Nearly two years ago, we wrote that 2024 would be the year AI gets practical. We predicted cautious but real adoption in banking, insurance, and healthcare. We warned about governance gaps and hallucination risks. Now it's time for an honest accounting: what did we get right, what did we get wrong, and what have we learned?
The Scorecard
Let's start with brutal honesty about our predictions:
✓ What We Got Right
- Document processing scaled successfully
- Human-in-the-loop patterns proved essential
- Governance became a board-level concern
- RAG emerged as the dominant enterprise pattern
- Prompt engineering became a real discipline
✗ What We Got Wrong
- Adoption was slower than we expected
- Governance proved harder than anticipated
- Cost management became a bigger issue
- Model stability was more problematic
- Skills gaps were more severe
Banking: The Cautious Leader
Banks moved on AI, but more cautiously than the hype suggested. Here's what actually happened:
Where AI Delivered
Fraud detection saw the clearest wins. AI models that analyze transaction patterns, flag anomalies, and adapt to new fraud vectors delivered measurable value. The key was augmentation: AI flags suspicious activity, humans make final decisions.
Document processing scaled as predicted. Loan applications, account opening documents, and compliance paperwork—AI extraction with human verification became standard practice at forward-thinking institutions.
Customer service augmentation worked where expectations were managed. AI assistants that help agents find information and draft responses delivered value. Fully autonomous AI that handles customer interactions directly? Still limited to simple, low-risk scenarios.
Where AI Struggled
Credit decisions remain contentious. Regulatory scrutiny around understandability and bias kept most banks from deploying AI in core lending decisions. Pilot projects abound; production deployments remain rare.
Regulatory compliance proved harder than expected. AI that interprets regulatory requirements and assesses compliance? The stakes are too high for current accuracy levels. Most compliance AI remains in advisory mode, not decision mode.
AI maturity varies significantly by use case — document processing leads, while high-stakes decisions remain cautious
Insurance: The Quiet Transformer
Insurance has been AI's quiet success story. Less flashy than banking, but arguably more progress:
Claims Processing Wins
AI-assisted claims processing delivered real ROI. Document intake, damage assessment from photos, fraud pattern detection, and automated routing—these capabilities moved from pilot to production across multiple carriers.
The pattern that worked: AI handles routine claims end-to-end, flags complex cases for human adjusters, and learns from human decisions to improve over time. Not full automation, but meaningful efficiency gains.
Underwriting Debates Continue
Using AI in underwriting decisions remains contentious. The potential for bias—in health insurance, life insurance, property coverage—keeps regulators watchful and carriers cautious. Pilot projects show promise; production deployment requires governance frameworks that most carriers are still building.
The Insurance Surprise
Claims photo analysis exceeded expectations. AI that assesses vehicle damage, property damage, or medical documentation from images proved more accurate than initially projected—and significantly faster than human review.
Healthcare: The Complex Landscape
Healthcare AI is a tale of two realities: administrative success and clinical caution.
Administrative AI Delivers
Scheduling, billing, medical coding, prior authorization—AI in healthcare administration delivered value without the regulatory complexity of clinical applications. These aren't glamorous use cases, but they're real cost savers.
Clinical AI Proceeds Carefully
AI that influences patient care remains heavily constrained—appropriately so. Diagnostic support, treatment recommendations, and clinical decision support tools face intense scrutiny from both regulators and clinicians.
The successful clinical AI deployments we've seen share a common thread: they augment clinician judgment rather than replace it, they're designed for specific narrow use cases rather than general diagnosis, and they underwent extensive validation before deployment.
The organizations that succeeded treated AI as a tool to augment human expertise, not replace it. Those who tried to automate high-stakes decisions without human oversight mostly failed—or worse, created risks they're still unwinding.
The Governance Gap
If we underestimated one thing, it was governance complexity. Organizations that moved fast on AI deployment often found themselves scrambling to build governance frameworks after the fact. The ones that succeeded invested in governance infrastructure before scaling deployment.
What Good Governance Looks Like
The mature organizations we work with have established:
- AI inventory and classification — They know what AI they have and how it's used
- Risk assessment frameworks — Different oversight for different risk levels
- Monitoring and observability — Visibility into AI behavior in production
- Incident response — Clear processes when AI produces problematic outputs
- Model lifecycle management — Versioning, testing, and controlled rollout
What's Still Missing
Even the leaders are still working on:
- Cross-functional coordination — AI governance that spans IT, risk, compliance, and business units
- Vendor management — Governance for AI embedded in third-party tools
- Regulatory alignment — Frameworks that satisfy multiple, evolving regulatory requirements
Lessons for the Next Year
What did successful organizations do differently?
1. They Started with Governance
Not as an afterthought, but as a precondition for deployment. The organizations now scaling AI confidently are those who built governance frameworks first.
2. They Focused on Augmentation
The wins came from AI that makes humans more effective, not AI that replaces humans. This isn't just about managing risk—it's about what actually works.
3. They Measured Ruthlessly
Clear success metrics defined before deployment. Willingness to kill projects that didn't deliver. Continuous monitoring of production performance.
4. They Built Skills Internally
The organizations most successful with AI invested heavily in upskilling existing teams, not just hiring AI specialists. Institutional knowledge matters.
5. They Managed Expectations
Realistic timelines. Honest communication about limitations. No promises of AI magic that couldn't be delivered.
Key Takeaways
- AI adoption in regulated industries is real but slower and more cautious than hype suggested
- Document processing and fraud detection delivered clear ROI
- Governance remains the biggest gap—most organizations underinvested
- The winners focused on augmentation, not replacement
- 2026 will be the year governance matures from afterthought to foundation
Looking Ahead to 2026
What comes next? Based on what we've seen:
Governance becomes table stakes. The organizations still treating AI governance as optional will face increasing pressure—from regulators, auditors, and boards.
Agentic AI enters regulated industries. Carefully, with extensive guardrails, but it's coming. The automation gains from AI that can pursue multi-step objectives are too significant to ignore.
Consolidation accelerates. Fewer, more integrated AI platforms rather than collections of point solutions. The integration tax is too high.
Skills gaps narrow. As AI tools become more accessible and training programs mature, the shortage of AI-capable talent will ease—though not disappear.
The next year will separate organizations that built solid foundations from those still experimenting. The foundation is governance, skills, and realistic expectations. The organizations that have these will accelerate; those without will struggle.
