Agentic Era · Core Pillar 2
Audience: Enterprise leaders, solution architects, and transformation owners wiring AI into real business pipelines.
The problem
Delegating high-level strategy to AI results in generic output because business logic remains trapped in unstructured silos-PDFs, slide decks, tribal Slack lore, and “ask Maria.” Models remix language; they cannot execute what was never modularized.
The ChunkMaps solution
Deconstruct enterprise complexity into modular, recursive chunks that can be ingested, audited, and executed seamlessly by automated workflows and internal AI pipelines. Chunk up for strategy, chunk down for procedures, chunk across for variants-then version the map like code.
Step-by-step example: Claims exception policy → AI pipeline
Scenario: An insurer wants an internal agent to recommend whether a claim exception can be auto-approved, escalated, or denied-using the same logic senior adjusters apply, not a vague “be careful and fair” prompt.
-
Chunk up: name the enterprise outcome.
L0 Outcome: Decide exception disposition with audit trail Options: AUTO_APPROVE | ESCALATE | DENY Non-negotiables: regulatory hold rules, fraud signals, authority limits
-
Chunk across: inventory logic silos. Interview + document sweep becomes sibling chunks:
L1 Knowledge domains ├── Policy wording (contracts) ├── Adjuster playbook (SOPs) ├── Authority matrix (who can approve what) ├── Fraud heuristics (risk signals) └── Customer promise rules (CX exceptions)
-
Chunk down recursively until each leaf is machine-testable. Example under Authority matrix:
Authority matrix ├── Auto-approve ceiling by claim type │ ├── Auto ≤ $2,500 (property) │ ├── Auto ≤ $1,000 (liability) │ └── Never auto if prior fraud flag ├── Escalation triggers │ ├── Sensitive claimant class │ └── Media / legal hold └── Dual-control rules └── Approvals > $10k require two roles -
Encode chunks as an auditable schema (pipeline-ready, not prose):
Chunk ID Type Rule / content Owner AUTH-01 threshold property auto ≤ 2500 Claims Ops AUTH-02 hard_block prior_fraud_flag ⇒ no auto SIU CX-04 override VIP retention path ⇒ ESCALATE CX REG-12 hold legal_hold ⇒ DENY auto path Legal -
Wire the AI step to consume chunks, not slides. The agent prompt becomes:
INPUT: claim_record + active_chunk_set[] PROCESS: 1. Evaluate hard_blocks first 2. Apply thresholds 3. Apply overrides 4. Emit disposition + cited Chunk IDs OUTPUT: {disposition, citations[], rationale_chunks[]} -
Govern like software. Change a dollar threshold? Version
AUTH-01, re-run regression cases, publish the new map. Strategy stays human-owned; execution stays modular.
What you get
- Business logic that AI can ingest without hallucinating “policy”
- Audit trails that cite Chunk IDs-not paragraph vibes
- Recursive decomposition that scales from strategy to SOP leaf
- A change-control surface your architects and risk teams can own
Full example prompt - copy and use
Note: You must edit the User Inputs section before running the example prompt. You may paste the EXAMPLES below into the prompt's USER INPUT area, or replace them with your real policy text, jurisdiction, and claim records.
USER INPUT (replace placeholders)
Use these slots in the paste-ready prompt. The EXAMPLES are a ready-made claims-exception policy trial for an AU personal-lines insurer.
## USER INPUT (replace placeholders) <<PASTE raw unstructured policy / SOP natural language>> <<PASTE jurisdiction / line of business>> <<PASTE system of record + audit requirements>> <<OPTIONAL: sample claim_record for a trial disposition>> EXAMPLES: jurisdiction: --- country: "AU" line_of_business: "personal lines property" regulator_notes: "Keep an auditable trail of rule citations for every exception disposition." --- system_of_record: --- systems: ["Guidewire ClaimCenter", "PolicyAdmin"] audit_requirements: - "Every AUTO_APPROVE|ESCALATE|DENY must cite one or more Chunk IDs" - "No disposition without evaluating hard_blocks first" - "Version every rule change; never silently overwrite AUTH-*" --- raw_policy_text: --- Claims Exception Handling (draft, unstructured) Adjusters may auto-approve property claim exceptions up to $2,500 where there is no prior fraud flag on the claimant. Liability exceptions may be auto-approved only up to $1,000 under the same fraud constraint. If the claimant is in a sensitive class, or there is a media or legal hold, do not auto-approve - escalate to a senior adjuster. VIP retention cases should escalate rather than auto-deny when commercially sensitive. Approvals above $10,000 require dual control (two authorized roles). Legal hold means the auto path is denied; route to Legal. Fraud heuristics: prior SIU referral or confirmed fraud indicator blocks all auto approval regardless of amount. --- sample_claim_record: --- claim_id: "CLM-20449" claim_type: "property" exception_amount_aud: 1800 prior_fraud_flag: false sensitive_claimant_class: false media_or_legal_hold: false vip_retention_path: false ---
The Policy-to-Pipeline Prompt
Paste the controller prompt below into an LLM (system message, or first user message). The USER INPUT section inside the prompt already includes the EXAMPLES so you can run a trial immediately - edit them first if you have real policy packs. The agent must compile versioned, citable logic chunks and (optionally) dispose a sample claim - not one flat policy summary.
SYSTEM / CONTROLLER PROMPT - ChunkMaps Policy-to-Pipeline (Claims Exception)
You are a ChunkMaps enterprise compiler. Do NOT summarize policy as prose paragraphs only.
You must deconstruct unstructured policy into modular, recursive, version-controlled logic
chunks with stable citation IDs, then optionally evaluate a sample claim_record against those chunks.
## Outcome (chunk up)
L0 Outcome: Decide exception disposition with audit trail
Options: AUTO_APPROVE | ESCALATE | DENY
Non-negotiables: regulatory hold rules, fraud signals, authority limits
## Method phases
### Phase 1 - DOMAIN INVENTORY (chunk across)
From raw_policy_text, list knowledge domains (siblings), e.g.:
Policy wording | Adjuster playbook | Authority matrix | Fraud heuristics | CX exceptions
### Phase 2 - RECURSIVE DECOMPOSITION (chunk down)
Decompose each domain until leaves are machine-testable
(boolean / threshold / override / hold). Reject fuzzy leaves.
### Phase 3 - ENCODE CHUNK TABLE
Emit rows:
| Chunk ID | Type | Rule / content | Owner | Source anchor | Version |
IDs MUST use stable prefixes (AUTH-, REG-, CX-, FRD-, POL-).
Never reuse an ID with a new meaning - bump version instead.
Suggested seed IDs if policy matches the trial text:
- AUTH-01 threshold property auto <= 2500
- AUTH-02 hard_block prior_fraud_flag => no auto
- AUTH-03 threshold liability auto <= 1000
- AUTH-04 dual_control approvals > 10000 require two roles
- CX-04 override VIP retention path => ESCALATE
- REG-12 hold legal_hold => DENY auto path
- FRD-01 hard_block SIU/fraud indicator => no auto
### Phase 4 - PIPELINE CONTRACT
Publish an evaluation function the runtime can execute:
INPUT: claim_record + active_chunk_set[]
PROCESS:
1. hard_blocks / holds first
2. thresholds
3. overrides
4. emit disposition + citations[] + rationale_chunks[]
OUTPUT: {disposition, citations[], rationale_chunks[], open_questions[]}
### Phase 5 - TRIAL DISPOSITION (if sample_claim_record provided)
Run the pipeline on sample_claim_record.
Return disposition with cited Chunk IDs only - no vibes.
## Hard rules
1. Do not invent rules absent from raw_policy_text; put gaps in open_questions.
2. Every disposition citation must be a real Chunk ID from your table.
3. Prefer typed rules (threshold|hard_block|override|hold|procedure).
4. Keep map_version semver; default "1.0.0" for first compile.
5. Output structured YAML/markdown tables before any narrative.
## Turn contract (emit each phase)
TURN CONTRACT
phase: Inventory|Decompose|Encode|Pipeline|Trial
may_write: [artifacts for this phase]
forbidden: [e.g. trial disposition before encode completes]
## USER INPUT (replace placeholders)
# Edit before running. EXAMPLES included for a trial run.
jurisdiction:
---
country: "AU"
line_of_business: "personal lines property"
regulator_notes: "Keep an auditable trail of rule citations for every exception disposition."
---
system_of_record:
---
systems: ["Guidewire ClaimCenter", "PolicyAdmin"]
audit_requirements:
- "Every AUTO_APPROVE|ESCALATE|DENY must cite one or more Chunk IDs"
- "No disposition without evaluating hard_blocks first"
- "Version every rule change; never silently overwrite AUTH-*"
---
raw_policy_text:
---
Claims Exception Handling (draft, unstructured)
Adjusters may auto-approve property claim exceptions up to $2,500 where there is
no prior fraud flag on the claimant. Liability exceptions may be auto-approved only
up to $1,000 under the same fraud constraint.
If the claimant is in a sensitive class, or there is a media or legal hold, do not
auto-approve - escalate to a senior adjuster. VIP retention cases should escalate
rather than auto-deny when commercially sensitive.
Approvals above $10,000 require dual control (two authorized roles). Legal hold
means the auto path is denied; route to Legal.
Fraud heuristics: prior SIU referral or confirmed fraud indicator blocks all auto
approval regardless of amount.
---
sample_claim_record:
---
claim_id: "CLM-20449"
claim_type: "property"
exception_amount_aud: 1800
prior_fraud_flag: false
sensitive_claimant_class: false
media_or_legal_hold: false
vip_retention_path: false
---
auto_run: true
## Start
Begin Phase 1 - DOMAIN INVENTORY with TURN CONTRACT.
Continue through Encode, publish Pipeline contract, then Trial disposition for CLM-20449.