The paradox of AI in 2025: Models are dramatically better, but deployment anxiety remains sky-high.
In ICONIQ’s latest State of AI report—300 AI company executives were surveyed about their biggest deployment challenges. The results reveal a fascinating contradiction that every AI builder needs to understand.
The good news: Hallucinations have objectively improved. GPT-4, Claude 3.5, and the latest models are orders of magnitude more reliable than GPT-3 was just two years ago.
The reality check: 39% of companies still rank hallucinations as their #1 deployment challenge. Not cost (32%). Not security (26%). Not even talent shortage (16%). Hallucinations.
Why This Matters More Than You Think
Here’s what 18 months of AI deployments have taught us: The technical problem is getting solved, but the trust problem is getting worse.
Think about it. When you’re building a search feature, 90% accuracy might be fine—users expect to refine queries. When you’re building an AI assistant that generates customer emails, 90% accuracy means 1 in 10 emails could be embarrassing or worse.
The stakes have risen faster than the reliability.
The Data Tells the Story
ICONIQ’s survey reveals the hierarchy of AI anxiety:
- 39% cite hallucinations as a top-3 challenge
- 38% worry about explainability and trust
- 34% struggle with proving ROI
- 32% stress about compute costs
- 26% concerned about security
Notice the pattern? The top 3 concerns aren’t about infrastructure or economics—they’re about reliability and trustworthiness.
The Training Makes All the Difference
Here’s the thing: For most B2B use cases, hallucinations shouldn’t be a huge issue at this point in 2025—if you train properly.
Our own SaaStr.ai has processed over 40,000 chats, trained on almost 20 million words of our content. With that level of domain-specific training, combined with daily QA monitoring, hallucinations have become relatively rare and generally immaterial.
When they do happen, they’re usually edge cases—someone asking about a company we’ve never covered, or a very recent event outside our training data. Not the kind of wild fabrications that plagued early AI deployments.
The key insight: Most companies worried about hallucinations haven’t invested enough in training specificity. They’re using general-purpose models for specialized tasks and wondering why the outputs are unreliable.
What High-Growth Companies Do Differently
The companies scaling AI successfully aren’t waiting for perfect models. They’re architecting around imperfection:
1. Domain-Specific Training Instead of hoping a general model will work, invest in training on your specific use case and content domain.
2. Human-in-the-Loop by Design 66% of companies use human oversight as their primary AI safety mechanism. Not as a fallback—as the foundation.
3. Confidence Scoring Advanced teams build confidence thresholds into every AI interaction. Low confidence = human review. High confidence = auto-execute.
4. Gradual Rollouts Start with internal tools where hallucinations are annoying, not disastrous. Build confidence before touching customer-facing workflows.
The Vertical Divide
Here’s a key insight from the data: Explainability and trust rank even higher for companies building vertical AI applications. Healthcare AI, legal AI, financial AI—these teams live in a different universe of liability.
If you’re building horizontal tools (coding assistants, content generation), you can often design around hallucinations. If you’re building vertical applications, hallucinations can literally be life-or-death issues.
But even in these high-stakes verticals, properly trained, domain-specific AI can achieve reliability levels that make hallucinations a manageable risk rather than a showstopper.
The Economic Reality
Despite the anxiety, companies are betting bigger on AI than ever:
- High-growth companies plan 37% of engineering focused on AI by 2026
- Internal AI productivity budgets are doubling year-over-year
- The average company uses 2.8 different models to optimize for different use cases
Translation: Teams are scared of hallucinations, but they’re more scared of falling behind.
The Three-Layer Strategy
The best AI teams I’ve talked to use a three-layer approach:
Layer 1: Model Selection & Training Choose models based on reliability for your use case, not just performance. Invest heavily in domain-specific training data. Sometimes GPT-3.5 with extensive fine-tuning beats GPT-4 raw for specific tasks.
Layer 2: System Design Build validation, guardrails, and feedback loops into your architecture. Assume hallucinations will happen and design graceful failure modes.
Layer 3: User Experience Set expectations correctly. Show confidence levels. Make it easy to report issues. Turn your users into your quality assurance team.
The Bottom Line
Hallucinations aren’t the existential threat they were in 2023. But they’re still a practical deployment blocker in 2025—mostly because teams aren’t investing enough in proper training and QA processes.
The companies winning aren’t the ones with perfect AI—they’re the ones with trustworthy AI systems built on solid training foundations. There’s a difference.
If you’re building AI products and not explicitly designing for hallucination management, you’re designing for production incidents. But if you’re still treating hallucinations as an unsolvable problem in 2025, you’re probably not training hard enough.
The meta-lesson: In AI, training specificity + reliability engineering matters more than model engineering. Build accordingly.
