AI Agents and Job Displacement in Quant Research and Risk Management

 

Prof. Hernan Huwyler, MBA CPA CAIO AI GRC Director · Academic Director, IE Law School

Published  · Last reviewed  

In short
  • Agents will absorb routine quant work first: tuning, backtest setup, stress configuration, data preparation. Judgment, objectives and accountability stay human.
  • Cheap testing multiplies false discoveries; finance lacks a proof checker, so firms need a trial ledger and locked test bench.
  • Junior build work shrinks, thinning future validators; train them with replication drills, and require explanations before approving any agent-built model.

On September 8, 2026, OpenAI said a group of roughly 10,000 concurrent agents had produced a solution to the Navier-Stokes problem in 88 hours. On November 20, a mathematician at Anthropic released a Lean library of about 17 million lines that the Lean and Mathlib communities certified as a correct proof of the Riemann Hypothesis. Both results came with a machine checker attached. In both cases, the human experts said that understanding the output would take far longer than producing it. Quant research and risk teams should treat these two events as a preview of their own next two years.

The direct answer is this. Agents will absorb the routine layer of quantitative work first: parameter tuning, backtest setup, stress scenario configuration, data preparation and first-draft documentation. That drives the cost of producing a research result toward zero and leaves the cost of verifying it where it was. Mathematics has a proof kernel that rejects bad arguments. Finance has nothing equivalent for a claimed edge, because a backtest that runs says nothing about whether the edge exists. So the control that matters most is a trial ledger that counts every variant an agent tried, paired with an evaluation bench the agent cannot edit.

This article is for heads of quant research, chief risk officers, model validation leads and internal auditors who will be asked to sign off on work that agents did most of. After reading it, you can write a trial-ledger policy, separate the generator from the judge, find the gap in current model risk guidance, and decide how to train validators when the junior work shrinks. Where I give numbers for a desk, they come from an illustrative scenario with stated assumptions, and I label it as such. Where I give facts about the world, they come from sources listed at the end, and I flag the ones that come from the companies selling the models.


 

Use Cases for Compliance Automation

 By Prof. Hernan Huwyler, MBA CPA CAIO AI GRC Director · Academic Director, IE Law School

Published 19  · Last reviewed 19  

In short
  • Compliance automation works only in order: clean, owned data first, then rules as code, continuous evidence, and governed AI agents.
  • Seven use cases show the current practices, from regulatory change and DORA registers to GDPR requests and pharma deviations.
  • Govern the automation itself: keep humans accountable for risk decisions, log everything, and pilot one process.

Every regulatory change, every control test, and every audit request costs a compliance team manual effort that grows faster than the team does. Analysts read new texts and map them to policies by hand. Control owners take screenshots that are stale the moment they are saved. Auditors sample a few dozen items and everyone hopes the rest looked the same. The business problem is not a shortage of dashboards. Assurance still depends on people moving information between systems, and people do that unevenly and late.

The practical answer today is to automate in four layers and to build them in order. First, fix the data that automation will read: a clean taxonomy of policies, risks, and controls, each with one named human owner. Second, express rules as code so a bad action is blocked before it happens instead of found at the next test. Third, collect evidence continuously into packages an auditor can re-perform. Fourth, add AI agents that draft and route work, with a named person approving anything that carries risk. Skipping a layer produces automated answers that are fast and wrong.

This article is for compliance officers, internal auditors, risk owners, and GRC platform owners who must show results this year. It starts with the problem and the value at stake, reviews the first-generation approaches most programs already run and where they stop, and then walks through seven use cases with their technical implications and implementation advice. It closes with how to govern the automation itself and a starting plan. Each section stands alone, so you can jump to the use case that matches your current backlog.


Roadmap for Compliance in IT Services

 By Prof. Hernan Huwyler, MBA CPA CAIO AI GRC Director · Academic Director, IE Law School

Published 19  · Last reviewed 19  

In short
  • Harmonize DORA, ISO 27001, ISO 42001 and NIST CSF 2.0 into one control catalog, scored for team workload.
  • Contract first, assess second, monitor third: bind vendors and their sub-suppliers before onboarding, using federated data licenses.
  • Test under realistic attack, and encode DORA’s four-hour reporting clock in a decision tree before an incident hits.

Paperwork never stopped a ransomware payload. An annual external audit proves you documented your policies on one specific Tuesday. It says nothing about whether you survive an automated exploit against a production database or a corrupted machine learning pipeline. You can satisfy every checkbox on the auditor’s clipboard and still lose days of operations. Compliance without engineering resilience is an expensive illusion.

Business managers now face two expanding fronts at once: new regulatory mandates and new algorithmic attack surfaces. IT risk management has moved from record-keeping to operational survival. The DORA EU Digital Operational Resilience Act, the EU Artificial Intelligence Act, the NIST Cybersecurity Framework 2.0, ISO 27001 and ISO 42001 form one interlocking set of legal and technical expectations. Together they require financial institutions, technology providers and enterprise operators to prove they can withstand, respond to, recover from and adapt to severe disruption. Restoring backups three weeks after a breach, while core transaction systems sit offline, fails every modern regulatory test. Resilience means staying stable while under active attack.

Fragmented governance fails under that pressure. Risk teams track obligations in separate spreadsheets, ticket queues and business units. Say a Chief AI Officer ships a generative model into a customer-facing workflow. The data scientists watch model weights and inference latency, the product owner chases feature velocity, and the security engineers watch the perimeter firewall. Nobody owns the end-to-end risk view. Third-party application programming interfaces (APIs), unmonitored training datasets and cloud misconfigurations then sit in the gaps between those teams, waiting to start a cascade.

A single standard leaves gaps of its own. DORA sets strict legal mandates for operational resilience, incident reporting, third-party ICT concentration risk and threat-led penetration testing. ISO 27001 specifies how to establish, implement, maintain and continually improve an Information Security Management System, and ISO 42001 does the same for AI Management Systems. NIST CSF 2.0 links governance, supply chain risk and technical execution across six functions: Govern, Identify, Protect, Detect, Respond and Recover.

Pick one of these and you will miss requirements in the others. Implement all of them separately and your engineers drown in duplicate audit requests and conflicting control language. The way out is a single IT compliance governance that translates overlapping legal citations into one automated set of technical controls. Build it from semantic harmonization at the top down to real-time telemetry and automated incident decision trees at the bottom. You get less audit fatigue and evidence that your resilience works.


Exposure Versus Retention Capacity Guide

  Prof. Hernan Huwyler, MBA CPA CAIO · AI GRC Director · Academic Director, IE Law School

Published 17 · Last reviewed 17

In short
  • Involve GRC professionals before business decisions are approved, so risks and control gaps can be addressed while changes are still possible.
  • Replace subjective control ratings with evidence of effectiveness and quantified exposure against the organization’s capacity to absorb losses.
  • Assign executive ownership, deadlines and formal risk acceptance to unresolved issues, with clear escalation and measurable GRC value.

A completed supplier risk questionnaire is often mistaken for evidence of control effectiveness. Procurement has a contract ready for signature, the supplier has submitted its security responses, and every control in the assessment carries a green rating. The risk team is asked to confirm that exposure is acceptable. Yet the ratings say nothing about whether the controls operate as claimed, whether the supporting evidence is reliable, or what financial loss the company could face if the supplier fails.

This is where GRC advisory should begin. A GRC professional must challenge control claims, test the evidence behind them, and quantify the exposure that remains before the organization commits. The assessment should inform decisions on contracts, investments, third-party onboarding, and asset activation while there is still room to change the terms, require remediation, or reject the proposal. Once the commitment is signed, the organization may have limited options beyond documenting the failure and managing its consequences. 

My answer to that fork is the point of this article. The advisory role of a GRC professional is to test the control attributes that protect a decision before the decision is made, then to compare the exposure that remains with what the organization can afford to retain, transfer, or price. I recommend doing this ahead of commitments, investments, contracts, third-party onboarding, and asset onboarding, because after signature the only tool left is a finding. This guide is written for risk managers, compliance officers, internal auditors, and control owners who want their opinion to change a decision instead of decorating a file.

The sections follow the order of the work. First, why paper compliance fails and how the Three Lines Model keeps advice separate from ownership. Then why control assessment is an evidence test with pass and reject rules, and where in the decision calendar it belongs. After that, why I stopped relying on residual risk and what I compare instead, followed by a worked example with illustrative numbers. The last sections cover self-assessments, deficiency reporting, formal risk acceptance, and what to do when leadership ignores the advice.


 

AI Journal Entries and SOX: AI Considerations for Financial Statement Auditors

  Prof. Hernan Huwyler, MBA CPA CAIO · AI GRC Director · Academic Director, IE Law School

Published 11 · Last reviewed 11

In short
  • Name who or what operates every AI-supported control step in the financial closing, then seed known errors into the review queue to prove a reviewer could actually catch mistakes.
  • Reconcile what enters the model and what leaves it against ERP records, because a polished AI summary is not audit evidence of completeness or accuracy.
  • Confirm vendor SOC reports cover the AI feature, give agents unique identities and posting limits, and approve prompt changes like code releases.

Consider a controller who opens a monthly report containing a significant number of AI-classified invoices, spends eleven minutes scrolling through the list, and signs the sign-off block. On paper, the control is executed and fully documented. In reality, it is incapable of detecting a single misclassification because the report displays no confidence scores, no anomaly flags, and no links to source documents. The moment an auditor asks what specific error that reviewer could have realistically caught, the illusion of control disappears.

This disconnect illustrates the core flaw in traditional human oversight of automated systems. When a manual check lacks the context, data sampling, and diagnostic tools to reveal operational errors, it produces audit documentation without providing actual risk mitigation.

Here is the short answer for SOX control owners and internal auditors. When AI touches financial reporting, treat it as a new operator inside controls you already have. Name the operator of each step and keep evidence that the review could change the result. The AICPA practice aid on AI considerations for financial statement auditors, issued on 13 October 2026, takes the same position: auditing standards are technology neutral, so the old questions still apply. The questions are the familiar ones: what could go wrong, and which control and evidence answer it.

This guide is for people who own or test controls over AI-supported processes. Think of AI-suggested journal entries, invoice extraction, close support, and agents that post into the ERP. The same logic covers tax determination and anomaly detection, though I use fewer examples there. The reader I have in mind is a SOX program manager or an internal auditor planning next year's scope.

The nine moves below follow the order you would work in. Find the AI, redraw the controls, test the human review, prove the data, control changes, restrict the agents, read vendor reports, widen the fraud and disclosure view, and package the evidence. Sources are the AICPA practice aid, a published audit firm guide on AI and automation in financial reporting, the COSO internal control framework, and for SEC issuers, PCAOB AS 2201 and Item 308 of Regulation S-K. Where a point is my recommendation and not a requirement, I say so.


Moves for Internal Auditors for AI Assurance

 By Prof. Hernan Huwyler, MBA CPA CAIO · AI GRC Director · Academic Director, IE Law School

Published 11 · Last reviewed 11

In short
  • Build the AI inventory yourself from procurement, gateway, discovery and repository data, and register each system with an owner and risk tier.
  • Treat vendor certificates as inputs: check scope, exclusions, date and auditor credibility, then test what you control, such as access, logging and change handling.
  • Audit the approved performance envelope instead of single outputs, and run the kill switch and a seeded bad output to prove oversight works.

Walk into an audit committee meeting, and you will almost certainly hear the board ask one uncomfortable question: who is actually verifying the AI systems already running in our business?

Usually, the head of audit hands over a single spreadsheet from the data science team listing a dozen active tools. Deep down, everyone in the room knows the real number is higher, but nobody can pinpoint the actual count. That exact blind spot is where a real AI audit begins.

Auditing artificial intelligence does not mean throwing out classic audit principles. The core foundation remains the same: risk-based scoping, objective benchmarks, clear evidence, and independent review. What changes is the tech itself. Unlike traditional software, an AI system adapts to its environment and can quietly drift out of alignment without anyone making a code update. Auditors simply need to apply their standard discipline with much sharper focus, concentrating on two core questions:

  • Where did the training data come from, and who owns the output when a model makes a mistake?
  • How quickly will the team spot a failing model before it causes operational or financial damage?

This guide walks through a nine-step framework for auditors, risk leaders, and board members preparing to scope or execute an AI assurance review. Running from initial discovery through to final board reporting, this method evaluates the full AI lifecycle across four primary lenses: governance, security, third-party reliance, and model risk.