AI Strategy for Public Sector: Hype vs Reality

Every vendor is selling AI. Every conference is about AI. Every minister wants an AI strategy.

Here's what most won't tell you: Most AI use cases in government don't work yet.

Let's talk about what does.

The Hype Cycle

Vendor pitch: "AI will transform your service delivery."

Reality: AI will help with some specific tasks, if you have the right data, the right skills, and realistic expectations.

The gap between pitch and reality: Expensive disappointment.

Where AI Actually Works in Public Sector

1. Document Classification & Routing

Use case: Automatically categorize incoming applications/cases.

Example: Benefits application arrives → AI classifies type → routes to right team → saves 20 minutes per case.

Requirements:

• ✅ Large volume of similar documents
• ✅ Clear classification rules
• ✅ Humans validate AI decisions

Reality: This works. ROI is clear. Implementation is straightforward.

2. Chatbots for Common Queries

Use case: AI chatbot handles FAQs, escalates complex questions to humans.

Example: "When is my bin collected?" → AI answers → 80% of queries handled.

Requirements:

• ✅ High volume of repetitive questions
• ✅ Well-documented answers
• ✅ Fallback to human for complex cases

Reality: Works if scope is narrow. Fails if trying to handle everything.

3. Fraud Detection

Use case: Flag suspicious patterns in benefits claims, tax returns, or procurement.

Example: Algorithm spots anomalies → human investigates → reduces fraud.

Requirements:

• ✅ Large historical dataset
• ✅ Clear fraud patterns
• ✅ Human oversight mandatory

Reality: Effective but requires careful governance to avoid bias.

Where AI Doesn't Work (Yet)

1. Complex Decision-Making

The pitch: "AI will make benefits decisions."

The reality: AI can't handle discretionary judgment, context, or compassion.

Example: Benefits eligibility isn't just rules. It's understanding circumstances, applying discretion, showing humanity.

Verdict: Don't automate complex public sector decisions. Use AI to assist, not decide.

2. "AI Will Fix Our Process Problems"

The pitch: "Implement AI and efficiency will improve."

The reality: AI automates existing processes. If your process is broken, AI makes it broken faster.

Fix process first. Then consider AI.

3. Generic "AI Strategy"

The pitch: "We need an AI strategy."

The reality: You need a strategy for specific problems where AI might help.

Better question: "Where do we have high-volume, repetitive tasks with clear rules where AI could add value?"

The AI Readiness Test

Before spending money on AI, answer these:

1. Data Quality

• ❓ Do you have clean, structured, labeled data?
• ❓ Is data current and representative?
• ❓ Can you access it easily?

If no: AI won't work. Fix data first.

2. Volume & Repetition

• ❓ High volume of similar tasks?
• ❓ Clear patterns to learn from?
• ❓ Enough data to train models?

If no: AI isn't cost-effective. Humans are faster.

3. Skills & Capability

• ❓ Internal data science capability?
• ❓ Budget for ongoing model maintenance?
• ❓ Understanding of AI limitations?

If no: You'll be dependent on vendors. Build capability first.

The Realistic Approach

Step 1: Identify High-Value Use Cases

Don't start with "Where can we use AI?"

Start with "What problems cost us the most time/money?"

Then ask: "Could AI help with this specific problem?"

Step 2: Pilot Small

• Start with ONE use case
• Small scope
• Clear success metrics
• 3-month pilot
• Measure ROI

Step 3: Learn & Scale

If pilot works:

• Document what worked
• Train internal team
• Scale to similar use cases

If pilot fails:

• Document lessons learned
• Move on (don't throw good money after bad)

Common Mistakes

Mistake 1: Skipping Process Improvement

Wrong: Implement AI on broken process. Right: Fix process, then automate with AI.

Mistake 2: Over-Ambitious Scope

Wrong: "AI will handle all citizen queries." Right: "AI will handle bin collection date queries."

Mistake 3: Ignoring Data Quality

Wrong: "We'll clean data as part of AI project." Right: "We'll fix data, then consider AI."

Mistake 4: Vendor Dependency

Wrong: "Vendor will build and run our AI." Right: "Vendor will help us build AI capability internally."

Data Sovereignty Challenges

Small jurisdictions specific:

Many AI services require cloud processing in US/EU. Data sovereignty rules might prevent this.

Options:

1. On-premise AI (expensive, complex)
2. Jurisdiction-specific cloud (limited AI services)
3. Anonymized data processing (if legally compliant)

Implication: AI strategy must account for regulatory constraints.

The Honest Assessment

AI will: ✅ Help with specific, high-volume, repetitive tasks ✅ Improve efficiency where data is clean and patterns are clear ✅ Work best when assisting humans, not replacing them

AI won't: ❌ Fix broken processes ❌ Work without good data ❌ Make complex discretionary decisions ❌ Be cheaper than you think

Our Approach

We help organizations:

1. Identify realistic AI use cases (not vendor-driven hype)
2. Assess AI readiness (data, skills, process)
3. Pilot small and learn
4. Build internal AI capability (not vendor dependency)
5. Navigate data sovereignty constraints

We won't: Sell you AI for the sake of AI.

We will: Tell you where AI adds genuine value and where it doesn't.

Related:

• Technology Strategy Service
• Fourth Domain: Why Contracts Matter