear confidence indicators flag where judgement is needed.
Non-Linear WorkingJump across sections in any order as client info arrives.
Amend the CaseUpdate inputs or narrative cleanly without losing progress.
Ecosystem SupportReuse intelligence seamlessly across CRM and downstream tools.
Design Outcome — 5-Stage AI Interaction Model Human-in-the-loop by design
01 · AIPreparesGathers docs, extracts data, and pre-fills structure.
02 · AIShows EvidenceLinks every figure and claim to its exact source.
03 · AIExplainsSurfaces rationale, gaps, and policy flags clearly.
04 · RMValidates / ChallengesRM interrogates, amends, and refines the case.
05 · RMDecidesRM owns final sign-off and credit submission.
The Solution

Three core interactions from the North Star prototype

01 · AI Launcher · Case Preparation

Proactive preparation inside the AI workspace

Launched alongside the credit application, the AI Launcher plans the case, pre-populates known client data, and surfaces policy recommendations so RMs can review and add context before submitting.

02 · AI Launcher · Document Intelligence

Financial & tenancy extraction with human sign-off

Inside the same AI workspace, raw client documents are extracted and compared against the Real Estate model—highlighting discrepancies for one-click RM approval or resolution.

03 · Narrative Creation

Evidence-backed narrative with inline AI editing

Generates a structured credit narrative complete with clickable source links for auditability and an inline prompt popover to weave in recent client context.

04 · AI Launcher · Multi-Agent Orchestration

Scalable hub-and-spoke AI workspace

Built on progressive disclosure—the AI Launcher keeps the active workflow focused in a collapsed rail and expands on demand to surface traceability tools and specialist sub-agents (e.g., Policy Agent, RM Assistant) that the lending orchestrator can invoke or be orchestrated by via a central RM orchestrator.

Incremental Roadmap

Turning the North Star into pragmatic delivery slices

Now

Document ingestion & extractionStandardise unstructured client PDFs into structured case data.
Source-linked narrative MVPDraft core credit sections with transparent audit links.

Next

Serviceability co-pilotAutomate financial spreading and variance commentary.
Pre-meeting RM briefSurface portfolio context ahead of client contact.

Later

Affordability automationExtend spreading into full repayment assessment.
Cross-sector scalingReuse CRE extraction models across wider commercial lending.
Predicted Impact

Modelled efficiency across the credit journey

-73%

Reduction in time from intent to proceed to first submission

-70%

Reduction in time from intent to last decision

80%

Of the jobs to be done suitable for an agentic AI single workflow


Validation

What users said

It’s more of a confirm, summarise, and oversight task rather than typing… the journey goes from 20 friction points down to three.”
Perfect accuracy and an audit trail… show me exactly where that data was extracted from.”
This removes that 18 months of learning curve for a new RM, because AI presents the context on a plate.”

Lead Designer Contributions

Contribution & reflection

01 · Systems & ArchitectureDesigned across experience & architecturePartnered with the BCB AI Lead and Solution Architect so human-in-the-loop UX and agentic architecture evolved together.
02 · Strategic PhasingNorth Star to incremental roadmapTranslated the long-term vision into Now / Next / Later slices anchored on data readiness.
03 · Reflective PracticeKnowledge transfer & retroTreated handover as part of design — running a “What We Learnt” retro so incoming squads kept full rationale and momentum.
Parallel Workstreams

Three parallel layers transforming commercial lending

Layer 1 Make the customer position visible

Bring facilities, limits and products into one coherent view, reducing system switching and manual rekeying.

5,000hours released annually so far
Layer 2 Automate straightforward decisions

Use information already held and agreed credit rules to automatically decide eligible lending requests.

£500Knew lending opportunity generated
Layer 3 · This Project Apply AI to complex preparation

Support evidence gathering, validation, financial analysis, serviceability and narrative creation for cases that still require human judgement.

>70%RM effort-to-decision opportunity
Not one building block after another: all three run in parallel, in a complex organisation.