visit-memo

Visit Memo

When the user dictates information on-site during a customer visit or requests to record visit notes, execute according to the following process.

Target Role

Role: Corporate relationship manager, risk manager
Use Scenario: Pre-loan due diligence on-site visits, customer field visits, risk investigation visits
Output Purpose: Quickly structure oral observations into records that comply with pre-loan due diligence standards
Decision Level: Basic materials for pre-loan investigation, providing a basis for due diligence reports and credit applications
Execution Frequency: Executed on-site at each visit, supporting multiple rounds of incremental updates

Data Sources

Required Data

Data Item	Source	Acquisition Method	Sensitivity Level
Relationship manager's oral information	On-site conversation	Speech-to-text or manual input	Internal
Customer basic profile	ECIF system	API: /api/customer/profile	Internal
Historical visit records	Credit system image archive	API: /api/notes/list	Internal
Risk warning signals	Risk management system	API: /api/risk/alerts	Confidential

Data Masking Rules

Customer ID number: show the first 3 and last 4 digits, replace the middle with *
Bank account number: show only the last 4 digits
Customer trade secret information (such as key customer lists): use "a certain customer" in records
Sensitive risk signals: mark only in internal records, do not disclose externally

Degradation Strategy

If the ECIF system is unavailable: mark "customer basic profile not verified" and continue recording oral information
If historical visit records are unavailable: mark "first visit, no historical records" and start recording from the beginning
If the risk warning system is unavailable: mark "risk warnings not included" and identify risks only based on on-site observations
If speech-to-text is unavailable: prompt the user to input manually or supplement later

Constraints

Faithful Recording: Record only the user's oral content, do not add inferences or evaluations, and mark facts and judgments separately
Restrained Follow-up: Ask at most 1-2 questions per round, prioritize risk dimensions, and do not interrupt the user's narrative rhythm
Incremental Merging: Merge new information with existing records, do not overwrite or lose it, and clearly mark corrections as "(corrected)"
Risk First: When risk signals are observed, mark them separately with ⚠️ at the end of the record, and do not let them be drowned out in the main text
Flexible Mode: When the user says "do not save yet", only display and do not save; when the user changes topics, respond normally without requiring an exit ritual
Data Traceability:Each record must indicate the information source (oral account/on-site observation/system query) and timestamp
Compliance Boundaries:No credit decision recommendations may be made in on-site records; only objective facts may be recorded
Skipping Steps Prohibited:Step 0 (Data Confirmation) and Step 4 (Risk Signal Scanning) must not be skipped; the complete process must be executed
Mandatory Enforcement of Red Lines:If any red line (R1-R3) is triggered, recording must stop immediately and the user must be prompted; it must not continue

Financial Compliance Red Lines (One-Vote Veto)

In the following situations, recording must stop immediately and the user must be prompted:

R1: The user requests recording of information suspected of financial fraud (compliance risk must be flagged)
R2: The user requests concealment of known risk signals (the obligation to record truthfully must be flagged)
R3: The user requests falsification of on-site observation data (must be refused and the compliance event recorded)

Audit Trail

Each visit record must automatically record the following information:

{
  "memo_id": "VM-20260520-001",
  "customer_id": "CUST-XXXX",
  "visit_date": "2026-05-20",
  "record_time": "2026-05-20 14:30:00",
  "relationship_manager": "张三（工号：10086）",
  "visit_type": "首次拜访/回访/尽调/风险排查",
  "update_count": 1,
  "risk_signals_count": 2,
  "data_sources": ["口述", "现场观察", "ECIF系统"],
  "saved": true,
  "save_time": "2026-05-20 14:45:00"
}

Structured Dimensions

1. Company Overview (on-site visible facts)

Factory/Office Environment: Estimated area, location (industrial park/standalone/shared), tidiness, maintenance status
Employee Scale and Status: Estimated number of people, dress code, morale, work atmosphere
Production and Equipment: Number of production lines, equipment brands and age, operating rate, automation level
Business Activity Level: Visible inventory (raw materials/work-in-progress/finished goods), logistics vehicles, visiting customers/suppliers

2. Financial Clues (oral accounts and on-site corroboration)

Revenue Situation: Annual revenue scale, recent trend (growth/stable/decline), seasonal characteristics
Profit and Costs: Gross margin estimate, main cost components (raw materials/labor/energy)
Fund Situation: Cash flow description, collection cycle, accounts receivable scale, funding gap
Debt Situation: Bank loans (institution/amount), signs of private financing, external guarantees
Bills and Settlement: Whether commercial acceptance bills are used, main settlement banks

3. Upstream and Downstream Relationships

Main Customers: Names/industries of the top three customers, concentration, cooperation stability, payment terms (days)
Main Suppliers: Sources of core raw materials, bargaining power, payment methods (prepayment/monthly settlement/bills)
Industry Position: Market share estimate, competitive landscape, technological or brand barriers
Bargaining Power in the Industry Chain: Whether there is pricing power, upstream-downstream payment term gap (reflecting working capital pressure)

4. Collateral and Assets

Real Estate: Factory/office building area, ownership nature (owned/leased), status of land certificate/property certificate
Equipment Assets: Core equipment brands, number of units/sets, years of use, whether already mortgaged
Other Assets: Inventory, accounts receivable, equity, and other pledgeable assets
Existing Mortgages/Pledges: To which institutions guarantees have been provided, whether there is remaining guarantee capacity

5. Risk Signal Identification

Operational Risks

Signs of work stoppage/production reduction (equipment not running, few employees, overgrown weeds)
Inventory accumulation (excessive finished goods piled up, reduced raw materials)
Signs of asset disposal (equipment being moved, factory premises rented out)

Financial risks

Stated figures clearly inconsistent with on-site scale (overstated revenue/assets)
Signs of multiple-source financing (multiple banks granting credit simultaneously with unclear use of funds)
Signs of private financing/usury (intense repayment pressure, opaque fund flows)

Related-party risks

Multiple companies at the same address (shell company/nominee shareholding risk)
Frequent entry and exit by unidentified persons
Actual controller's introduction inconsistent with business registration information

Legal and compliance risks

Litigation/arbitration notices posted on site
Signs of enforcement/seizure (seals affixed to equipment)
Environmental/safety issues (abnormal emissions, storage of hazardous materials)

6. Impressions of people

Actual controller/legal representative: personality traits (steady/impulsive/slippery), professionalism (knows the business/an outsider), attitude toward the bank, willingness to cooperate
Finance head: familiarity with accounts, cooperativeness, whether they avoid sensitive questions
Other key persons: observations of key figures such as shareholders, professional managers, guarantors, etc.

7. Results of this visit (to be supplemented later)

Whether the preset objectives were achieved
List of key materials obtained (copies of statements/contracts/credentials)
Key commitments and agreements (provide materials next time/arrange special inspection)
Relationship manager's judgment and preliminary opinion

Execution workflow (Workflow)

Interaction mode: Mode D - Dialogue-Assisted

When the user dictates information at a client visit site or asks to record visit notes, execute the following workflow:

Step 0: Data confirmation and initialization (read first, then write)

Confirm client identity: extract client name or ID
Read pre-visit plan (if any): references/visit-memo-templates.md visit objective list in
Read historical visit records (if any): check the time of the last visit and key to-dos
Confirm visit type: first visit/return visit/due diligence/risk screening
Initialize record template: call assets/visit-memo-template.md

📋 Data source:system_api(ECIF system, credit system image archive) 📋 Executing entity:ai 📋 Confirmation mechanism:none ⚠️ Mandatory instruction: do not skip the historical reading step; you must confirm whether there are to-do items that need follow-up

Step 1: First input record (data_source: user_upload, executor: ai)

Extract user dictation information and classify it into 7 structured dimensions (see references/industry-observation-guide.md)
Mark the source of each piece of information: (enterprise dictation)/(on-site observation)/(CM judgment)
Figures must include units (10,000 yuan/100 million yuan/square meters/days)
Output structured records and call the note-saving tool to persist them

Integration prompt: map to your bank's client note-saving interface (such as save_client_note), passing in the client ID and structured content ✅ Record successful → proceed to Step 4 (risk scan) ❌ Save failed → mark as "not saved", continue the conversation, and retry later

Step 2: Subsequent incremental merge (data_source: context, executor: ai)

Read existing records (credit system image archive)
Merge new information into the corresponding dimension, without overwriting or losing anything
Mark new content with "新增:" and corrected content with "已更正:"
Update and save

Integration note: map to your bank's note-reading interface (such as read_note), retrieve existing records by customer ID, then append and update ✅ Merge successful → proceed to Step 4 (risk scan) ⚠️ Historical records unavailable → mark "首次拜访,无历史记录" and start recording from scratch

Step 3: Information correction (data_source: user_upload, executor: ai, confirmation: inform)

When the user says "刚才说错了" or "更正一下", locate the specific field and correct it
Mark "(已更正)" and retain the original value
Do not lose other content

✅ Correction successful → proceed to Step 4 (risk scan) ❌ Field location failed → prompt the user to clarify the correction target

Step 4: Risk signal scan (data_source: context, executor: ai)

After each organization, proactively scan the risk signal dimension (5th dimension, refer to references/risk-signal-catalog.md)
When an anomaly is found, mark it separately with ⚠️ at the end of the record
Do not skip the risk scan just because the information is positive

✅ No risk signals → output "本次拜访未发现明显风险信号" ⚠️ Risk signals found → list them one by one and mark the risk type (operational/financial/related party/legal compliance)

Step 5: Missing information follow-up (data_source: context, executor: human)

Check for gaps in key dimensions
Follow up in the following order of priority:
- First priority: risk signal dimension (most important)
- Second priority: collateral status (core to credit decisions)
- Third priority: missing core indicators in financial clues
Ask at most 1-2 questions per round, using references/visit-memo-templates.md the follow-up phrasing in

⚠️ Mandatory instruction: do not ask broad questions (such as "还有什么要补充的吗"); you must point precisely to the missing dimension

Step 6: Do-not-save mode (data_source: user_upload, executor: ai)

When the user says "先不保存" or "只看不记", only display the structured content in the conversation
Do not call the save tool
When the user says "现在保存", execute Step 1 or Step 2

Step 7: Natural interaction and conclusion (data_source: user_upload, executor: ai)

The user may add information at any time; respond normally when the topic changes
When the user says "记录完了" or "拜访结束":
- Output the complete version
- Remind the user to record the results of this visit and the next-step agreement (7th dimension)
- Generate an audit log

✅ User confirms completion → output the final version + audit log ⚠️ User has not confirmed the results → remind: "请补充本次拜访成果与下一步约定"

Interaction Pattern

Pattern D - Dialogue-Assisted

Dialogue stage

Stage	Trigger condition	Behavior	Exit condition
Start	The user inputs visit information for the first time	Execute Step 0 (data confirmation) and initialize the record template	Template initialization complete
Middle	The user continues to add information	Execute steps 1-5 (record → merge → correct → risk scan → follow-up questions)	The user says "recording finished" or changes the topic
End	The user says "recording finished"/"visit ended"	Execute step 7 (output the complete version + audit log)	The user confirms completion

Context Management

Context accumulation rules:

Each round of dialogue accumulates information into the corresponding structured dimension, without overwriting historical content
When correcting, retain the original value and mark it "(corrected)"
When the context window is full, automatically summarize the content of the first 3 rounds, retaining key figures and risk signals

Context reset conditions:

The user explicitly says "restart recording"
Switch to another customer (customer ID changes)
After the visit ends and is saved, automatically clear the current session context

Multi-round dialogue limit:

A single visit record supports at most 20 rounds of incremental updates
When more than 20 rounds, remind the user: "There are currently many recording rounds; it is recommended to save the current version first"

Output Format

This output can be parsed and used by the credit-due-diligence Skill and submit-credit-application Skill

Use the assets/visit-memo-template.md template. The output must contain the following 7 sections:

Section	Content	Downstream parseable fields
1. Company Overview	Factory site/employees/production/business activity level	`factory_area`(string)、`employee_count`(int)、`production_lines`(int)
2. Financial Clues	Revenue/profit/funds/liabilities/notes	`annual_revenue`(number)、`gross_margin`(percent)、`cashflow_status`(enum)
3. Upstream and Downstream Relationships	Customers/suppliers/industry position/bargaining power	`top3_customer_concentration`(percent)、`payment_terms`(string)
4. Collateral and Assets	Real estate/equipment/other assets/existing mortgages	`real_estate_ownership`(enum)、`existing_mortgage`(array)
5. Impressions of People	Actual controller/finance lead/other key persons	`controller_profile`(string)、`cooperation_level`(enum)
6. Results of This Visit	Goal achievement/document checklist/commitments and agreements	`goals_achieved`(bool)、`materials_obtained`(array)
7. Risk Signals	Operational/financial/related-party/legal and compliance	`risk_signals`(array)、`risk_level`(enum)

Output requirements:

Each piece of information must be labeled with its source: (stated by the company)/(on-site observation)/(CM judgment)
All figures must carry units (ten thousand yuan/one hundred million yuan/square meters/days)
Risk signals must be marked separately at the end with ⚠️ and must not be buried in the main text
At most 2 follow-up questions, precisely targeting missing dimensions

Disclaimer: This record is only basic pre-loan investigation material and does not constitute a basis for approval. All records must be confirmed by the account manager before they can be used for subsequent due diligence report generation.

Gotchas

#1: Confusing oral information with on-site observations

Symptom: Mistaking the account manager's subjective judgment for objective fact
Cause: Failure to strictly distinguish between "the customer said" and "I saw"
Solution: All information must be labeled with its source—(stated by the company), (on-site observation), (CM judgment)

#2: Risk signals are buried in the main text

Symptom: Important risk signals (such as private financing) are mixed into the financial clues paragraph and are not easy to identify
Cause: Not separately marked with ⚠️ at the end of the record
Solution: After each compilation, the risk signal dimensions must be scanned and listed separately at the end

#3: Overwriting historical content during incremental updates

Symptom: New information replaces old information, resulting in data loss
Cause: The "incremental merge" rule was not followed
Solution: Any update must retain the original value; new content is marked "New:", and corrected content is marked "Corrected:"

#4: Follow-up questions are too broad

Symptom: Asking "Is there anything else to add?" leaves the user unsure where to start
Cause: The question was not targeted at a specific missing dimension
Solution: Use the follow-up question reference table to precisely target the missing dimension (e.g., "You mentioned a 15% gross margin. What are the main cost components?")

Examples

Example 1: First visit record

User input：

我刚拜访了明泰铝业，记录一下：
厂区大概50亩，在工业园区内，环境整洁。目测员工100人左右，都在正常生产。
有3条生产线，设备比较新，应该是近3年买的。开工率大概80%。
仓库里原材料和成品都不少，物流车来来往往挺忙。
张总说去年营收大概3亿，今年预计增长20%。毛利15%左右。
主要客户是汽车零部件厂商，前3大客户占比50%。
厂房是自有的，已经抵押给工行贷款2000万。
张总人很务实，对行业很懂行，配合度很高。

AI output：

## 拜访记录 — 明泰铝业有限公司
> 记录时间：2026-05-20 14:30 | 拜访人：[待补充] | 拜访类型：首次

### 企业概况
- 厂区面积约50亩，位于工业园区内，环境整洁（现场观察）
- 目测员工约100人，着装规范，工作氛围正常（现场观察）
- 3条生产线，设备较新（近3年购置），开工率约80%（现场观察）
- 库存充足（原材料+成品），物流车辆频繁，经营活跃（现场观察）

### 财务线索
- 年营收约3亿元（2025年），2026年预计增长20%（企业方口述）
- 毛利率约15%（企业方口述）
- 主要成本构成：待补充（企业方口述）

### 上下游关系
- 主要客户：汽车零部件厂商（企业方口述）
- 前3大客户占比50%（企业方口述）
- 主要供应商：待补充

### 担保物与资产
- 厂房自有，已抵押给工行贷款2000万元（企业方口述）
- 设备资产：3条生产线（近3年购置）（现场观察）

### 人物印象
- 实控人张总：性格务实，专业度高（懂行），配合意愿强（CM判断）

### 本次拜访成果
- [待拜访结束后补充]

---
⚠️ 风险信号：本次拜访未发现明显风险信号

❓ 建议追问：
1. 您提到毛利15%，主要成本构成是什么？（原材料/人工/能源占比）
2. 除了工行2000万贷款，还有其他银行授信吗？

Example 2: Risk signal record