AI Governance: What It Takes to Scale AI with Confidence

I spend much of my time where technology projects have to be justified: in boardrooms, investment committees and conversations about cost, risk and return.

From that perspective, I’ve noticed something curious about artificial intelligence. Few technologies have been approved so readily for testing, yet so many organisations still struggle to scale them.

A Gartner study published in September 2026 found that only 4% of insurers were at the most advanced stages of AI maturity for scaling AI to increase business value. When I discussed this with Dennis Milan, CIO and CTO of Digital Innovation at Tokio Marine Seguradora, on The IT Leaders’ Brief, he made a distinction I recognise: experimenting with generative AI at home is one thing; using it in a company responsible for millions of policyholders is another.

The gap between pilot and scale is not simply technical. It is also a question of trust. Governance helps build that trust.

Boards approve risks they understand

Governance is often treated as a brake on innovation: another committee, policy or form to get through. From the perspective of people approving investments, I see it differently.

A board does not approve a technology simply because it looks promising. It needs to understand what is at stake: who is accountable, what risks are involved, how much the project may cost and what controls are in place—including whether the operation can be stopped if necessary.

AI projects without clear answers can become costly to leave in limbo. They may be too uncertain to scale, but lack a clear enough case for anyone to stop them. That is how pilots accumulate.

Dennis’s point was that the company remains responsible for the responses it gives its customers, even when a model provider supplies the technology. Accountability cannot be outsourced with the technology. When a company accepts that responsibility and defines how it will manage it, governance becomes part of what allows leaders to approve AI with confidence.

AI can expose weaknesses in existing processes

One example Dennis shared showed how introducing AI can uncover problems that were already present.

While preparing documents to train AI models, his team found that some contracts and procedures were difficult for the AI to interpret. On investigation, they found people were struggling with them too. Rewriting the material improved how the AI handled it and made the documents easier for brokers and customers to understand.

I see the same pattern across organisations. AI can expose ambiguities that teams have learned to work around through experience, informal conversations and knowledge held by a few people. A model cannot rely on those shortcuts; it depends on the information it receives.

That means some of the value of an AI programme can emerge before an agent enters production: clearer processes, more consistent documentation and business rules that are finally written down. These benefits may not appear in the initial business case because the underlying problems were not visible. They should at least be considered as potential outcomes.

Measure AI through business results

The question I hear most often in investment committees is how to measure AI’s return. It is a fair question, but measuring every agent individually can lead to the wrong conclusions.

As organisations introduce more agents, agent-by-agent measurement becomes harder. It can also favour what is easiest to count over what matters to the business.

Instead, assess AI through its effect on processes, customer experience and operational and financial results. In insurance, that might include:

• How long it takes to produce a quotation
• Whether brokers can sell products that previously required specialist knowledge
• How many customer enquiries are resolved without escalation

Dennis described a commercial insurance example in which a broker enters a company’s business activity code and receives a quotation with suitable cover based on historical data. The business measure is not the model’s performance in isolation; it is whether more brokers can sell a product they could not sell before.

The wider market figures point to the same challenge. BCG’s analysis of AI adoption and value in property and casualty insurance reports that P&C insurers expect AI spending as a share of revenue to triple in 2026, while only 38% are generating value at scale from AI in core workflows. More investment does not automatically produce better results. The work is connecting AI to strategic priorities, business processes and measurable outcomes.

As Dennis put it in our conversation, “AI isn’t a technology issue, it’s a business issue.”

Scaling AI needs ownership and cost visibility

Dennis said that Tokio Marine had more than 1,200 agents in operation at the time of the conversation. He described how the company was creating a catalogue and assigning AI business partners to help business units decide where AI made sense—and where a simpler option, such as automation, process change or a manual procedure, would work better.

Every agent can bring value, but it also creates recurring costs, dependencies and risks to manage. Organisations need to know who owns each agent, what it costs to maintain, what results it delivers and when it was last reviewed. Without that visibility, it is difficult to decide what to keep, improve or discontinue.

This discipline also helps address the cost of AI. Adding an AI layer to a digital service that already works well can add expense without adding value. Knowing when not to use AI—and having clear reasons for that decision—is part of a sound AI strategy.

AI governance must include business partners

Governance discussions often focus on what happens inside the organisation. But AI can also affect the partners and intermediaries that connect a business to its customers.

Dennis described Tokio Marine’s network of independent brokers and the different levels of technological maturity across that network. If partners cannot adopt or work with new systems, moving ahead internally may not deliver the intended result.

For organisations that rely on intermediaries, this calls for clear integration rules, shared data standards and a deliberate approach to supporting the distribution channel. Governing AI across an ecosystem is harder than governing it within one organisation, but it is also a business decision—not only a technical one.

Leaders need to manage today and prepare for what comes next

Dennis used the word “ambidexterity” to describe managing day-to-day operations while continuing to make room for innovation. He also emphasised the importance of culture: identifying people who are comfortable with AI and building a network that can share knowledge and support adoption.

I agree. I would add a question I bring to conversations about AI with executives: Who in your organisation is authorised to say yes?

If nobody can answer, AI may remain in pilot mode. If everyone can answer yes, adoption may run ahead of oversight. Effective governance makes clear who decides, what information informs that decision and what limits apply.

The board does not need to master every technical detail of every model. It needs to understand the risks, responsibilities and expected results, and trust the processes that guide decisions about AI.

Building that trust is one of the less visible—but essential—tasks in moving AI from proof of concept to business results.

Explore Getronics’ Artificial Intelligence services to see how governance can support the move from use-case discovery to deployment and ongoing improvement.