From 8 Work Lists to a 2-Week AI MVP: A Complete Enterprise Requirement Mining Playbook

Step 1: Business Inventory — Turn Daily Repetitive Work into a Structured List

The process starts at the business unit's desk, not in a meeting room. Interviewers ask four concrete questions:

"If a digital employee could do one thing for you, what would it be?"

"What do you repeat three or more times per day/week?"

"Which task causes errors when you do it?"

"Which task would you like to never do again?"

Each answer is recorded in a High-Frequency Work List with six mandatory fields: Department, Task Name, Frequency, Single Duration, Current Pain Point, Submitter. The article provides a real excerpt from five departments (Sales, HR, Finance, Customer Service) showing eight tasks:

Sales: Create Quotation — 5×/day, 15 min, price errors, format inconsistency

Sales: Enter Customer Info into CRM — 10×/day, 3 min, duplicate entry, missing fields

Sales: Weekly Report Summary — 1×/week, 45 min, collecting member reports time-consuming

HR: Issue Offer Letters — 3×/week, 10 min, manual template filling, missing info

HR: Resume Screening — 10/day, 5 min each, time-consuming, inconsistent standards

Finance: Invoice Verification — 20/day, 3 min each, manual website login, repetitive

Finance: Bank Reconciliation — 1×/month, 2 hours, manual matching error-prone

Customer Service: Answer Common Questions — 50/day, 2 min each, repetitive, consumes capacity

Three common pitfalls are highlighted:

Granularity too coarse — "Marketing needs AI marketing" is not a scorable requirement; "Create Quotation, 5×/day, 15 min, price errors" is.

Missing duration data — without "5×/day × 15 min" there is no basis for prioritization.

Letting IT fill the list — information distortion makes business owners unable to sign off during review.

Goal: produce 10–30 qualified rows; fewer means insufficient coverage, more is fine because the next step narrows the set.

Step 2: Pain Point Scoring — Prioritize by Formula, Not Feeling

All candidate tasks are scored with a fixed formula:

Total Score = Frequency Score × 30% + Duration Score × 30% + Value Score × 40%

Scoring rubrics are fixed so anyone calculates the same result:

Frequency Score : multiple times/day=10, once/day=7, several times/week=4, several times/month=1

Duration Score : >30 min=10, 10–30 min=7, <10 min=4, <5 min=1

Value Score : direct customer/revenue impact=10, internal efficiency=7, reduce errors/rework=4, experience only=1

Example calculations from the article:

Invoice Verification : Freq 10 ×30% + Dur 4 ×30% + Val 7 ×40% = 7.0

Bank Reconciliation : Freq 1 ×30% + Dur 10 ×30% + Val 7 ×40% = 6.1

Answer Common Questions : Freq 10 ×30% + Dur 1 ×30% + Val 10 ×40% = 7.0

Decision thresholds: ≥7.5 = prioritize, 5–7.5 = consider, <5 = defer. Only the top 5–8 tasks proceed, because evaluation capacity is limited.

Step 3: Feasibility Assessment — This Step Explicitly Kills High-Scoring but Infeasible Scenarios

Pain score indicates "worth doing"; feasibility indicates "can be done." A three-dimension, nine-point scale is used:

Feasibility Total = Data Availability (0–3) + Rule Clarity (0–3) + Fault Tolerance (0–3)

Scoring rubric:

Data Availability : 3 = stable, uniform source; 1 = manual cleanup, mostly stable; 0 = no data / chaotic format

Rule Clarity : 3 = clear, unambiguous logic; 1 = few exceptions, need human judgment; 0 = relies heavily on veteran experience

Fault Tolerance : 3 = errors have no severe consequence; 1 = medium risk; 0 = high risk, unacceptable

Decision matrix: ≥7 = MVP ready; 4–6 = resolve blockers first; ≤3 = defer.

The article walks through the "Answer Common Questions" case: Pain 7.0 but Feasibility = Data 1 (scattered docs) + Rules 1 (inconsistent answers) + Tolerance 0 (customer-facing errors) = 2 → Defer .

For 4–6 point "blocked" scenarios, targeted mitigations are provided:

Poor data → spend 1–2 weeks standardizing formats or exporting reports.

Low tolerance → AI alerts + human confirmation, no auto-execution.

Unclear rules → cover 80% standard cases, escalate exceptions.

Heavy tech dependency → simulate with Excel import/export before integrating APIs.

Step 4: MVP Selection — A Small Working Scene Beats a Big Stalled Project

Final priority combines both scores:

Composite Priority = Pain Score × 0.6 + Feasibility Score × 0.4 (normalized to 0–100). Weights can be adjusted (higher pain weight for speed, higher feasibility for compliance) but must be agreed with business stakeholders on the spot.

The chosen scenario is documented in a six-field execution template. Example from the article for "Inspection Report Entry":

Scenario Name : Verb-object + object, e.g., "Inspection Report Entry"

MVP Goal : Read inspection images, auto-extract 8 key fields, generate structured ledger

Success Criteria : Run 1 week continuously, field accuracy ≥95%, matches manual entry

Estimated Effort : 2–3 hours (low-code/agent) or 3–5 person-days (custom dev)

Owners : Tech lead + Business lead, both must be assigned

Deadline : Within 2 weeks, specific date

A manufacturing company followed this flow: 10–30 candidates from five departments → three-dimension scoring → "Inspection Report Entry", "Daily Production Summary", "Invoice Verification" made the top list → first MVP was "Inspection Report Entry" because data was internal, rules fixed, tolerance high. It went live in two weeks and was then replicated to other production lines.

Iron rule: Better a small success than a big failure.

After Launch: Closed Loop and Accountability

MVP launch is not the end. A five-step closed loop runs: Detect Issue → Analyze Root Cause → Generate Recommendation → Human Review/Execute → Verify Outcome . Low risk = auto-execute; medium = suggest + confirm; high = alert only. The loop aims for business value closure, not just technical closure.

Review presentations rely on four metric categories with evidence:

Core business KPIs (e.g., MTTR from hours to minutes, with before/after comparison)

AI capability metrics (accuracy on test set)

Operational quality (availability, false positive rate)

Data source & verification — results without evidence indexes are not admissible .

Narrative format: "Originally … Now …" — objectively describe the old way, let the improvement speak for itself.

Closing Summary

Break fuzzy "AI transformation" into 10–30 real tasks with frequency, duration, pain points.

Use two fixed scoring systems to turn "I think this matters" into "data says do this."

Enforce one iron rule: a two-week runnable small scene trumps a three-month big project with no visible result.

The scarce resource in the AI era is not models, but judgment on what AI should do — and that judgment can be process-ized.