Sample deliverable. Illustrative client, fictional names.

Mortgage Operations: from event log to live agent

A walkthrough of what a full Pi Mapper engagement delivers, based on a composite of specialist-lender work. Scenario: a UK specialist mortgage lender with a 140-person underwriting and completions operation, a core origination platform, a CRM, a document management system and a great deal of email.

63distinct process variants surfaced from the event log

5systems mined and reconciled

41%of case handling time inside the first agent's scope

0customer-facing decisions taken by the agent without a human gate

Chapter 1 of 4

Discovery: the process as it actually runs

Event data from five systems, eleven people shadowed, 18 workflows reconstructed. The documented process had seven steps. The real one had 63 distinct variants.

Engagement timeline

Every shadowing session names the operator and the systems they touched during it.

Stage 1: extract and shadow

Kickoff and scope with Head of Operations and Head of Underwriting

Origination platformCRM

Event log extraction begins; case IDs reconciled across systems

Origination platformDoc managementCRMEmailShared drives

Shadow: application intake and document checking (R. Okonkwo, Case Handler)

Origination platformDoc managementEmail

Shadow: income verification and affordability (J. Whitfield, Senior Underwriter)

Origination platformExcelOpen banking portal

Shadow: referrals and exceptions desk (P. Nair, Underwriting Lead)

EmailExcelTeams

Stage 2: reconstruct

Shadow: valuation ordering and chasing (S. Reilly, Completions Analyst)

Origination platformValuer portalEmail

Shadow: offer production and conditions (A. Desai, Completions Lead)

Origination platformDoc management

Variant analysis: 63 variants across 4,800 cases, rework loops identified

Exception pattern analysis: 1,140 cases resolved partly or wholly outside the platform

Time allocation study across 11 operators, validated against system timestamps

Stage 3: rank and recommend

Agent-suitability scoring for 14 candidate activities

Business case for recommended first agent; playback with sponsors

Discovery pack handed over; decision to proceed to redesign

Where the time goes

Share of case-handling effort by activity, from 4,800 cases over six months. Amber marks work that the documented process says does not exist.

Income and affordability verification

24%

Document checking and re-requests

17%

Chasing valuers, brokers and solicitors by email

15%

Underwriting decision and conditions

14%

Re-keying between platform, CRM and spreadsheets

11%

Offer production and checking

Referrals resolved by Teams message or phone call

Everything else

Systems and what we found

System	Role	Finding
Origination platform	System of record	Reliable timestamps Case status accurate; 31% of status changes lag the real event by more than a day.
Document management	Evidence store	Action needed Document types inconsistently tagged; 22% of re-requests were for documents already on file.
CRM	Broker contact	Action needed Broker notes re-keyed by hand from the platform; no integration.
Email and Teams	Exception channel	Action needed 1,140 cases carried decisions that never reached the platform. Effectively a shadow workflow.
Excel (affordability models, chase lists)	Real rules engine	Action needed Four versions of the affordability calculator in use. Two disagree on treatment of variable income.

Recommended first agent

Candidate activity	Effort in scope	Decision complexity	Error cost	Verdict
Document completeness and re-request	17%	Low	Low	First agent
Income verification pre-assessment	24%	Medium	Medium	Second agent
Third-party chasing	15%	Low	Low	Automate with existing workflow tool
Underwriting decision	14%	High	High	Stays with people
Platform-to-CRM re-keying	11%	None	Low	Integration, not an agent

Not every finding needs an agent. Two of the five biggest time sinks were solved with an integration and the workflow tool the lender already owned.

Next: Redesign

Chapter 2 of 4

Redesign: people, agents and existing software

The document completeness process redrawn with every task owned by a person, an agent or an existing system, and every handoff made explicit.

Task ownership, before and after

The agent takes reading, matching and drafting. People keep anything that changes the customer's outcome.

Task

Before

After

Receive application pack from broker

SystemSystem

Identify which documents are present and classify them

PersonAgent

Check each document against the product's evidence requirements

PersonAgent

Detect expired, illegible or mismatched documents

PersonAgent

Draft the re-request to the broker, listing exactly what is missing

PersonAgent

Approve and send the re-request

PersonPerson

Chase broker after 3 and 7 days

PersonSystem

Decide whether a non-standard document is acceptable

PersonPerson

Update case status and CRM

PersonSystem

Mark pack complete and route to underwriting

PersonAgent

Designed handoffs and stops

The agent has four conditions under which it must stop and hand to a person. These are designed, not discovered in production.

Non-standard evidence

A document type not in the product's evidence list, or a document the agent classifies with low confidence.

Stop

Conflicting information

Name, address or date of birth differs across documents in the pack.

Stop

Vulnerability indicators

Any language in the pack or broker notes that suggests a customer may be vulnerable.

Stop

Third re-request

A pack that has already been re-requested twice is escalated to a lead rather than chased again.

Stop

Controls mapped to the risk framework

Control	Design	Owner
Human approval of every outbound communication	Agent drafts; case handler approves in one click; edits are logged and fed back into evaluation	Operations
Sampling	10% of agent-completed packs reviewed by a senior handler at launch, stepping down to 3% as confidence builds	Quality
Audit trail	Every classification and check recorded with the evidence and rule that produced it	Second line
Kill switch	Operations lead can route all new packs to the manual queue instantly, no engineering involved	Operations

What changes for the team

Case handlers stop being document clerks and become pack approvers and exception resolvers. The role gets narrower and more expert. Two handlers moved to the exceptions desk, which had been under-resourced for a year. No headcount reduction was planned or needed; the lender's growth plan absorbed the capacity.

A capability plan covered the two skills the team asked for most: reading agent evidence trails quickly, and giving the feedback that improves the agent rather than just overriding it.

Back: Discovery Next: Context pack

Chapter 3 of 4

Context pack: what the agent needs to know

214 rules traced to 31 sources, 19 contradictions resolved with the people who own them, and 86 precedent cases selected to show the agent how the hard ones are handled.

Where the knowledge lived

Sources inventoried for the document completeness process. The number of contradictions is the point.

Source	Count	Contribution
Product evidence requirement matrices	6	Core rules. Three were out of date against the current product range.
Lending policy and underwriting manual	2	Boundaries and mandatory checks.
Procedure documents and checklists	9	Steps, mostly aspirational. Two contradicted the evidence matrices.
Team spreadsheets and "cheat sheets"	7	The real rules, including how to treat 14 document edge cases nobody had written down elsewhere.
Precedent cases (historical packs)	86 selected	Worked examples of each edge case, labelled by senior handlers.
Interviews with senior handlers and underwriting leads	7 people	Resolution of contradictions; tone and boundary guidance for broker communications.

Contradictions resolved

A sample. Every resolution names the decision owner and the date, so the pack is auditable.

Conflict	Sources	Resolution
How many months of bank statements for self-employed applicants	Matrix says 3; cheat sheet says 6; underwriters used 6	6 months. Matrix corrected. Owner: Head of Underwriting.
Whether a screenshot of a payslip is acceptable	Procedure silent; practice varied by handler	Not acceptable; agent requests original. Owner: Head of Underwriting.
Treatment of documents in a language other than English	No source	Stop condition; person decides. Owner: Operations Lead.
Tone of third re-request to broker	Two templates in circulation	Single template, escalated by a person. Owner: Broker Relations.

Structure of the pack

Versioned and loaded directly into the lender's agent framework. Each file has an owner and a review cadence.

Component	Contents	Owner
Rules	214 evidence and validation rules, each traced to a source and a decision owner	Underwriting policy
Resolution priors	What to do when rules conflict or evidence is ambiguous; the order of precedence	Head of Underwriting
Exception playbooks	14 edge cases with worked examples and the required action	Operations Lead
Boundaries	What the agent must never do or say; vulnerability and complaint triggers	Second line risk
Communication guidance	Tone, structure and mandatory wording for broker re-requests	Broker Relations
Precedent set	86 labelled historical packs used for grounding and for evaluation	Quality

Back: Redesign Next: Go-live assurance

Chapter 4 of 4

Go-live assurance: the evidence behind the decision

An evaluation set of 620 real historical packs, six gates agreed with risk and compliance, two failed on the first run, both fixed at the context layer rather than the model.

Gates and results

Thresholds were set by the process owner and second line before the first run, not after.

Document classification accuracy

Threshold 98%. Measured against 620 expert-labelled packs.

99.1%

Completeness decision agreement with senior handler

Threshold 96%.

97.4%

Stop-condition recall

Threshold 100% on vulnerability and conflict cases. No misses permitted.

100%

Re-request precision

Threshold 95%: the agent must not ask for documents already on file. Run 1: 91%. Fixed by adding document-type aliases to the rules.

97.8%

Conformance to redesigned process

Threshold 99%. Run 1: 96%, agent skipped the CRM update on one pathway. Fixed in the resolution priors.

99.6%

Communication quality

Broker re-requests rated by Broker Relations against the agreed template.

Pass

Controlled release

Shadow mode

Agent runs on every new pack; people work as before; outputs compared daily. 1,020 packs.

Assisted mode

Agent output presented to handlers for approval. Override rate tracked; every override reviewed and fed back into the pack.

Live with sampling

Agent completes packs; sampling starts at 10% and steps down to 3% as confidence builds. Monitoring dashboard owned by Operations, evidence trail owned by second line.

What production monitoring watches

3.1%override rate in live running (target below 5%)

11%of packs stopped and handed to a person

2.4dmedian time to complete pack, down from 6.8

0complaints attributable to the agent

Drift alerts fire on classification confidence, stop rate and override rate. A shift in any of them triggers a review of the pack before it triggers a review of the model.

What the sponsor signed

A four-page go-live pack: the six gates with results and evidence, the release plan, the monitoring commitments, the named owners of the context pack, and the kill-switch procedure. The CRO's question was "how will you know when this stops being true?" The answer was on page three.

The second agent, income verification pre-assessment, started discovery next, reusing 40% of the context pack.

Back: Context pack Talk to us about your process