The next AI infrastructure will not only be computational. It will be built on trust.
By Paulo Carvalho
The discussion around Artificial Intelligence has entered a new phase.
Over the past few years, governments and companies have focused much of their attention on models, chips, data centres, productivity and automation. But a new report from the Tony Blair Institute for Global Change places another element at the centre of the debate: the need to build AI assurance ecosystems as essential infrastructure for the safe and confident adoption of the technology.
The paper’s thesis is direct: AI is already becoming a fundamental technology for economic growth, public-service transformation and national competitiveness. But growing concerns around safety, legal liability and governance are slowing investment, procurement and the deployment of AI systems in sensitive areas such as health care, finance, critical infrastructure, national security and government operations.
In other words: AI may be ready to scale, but institutional trust is not.
The bottleneck for AI is not only technical. It is trust.
The most important point in the report is that there is no real choice between innovation and governance.
The narrative that regulating, auditing or certifying AI slows innovation is simplistic. The study argues the opposite: without trust, adoption slows down. Companies delay investment. Governments hesitate to procure. Citizens resist its use. Critical sectors avoid deploying systems they cannot explain, test or monitor.
The report defines AI assurance as the process of testing, evaluating and monitoring AI systems, while communicating evidence that justifies confidence in their reliability. An AI assurance ecosystem brings together standards, accreditation, incentives, supporting institutions, auditing, certification and accountability mechanisms into a coherent structure.
This distinction is central.
It is not enough for a company to say its AI is safe.It is not enough for a supplier to present a slide deck of good practices.It is not enough for a model to “appear to work” in a demonstration.
Trust must be demonstrated through evidence, testing, evaluation, documentation, auditing and continuous monitoring.
AI needs the institutional equivalent of aviation, medicine and food safety.
The report makes a powerful comparison: food, medicines, aircraft, vehicles and consumer products all have assurance infrastructures, tests, inspections, standards, audits and certifications. If we accept this type of infrastructure for high-impact sectors, why should AI — a technology that will touch almost every aspect of daily life — operate without a similar ecosystem?
This comparison raises the level of the conversation.
AI can no longer be treated merely as software.
When an AI system helps decide credit, medical triage, hiring, policing, access to public benefits, industrial operations or cyber security, it stops being just a digital tool. It becomes decision-making infrastructure.
And decision-making infrastructure needs assurance.
The political fact: lack of trust is already costing money.
The paper presents two figures that should concern any government or board of directors.
According to the TBI, 38% of adults in the United Kingdom cite lack of trust in AI as the single greatest barrier to using the technology. In addition, 61% of organisations report having delayed or reduced planned AI investments due to concerns about reliability risks.
These numbers show that trust is no longer just a reputational issue. It has become an economic variable.
When there is no trust, AI remains stuck in pilots.When there is no trust, legal departments slow things down.When there is no trust, buyers hesitate.When there is no trust, regulators react.When there is no trust, citizens reject it.
The productivity promise of AI depends on an infrastructure capable of turning risk into verifiable evidence.
AI assurance could become a new industry.
The report also presents AI assurance as an economic opportunity.
Testing, evaluation, audits, certifications, accreditation, insurance, specialist consulting, monitoring tools, benchmarks, red teaming, model validation and risk assessment could form a new global industry of trust services for AI. Countries that build these ecosystems early will not only be able to adopt AI more safely, but also influence international standards and capture part of this emerging market.
This vision is especially relevant for countries that do not lead the race for foundation models.
Not every country needs to build the next GPT, Claude, Gemini or Llama.
But every country will need to know how to assess, procure, audit, certify, monitor and hold AI systems accountable.
This could become a strategic agenda for middle powers, developing countries, regional blocs and regulated markets.
The role of government: creating demand for trust.
The report argues that governments do not need to perform all assurance services themselves. The public role is to create the conditions for the ecosystem to emerge: establishing clear expectations, creating demand for assurance, investing in infrastructure where the market alone fails, and coordinating institutions so that standards, certifications and audit practices work in an interoperable way.
The TBI’s central recommendations are organised around three areas.
The first is to make AI assurance a national strategic priority. This includes using risk-proportionate regulation and public procurement to require evidence of reliability from suppliers throughout the lifecycle of AI systems.
The second is to establish legal and market foundations, including civil-liability frameworks and insurance markets for AI. The idea is that responsibility should be allocated to those best placed to manage the risk, creating permanent incentives for testing, evaluation and risk management.
The third is to develop technical and human capabilities: assessment tools, testing environments, secure access to models and data, qualified professionals, accreditation, certification and trusted directories of assurance providers.
AI insurance: the market as a risk supervisor.
One of the most provocative points in the report is the case for mandatory insurance for AI systems capable of causing material harm to individuals, organisations or society.
The logic is simple: insurers have a direct financial interest in understanding risk. If they are asked to price AI systems, they will begin to demand tests, evidence, controls, audits and robust practices before offering coverage. This could turn insurance into one of the strongest commercial mechanisms for driving real assurance — not just superficial compliance.
This thesis matters because it moves part of AI governance from the purely regulatory field into the economic field.
Insurance would ask objective questions:
What is the real risk of this system?Has it been tested?Is there monitoring?Has there been independent auditing?Who is responsible if it fails?Which controls reduce the probability of harm?
If AI is going to operate in critical areas, the insurance market may become one of the new frontiers of algorithmic governance.
Brazil should pay attention.
This debate has direct implications for Brazil.
The country is discussing AI regulation, public-sector use of AI, business automation, data governance, LGPD, liability, transparency and social impact. But one structuring question is still missing: what will Brazil’s AI assurance ecosystem look like?
Who will test critical systems?Who will certify models?Who will audit corporate use?Who will train AI assurance professionals?How will public procurement require evidence?How will insurers assess algorithmic risk?How will international standards be incorporated?How will medium-sized companies access reliable assessment providers?
Without this infrastructure, the trend will be to repeat a familiar pattern: many pilots, little production; many promises, little trust; plenty of showcase innovation, little demonstrable governance.
The link with ISO 42001, the EU AI Act and corporate governance.
The TBI report also mentions the fragmentation of international standards, citing frameworks such as the NIST AI RMF, ISO/IEC 42001, CEN-CENELEC and emerging sectoral approaches. The challenge is to prevent this multiplicity from increasing compliance costs and weakening international interoperability.
This point is essential for global companies.
AI governance will not be solved with a single internal document. It will require integration between risk management, technical documentation, auditing, monitoring, international standards, regulatory requirements, supplier controls, information security, data protection and human oversight.
ISO/IEC 42001, the EU AI Act, risk frameworks, sectoral audits and future AI certifications all appear to be converging in one direction: demanding evidence of control throughout the entire AI lifecycle.
My reading.
The report is right to argue that AI assurance should not be seen as a brake, but as infrastructure.
Large-scale adoption will only happen when organisations can answer, with evidence, simple questions:
Is this system reliable?Under what conditions was it tested?Which risks were identified?Who validated it?How is it monitored?What happens if it fails?Who is liable for harm?Is there human review?Are there logs?Is there documentation?Has there been independent auditing?
Without these answers, AI will continue to shine in presentations and stumble in real-world adoption.
Conclusion.
The next stage of Artificial Intelligence will not be defined only by those who build the most powerful models.
It will be defined by those who build the institutions, standards, professions, audits, insurance mechanisms, certifications and trust infrastructures that allow those models to be used safely.
The Tony Blair Institute report sends a clear message: the AI era requires a new layer of public and private infrastructure.
Not only data centres.Not only chips.Not only models.Not only APIs.
But measurable trust.
AI that cannot be tested, assessed, monitored and explained will struggle to scale in critical environments.
AI that can demonstrate safety, reliability, responsibility and value will have a clear path towards broad adoption.
The future of AI, therefore, will not only be a race for intelligence.
It will be a race for trust.
Comments