ArkoliteArkolite
Our POV · POV 05 · AI risk

Which AI risks should you manage first?

Eugene Yap, Managing Partner4 min read
The short answer

The ones that match the parts of AI you actually use, and the size of your business. Every AI system has three layers: the model, the harness that sets it up for a job, and the infrastructure it runs on. Each part carries its own risk and needs a named owner. Start with the parts closest to your customers and your data, then add the rest as you grow.

831

separate AI risk controls, counted by MIT across 13 frameworks. Nobody needs all of them on day one.

97%

of organisations breached through AI lacked proper AI access controls.

80%

of organisations say their AI agents have taken actions nobody intended.

Sources: MIT AI Risk Initiative, 2025. IBM Cost of a Data Breach, 2025. SailPoint, 2025.

The chatbot is you

In 2022 Air Canada’s website chatbot told a grieving passenger he could claim a bereavement discount after flying. The policy did not exist. When he took the airline to a tribunal, Air Canada argued the chatbot was a separate legal entity responsible for its own actions. The tribunal disagreed and ordered the airline to pay. The sum was small. The lesson is large: whatever your AI says or does, your business owns it.

The risk map

Every part of AI carries its own risk.

Every AI system has three layers. The model is the AI you buy, the hire. The harness is how you set it up for a job. The infrastructure is the data and systems it runs on, the company it joins. Choose your size to see where to start, then select any part.

Eleven parts, eleven risks.

Choose a size to see where to start. These priorities are our judgement, not a standard.

    Accountability · one named owner for all of it
    The modelThe hire

    It gets things wrong, and you are tied to it.

    The harnessHow you set them up

    It does the wrong thing, or the right thing for the wrong person.

    The infrastructureThe company they join

    It exposes what it touches.

    Select a part

    Each part shows what can go wrong, a real example, the control and who owns it.

    Where the damage lands
    Legal and complianceCustomers and reputationMoney and operations
    All eleven risks as a list
    • The hire (model): Wrong answers, locked in. Confident errors and hidden bias. The vendor updates or retires the model under you, or you are tied to one provider when better or cheaper models appear. Control: Test on your own questions before launch and after every model update. Keep switching providers possible by design. Owner: Technology lead, with the business owner for accuracy.
    • The job description (instructions): Hijacked instructions. Hidden text in an email, web page or document tells the AI to ignore its rules. Security specialists call it prompt injection, and OWASP ranks it the top risk for AI applications. Control: Treat outside content as untrusted, keep sensitive actions behind a separate check, and test with deliberate attacks. Owner: Business owner of the AI use, with security.
    • The briefing pack (context): Wrong or leaked context. Out-of-date documents produce wrong answers, and confidential files surface for people who should never see them. Control: Curate the sources, keep them current, and give the AI only the document permissions the person already has. Owner: Business owner, with the data lead.
    • System access (tools): Wrong actions, unverified agents. Agents send, pay, delete or change things they should not, often using a borrowed human login. 80% of organisations say their agents have taken unintended actions, and 23% saw agents tricked into revealing credentials. Control: Four questions for every agent. Who is it? Its own identity, never a borrowed login. Who sent it? The person or customer it acts for. What can it do? Limited, time-bound access, with sign-off above a threshold. What did it do? Every action logged, with an off switch. For outside agents, check them against emerging Know Your Agent standards. Owner: IT or security lead, with the business owner.
    • A notebook (memory): Remembers too much. Personal data kept longer than needed, or details from one customer surfacing for another. Control: Decide what it may remember and for how long, and keep each customer’s memory separate. Owner: Data protection officer.
    • A manager who reviews (checks): Rubber-stamping. People approve AI work without truly checking it. In Wharton’s 2026 study, people who consulted an AI followed its wrong answers about 80% of the time. Control: Human sign-off where the stakes are high, with the time to check properly and credit for catching errors. Owner: Business owner of the AI use.
    • Accurate records (data): Data used beyond consent. Personal or confidential data goes into AI, or is reused in ways nobody agreed to. Control: A one-page rule on what data may go into which AI tools, with consent and purpose checks under the PDPA. Owner: Data protection officer.
    • Connected systems (integration): Broken processes. An AI change breaks a process that worked, or one error spreads across connected systems. Control: Roll out in stages, test end to end, and keep a manual fallback for critical processes. Owner: Operations lead.
    • The right permissions (security and access): Shadow AI and leaks. Staff paste company data into unapproved tools, and attackers target AI apps. 49% of employees use AI tools their employer has not approved. Of organisations breached through AI, 97% lacked proper AI access controls. Control: An approved-tools list, access controls on every AI app, and vendors contracted to keep your data out of their training. Owner: IT or security lead.
    • IT that works (compute and platform): Runaway cost, one provider. Usage-based bills spike, a provider goes down, or the business depends on a single platform. Control: Usage budgets and alerts, and a fallback provider for critical processes. Owner: Technology lead, with finance.
    • Oversight (monitoring): Flying blind. No list of what AI is in use, no audit trail, and quality drifting quietly. Control: An inventory of AI in use, logs of what it did, regular re-testing and a simple incident plan. Owner: COO or risk lead.
    Readiness check

    Eight questions. Two minutes.

    A general gauge, not an audit. Your weakest parts light up on the risk map.

    1

    Do you have a list of every AI tool your people use, including free ones?

    2

    Is there a written rule on what data can go into AI tools?

    3

    Is one named person accountable for how the business uses AI?

    4

    Are AI answers that reach customers tested before launch and after model updates?

    5

    Does each AI agent have its own identity and limited permissions, rather than a person’s login?

    6

    Before buying an AI tool, do you check where the vendor stores and uses your data?

    7

    Is high-stakes AI work reviewed by someone with the time to check it properly?

    8

    If an AI tool fails, can the process carry on manually?

    You cannot manage every AI risk at once. You can manage the right ones first.
    What to do next
    1. 1

      Know what you run.

      List every AI tool and agent in use, including the free ones people found for themselves.

    2. 2

      Name an owner for each layer.

      One person for the model, one for each AI use, one for the infrastructure, and one accountable for all of it.

    3. 3

      Start where you are.

      Cover the parts closest to your customers and your data first, then add the rest as you grow.

    Questions leaders ask
    What are the main risks of using AI in business?

    They sit in three layers: the model (wrong answers, lock-in), the harness that sets it up (hijacked instructions, wrong actions by agents, unchecked outputs) and the infrastructure (data use, security, integration, cost and oversight). Legal, customer and financial damage are the consequences.

    Who should be responsible for AI risk?

    One named executive overall, with an owner per layer: the technology lead for the model, the business owner of each AI use for the harness, and IT, security and the data protection officer for the infrastructure.

    How do you verify an AI agent’s identity?

    Give every agent its own identity, know who it acts for, limit what it can do, and log what it did. Payment networks are developing shared Know Your Agent standards for agents acting on behalf of customers.

    What AI rules apply in Singapore?

    There is no AI-specific law yet. The PDPA governs personal data used in AI, and IMDA’s Model AI Governance Framework for Generative AI and the AI Verify testing toolkit offer voluntary guidance.

    Sources
    1. MIT AI Risk Initiative (2025). Mapping AI Risk Mitigations: evidence scan and draft taxonomy.
    2. NIST (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0).
    3. IMDA and AI Verify Foundation (2024). Model AI Governance Framework for Generative AI.
    4. IBM (2025). Cost of a Data Breach Report.
    5. SailPoint (2025). AI agents: the new attack surface. 353 IT and security professionals.
    6. BlackFog (2026). Shadow AI research. 2,000 employees in the UK and US.
    7. OWASP (2025). Top 10 for Large Language Model Applications.
    8. Moffatt v. Air Canada, 2024 BCCRT 149.
    9. Chicago Sun-Times (October 2026). McDonald’s sued over AI tool that recommends prices.
    10. PYMNTS (September 2026). Visa and Mastercard team with Ant on Know Your Agent framework.
    Put it to work

    Strategy for the AI-first leader.

    Our Advisory partners map AI risk to the parts of your system and put proportionate controls in place, so the business can move fast with confidence.

    Next brief
    POV 01AI strategy3 min

    Which job should AI do in your business?

    Match the role of AI to the stability of the decision. Two questions place any task in one of four roles: Builder, Agent, Researcher or Sounding Board.

    Read the brief