AI’s Growing Pains

The 2025 State of AI Report marks something of a watershed moment. After years of racing to see what AI could accomplish, we’re finally asking the tough questions about whether we should trust it to do those things at all.

Just because something is powerful doesn’t make it trustworthy or reliable. Today’s AI systems are remarkable. They write, they code and they converse, but these are all predictions that don’t come with an explanation as to why they did what they did. And if you’re running a financial institution, an insurance company, or an accounting firm, that’s a big compliance problem.

When you can’t explain how a system reached its conclusion, you can’t defend that decision to regulators, customers, or your own board. The good news is we seem to be moving past the era of taking AI’s word for it and into one where we need answers we can trust.

What’s Changed

This year’s report reveals several trends that all point in the same direction: trust has become the make-or-break factor for enterprise AI adoption.

Regulation is no longer playing catch-up. The EU AI Act and US financial guidelines aren’t suggestions anymore. Boards want to know not just does it work? but can we prove it is correct? For the first time, regulators are getting ahead of the curve, not chasing it.

Highly regulated industries need explainability. Banks, insurers, auditors, healthcare providers, the sectors that need AI most urgently, also face the strictest requirements. They have to demonstrate their automated decisions are fair, consistent, and compliant. Building AI that has these attributes has not been easy, but necessary to close the gap from PoC to production. In short, it’s where the real value lies.

Agentic AI has emerged as the go-to architecture, adopting the principle of breaking down complex problems to be tackled by smaller agents, all with their own specialist function. This has introduced new risks. 

Everyone’s excited about the potential of AI agents that are provided with the “agency” to take action, but rightly remain concerned about the implications. The probabilistic nature of the Large Language Models (LLMs) that power such agents means that the institutional knowledge which is prompting them is not a precise programming instruction. Common approaches like prompting, RAG and Graph-RAG lack engineering precision. So although an agentic approach looks to simulate a logical process, it isn’t. Each agent still suffers the innate limitations, lacking precision, determinism and auditability.

Perhaps the biggest shift is conceptual. 

A knowledge-first approach beats a data-first approach. 

Instead of training models on historical data, which incidentally typically results from a documented human process, we can take knowledge sources and use them to build “world models” that describe the underlying principles of decision-work. The key to quality decisioning is being able to scale institutional knowledge, and make that a “first-class citizen” in AI systems, leverageable with precision. 

The future belongs to AI that can logically reason over the world, not just make predictions based on publicly trained data.

The Fundamental Problem

Current Gen AI systems share one critical flaw: they don’t know when they’re wrong. They generate answers that sound right because they’re statistically probable, not because they’re logically calculated.

For creative work, that’s fine. LLMs are ideal for creating marketing content for example, but not for ensuring that marketing content meets compliance obligations. For high-stakes decisions: loan approvals, insurance claims, tax filings, medical eligibility, it’s unacceptable. In high-stakes applications you need precision, consistency and an audit trail that describes how that decision was reached. 

We founded Rainbird on a simple principle: if a system can’t explain its reasoning, it can’t be trusted where it matters. 

You need systems that reason over what’s important to you, your institutional knowledge, without being knocked off course by publicly trained data. You need to be able to generate the same answer from the same inputs, every single time. Determinism matters! And finally you need to understand the reasoning, not an ad-hoc after-the-event prediction as to what might have happened, but the logic that led to an outcome. This is the difference between believing its the right answer and being able to prove it’s the right answer.

A Different Approach

Our approach combines three elements. 

First, the modelling of knowledge as graph-based world models that represent the rules, regulations, and expertise required for a specific decision domain. 

Second, a powerful symbolic reasoning engine that can process knowledge with the same mathematical precision that Excel processes numbers. A deterministic engine that produces consistent, auditable results with a clear trail showing how each conclusion was reached.

Third, LLMs but only where they are strong: understanding natural language and extracting knowledge, but not as a proxy for reasoning where they are inherently weak.

This gives enterprises what they actually need: Gen AI benefits but with none of the risks. Regulated organisations can deploy it readily in decision-intensive processes that are knowledge-dense and there are commercial or regulatory consequences of error. 

Why Trust Matters for Business

Trust isn’t just about doing the right thing, it’s an economic necessity. Our research shows the next trillion dollars in AI value will come from areas where precision, consistency, and auditability aren’t optional: financial crime prevention, tax and audit automation, insurance underwriting, claims, etc.

In these fields, companies don’t just want faster poor decisions, they want better quality decisions that are fully auditable and therefore justifiable. As the inevitable adoption of AI expands, trust becomes more critical than ever, and as much of a competitive advantage as any feature or price.

That’s why the most regulated institutions have started asking “how certain are you that your model is right?” That question will define the next decade of AI adoption. It’s the question we built Rainbird to answer. We didn’t pivot to trust when it became trendy, we started there.

Looking Forward

Ben Taylor and I founded Rainbird in 2013 on what seemed like a contrarian bet: that AI’s future would depend more on being able to make judgements, not just predictions. Twelve years later, the rest of the world has arrived at the same conclusion.

2025 will be remembered as the year AI matured somewhat, when the industry accepted that with power comes responsibility, and that trust in AI is an unquestionable necessity. 

The choice for enterprises and regulators is straightforward: adopt AI that is trustworthy by design, or risk being audited and fined for AI you can’t explain. 

Our goal has always been to help our customers to scale their organisational knowledge to machine levels to deliver consistent, auditable intelligence-led products and services that customers can rely on. The future won’t belong to whoever generates the most content, but to those who can architect AI to deliver trusted solutions. AI that isn’t just powerful, but provable.