Direct Answer
AI process controls are the operating rules that bring an AI-enabled step inside your management system — a defined owner, defined inputs and outputs, operating criteria, validation where the output cannot be fully verified, competent human supervision, records of what the system decided, and a review cadence that catches drift. They are not a new standard. Nearly every AI process control an organization needs is already required by ISO 9001; what changes in 2026 is that the requirements now apply to a decision-maker that is not a person.
The Control Gap
Why AI Process Controls Decide Whether Automation Helps
Define. Control. Prove.
There is a moment in every AI deployment when the risk quietly relocates, and most organizations miss it. It happens when a model stops recommending and starts deciding.
Up to that moment, AI is an input. A person reads what the model produced, applies judgment, and owns the outcome. Past that moment — when the system routes the work, adjusts the parameter, approves the record, or triages the complaint on its own — the model has become part of how the work gets done. It is inside the boundary of your management system whether anyone wrote it down or not. And every requirement that governs a process now governs it.
That is the entire argument for AI process controls. Not caution for its own sake, and not a reason to slow adoption. The organizations that lose ground in 2026 will not be the ones that moved fast — they will be the ones that automated a step nobody owned, could not explain what the system decided when an auditor asked, and had no evidence the decision was ever reviewed.
MSI client experience suggests the most expensive AI mistake is rarely a bad model. It is automating a process that was never stable to begin with. Automation does not remove a defect; it reproduces the defect faster and at greater volume. Sound AI process controls are what let a leadership team tell the difference between a process ready to be automated and one that is merely tiresome.
The Symptom Leaders Notice First
The gap usually surfaces during an internal audit, not a board meeting. An auditor asks a simple question about a routine output — where did this number come from, who approved this disposition, why was this batch released — and the answer is that a tool produced it. Then the follow-up question lands: which tool, verified how, reviewed by whom, and against what criteria? Organizations without AI process controls discover at that point that a meaningful part of their operation has no owner and no record. The finding is rarely about the technology. It is about the absence of the ordinary controls that every other process in the building already has.
Definition
What Are AI Process Controls?
Rules. Records. Review.
AI process controls are the documented, operating-level rules that govern an AI-enabled step inside a management system: what the step is allowed to decide, on what inputs, against what criteria, with what human oversight, producing what records, monitored how, and corrected by whom when it fails. They sit one layer below AI policy and one layer above the tool itself.
The phrasing matters. A control is not a prohibition and not an aspiration — it is a rule that operates. “We use AI responsibly” is a value. “The model may propose a complaint category; a trained reviewer confirms or overrides it before the record closes, and the override is logged” is an AI process control. The first cannot be audited. The second can be audited on a Tuesday afternoon by someone who has never met the person who wrote it.
This is also why AI process controls tend to be far less exotic than the surrounding conversation suggests. Organizations already know how to control a process whose output cannot be fully verified by later inspection — it is the same discipline that governs welding, sterilization, and additive manufacturing, where MSI's guide to ISO for additive manufacturing shows validation doing exactly this job. A model is a new instance of a familiar problem, not a new category of problem.
The Three Postures That Determine the Control Set
Before writing any AI process controls, classify the step by what the system is permitted to do. The control burden scales with the posture, and mismatching the two is the most common design error.
Assistive. The system drafts, summarizes, or suggests; a person reviews everything before it takes effect. Light controls: identify the tool, train the user, and make clear that the human signature carries the accountability.
Supervised-autonomous. The system acts, and a person reviews on a defined sample or against defined exception rules. This is where most operational AI actually sits, and where AI process controls do the heaviest lifting — sampling rate, escalation triggers, reviewer competence, and override records all have to be specified.
Autonomous. The system acts with no routine human review. This posture demands validation, continuous monitoring, and a documented rationale for why unattended operation is acceptable. In regulated contexts it is frequently the wrong answer, and deciding that deliberately is itself a control.
Two Different Jobs
AI Process Controls vs. AI Governance: Where the Line Sits
Policy. Practice. Proof.
AI governance decides; AI process controls execute. Governance is the leadership-level policy that determines where the organization will and will not use AI and who remains accountable. AI process controls are the operating rules that make a specific automated step behave the way the policy says it should — and produce the evidence that it did.
Organizations get into trouble by doing one without the other. A policy with no controls beneath it is a document nobody operates — the classic shelf artifact. Controls with no policy above them produce a patchwork where one department forbids what another has automated, and nobody can explain the inconsistency to a customer.
If your organization has not yet made the policy-level decisions — the red lines, the accountability map, the use cases you will never automate — start there. MSI's guide to AI governance for business covers that layer, including the executive questions worth answering before any policy is written. This article assumes those decisions exist and addresses what happens next: turning them into controls that operate.
One clarification, because the standards landscape invites confusion. ISO/IEC 42001 is the international standard for artificial intelligence management systems, and it is the reference point most often cited for what an AI policy should cover. MSI references it as landscape rather than service — MSI's work is building the mature ISO management systems that give an organization the discipline to make any AI policy operate in practice. The AI process controls described below live inside those systems.
The Control Set
The Nine AI Process Controls Every Quality System Needs
Own. Operate. Evidence.
The nine essential AI process controls are process definition, operating criteria, validation, tool verification, competence, records, monitoring, management review, and nonconformity handling. Each maps to a requirement an ISO 9001 system already places on the organization — which is why building them is a matter of extending existing discipline rather than inventing a parallel framework.
1. Process Definition — Name the Step and Give It an Owner
ISO 9001 Clause 4.4 requires the organization to determine its processes, their sequence and interaction, the criteria and methods needed for effective operation, and the responsibilities for them. An AI-enabled step is a process. It needs a name, a position in the sequence, defined inputs and outputs, and a named human owner — not the vendor, not “IT,” and not the person who happened to configure it. The first of the AI process controls is simply refusing to let an automated step exist unnamed. In practice this single control resolves a surprising share of the confusion, because most of what feels ungovernable about AI is really the absence of an owner.
Where the step touches an existing documented procedure, amend the procedure rather than writing a separate AI annex. A parallel AI document creates the same handoff gaps that parallel corrective-action procedures create — a failure pattern MSI describes in its guide to writing a corrective action procedure that holds up in practice.
2. Operating Criteria — State What “Correct” Means
Clause 8.1 requires operating criteria for processes and control implemented in accordance with those criteria. For an AI-enabled step this means writing down the acceptable performance boundary before deployment, not after the first incident: what confidence threshold triggers human review, what input conditions the step is not authorized to handle, what happens when the system encounters a case outside its intended use. AI process controls without stated criteria cannot fail visibly, which sounds convenient and is the opposite of convenient — an undetectable failure keeps running.
3. Validation — Prove It Works Where Inspection Cannot
Clause 8.5.1 addresses processes whose resulting output cannot be verified by subsequent monitoring or measurement — the classic special-process problem. Many AI-enabled steps qualify, because you cannot inspect a routing decision or a risk score after the fact and conclude it was correct. Validation for these AI process controls means defining the qualification method, the acceptance criteria, the review and approval, and the requirement for revalidation when anything material changes. A model update is a change. A change in input data source is a change. Re-scoping the step to a new product line is a change.
4. Tool Verification — Treat the Model Like Measuring Equipment
Clause 7.1.5 requires that monitoring and measuring resources be suitable and maintained, with confidence in the validity of results. When a system produces the number a decision rests on, that system is a measuring resource in everything but name. The corresponding AI process controls establish which tool version is in use, what verification was performed against a known reference, how often it is rechecked, and what happens to prior results when a check fails. Organizations comfortable with calibration recall already understand this: if the gauge was out, the parts it passed are in question. The same logic applies to a drifted model and the decisions it produced.
5. Competence — Someone Must Be Qualified to Overrule It
Clause 7.2 requires determining necessary competence and ensuring people are competent on the basis of education, training, or experience. Supervision of an automated decision is a competence requirement, and it is the one organizations most often skip. A reviewer who cannot independently reach a judgment about the case in front of them is not supervising; they are approving. These AI process controls define what the reviewer must know, how that is evidenced, and — critically — that the reviewer has enough time and authority to override. Related reading on how AI reshapes competence requirements appears in MSI's analysis of AI skills-based hiring.
6. Records — Capture What the System Decided and Why
Clause 7.5 governs documented information: creation, identification, control, protection, and retention. AI process controls at this layer specify that the record shows what was decided, by which system and version, on what inputs, whether a human confirmed or overrode it, and when. One verification is worth performing before you trust any platform: confirm the audit trail cannot be disabled by the person making the entry. If it can, records are not protected against loss of integrity regardless of what else the system does well. MSI's guidance on document and records control sets out the decisions that should be made before software is configured, not after.
7. Monitoring — Watch for Drift, Not Just Downtime
Clause 9.1 requires determining what needs to be monitored and measured, the methods, and when results are analyzed and evaluated. A model that has stopped working is easy to notice. A model that has quietly become less right is not. Useful AI process controls here track override rate, exception volume, distribution shifts in the inputs, and outcome quality sampled against an independent standard. A rising override rate is the single most informative early indicator available, because it means the humans closest to the work have already lost confidence — usually months before anyone measures a defect.
8. Management Review — Put AI Performance in Front of Leadership
Clause 9.3 gives the organization a required forum where process performance, trends, and resource adequacy are examined by top management. AI-enabled steps belong on that agenda as a standing input: where AI operates, how those steps performed, what drifted, what was overridden, and what resourcing the oversight requires. Handled well, this converts AI from a topic that surfaces only after an incident into a governed part of the operating rhythm — and it satisfies the accountability that the policy layer above these AI process controls asserts.
Give AI Oversight a Standing Agenda Slot
The Management Review That Actually Produces Decisions
Most management reviews collect inputs and adjourn without a decision on the record — which is exactly how automated steps go unexamined for a year. MSI's ISO Management Review Toolkits give you the agenda, input templates, trend tables, and minutes format auditors expect, built standard by standard from 200+ audits attended — so performance data, including from AI-enabled steps, produces documented decisions instead of a folder nobody reopens.
9. Nonconformity and Corrective Action — Define Failure in Advance
Clause 10.2 requires reacting to a nonconformity, evaluating the need to eliminate its cause, implementing action, and reviewing effectiveness. The last of the AI process controls answers a question most deployments never ask: what counts as a nonconformity here? A wrong output? An output outside the intended use? An unreviewed autonomous action that policy required a human to confirm? Deciding in advance is what allows the corrective-action system to engage at all. Without it, AI failures get handled informally as “tuning” and never enter the improvement loop, which is precisely how the same failure recurs. MSI's analysis of why AI alone fails at corrective action examines the related trap on the drafting side.
Controls only hold inside a system that requires them. MSI's ISO consulting work turns each one into an auditable habit — here is where to go deeper:
- Process Optimization Skills: Why Proven Methods Always Win — the leadership competencies that decide whether automation compounds or backfires.
- Free Process Optimization Scorecard (ISO 9004) — a 20-statement self-assessment that ranks the five skills and names the one to fix first.
- Document Control: Why FINAL_final_v7 Never Wins — the version discipline every AI record inherits.
- Internal Audit Planning: Why Proven Methods Always Win — how to build an audit program that actually examines automated steps.
- Continual Improvement: The Proven Engine ISO 9001 Demands — where AI findings have to land to produce change.
- The Future of Risk: AI, Ethics, and Emerging Threats — the risk categories that sit above any single control.
The Revision
What ISO 9001:2026 Changes for AI Process Controls
Less. Than. Promised.
ISO 9001:2026 publishes on September 16, 2026 and does not create a dedicated artificial-intelligence clause. There is no new requirement that says “control your AI.” Every one of the nine AI process controls above rides on requirements that already exist — which means organizations waiting for the revision to tell them what to do are waiting for an instruction that is not coming.
A great deal of commentary published ahead of the release promised sweeping new technology requirements. The revision, prepared by ISO/TC 176/SC 2, is evolutionary: a clearer separation of risks from opportunities, a stronger treatment of culture and ethics, climate-change provisions formalized into the context clauses, and a substantially expanded informative annex to aid interpretation. Read against the hype, that looks like a disappointment. Read correctly, it is a useful signal — the committee concluded that the existing requirements already reach AI, and it is right.
Two changes do sharpen how AI process controls should be written. The first is the risk-and-opportunity separation. Deploying a model to shorten cycle time is opportunity management; determining what happens when that model drifts or receives data it was never intended to handle is risk management. Treating them as one exercise produces an enthusiastic rollout with no control plan behind it, which is the single most common pattern MSI sees. The second is the strengthened treatment of culture and ethics, which gives the human-oversight controls a home in the standard rather than leaving them as good intentions. MSI's coverage of the ISO 9001:2026 ethics and culture update works through what that means for leadership.
Practical consequence for anyone planning a transition: do not wait for September to begin. The AI process controls described here are buildable today against ISO 9001:2015, and everything built now carries forward. Organizations planning both the revision transition and an AI rollout in the same window benefit from sequencing them deliberately rather than colliding them, which is a conversation worth having in a structured planning session rather than mid-project.
The Regulatory Floor
What Changed in 2026 — And Why Last Year's Guidance Is Wrong
Deferred. Not. Repealed.
For more than a year, August 2, 2026 was treated across the industry as the enforcement cliff for high-risk obligations under the EU Artificial Intelligence Act. Under the Digital Omnibus amendments, obligations for stand-alone high-risk systems listed in Annex III moved to December 2, 2027, and those for AI embedded in regulated products under Annex I moved to August 2, 2028. Most transparency obligations were not deferred.
The reason for the deferral matters more than the dates. The timeline moved because national authorities and harmonized technical standards were not ready — not because the requirements were reconsidered. Organizations reading this as permission to defer their own AI process controls are misreading it. Anyone with European market exposure should confirm current applicability against the consolidated regulation text rather than against articles written in 2025.
For a structure to reason with in the meantime, the NIST AI Risk Management Framework is voluntary, non-regulatory, and pairs cleanly with a management system — it helps determine how much oversight a given use case warrants, which is exactly the judgment the three postures above require. Certification and accreditation questions in this space route through Global ACI, which replaced the prior international accreditation arrangements effective January 1, 2026.
Higher Stakes
AI Process Controls in Regulated Environments
Device. Environment. Care.
The nine controls are universal, but the evidentiary bar rises sharply in regulated settings, and the sequencing changes with it.
Medical Devices
Under ISO 13485, software used in the quality system carries an explicit validation expectation, and the FDA's Quality Management System Regulation — effective February 2, 2026, incorporating ISO 13485:2016 into 21 CFR Part 820 — carries that expectation into US enforcement. AI process controls in this environment are documented before use, not retrofitted after. A note on vocabulary worth getting right: ISO 13485 predates the harmonized structure shared by ISO 9001, ISO 14001, and ISO 45001, and it contains no “risks and opportunities” construct. Importing that phrasing from ISO 9001 into a device quality system is a common and avoidable error. MSI's ISO 13485 work starts from the standard's own language.
Environmental Management
Two 2026 publications changed this ground. ISO 14001:2026 published on April 15, 2026 with a transition deadline of April 30, 2029, restructuring risk and opportunity requirements, strengthening the life cycle perspective, adding explicit change-planning requirements, and broadening operational control from outsourced processes to externally provided processes, products, and services — the last of which reaches directly into how you control a vendor-supplied model. Separately, ISO 19011:2026 published on May 27, 2026 and withdrew the 2018 edition with no transition period, changing the auditing guidance behind every internal audit program overnight. There is also a substantive point most organizations miss: the energy demand created by analytics and AI workloads is a genuine environmental aspect under a life cycle perspective, not an IT line item. MSI's ISO 14001 work covers the transition in full.
Healthcare
ISO 7101, published in October 2023, brings healthcare quality management under a formal management-system structure, and AI process controls land here with unusual weight because the affected party is a patient rather than a batch. Triage support, documentation generation, and scheduling optimization all touch service-user experience directly, and the competence control — a clinician with the time and standing to override — is the one that carries the most risk when it is treated as a formality. MSI's ISO 7101 healthcare quality practice is an expanding focus area for exactly this reason.
Execution
How to Build AI Process Controls Without Stalling Adoption
Inventory. Classify. Control.
Build AI process controls in five steps: inventory where AI already operates, classify each use by posture, control the highest-consequence step first, write the controls into the existing procedure rather than a separate document, and put performance on the management review agenda. Most organizations can complete the first two steps in a week.
Step 1 — Inventory what is already running. This is invariably the surprise. AI has usually entered the organization through individual initiative rather than a program: a scheduler here, a drafting assistant there, a vendor feature switched on in an update nobody reviewed. Ask each function a concrete question — where does software make or materially shape a decision that used to be made by a person — and you will find more than expected. An inventory nobody has built is not a governance failure; it is the normal starting condition.
Step 2 — Classify by posture and consequence. Sort each use into assistive, supervised-autonomous, or autonomous, then rate the consequence of a wrong output: internal inconvenience, customer impact, regulatory exposure, or harm. The combination tells you where AI process controls are urgent and where light-touch treatment is proportionate. Applying heavy controls uniformly is how a control program dies of its own weight.
Step 3 — Control the highest-consequence step first, completely. One step with all nine controls in place teaches the organization more than nine steps with one control each. It also produces a working pattern the rest of the organization can copy, which is how this scales without a central team writing everything.
Step 4 — Write controls into existing procedures. Amend the complaint-handling procedure, the release procedure, the supplier evaluation procedure. Resist the separate “AI Policy Manual” that lives beside the system rather than inside it — the shelf artifact is the predictable outcome. This is where a strong procedure set earns its cost, because amending a well-built procedure is a small edit and amending a weak one means rewriting it.
Step 5 — Route performance into the management review. Without this step the other four decay, because nothing that is never reviewed stays maintained. Bring override rates, exception volumes, and incidents to the same forum that reviews every other process. MSI's internal audit work is designed to make automated steps genuinely examinable rather than nominally in scope.
A note on tooling, because the sequencing error here is expensive. Software can carry out AI process controls; it cannot decide them. A platform configured before the organization has settled its posture classification, its review thresholds, and its record requirements will encode a set of vendor defaults, and those defaults quietly become policy. Decide first, configure second. That sequencing judgment is a leadership skill rather than an IT one — ISO 9004 names it technology enablement, and you can score your team's technology enablement maturity against the other four before the platform conversation starts.. MSI's guide to ISO compliance automation — and the MSI alliance with CAQ AG Factory Systems — is built on exactly that order of operations. Organizations further along in digital deployment may also find MSI's coverage of AI in quality management and change-management automation useful as background.
“Automating an unstable process does not remove the defect. It produces the defect faster, at greater volume, with fewer people watching. AI process controls are what tell you which processes are ready.”
— MSI Consulting Practice
The MSI Difference
How ISO Consulting Turns AI Process Controls Into Habit
Proven. Practical. Permanent.
The organizations that implement AI process controls successfully are almost never the ones with the most sophisticated technology. They are the ones with a management system mature enough that adding a new kind of process is a routine act rather than a special project.
Management Systems International (MSI) has spent 28 years building systems that make that true. MSI's track record includes 80+ certifications supported, 200+ audits attended, and 600+ professionals trained across manufacturing, technology, medical device, government, healthcare, and other regulated industries — which is why MSI's view of AI process controls is unromantic. Sitting in 200+ audits teaches you that the control that fails is almost never the clever one. It is the ordinary one nobody assigned.
That measurable authority matters here because AI claims are cheap and verification is rare. An ISO 9001 system makes the controls examinable: the step is defined, the criteria are written, the records exist, and the management review forces leadership to confront whether oversight actually happened. For organizations weighing whether ISO 9001, ISO 13485, ISO 14001, or ISO 45001 belongs in their operating model — or how to sequence an AI rollout against a revision transition — a structured planning session early is the most reliable way to scope the work and avoid rework. To talk through fit, call MSI at 760-434-9141. Background on MSI's approach is available through the ISO overview, the integrated management systems practice, and the industries MSI serves.
Take the Next Step
Controls Live in Procedures. Start From Ones That Already Work.
Every control in this article ends up as a clause inside a procedure someone follows on a Tuesday — and amending a well-built procedure takes an afternoon while rewriting a weak one takes a quarter. MSI's ISO Procedure Templates & Guides cover 15 procedure topics across five standards and combinations, in editable Word, with the judgment calls on scope, evidence, and record retention already made by a consultant who has attended 200+ audits. You edit. You do not architect.
See the ISO Procedure Templates & Guides →
Still working out the policy layer above these controls? Start with AI governance for business. Already past the decision phase and want the system built? Explore SurePath, MSI's turnkey ISO certification program, or SureResults for year-round maintenance. Prefer to talk first? Call 760-434-9141.
Questions Answered
AI Process Controls: Frequently Asked Questions
Ask. Answer. Apply.
What are AI process controls in a quality management system?
AI process controls are the operating-level rules governing an AI-enabled step: what it may decide, on what inputs, against what criteria, with what human oversight, producing what records, monitored how, and corrected by whom when it fails. The nine essential ones are process definition, operating criteria, validation, tool verification, competence, records, monitoring, management review, and nonconformity handling.
Does ISO 9001:2026 require AI process controls?
Not by that name. ISO 9001:2026 publishes on September 16, 2026 with no dedicated artificial-intelligence clause — the revision covers a clearer risk-and-opportunity separation, culture and ethics, climate provisions in context, and an expanded informative annex. AI process controls are nonetheless required in substance, because an AI-enabled step is a process and the existing requirements for processes apply to it.
How do AI process controls differ from AI governance?
Governance is the leadership policy layer — where the organization will and will not use AI, and who is accountable. AI process controls are the operating layer that makes a specific automated step behave as the policy requires and produce evidence that it did. Policy without controls is a shelf document; controls without policy produce inconsistency nobody can explain to a customer.
Do AI process controls require ISO/IEC 42001 certification?
No. ISO/IEC 42001 is the international standard for artificial intelligence management systems and a useful reference for what an AI policy should cover, but AI process controls can be built entirely inside an existing ISO 9001, ISO 13485, ISO 14001, or ISO 45001 system. MSI references ISO/IEC 42001 as landscape rather than a service line; MSI's work is building the mature management systems that let any AI policy operate.
Where should an organization start with AI process controls?
Start with an inventory of where AI already operates, then classify each use as assistive, supervised-autonomous, or autonomous and rate the consequence of a wrong output. Apply the full set of AI process controls to the highest-consequence step first and completely, write them into existing procedures rather than a separate manual, then route performance into the management review.
What is the most common AI process controls mistake?
Automating a process that was never stable. Automation reproduces an existing defect faster and at greater volume rather than removing it, and sound AI process controls exist partly to distinguish a process ready for automation from one that is merely tiresome. The second most common mistake is configuring a platform before deciding the review thresholds and record requirements, which lets vendor defaults become policy by accident.
How does ISO consulting help with AI process controls?
Experienced ISO consulting installs the process, records, and review discipline that makes adding a new kind of process routine rather than a special project — which is what AI process controls require. MSI's 28 years, with 80+ certifications supported, 200+ audits attended, and 600+ professionals trained, is built around leaving that capability behind so the team can run the system after the engagement ends.
References & Authoritative Sources
1. International Organization for Standardization — ISO 9001 Quality Management, and ISO/TC 176/SC 2.
2. ISO/IEC — ISO/IEC 42001, Artificial intelligence management systems.
3. ISO — ISO 13485 Medical Devices; eCFR — 21 CFR Part 820.
4. ISO — ISO 14001:2026 publication announcement, and ISO 14001 Environmental Management.
5. ISO — ISO 19011, Guidelines for auditing management systems, and ISO 45001.
6. European Commission — Regulatory framework for AI; EUR-Lex — Regulation (EU) 2024/1689.
7. NIST — AI Risk Management Framework, Baldrige Performance Excellence Program, and Manufacturing Extension Partnership.
8. Global ACI — international accreditation arrangements.
9. ASQ — Continuous improvement resources.
About Management Systems International (MSI)
Diana Lynn is President and Principal ISO Consultant at Management Systems International (MSI), a consulting firm she co-founded in 1998. With 28 years of experience including extensive AS9100 work in MSI's early years, MSI's track record includes 80+ certifications supported, 200+ audits attended, and 600+ professionals trained across manufacturing, technology, medical device, government, healthcare, and other regulated industries. Today MSI implements ISO 9001, ISO 13485, ISO 14001, and ISO 45001, with an expanding focus on ISO 7101 healthcare quality.
msi-international.com · 760-434-9141
