TLDR: A new article advocates for a cautious, incremental approach to integrating AI into business operations, likening AI to an intern rather than a full partner. Authored by Philip Hardy and Chris Baker of Ashurst Risk Advisory, the piece emphasizes starting with narrow, verifiable tasks and gradually expanding AI’s responsibilities only after proven accuracy and reliability. It highlights the risks of unverified AI outputs, citing a government report that included AI-generated non-existent sources. The authors propose a three-stage roadmap—data extraction, pattern spotting, and predictive alerts—each with specific ‘promotion gates’ and continuous monitoring to ensure accountability, auditability, and effective risk management.
The article, authored by Philip Hardy and Chris Baker, Partners in Ashurst Risk Advisory, argues for a measured and incremental integration of Artificial Intelligence (AI) into modern firms, suggesting that AI should be treated as a ‘smart intern’ rather than a ‘full partner.’ This perspective stems from the understanding that while AI excels at high-volume tasks like reading, summarising, and cross-referencing, it still ‘guesses,’ making its outputs unsuitable for critical decisions without rigorous verification.
The authors cite a recent incident reported in the Australian Financial Review in August, where a government-commissioned report contained non-existent sources and an invented legal quote, directly linked to the use of generative AI. This serves as a stark reminder that ‘if you can’t prove accuracy and review, you will bear the consequences.’
To mitigate such risks, the article proposes a ‘roadmap for incremental progress’ with defined ‘promotion gates.’ This approach mirrors how a talented junior employee would be managed, with limited scope, tested outputs, and catalogued misses during a ‘probation period.’ The core principle is to ensure AI is ‘fast where speed helps, cautious where judgment matters and auditable all the way through.’ A crucial distinction is made: unlike human juniors who generally improve, AI systems can ‘backslide’ due to changes in data, upgrades, or context shifts, necessitating ongoing review and revalidation.
The proposed three-stage roadmap for AI integration includes:
Stage 1: Data Extraction (Read, Label, File): Begin with narrow, verifiable tasks such as reading large volumes of material, extracting facts, classifying, tagging, and building registers. The benefits here are speed and consistency, transforming unstructured content into structured data, freeing human employees for judgment-based tasks.
Promotion Gate: Achieve a consistently high accuracy bar (e.g., mid-90s on a held-out test set). Regular monthly rechecks for drift are essential, with auto-publish features disabled if accuracy drops significantly. All extracted values must be traceable to source snippets to ensure auditability.
Stage 2: Pattern Spotting (From Summaries to Signals): Once data extraction is dependable, AI can be advanced to group similar complaints, flag inconsistent language, compare market changes, and identify anomalies in KPIs or control data. This stage focuses on early warning and triage, directing human attention to critical issues.
Promotion Gate: Demonstrate through back-testing that signals correlate with important outcomes (e.g., reduced losses or churn). Systems should be intentionally ‘broken’ by feeding in exclusions and clashing rules to identify weaknesses and add guardrails. Continuous light human review is necessary to track agreement with the system’s calls, especially in complex domains where subtle errors can be missed. Transparency is key for explainability and defensibility.
Stage 3: Predictive Alerts (Signals to Next-Best Actions): At the highest stage, AI can suggest actions, such as identifying risky supplier trends, potential control failures, filing date risks, or client re-pricing opportunities. This aims to shorten reaction times and standardize escalation processes.
Promotion Gate: Maintain a decision log for each alert, detailing inputs, rationale, and human decisions. Implement a ‘noise budget’ to prevent alert fatigue; if more than 10% of alerts are dismissed, sensitivity should be reduced. Quarterly reviews are vital to assess if alerts effectively reduced incidents, costs, or resolution times. The ability to explain and defend actions (or inactions) based on AI flags is paramount, as ‘alerts change behaviour and behaviour carries risk.’
The authors conclude with actionable advice for businesses:
1. Firm up your scope: Clearly define what AI can and cannot do, and assign ownership for decisions when AI errs.
2. Show your work: Ensure all AI-informed outputs include sources, rationale, and reviewer stamps.
3. Run regular drills: Periodically test the riskiest AI-managed workflows and share lessons learned.
4. Grant privileges incrementally: Increase AI’s responsibilities only after 90 days of meeting quality standards, and be prepared to scale back if quality declines.
Also Read:
- Organizations Redefine Governance to Mitigate AI Risks and Drive Prosocial Innovation
- AI’s Transformative Impact: Small Banks Face Strategic Risks in Evolving Financial Landscape
The overarching message is that ‘AI will make you faster, but speed without evidence is tomorrow’s problem; and sometimes tomorrow’s headline.’ By treating AI as a smart intern that earns trust and promotion through rigorous testing and transparent record-keeping, organizations can harness automation’s benefits without jeopardizing their reputation.
