Choosing PoCs by Department
"We want to start using AI at the company. Which department for the first PoC?" The first choice decides AI-adoption success. Here's how to find work where effect shows easily.
4 Conditions for an Effective PoC
1. High Repetitiveness
Work doing the same thing daily/weekly. Effect multiplies as templatization advances.
2. Digitally Completable
Paper/phone-centric work is hard to AI-ify. Target work completable in email/documents/internal systems.
3. Small Failure Impact
Be cautious with direct customer contact/contracts/accounting close. Start with internal-facing work.
4. Easy Effect Measurement
Has quantifiable metrics like "time saved," "count increase," "accuracy improvement."
PoC Candidates by Department
Customer Support (★★★★★)
- Auto-generate FAQ answers (CSV into internal GPT)
- Customer-email sentiment analysis/priority classification
- Past-case search (RAG)
- Effect: shorter response time, more handled cases, higher satisfaction
Sales (★★★★☆)
- Auto-create minutes from deal notes
- Proposal draft generation
- Customer analysis/prioritization
- Effect: less pre-visit prep time, better proposal quality
Marketing (★★★★☆)
- Mass-produce articles/social posts
- Creative A/B test ideas
- Market/competitor research
- Effect: more content, doubled measure cycle speed
HR (★★★☆☆)
- Recruiting: generate job posts/scout messages
- Organize post-interview feedback
- Generate internal-training content
- Caution: don't use for evaluation (fairness concern)
Accounting/Finance (★★★☆☆)
- Journal automation (existing rule-based + AI)
- Monthly report generation
- Anomaly detection (fraud/misentry)
- Caution: pre-audit final processing needs human confirmation
Legal (★★★☆☆)
- Contract issue extraction
- Past-contract search (RAG)
- Regulation research
- Caution: final judgment is a lawyer/legal staff
IT/Development (★★★★★)
- Coding support (Copilot, Cursor, Claude Code)
- Operations automation
- Internal chatbot
- Effect: doubled dev speed, less routine ops
PoC-Selection Flowchart
- List recurring work: for 1 week, 5 "daily tasks" per department
- Measure time consumption: average time per task × frequency
- Score AI-ifiability: rate each task /5 on the 4 conditions above
- Pick the top 3 by total score
- A "PoC hypothesis" for each: state what success means
PoC Scale
- Period: 2-4 weeks
- Participants: 5-10 people
- Budget: ~USD 1,000-5,000 (tool subscription + labor)
- KPI: 30%+ time saving, or 20%+ count increase
PoCs to Avoid
- "Company-wide at once": big failure impact, pilot first
- Too confidential: don't handle personal/financial data first
- Qualitative work: hard to measure effect
- Just "make a PowerPoint": results hard to see
Common Points of Success Cases
- Have a KPI you can be confident is measurable before choosing the work
- Field people (task owners) participate
- Make a "usable-level" prototype in 2 weeks
- Don't fear failure; run multiple PoCs in parallel
Next Step
Once a PoC succeeds, next is the "Pilot → Company-Wide Rollout Roadmap."
