Crafting Clarity in a
World of Complexity
Learn more about
Camdor Global Advisors
Camdor Global Advisors evaluates accuracy, depth and strategic insight
AI adoption is accelerating. Decision makers, researchers and policy professionals increasingly use AI-generated output to drive strategic assessments, policy advice and investment discussions. The question is no longer what AI can do, but how much we can trust it. AI adopters should be cautious and consider the risks regarding accuracy, credibility and reliability. While we acknowledge AI significantly accelerates research, it has not removed the need for human oversight.
At a high level, we note that AI excels at speed, structure and coverage, but falls short on policy alignment and judgement. Additionally, a few of the models can reach near-human reasoning through multi-step prompting, but continue to produce hallucinated references and generic outputs. The UK’s competition regulator and other authorities flag hallucination/unreliability as known risks – demonstrating that AI still fails to address regulatory concerns.
Camdor Global Advisor’s (CGA) Test :
We designed a test to evaluate AI-generated outputs against human analysts’ prompt responses using three criteria: Factual Accuracy, Analytical Depth and Policy & Strategic Insight.
Prompt topics:
We employed distinct prompts to make our experiment robust, while testing the AI models’ ability to extract data accurately and provide analytical insights.
Samples From Our Testing
Prompt 1: Historical local authority capex data: Testing Factual Accuracy
What AI did and responded with:
Several AI models extracted £25.9bn from the Department for Levelling Up, Housing and Communities (DLUHC) 2018-19 outturn report for the FY2018-19 figure and then applied a GDP inflator (a few hallucinated the inflation level to apply) to arrive at a standardised 2022-23 price. The models failed to notice that DLUHC in their 2022-23 report published the real term figure of £29.6bn. The models needlessly introduced complexity and arrived at an incorrect figure, demonstrating inconsistency.
Risk summary: Even with a precise and well-framed prompt, AI models selected an outdated price base and generated inaccurate comparisons. This highlights an issue with data retrieval and inflation-adjustment logic.
VS
What a human analyst would respond with:
An analyst would accurately identify the correct DLUHC dataset, verify the price base, and correctly cite £29.6bn (real, 2022-23 prices) for 2018-19 and £27.5bn for 2022-23. They also clarified that the £25.9bn remains valid only in 2018-19 prices, adding transparency and eliminating ambiguity.
Stakeholders should be aware that hallucinations and other errors are not isolated incidents. A recent example was highlighted by the Guardian where a company used AI to produce a report for a local Australian government containing errors and hallucinated references to non-existent sources.
Prompt 2: Local authorities investing responsibly: Testing Analytical Depth
What AI did and responded with:
The response is partially surface level, remaining very general. It is not fully tailored to local authorities or pension funds, presenting strategies in a mainly institutional context rather than addressing the specific audience. While it does offer strategies and practical steps, it requires significant prompting to produce a response approaching the depth and relevance of a human analysis.
Risk summary: AI struggles to tailor analysis to specific audiences and can ignore key principles.
VS
What a human analyst would respond with:
A human analyst would refer to key policies such as LGPS Frameworks, Task Force on Climate-related Financial Disclosures (TCFD) and the UN’s six PRI principles etc. They would analyse how each can be applied to real-world scenarios associated with local authorities and pension funds. Drawing on experience with clients, a human can tailor the analysis to specific contexts, adding valuable insights that enhance understanding. Human insight provides additional depth and practical relevance, resulting in analytical reasoning that is more comprehensive.
A recent example reiterates our findings. CNET published AI-generated articles on topics such as personal finance, but many of these pieces were later found to contain numerous factual and analytical errors.
Prompt 3: Responsible Investment policy frameworks: Testing Policy & Strategic Insight
What AI did and responded with:
When prompted to explain financing mechanisms for UK local authority climate projects, several AI models, particularly Copilot and Perplexity, confidently stated that “local councils can issue sovereign green bonds to fund energy-efficiency projects” or cited the EU Green Bond Standard as the relevant framework. Both claims are incorrect in the UK context.
Risk summary: The output showed limited awareness of UK’s financing architecture and local borrowing limits, underscoring the risk of over-relying on unverified AI in policy advice.
VS
What a human analyst would respond with:
An analyst would recognise that Councils can finance capital via PWLB and the UK Municipal Bonds Agency, as per the Local Government Act 2003 and CIPFA Prudential Code and not by issuing sovereign debt. Additionally, the analyst would propose i) place-based climate infrastructure debt; ii) green municipal or UK municipal-style bonds; iii) retrofit outcome partnerships based on the client’s requirements. The response would demonstrate nuanced understanding of fiscal limits, policy coherence and implementation pathways – ensuring the recommendations are both legally compliant and investment-ready.
CGA’s Experiment Conclusion
We clearly find that AI excels at speed, structure and breadth, producing solid first-pass insight and analysis. A positive example of this efficiency benefit is indicated in Visual capitalist, allowing professionals to focus more on analysis and decision-making rather than routine work. However, AI models continue to be plagued by hallucinated sources, contextual gaps, factual inaccuracy and other errors. If purely AI generated reports or highly AI dependent opinions were presented to councillors/local government officers, it could potentially lead to incorrect decision-making, non-compliant financial strategies or reputational risks.
Human analysts will produce accurate, nuanced and contextually grounded outputs by applying their experience/expertise in institutional and policy awareness that AI still lacks. These qualities remain indispensable when precision, accountability and policy alignment underpin credibility.
At Camdor Global Advisors (CGA), we recognise these challenges and address them through extensive research and expert analysis. We also acknowledge that AI is improving, so these limitations found in the experiment may gradually be mitigated over time.
Implications for Firms
AI tools are invaluable for drafting, summarising and early-stage research, but their insights must be reviewed, verified and contextualised by human experts. At CGA, we view AI as a strategic augmentation tool, not an autonomous analyst. The combination of human judgment and AI efficiency offers the most credible and compliant path forward for regulated and advisory environments. Many AI solution providers frame automation as a path to cost reduction. Yet what is gained in speed is often lost in accuracy, context and accountability. At CGA, we view such shortcuts as false economies. Precision, policy alignment and informed judgment cannot currently be automated, they require expertise. For investors and public institutions, the path forward lies not in replacing human expertise but in integrating AI responsibly, as a tool for research and insight generation, supported by human validation.
*Note: This experiment was conducted in late 2025. Given the rapid pace of AI development, some findings may not fully reflect the capabilities of newer models. A follow-up experiment will be conducted in the future to reassess results against updated models.