
Enterprise-Wide AI Risk Management Implementation and Operationalisation
The Operational Blueprint for Enterprise-Wide Artificial Intelligence Governance, Controls, and Resilience
by Markus Krebsz
Moving beyond foundational theory, this volume provides the definitive architectural blueprint for operationalising the Enterprise-Wide AI Risk Management (EW-AiRM) framework across modern organisations. Designed for immediate operational deployment, this manual details how institutions can establish named accountability, embed the seven-dimension HAiPECR ethical filter, and manage complex cross-border regulatory obligations. Readers are guided through scaling defensive architecture across organisational tiers—from streamlined teams-of-one in EW-AiRM Core to federated, multi-subsidiary global enterprises in EW-AiRM Full. Learn how to construct robust AI risk registers mapped against the MIT AI Risk Repository’s seven domains and 24 subdomains. Implement practical mitigants utilising the MIT AI Risk Mitigation Database’s 831 evidence-based controls, the EW-AiRM Quantum Risk Control Set, and the master mapping architecture. Across high-stakes sectors including financial services, healthcare, and critical infrastructure, this guide delivers concrete operational controls against agentic cyber threats, Model Context Protocol vulnerabilities, and multi-agent coordination collapse. Integrate live tooling—including the companion platform at www.ewairm.com, the knowledge bot, and the digitally sovereign 'Ask AiRisk' agent at my.airisks.net. Prepare your enterprise for harvest-now-decrypt-later vectors and black swan disruptions, transforming regulatory compliance from a defensive overhead into an enduring strategic advantage.
- Business & Entrepreneurship
- Science & Technology
- Educational & Academic
- Corporate Leadership
- Business Strategy
- Artificial Intelligence
Operationalising the Non-Negotiables: The Baseline for Any Scale
Governance failure in artificial intelligence is rarely a failure of documentation. Walk into almost any organisation after an AI incident and you will find a policy. You will find a framework document, perhaps a slide deck from a training session eighteen months ago, maybe even a committee terms of reference with an impressively long list of attendees. What you will not find, in the overwhelming majority of cases, is a single person who can say, without hesitation, ‘That system was mine. I owned it. I should have caught this.’ The paperwork survives the incident. The ownership does not, because in most organisations it never existed in the first place.
This chapter is about building the floor beneath that paperwork: the operational baseline that has to exist before an organisation has any credible claim to be governing its AI responsibly, regardless of whether it employs three people or three hundred thousand. It does not matter whether the organisation runs a single customer service chatbot or a portfolio of forty production models. It does not matter whether it operates in financial services, manufacturing, or a two-partner accountancy firm above a shop. The five non-negotiables that follow apply in full, and none of them is an optional extra to be phased in once the organisation feels ready.
The Five Non-Negotiables
EW-AiRM sets out five baseline requirements that sit beneath every tier of the framework, whether an organisation is operating EW-AiRM Core as a team of one or EW-AiRM Full as a federated multinational. They are: a named human accountable for every deployed AI system; a completed HAiPECR pre-deployment assessment before that system goes live; a tested human override capability with a defined response time; an explicit, signed acceptance of residual risk; and a protected channel through which anyone in the organisation can raise an AI-related concern without fear of retaliation. None of these five requirements is expensive, requires specialist technical infrastructure, or demands a chief AI officer, a model risk committee, or a six-figure governance platform. What they require is a decision, made by someone with the authority to make it, that these five things will exist before the next AI system is switched on, and that none of the organisation’s existing AI systems will continue running without them for more than a brief, deliberately bounded transition period.
The reason these five sit below everything else in the EW-AiRM architecture is that they address the accountability vacuum directly, at its source, rather than attempting to compensate for it after the fact with more process. A sophisticated model inventory does nothing if nobody is named against each entry. An elaborate risk taxonomy does nothing if the organisation has no mechanism for switching the system off when the taxonomy’s warnings turn out to be correct. Everything else in this book, the tiered scaling in the chapters that follow, the sector-specific controls, and the agentic and quantum resilience work in the later chapters assume that these five things are already in place. If they are not, the right next step is not to read further. It is to close this book for an afternoon and put them in place.
Non-Negotiable One: A Named Human Being
The first non-negotiable sounds almost too simple to need a chapter section of its own: a named individual is accountable for every deployed AI system, with documented responsibility for its governance from deployment through to decommissioning. The complication is not in the statement. It is in the organisation’s instinct to dilute it. Ask a typical risk committee who owns the customer-facing chatbot, and you will often get an answer like, ‘That sits with the digital transformation steering group,’ or, ‘IT and marketing jointly manage that one.’ Both answers are a way of saying that nobody owns it.
A committee can discuss a system. A committee can approve a system. A committee cannot be woken at 2am and asked to make a judgement call about whether to pull the plug on a model that has started generating outputs nobody anticipated. A named person can. That asymmetry is the entire argument for named accountability, and it holds regardless of how collaborative, consultative, or multidisciplinary the organisation’s culture otherwise is. Committees remain useful for input, challenge, and shared visibility; what they cannot do is substitute for the single point of accountability that the three-lines model has relied on for decades in other risk domains, where first-line ownership, second-line oversight, and third-line independent assurance each has a named head, not a rotating cast.
In practice, naming an owner means more than writing a name in a spreadsheet cell. It means that the name is attached to a short, specific set of responsibilities covering the system’s entire lifecycle, not just its launch. The named owner signs off on deployment, holds responsibility for monitoring outputs against the behaviour the system was approved for, is the first point of contact when something goes wrong, and remains accountable until the system is formally decommissioned or the ownership is formally transferred to someone else, in writing, on a specific date. If a named owner leaves the organisation, resigns, or moves to a different role, EW-AiRM treats the system as temporarily ownerless, which is itself an escalation trigger. A system without a current, named, human owner should not continue running past a short grace period, typically no more than a week, before a replacement owner is confirmed.
A compliance officer at a mid-sized organisation once described the exercise of assigning named owners as ‘the single most uncomfortable afternoon of the whole rollout’. People do not enjoy having their name attached to something they do not fully understand, and the instinct to push back is strong. ‘I don’t even know how the model works,’ is a common objection, and it is worth answering directly: named accountability under EW-AiRM does not require the owner to be a data scientist. It requires the owner to understand what the system is supposed to do, what it is not supposed to do, how to recognise when it has drifted from that boundary, and who to call. That is a governance competency, not a technical one, and it can be taught in an afternoon to someone who has never written a line of code.
Committee Oversight Versus Named Accountability
| Dimension | Committee Oversight | Named Accountability |
|---|---|---|
| Who is called at 2am when the system fails | Nobody, until a chair is located | One identified individual, with a documented backup |
| Who signs the deployment approval | The committee, collectively, often unminuted | The named owner, personally, on the record |
| What happens when the owner leaves | Nothing changes, because no individual was ever tied to the system | An escalation trigger fires until a replacement is confirmed |
| Where responsibility sits after an incident | Diffused across the group and difficult to trace | Clear, traceable, and auditable |
None of this is an argument against multidisciplinary involvement, which remains valuable and is expanded considerably once an organisation scales beyond Core into the Standard and Full tiers covered later in this book. It is an argument against mistaking multidisciplinary involvement for accountability. The two are not substitutes for each other, and an organisation that treats its AI governance committee as the owner of its AI systems has, in practice, no owner at all.
Non-Negotiable Two: The Core HAiPECR Assessment
HAiPECR operates as a continuous ethical filter across every layer of the framework, asking the same seven questions at every significant AI decision point, from initial design through to retirement. At the Core tier, before a small organisation attempts to run the full seven-dimension assessment with the rigour expected of a larger enterprise, three of those questions do almost all of the practical work, and they deserve to be treated as mandatory even for the smallest deployment.
The first inquiry asks whether the system is necessary and proportionate: is this AI tool actually solving a problem that justifies its use, or has it been adopted because a vendor made a persuasive sales pitch, because a competitor has one, or because a junior team member discovered a free tool online and nobody thought to ask whether it should be there. This is the practical, Core-tier expression of HAiPECR’s Conduct risk (C) dimension, and answering it honestly often does more to reduce risk than any amount of subsequent monitoring, because a system that should never have been deployed cannot be made safe by better oversight. It can only be made safe by not deploying it.
The second inquiry examines behavioural incentive modification, which is a formal way of asking an ordinary question: what does this system make people do differently, and is that different behaviour something the organisation actually wants? A collections system that scores customers and nudges staff toward more aggressive recovery tactics is modifying behaviour. A recruitment screening tool that quietly teaches hiring managers to trust the algorithm’s shortlist without checking it themselves is modifying behaviour. Neither of these effects shows up in a technical specification sheet, and both are exactly the kind of risk that a small organisation, moving fast and without a dedicated model risk function, is most likely to miss. This inquiry sits alongside necessity and proportionality within HAiPECR’s Conduct risk dimension and should never be treated as a one-off question asked only at launch; incentives shift as staff become more comfortable with a system, and a Core-tier organisation should expect to revisit this question at least annually, or sooner if the system’s use case changes.
The third inquiry concerns explainability and auditability, the practical heart of HAiPECR’s E dimension: can the system’s output be explained to a non-technical person in terms they can actually understand and challenge, and is there a record good enough to support that explanation six months after the fact, when memories have faded and the original implementation team may have moved on. A small organisation does not need a model explainability platform to satisfy this requirement. It needs a short, plain-language description of how the system reaches its outputs, kept somewhere the named owner can retrieve in minutes, and a log of decisions detailed enough that a regulator, a customer, or an internal auditor asking ‘why did the system do that’ eighteen months from now can be given a real answer rather than a shrug.
A Core-tier organisation running this three-question assessment honestly, on paper no longer than two or three pages per system, has done more serious ethical and risk work than many much larger organisations manage with a formal model governance committee and a backlog of unreviewed model cards. The length of the document is not the point. The honesty of the answers is.
Non-Negotiable Three: The Human Override, Tested and Timed
A documented override capability is worth very little if nobody has ever tried it. EW-AiRM requires that the override, whatever form it takes for a given system, cutting off access, halting outputs, or isolating the system from the networks and workflows it touches, is tested and verified to work within a defined response time, which should be no more than four hours regardless of whether the failure occurs during business hours or at two o’clock on a bank holiday Sunday morning.
The 4-hour figure is deliberately conservative rather than aspirational. It is not a target to work towards eventually; it is the outer limit beyond which the organisation has, for practical purposes, lost control of the system for an unacceptable length of time. Four hours is long enough for a misbehaving recommendation engine to serve tens of thousands of bad suggestions, send several rounds of incorrect automated correspondence, and decline or approve a meaningful volume of credit applications on faulty logic. It is not, by contrast, so short that it demands infrastructure beyond the reach of a small organisation. A documented, rehearsed process involving a single named person with the authority and the access credentials to disable a system can meet a 4-hour target without any specialist tooling at all.
What actually has to be in place for this non-negotiable to be real, rather than theatre, is straightforward to list but easy to neglect:
- A documented, specific mechanism for disabling each system, whether that means revoking an API key, disabling an account, removing network access, or pulling a physical switch; vague language such as ‘IT will handle it’ does not meet the standard.
- A named person, matching the owner identified under the first non-negotiable, who holds both the authority and the technical access required to execute the override without needing to ask permission from someone else first.
- A documented backup, a second named person who can act if the primary owner is unreachable, on leave, or otherwise unavailable, because override capability that depends entirely on one person’s phone being switched on is not resilient.
- A tested escalation route that has actually been rehearsed, not merely written down. Reading a procedure is not the same as executing it under time pressure.
- Separate verification of out-of-hours response, because many organisations discover, only when they actually test this, that their override process assumes office hours, daytime staffing, and access to systems that are locked down overnight for unrelated security reasons.
Run the test properly and you will usually learn something uncomfortable. A compliance officer testing an out-of-hours override for the first time described the experience this way: ‘We assumed it would take twenty minutes. It took three hours and forty minutes, almost all of it spent trying to reach someone with the right system access, because the person who built the integration had left the company and nobody had updated the access list.’ That is precisely the kind of failure the test is designed to surface while the stakes are a rehearsal rather than a live incident. An organisation that has never tested its override arrangement has, in effect, no override arrangement; it has only a belief that one exists, and belief is not a control.
| Testing scenario | What it checks | Common failure point |
|---|---|---|
| Business-hours disablement | Can the named owner execute the override while fully staffed and reachable | Access credentials shared informally and now expired or mislaid |
| Out-of-hours disablement | Can the backup execute the override with reduced staffing and no immediate IT support | Backup lacks the system permissions the primary owner holds |
| Owner unreachable scenario | Does the escalation route reach a second person within the 4-hour window | No documented secondary contact, or contact details are out of date |
Non-Negotiable Four: Explicit Risk Acceptance, Not Passive Assumption
Every AI system carries residual risk once the available controls have been applied, and the critical governance question is not whether that residual risk exists but whether anyone in the organisation has actually agreed to accept it. Assessment Dimension 3 within EW-AiRM draws a sharp line between explicit and implicit acceptance, and that distinction matters more at the Core tier than almost anywhere else in the framework, because small organisations are particularly prone to implicit acceptance disguised as a decision that was never really made.
Implicit acceptance looks like this: a system goes live because nobody objected, a deadline passed, and the absence of a formal ‘no’ was quietly treated as a ‘yes’. No executive signed anything. No one can point to the moment the decision was taken, because there wasn’t one; there was only drift. This is the pattern behind a large proportion of AI governance failures that later get described, inaccurately, as a failure of controls. The controls were often perfectly adequate on paper. What failed was that nobody with the authority to accept the residual risk ever actually did so, in writing, with their name attached.
The fourth non-negotiable requires that this gap be closed with a formal risk sign-off, a short document that forces the moment of decision into the open rather than letting it dissolve into organisational silence. A usable template for a Core-tier organisation needs only a handful of fields, but each one should be completed in full before the system goes live:
- System name and named owner, matching the record created under the first non-negotiable.
- A plain-language summary of the residual risk that remains after controls have been applied, written so that someone outside the project can understand what could go wrong.
- The specific controls already in place to reduce that risk, including the override arrangement tested under the third non-negotiable.
- A statement of who is accepting the residual risk on the organisation’s behalf, and their job title, because the authority to accept risk should scale with the seriousness of the risk being accepted.
- A signature and date, with a review date no more than 12 months ahead, after which the acceptance expires and must be renewed, not assumed to continue indefinitely.
The signature matters more than its brevity suggests. It converts a vague organisational comfort with a system into a specific, attributable decision made by a specific, identifiable person. If that person later has to explain their decision to a board, a regulator, or a court, they will be explaining something they actually chose, rather than something that simply happened around them. For a team-of-one risk function, this often means the sign-off goes to the managing director or the most senior executive available, not because the risk manager lacks the competence to judge the risk but because they frequently lack the organisational authority to accept it on the organisation’s behalf. Separating the judgement from the authority is itself part of the discipline.
Non-Negotiable Five: The Protected Escalation Channel
The fifth non-negotiable addresses a problem that has nothing to do with technical architecture and everything to do with culture. An incident reporting pathway must exist for AI-related concerns, and it must come with documented protection against retaliation, consistent with the organisation’s existing whistleblowing protections and applicable jurisdictional requirements, whether the UK’s established statutory protections for workers raising concerns in the public interest, the European Union’s Whistleblower Protection Directive, or an equivalent regime elsewhere.
Psychological safety does the work here, because a channel that exists on paper but is never used is governance theatre with a reporting form attached. The person most likely to notice that an AI system is behaving strangely is rarely the named owner, who may be looking at dashboards rather than frontline outcomes. It is far more often the junior analyst who notices that the credit model keeps flagging the same postcode, the customer service agent quietly overriding the chatbot’s suggested replies because they sound wrong, or the operations assistant who has worked out that a scheduling tool is systematically disadvantaging staff on certain shift patterns. These are exactly the people least likely to raise a concern through formal channels if they believe doing so puts their job at risk, which is why the protection against retaliation is not a legal footnote. It is the mechanism that determines whether the channel produces any useful information at all.
For a small organisation, building this fifth non-negotiable does not require a dedicated hotline operated by a third party, although many larger organisations will eventually move in that direction as they scale into the Standard and Full tiers. What it requires at the Core level is a documented route, communicated clearly to every member of staff, that specifies exactly how to raise an AI-related concern, who receives it, what happens next, and what protection applies to the person raising it. One compliance officer summarised the minimum bar well: ‘If a new starter can’t tell me, within their first week, exactly who they’d go to if the AI tool did something that worried them, the channel doesn’t really exist yet, whatever the policy document says.’ That is a useful test, and it costs nothing to apply.
It is worth pairing this channel explicitly with the organisation’s existing grievance and whistleblowing framework rather than inventing a parallel structure, because employees are more likely to trust and use a mechanism they already understand than a new one invented specifically for AI concerns and bolted on top. The novelty here is not the mechanism. It is the explicit instruction that AI-related concerns fall within its scope, stated clearly enough that nobody has to guess.
Thirty Days, One Person: Building the Floor
None of the five non-negotiables, taken individually, demands significant resource. Taken together, under time pressure, they can feel like a substantial project to a solitary risk manager or compliance officer who may be handling AI governance as one responsibility among many, alongside data protection, financial crime, health and safety, or whatever else their job description has accumulated over the years. EW-AiRM’s Phase 1 treats this as an assessment period running from day one to day thirty, and that window is realistic precisely because the five non-negotiables do not require new systems, new budgets, or new headcount. They require decisions, documentation, and one or two uncomfortable conversations with senior leadership about authority and sign-off.
A practical thirty-day sequence for a solitary risk function might look like this:
- Days 1 to 5: Inventory and naming. List every AI system actually in use, including the shadow tools staff have adopted on their own without formal approval. Against each one, write a name, not a department, not a committee. If no obvious owner exists, assign one immediately and tell them.
- Days 6 to 12: Core HAiPECR assessment. Run the three fundamental inquiries against every listed system. Expect this to surface at least one tool that fails the necessity and proportionality question outright. Decommissioning that tool is itself a legitimate, and often the most valuable, outcome of the thirty days.
- Days 13 to 18: Override design and testing. For each surviving system, document the specific disablement mechanism, confirm the named owner has the access required to execute it, name a backup, and run at least one test during business hours.
- Days 19 to 22: Out-of-hours override test. Repeat the test outside normal working hours, ideally involving the backup rather than the primary owner, to confirm the arrangement does not quietly depend on daytime staffing.
- Days 23 to 26: Risk sign-off. Draft the sign-off document for each system, present the residual risk honestly to the appropriate level of seniority, and secure signatures with review dates no more than 12 months out.
- Days 27 to 30: Escalation channel communication. Confirm the protected reporting route exists, aligns with existing whistleblower protections, and is communicated to all staff in plain language, with a specific test of whether a sample of employees can correctly describe how to use it.
A risk manager working through this sequence alone described the experience in terms that will be familiar to anyone who has tried to retrofit governance onto a system that has been running quietly for years: ‘The hardest part wasn’t writing anything down. It was finding the person who actually knew how to switch the thing off, because the person who built it left eighteen months ago and nobody had thought to ask.’ That observation captures something important about the entire exercise. The five non-negotiables are not primarily a documentation project. They are a project of locating, naming, and testing capabilities that the organisation often assumes exist until someone actually goes looking for them.
It is worth being honest about what thirty days buys an organisation and what it does not. It buys a floor, not a ceiling. An organisation that completes this sequence has closed the accountability vacuum for its current AI footprint and established the minimum conditions under which further scaling, further sector-specific controls, and the more advanced resilience work covered later in this book can be built without the structure collapsing under its own weight. It has not yet built a mature risk register, has not yet mapped its controls against the wider repository of risks and mitigations that later chapters introduce, and has not yet addressed the agentic, quantum, and systemic threats that sit ahead in the advanced operational chapters. Those are legitimate next steps, not omissions from this chapter. The point of the floor is that everything built afterwards needs somewhere solid to stand.
Where the Gaps Usually Show Up
Having walked many organisations through this exact sequence, a handful of patterns recur with enough regularity that they are worth naming explicitly, so that a reader working through their own thirty days can watch for them rather than discovering them the hard way.
The first is the discovery of shadow AI during the inventory step. Almost every organisation that conducts an honest inventory for the first time finds tools nobody formally approved: a free writing assistant installed by a marketing contractor, a spreadsheet plug-in with embedded predictive features that nobody realised was AI at all, a chatbot trial that was never switched off after the pilot ended. These tools frequently fail the necessity and proportionality inquiry immediately, simply because nobody can articulate why the organisation needs them, and the correct response is usually decommissioning rather than retrospective governance. A tool nobody can justify does not deserve a named owner. It deserves to be switched off.
The second recurring gap is access concentration. Override testing frequently reveals that only one person in the entire organisation holds the technical permissions needed to disable a given system, and that person is often the most junior member of staff involved in the original deployment, someone whose departure would leave the organisation with a documented override procedure it has no practical way to execute. Naming a backup under the third non-negotiable exists specifically to close this gap, and it should be treated as a genuine finding, not a box-ticking formality, whenever the test surfaces it.
The third is sign-off inflation, where every system, regardless of its actual risk profile, gets routed to the most senior executive available because nobody has thought about proportionality in the sign-off process itself. A simple internal scheduling tool with a limited blast radius does not need the same level of signature as a system making customer-facing credit decisions. Matching the level of sign-off authority to the seriousness of the residual risk keeps the fourth non-negotiable credible rather than becoming a rubber stamp that senior leaders apply without genuinely engaging with the risk in front of them.
The fourth is silence in the escalation channel during the first few months after launch. An organisation that introduces a protected reporting route and receives zero reports in the following quarter should not read that silence as reassurance. It is far more likely to indicate that staff do not yet trust the channel, do not know it exists, or do not believe raising a concern is safe. A small organisation can test this directly by asking a sample of employees, informally and without warning, whether they know how to report an AI concern and what would happen if they did. If the answers are vague, the channel needs more work regardless of how clean the incident log currently looks.
Carrying the Floor Forward
The five non-negotiables established in this chapter are deliberately scale-agnostic, which is precisely why they sit beneath everything that follows in this book. The next chapter builds the full EW-AiRM Core blueprint for the team-of-one organisation on top of this floor, extending it into shadow AI management and a fuller treatment of override protocols for the smallest organisational footprint. Later sections scale the same architecture into mid-market and enterprise settings, where named accountability becomes layered rather than singular, where the three Core HAiPECR inquiries expand back out to the full seven-dimension assessment, and where override testing moves from a manual, individually executed process to something closer to automated, continuously monitored infrastructure. None of that later complexity replaces the five non-negotiables. It is built directly on top of them, and an organisation that tries to skip ahead to sector-specific controls or agentic threat resilience without first securing this floor will find that the more advanced material has nothing solid to attach to.
Before moving on, it is worth sitting with a short set of questions that any compliance team, regardless of sector or size, should be able to answer without hesitation. Can every member of staff name the specific regulatory frameworks that apply to the AI systems they use day to day, and recognise which would be classified as high-risk under the EU AI Act? Is there a documented process for monitoring regulatory change across every jurisdiction the organisation operates in, rather than relying on someone noticing a headline? Has anyone in the organisation considered what happens when an AI system stops behaving as a single, governable tool and starts acting as an autonomous agent capable of taking actions without a human reviewing each step? Has the organisation considered, even informally, what a migration to post-quantum cryptography will mean for the AI systems it depends on today? And does anyone understand how dependent the organisation’s own risk taxonomy is on external classification systems that may themselves change, be withdrawn, or be superseded? These are not questions this chapter answers. They are the questions the rest of this book exists to address, and an organisation that cannot yet answer them honestly has, at minimum, found the right place to start.
The Team-of-One Blueprint: Implementing EW-AiRM Core
Priya runs compliance, HR and a good portion of operations for a forty-person logistics scheduling firm, and on the Tuesday she finally sat down to tackle AI governance properly, she had six browser tabs open, a half-finished spreadsheet from eighteen months earlier, and a growing suspicion that she had no idea how many AI tools her own staff were …

