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.


What Business Problem Does Compliance Automation Solve?

Compliance automation is the use of software, rules, and AI to carry out compliance tasks and to produce evidence that they were carried out, with little manual effort. The problem it solves is the gap between the number of obligations and the number of people who can read, map, test, and prove them. Obligations arrive from regulators and from internal policy owners at the same time, and they keep changing after publication. The value of automation is easiest to defend when it is tied to four numbers the business already cares about. Time to decision on a regulatory change shows how fast the organization adapts. The share of controls tested on every event, rather than by sample, shows how much risk is watched. Hours spent answering one audit request show the cost of proof. Cost per control test shows whether the budget goes to judgment or to collection. Pick two of these before you buy anything, record the baseline, and report the same two after each release. A program that cannot name its baseline cannot show that it saved anything.

Start with where your own analysts spend their hours. Ask each of them to log one month of work split between collecting and organizing information and judging what it means. If the first category dominates, that split decides where automation pays first. Collection tasks such as pulling configuration data, matching a regulation to a policy, or chasing an owner for a screenshot are rule-based and repeatable. Judgment tasks such as accepting a risk or deciding whether a pattern is misconduct are not. The aim is to move analyst hours from the first category to the second, and the four numbers above show whether that happened.

Outside pressure is moving the same way. The US Department of Justice revised its guidance for evaluating corporate compliance programs in September 2024, and prosecutors now ask how a company manages the risks of AI and other new technologies, and whether the compliance function uses data analytics to measure how well its own program works, as Orrick's summary of the guidance describes. In EU financial services, the Digital Operational Resilience Act has applied since 17 January 2025 and requires a register of every contractual arrangement with providers of ICT services, as the European Banking Authority explains. Neither instrument tells you to buy a tool. Both ask for evidence that you know your risks and suppliers and that you test your own program with data. Automation is a sensible way to produce that evidence at the frequency they imply.

Which First-Generation Approaches Still Earn Their Keep?

Programs that started automating compliance before 2025 rely on five approaches. Each solved a real problem, and each still has a place. Each also stops at a point where a human has to take over, and that point is where 2026 practice has moved. The paragraphs below show what to keep, what to replace, and what the change means for your data and your people. A comparison table closes the section so you can map your own program against it.

The first approach is to ingest regulatory content from licensed intelligence providers or directly from regulators, triage it by keyword, and map it to an internal library of policies, risks, and controls. The technical requirement is a stable internal taxonomy and a mapping table that links each external obligation to an internal owner. This approach cuts noise and is worth keeping as the intake layer. It stops at reading: a person still decides whether the change applies, what it affects, and what to rewrite. A feed with weak mapping delivers the same overload, only faster. Judge it by the share of alerts that reach the right owner without manual forwarding.

The second approach applies machine learning to the control library to find orphan controls that map to no risk or regulation, duplicates raised by different teams, controls tested too often or too rarely, and controls whose evidence nobody retains. It needs a clean mapping dataset, and its output is a list of candidates, not a decision. Used once a year as a hygiene step, it saves testing effort and exposes gaps. Used blindly, it removes coverage that nobody documented. In 2026 the more useful question is which detective controls can become preventive, a point developed in the compliance as code section below. Keep the analysis, and add a human review of every control proposed for retirement.

The third approach replaces static policy portals with intent-based search and chatbots that answer employee questions. The strength is speed: an employee types travel and sees the gift and hospitality rules together. The weakness is a common belief that the assistant improves by learning from what users search. In compliance that is a risk, because an assistant that adapts on its own can drift and invent a limit that no policy contains. A safer design retrieves from an approved library and answers only from the retrieved text, as described in the policy answers section below. Keep the search, and remove any live learning from user behavior.

The fourth approach uses AI to read supplier compliance reports and redline anomalies before an analyst reviews them. It speeds up reading, and it still reads a document that describes a past period. An annual attestation says little about last month. Regulators now expect registers, ongoing monitoring, and records of subcontractors, so the document is one input and not the control. Keep the reading step as a triage aid, and add live data for the suppliers that matter most.

The fifth approach connects a central GRC platform to ERP, HR, and finance systems through APIs, pulls sample data for continuous control monitoring, and lets business users configure forms and dashboards without IT. The benefit is a single view across functions. The weak points are periodic pulls, which show the past, and the copy of personal and transaction data in a central store, which adds residency, access, and breach exposure. Where a rule can run next to the data and return only a result, the copy is unnecessary. Keep the platform as the register of obligations, owners, and results, and avoid making it the warehouse for every record.

The table maps each function to its first-generation approach and to the 2026 practice that the rest of this article develops.

FunctionFirst-Generation Approach2026 PracticeSource To Read
Regulatory changeFeeds, keyword triage, alerts to humansGoverned agents that map impact and draft updates for approvalEuropean Commission AI Act guidance; DOJ guidance
Policy managementProse library with search and chatbotRules as code, retrieval with citationsNIST OSCAL; Open Policy Agent
Control testingPeriodic data pulls, sample testingEvidence collected on every run, preventive rulesNIST OSCAL
Third partiesAnnual reports read by AIRegister held as data, supplier software inventories, live signalsDORA register guidance; CISA SBOM minimum elements
Data integrationCentral copy through APIsChecks run near the data, results returnedData protection law for your jurisdictions
AI in complianceEfficiency toolRegulated asset with an owner, an inventory, and monitoringISO/IEC 42001; DOJ guidance; EU AI Act

Reading the table by row shows a pattern. The older approaches mostly help people read faster, while the newer ones change who or what does the work. Reading faster has a ceiling, because the volume of obligations keeps rising. Changing the work has a different cost curve, since a rule written once runs on every event. That is why the rest of this article spends less time on dashboards and more time on data, rules, and evidence.

How Do You Prepare the Ground Before Automating Anything?

Automation multiplies whatever it reads. If two policies contradict each other, a workflow will apply both and produce exceptions at machine speed. If a process has no documented owner, an automated alert will land in a shared mailbox where nobody acts on it. Run a pre-automation hygiene sprint of two to four weeks before you switch on any tool. Its job is to make the inputs consistent and owned. Teams that skip it find the gap when the first month of alerts goes unanswered.

Start with the taxonomy. Give every policy, risk, control, system, and regulatory obligation a unique identifier, and store the links between them as data, in a table or graph, and not inside document text. Adopt a recognized open control framework as the naming baseline so your control wording does not become a private dialect. Version each record and keep the history. These records feed every later step, including machine-readable control formats such as OSCAL, which expect unique identifiers and explicit relationships. A mapping that exists only in a person's head, or in a paragraph of a Word file, cannot be queried.

Assign one human owner to every policy and control, and record the owner as a role with a named deputy. A shared inbox cannot own a remediation. Decide what happens when the owner is absent, because automated tasks do not wait. Review the owner list each quarter against the organization chart, since reorganizations break routing silently. Treat an unowned control as a finding and not as a data quality issue. Owners need the authority to accept or reject an automated result, or the workflow becomes a notification system.

Check what you already pay for before you buy. Most enterprises hold workflow, approval, document, and collaboration tools that can run a policy attestation, a conflict of interest disclosure, or an exception approval within days. Build one workflow on those tools, measure the hours saved, and use the number to justify a dedicated platform if one is still needed. This route fits a limited budget and produces evidence for the business case. It also shows early where the data is weak. Do not extend a makeshift workflow into a permanent system without a security and records review.

Bring your external and internal auditors in before go-live. Tell them which controls you plan to automate, show the evidence each run will produce, and agree what they will accept as proof. An auditor who sees a green indicator on a dashboard will ask how you know the data feed did not drop records or the script was not edited. Settling that question in advance prevents rework at year end. Choose the first process where volume is high and judgment is low, and where the rules can be written as if-then statements. Gift approvals, access reviews, and training attestations usually pass that test.

Use Case 1: How Do You Turn a Regulatory Change Into an Approved Action?

The problem is the delay between a legal change and a corrected plan. A new or amended text appears, an analyst reads it, a policy owner is found, a ticket is raised, and weeks pass before anyone confirms what changed in practice. The 2026 answer is a governed agent workflow that does the reading, mapping, and drafting, and stops at a named human for approval. An agent here means software that takes a goal, plans several steps, and acts through the systems you connect to it. The value is a shorter time to decision, measured from publication of the change to an approved action.

Build it as a chain with explicit handoffs. An intake step watches official sources and your licensed feeds and records each new item with its source link and date. A mapping step compares the item with your obligation register and lists the policies, controls, systems, and owners that the change touches. A drafting step prepares proposed edits to the policy text and, where rules exist as code, to the rule files. A review step sends the package to the owner with a comparison of old and new wording, the sources used, and the confidence the model assigns to each mapping. The owner approves, edits, or rejects, and the decision is logged. No agent publishes a policy or changes a production rule on its own.

The staggered transition rules of the EU AI Act (Regulation (EU) 2024/1689) show why this pays. While high-risk AI systems embedded in regulated products (Annex I) and legacy general-purpose AI models placed on the market before August 2025 have until 2 August 2027 to comply, standalone high-risk systems (Annex III) and the transparency duties in Article 50 apply a full year earlier, on 2 August 2026. Article 111 adds another wrinkle: high-risk systems already in use before 2 August 2026 are exempt unless they undergo a "significant change in their design." A team that applies a single blanket deadline across its AI inventory will either miss an active 2026 transparency obligation or waste budget retrofitting a legacy system that was grandfathered in. A mapping step that links each article to a dated plan item, and that flags the partial exceptions for a person to read, catches this. The lesson is that agents reduce reading time, and they do not remove the need for a lawyer to read the exceptions.

Several technical points decide whether this works. Keep the obligation register as structured data with stable identifiers, so the mapping step compares like with like. Give each agent least privilege access: read access to registers and repositories, write access only to a drafts area. Log every prompt, every retrieved source, and every output, because the log is your evidence if a regulator asks how you decided. Test the mapping step against a set of past changes where you already know the right answer, and track how often it misses an affected policy. Retest after any model change. A missed mapping is the failure that costs most, so measure recall before you measure speed.

Use Case 2: How Does Compliance as Code Block Violations Before They Happen?

The problem is that a policy written as prose can only be checked after the fact. Someone deploys an unencrypted database, a payment above an approval limit goes through, or an access right stays after a role change, and the control that should have stopped it finds it at the next test. Compliance as code writes the rule in a machine-readable form and runs it at the point of action. A violating change is stopped, or flagged for approval, before it takes effect. This converts a detective control into a preventive one, and the preventive version runs on every event at nearly no marginal cost.

Two open building blocks matter here. The first is NIST's Open Security Controls Assessment Language, known as OSCAL. It is a set of machine-readable formats in XML, JSON, and YAML that describe control catalogs, baselines, implementations, and assessment results, and NIST's project page presents it as a move away from word processor and spreadsheet documents. The second is a policy engine. Open Policy Agent is a general-purpose, open source engine that evaluates declarative rules, written in a language called Rego, against input such as a deployment request, and the CNCF project page lists it as graduated since January 2021. You can use another engine. The point is that rules live in version control, are reviewed like any change, and run in the pipeline.

Start small and in the right order. Take one control with a plain yes or no test, such as encryption at rest for production data stores, and write the rule. Run it in audit mode for a month, where it logs what it would have blocked and blocks nothing. Review the false positives with the control owner, correct the rule, and only then switch it to enforcement for new deployments. Keep an exception path with an expiry date and a named approver, because a rule without an exception path gets bypassed. Store the rule, its tests, and its approval record together so an auditor can see which version ran on which date.

A decision rule helps with scope. Write as code the rules that are objective, frequent, and cheap to get wrong. Leave to people the rules that depend on context, such as whether a gift is reasonable in a given culture. Where a rule stays in prose, map it to the code rules that support it so the link is visible. Resist converting the whole policy library at once, since many policies mix testable statements with statements of intent. A realistic first target is five to ten rules in one technology domain, owned jointly by compliance and engineering.

Use Case 3: How Do You Collect Evidence Once for Many Frameworks?

The problem is audit fatigue. Teams answer the same question for several frameworks, in several formats, at several times a year, and engineers spend days producing screenshots that are out of date when saved. The approach is to collect evidence once, automatically, from the system of record, and to link that single record to every requirement it supports. The value shows in a metric you can track: hours spent answering one audit request, and the number of times the same evidence is requested.

Technically, you connect read-only service accounts to the systems that hold the facts: cloud configuration, identity management, source control, endpoint management, and the ERP. A scheduled job runs a defined query, such as which production databases lack encryption or which active users lack multi-factor authentication. The job stores the raw output with a timestamp and a hash. A mapping table then links the query to each control and framework requirement that it supports. OSCAL's assessment results model suits this, since it records observations, risks, and findings, and an observation can hold evidence produced by a machine, as NIST's OSCAL FAQ explains.

Auditors accept automated evidence when they can re-perform it. Build an evidence package for each automated control with three parts. The first is the configuration baseline: the version of the query or rule that ran, held in version control. The second is a completeness and accuracy record: the count of items the source holds, the count the job retrieved, and the hash or log that shows nothing was filtered out. The third is the exception trail: for every item the automation flagged, the record of who reviewed it, what they decided, and when. Ask your auditors to test one package before you scale, and write down what they asked for.

Watch three failure modes. A job that fails silently produces a gap that looks like compliance, so alert on missing runs and not only on failed checks. A query that is too narrow passes controls that should fail, so review the population definition with the control owner. A service account with broad rights becomes a risk of its own, so scope it to read access, rotate its credentials, and review it with the rest of your privileged accounts. Treat each automated test as a control that itself needs a test.

Use Case 4: How Do You Monitor Third Parties Beyond Annual Reports?

The problem is that an annual attestation describes a past period, and a supplier can change its subcontractors, software, or location the week after it is signed. Regulators have moved toward registers and continuous knowledge of the supply chain. Under DORA, financial entities must keep a register of information on all contractual arrangements with providers of ICT services, and for services that support critical or important functions the register must also name the subcontractors involved, as BaFin's guidance describes. BaFin also notes that identification codes for entities and providers were often wrong or missing in early submissions, which is a data quality problem and not a legal one.

Treat the register as a living dataset. Each row links a contract, a service, a business function it supports, the provider, and its subcontractors, with unique identifiers such as legal entity codes. Feed it from procurement and contract systems, and reconcile it with accounts payable, since a supplier that is paid but absent from the register is a finding. Where a supplier delivers software, request a software bill of materials, which is a machine-readable list of the components in that software. In July 2026, CISA, the NSA, the FBI, and international partners published the 2026 Minimum Elements for an SBOM, replacing the 2021 guidance from the US Department of Commerce's NTIA, according to the joint publication. Use that document as your baseline when you write the contract clause.

Apply the same logic to AI components. A sensible practice is to ask suppliers for an inventory of the models, datasets, and third-party services behind any AI feature they sell you, and to treat a missing inventory like a missing SBOM. No single binding format for this has been confirmed in the sources I reviewed, so write the clause around the information you need and not around a format name. Ask for the model provider, the model version, whether customer data is used for training, and where processing takes place. While regulators have not yet mandated a single statutory schema, open standards such as OWASP CycloneDX (ML-BOM) and SPDX 3.0 (AI and Dataset profiles) already provide machine-readable structures for this. Write your contract clause around the substance you need, whether delivered in CycloneDX, SPDX, or structured JSON: the model provider, the model version, training data provenance, whether your data is used for training, and where processing takes place.

Prioritize by dependence. A supplier whose failure would stop a core service deserves live signals, such as status data, vulnerability alerts on its software, and changes in its subcontractor list. A supplier of office supplies does not. Keep the document review step from the first-generation approach for the long tail, and use it only to triage. Define in advance what a signal triggers: a call to the supplier, a risk re-score, or a review of your exit plan. Test the exit plan for your most dependent suppliers at least once, because a plan that has never been run is an assumption.

Use Case 5: How Do You Detect Conflicts of Interest Without Annual Surveys?

The problem is that annual self-declarations rely on the honesty of the people who are least likely to be honest. A conflict of interest or a bribery scheme usually leaves traces in data the company already holds, in different systems that nobody joins. The approach is to compare data sets that already exist for a legitimate control purpose and to route unexplained matches to a human reviewer. It is an analytics control, and it is worth the effort only if it is designed with privacy and fairness in view.

Here is an illustrative scenario, with assumptions stated. A mid-sized manufacturer with 2,400 employees and 3,000 active suppliers wants to test whether any approver has a personal link to a supplier. It compares the supplier master file against the declared outside interests and the registered addresses in the HR system, and it compares purchase approvals against expense reports. Suppose the first run produces 41 matches. The reviewer finds that 33 are harmless, such as a shared postcode in a small town, 6 need a follow-up conversation, and 2 go to investigation. The metric is the rate of confirmed findings per match, here 2 of 41, or about 5 percent. The number is invented for illustration, and your own rate will depend on the quality of your data and the tightness of your matching rules.

The technical design needs care. Normalize addresses, names, and bank identifiers before matching, and record the match logic as a versioned rule so that results can be reproduced. Score matches by strength, so that a shared bank account ranks far above a shared street name. Keep the reviewing team small, and log every access to the underlying personal data. Limit retention of non-matches, and delete them on a schedule.

Involve legal, privacy, and employee representatives before the first run. Using HR data for monitoring needs a documented lawful basis, a data protection impact assessment in jurisdictions that require one, and, in many countries, consultation with works councils. Tell employees what categories of analysis you run. A program that surprises people damages trust, and a model that produces false accusations damages individuals. A match is a question for a human to ask, and never a conclusion. Train reviewers to open a conversation with the facts, to document the answer, and to close the case if the explanation holds.

Use Case 6: How Do You Answer Data Subject Requests Within the Legal Deadline?

The problem is that a request to access or delete personal data triggers a search across many systems, and each system has a different owner. Under the GDPR, the controller must act on the request without undue delay and within one month of receipt, and it may extend by two further months where necessary because of complexity or the number of requests, provided it tells the person within the first month and gives reasons, as the Irish Data Protection Commission explains. Notes that circulate in the field sometimes quote 30 to 45 days, which comes from other regimes. For GDPR requests, plan around one calendar month and treat the extension as a notice you must send on time.

A workable automation has five stages. A portal or mailbox rule logs each request with its receipt date, which starts the clock. An identity check confirms who is asking, using the least data needed, and the clock handling follows the law of your jurisdiction. A discovery step queries each connected system by a stable person identifier and returns a list of locations and records. A review step lets a privacy specialist apply exemptions, such as the data of other people or material under legal hold, before anything is released or deleted. A fulfilment step exports the data or removes it, then records what was done in each system. The log of that work is your evidence.

Keep a human at the decision points. Automation finds and assembles; a person decides what may be released and what must be kept. Deletion needs extra care, since a legal hold, a tax retention rule, or a fraud investigation may require you to keep some records, and backups often cannot be edited record by record. Write the rule for backups in advance, for example that deleted data ages out when the backup expires and is not restored. Some descriptions promise that a deletion completes across all systems in minutes. That claim is not credible for estates with backups and archives, and I would not promise it to a regulator.

Measure the process with three numbers: the median days to close, the share of requests closed inside the first month, and the number of systems that needed manual search. The third number tells you where to invest. Each system that cannot be queried by person identifier is a gap in your record of processing, and fixing it improves more than this one process.

Use Case 7: How Do You Triage Pharmaceutical Deviations Faster?

The problem is that every departure from an approved procedure in manufacturing or a clinical study needs a documented investigation, and the volume of minor deviations can crowd out the serious ones. Quality teams spend time finding similar past events, classifying severity, and drafting the first version of the report. Those steps are repeatable, while the decision on root cause and the sign-off are not. This is a natural fit for an assistant that prepares, with a qualified person who decides.

The workflow starts when a deviation is logged. A retrieval step searches the historical deviation records for events with the same equipment, product, batch family, or failure pattern, and lists them with links. A classification step proposes a severity and a category using the company's own definitions, and shows which text in the record drove its proposal. A drafting step fills the investigation template with the facts already in the record, and marks every sentence that the model inferred and did not copy. The quality lead reviews the draft, corrects it, and signs. The system keeps the original draft, the edits, and the signature as separate records.

Pharmaceutical quality systems operate under strict rules for electronic records and for data integrity, so the validation of this assistant is the real work. Define the intended use in writing, test the classification step against a set of past deviations that quality staff have already rated, and set a threshold for acceptable agreement before go-live. Track false negatives, meaning serious events the assistant rated minor, more closely than false positives. Restrict the assistant to the records it needs, and keep patient and trial participant data out of its prompts unless your privacy review has approved it. Re-validate after any change to the model or the prompts.

The same pattern fits other review-heavy work in regulated sectors, such as checking draft promotional material against approved product information before a legal reviewer sees it. The design principle does not change. The assistant finds, compares, and drafts, and a named qualified person decides. Set the success measure as the time a qualified reviewer spends per case at equal or better error rates, and do not claim a figure until you have measured it in your own process.

How Do You Replace Policy Chatbots With Cited Answers?

A policy assistant should answer from your approved documents and say so, or decline. Many first-generation chatbots were described as improving by learning from what employees ask. In a compliance setting, that is a design to avoid, because an assistant that adapts to user behavior can drift and can state a limit or an exception that no policy contains. The safer design is retrieval from a controlled library: the assistant finds the relevant passages in approved documents and writes its answer only from those passages.

Set four rules in the design. Every answer must cite the document, version, and section it came from, with a link to that passage. If the library holds no relevant passage, the assistant must say it cannot confirm the answer and route the question to a named owner. The library has an owner, and a document enters it only through a change process with approval and an effective date. Use settings that reduce variation in the model's output, and log each question, the retrieved passages, and the answer so you can review samples.

Test it before you open it to employees. Write fifty questions with known answers, including some that the policies do not cover and some that are traps, such as a question that mixes two policies. Score the assistant on whether it cites correctly and whether it declines when it should. Review a monthly sample of live questions, and feed the gaps back to policy owners, since a frequent question the library cannot answer is a signal that a policy is unclear. Place the assistant where people already work, such as the chat or approval tool they use daily, because few employees open a separate compliance portal on their own.

Assign ownership of the assistant itself. The compliance function owns the library and the rules for declining, the technology team owns the platform and its access controls, and the privacy officer reviews what the logs contain, since employee questions can reveal personal circumstances. Set a retention period for the logs and restrict who can read them. An employee who asks about a conflict of interest or a harassment procedure should not find that the question became a monitoring record. Publish a short notice that tells staff what the assistant does with their questions, and review it whenever the design changes.

How Do You Stop Alert Fatigue From Continuous Monitoring?

Moving from sampling to testing every event changes the volume of exceptions. A control that once produced ten findings a quarter from a sample of twenty-five can produce thousands of exceptions a month when it tests the full population. If every exception goes to an analyst, the team drowns and real problems hide among trivial ones. The answer is tiering: define in advance what the system may resolve by itself, what it groups into patterns, and what a person must review.

Here is an illustrative scenario with assumptions stated. A company runs full-population testing on expense reports and training deadlines and gets 1,200 exceptions in a week. Policy owners agree that 1,000 low-value cases, such as a small overage on a meal or a training deadline missed by a day, may be handled automatically with a notice to the employee and a record. The system groups 140 further exceptions by employee, counterparty, and amount pattern into 22 cases. The remaining 60 are medium or high severity and go to a reviewer one by one. Human review drops from 1,200 items to 82 items, which is 22 cases plus 60 single items, or about 7 percent of the raw volume.

Several controls make this safe. The thresholds are set by the policy owner and approved, not tuned by the analytics team to reduce workload. Automated handling must be reversible, and each action must be logged. Pattern rules look for threshold avoidance, such as many amounts just under an approval limit, since a low-value exception repeated is a different risk. Sample the auto-handled population each month to check that nothing serious was classified as trivial. Report to management the volume by tier, the handling time, and the sample error rate, so the tiering itself is under control.

Expect the first months to be noisy. Rules written from policy text alone will misfire on real data: an expense category that the policy never mentions, a subsidiary with a different approval limit, a training course with two names. Hold a weekly review with the policy owner during the first two months, fix the rules that misfire, and record each fix with its reason. Resist the urge to widen the auto-handling tier early to cut the workload, since the sample error rate is what justifies each widening. When the sample error rate stays low for a quarter, propose the change to the owner in writing and keep the approval on file. That record is what shows an auditor that the thresholds were chosen with care.

How Do You Govern the Automation Itself?

The AI and rules that run your compliance work are assets with their own risks, and regulators and auditors increasingly ask about them. The DOJ guidance on evaluating corporate compliance programs asks prosecutors to consider how a company manages the risks of AI and other new technologies, including in the compliance program, and whether it monitors those technologies for trustworthiness and reliability, according to the Baker McKenzie summary. The first-generation view of AI as a pure efficiency gain no longer fits that expectation.

Run an inventory of every AI system and automated rule that touches compliance work, with an owner, a purpose, the data it uses, the decisions it influences, and the date of its last test. ISO/IEC 42001, the international standard for AI management systems, gives a structure for this: it specifies requirements for establishing, implementing, maintaining, and continually improving such a system, and certification is voluntary, as ISO describes. You can adopt its practices without certifying. Add a review step for every change of model, prompt, or rule, and keep a test set so that you can compare results before and after.

Decide where human oversight is mandatory and write it down. The high-risk requirements of the EU AI Act include human oversight, and any decision that affects an employee, a customer, or a legal position should have a named person who can understand, question, and override the output. Check that the reviewer has the time and the information to do so, because a person who approves hundreds of items an hour is rubber-stamping. Measure the override rate. A rate of zero over several months is a warning sign and not proof of accuracy.

Look for unmanaged automation as well. Business teams create their own AI agents, connectors, and workflows that reach company data, often without telling compliance. Use your identity and access management and your network controls to discover them. Treat each agent as a non-human identity with an owner, a defined scope, and an expiry date, and bring those you find into the inventory. A compliance program that governs its own tools and ignores the ones that spread through the business has a gap in exactly the place regulators are looking.

What Does a 90-Day Start Look Like?

Choose one process that is frequent, rule-based, and disliked, such as gift approvals, access evidence, or policy attestations. Spend the first thirty days on mapping every step, every system, and every handoff, and on cleaning the rules until they can be written as if-then statements. Name the owner and the deputy, record the baseline for two metrics, and agree with your auditors what evidence they will want. Do not buy software yet. The output of this phase is a one-page process map, a rule list, an owner list, and a baseline.

In days thirty-one to sixty, build the smallest working version on tools you already own. Run it in audit mode alongside the manual process so you can compare results. Write the evidence package from the start, with the baseline, the completeness record, and the exception trail. Test the failure paths: a missing data feed, an absent owner, a rule that is wrong. Ask your auditors to review one package. Fix what they raise before you extend the scope, because their questions show you what the final design must carry.

In days sixty-one to ninety, switch the process to live use for a limited population and measure the two metrics against the baseline. Report the result in hours, error rates, and exceptions found, and include what did not improve. Take the numbers to the budget holder with a proposal for the next process and for any platform your tests showed you need. Review the governance items at the same time: add the new automation to the inventory, confirm its owner, and schedule its first retest. A second process built on the same data and evidence design costs much less than the first, and that is the point at which the program starts to scale.

Plan the budget conversation with the same discipline as the technical work. Present three figures: the hours saved in the pilot, the cost to run it, and the cost to extend it to the next two processes. Include the time of the control owners and of the auditors who reviewed the evidence, since leaving it out makes the pilot look cheaper than it was. If the pilot showed that a dedicated platform is needed, say which requirement the native tools failed to meet. If it showed they are enough, say that too. A business case that admits what the pilot did not prove is easier for a finance director to approve than one that promises everything.

Final Perspective

The next decision on your desk is probably which process to automate and with what. The sequence in this article gives you a test for any proposal: does it read clean data with an owner, does it express the rule so it can run at the moment of action, does it leave evidence that an auditor can re-perform, and does it keep a named person accountable for the decisions that carry risk? A proposal that cannot answer all four will produce faster work and weaker assurance. Dates and rules will keep moving, as the EU AI Act timeline did this year, so build the register and the mappings that let you absorb a change in days.

The practical takeaway is small. This month, pick one process, count the hours your people spend collecting information for it, and write down who owns each rule. If you cannot name the owner of a rule, you have found your first automation project, and it needs no software. Which of your controls would you be comfortable letting a machine enforce at midnight with no one watching, and which would you not?

References

  1. European Commission, AI Act frequently asked questions, https://digital-strategy.ec.europa.eu/en/faqs/navigating-ai-act
  2. Orrick, DOJ Updates Corporate Compliance Guidance, Takes Aim at AI, Emerging Technologies, and Data Analytics, https://www.orrick.com/en/Insights/2024/09/DOJ-Updates-Corporate-Compliance-Guidance-Takes-Aim-at-AI-Emerging-Technologies-and-Data-Analytics
  3. Baker McKenzie, US Department of Justice updates compliance program evaluation guidance with focus on artificial intelligence, https://insightplus.bakermckenzie.com/bm/investigations-compliance-ethics/united-states-us-department-of-justice-updates-compliance-program-evaluation-guidance-with-focus-on-artificial-intelligence-and-evolving-policy_1
  4. European Banking Authority, Preparations for reporting of DORA registers of information, https://eba.europa.eu/activities/direct-supervision-and-oversight/digital-operational-resilience-act/preparation-dora-application
  5. BaFin, DORA register of information, https://www.bafin.de/ref/19860272
  6. NIST, Open Security Controls Assessment Language (OSCAL), https://csrc.nist.gov/projects/OSCAL
  7. NIST, OSCAL frequently asked questions, https://csrc.nist.gov/Projects/open-security-controls-assessment-language/faqs
  8. Cloud Native Computing Foundation, Open Policy Agent project page, https://cncf.io/projects/open-policy-agent-opa
  9. CISA and partners, 2026 Minimum Elements for a Software Bill of Materials (SBOM), https://www.ic3.gov/CSA/2026/260729.pdf
  10. Data Protection Commission (Ireland), How long does an organisation have to respond to my access request, https://www.dataprotection.ie/en/faqs/access-and-rectification/how-long-does-organisation-have-respond-my-access-request
  11. ISO/IEC, ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system, https://www.iso.org/standard/42001