AI Governance Without Killing Innovation

Episode transcript

Transcript provided for accessibility and search. Minor transcription errors may remain.

AI Governance Without Killing Innovation

SPEAKER_00 0:02

There's a fairly reliable way to create shadow AI in an organization. Make the approved AI process sufficiently painful. If using the company approved AI platform requires three forms, two committees, legal approval, and a meeting next Thursday, employees will discover ChatGPT remarkably quickly. And honestly, can you blame them? Over the last four episodes, we've talked about some very real AI risks data, identity, permissions, accountability, agents, delegated authority, prompt injection. And we've only just got started. So the natural response is governance, policies, standards, risk assessments, committees, approval processes, controls, all sensible things. But there's a danger here. If we're not careful, we can create an AI governance model that's technically very safe because nobody can actually get anything through it. That's not governance. That's obstruction with a PowerPoint deck. I'm Steve, and this is AI Security Multipliers. Today we're going to talk about how to govern AI without killing the very innovation we're trying to enable. Because good governance shouldn't make safe AI harder to use. It should make the safe path the easiest path. Governance

Governance Has a Branding Problem

SPEAKER_00 1:23

has a bit of a branding problem. Tell a room full of people that you're introducing governance and you'll rarely see spontaneous applause. Nobody says fantastic, I was hoping for more oversight. Governance tends to conjure images of committees, policies, approvals, and people asking whether you've completed section 7 of a form nobody remembers creating, but governance itself isn't the problem. Bad governance is. Good governance answers some very practical questions. What are we trying to achieve? Who is accountable? What risks matter? Which decisions require approval? What controls do we need? And how do we know whether any of this is actually working? Those are useful questions. The problem starts when governance becomes disconnected from the work. We build a process designed around the organization chart rather than around how AI is actually being used. Then every AI use case gets sent through exactly the same machinery. A developer using AI to explain a piece of non-sensitive code gets treated like somebody building an autonomous agent that can approve financial transactions. They're both AI. But they're clearly not the same risk, and if our governance process can't tell the difference, the problem may not be the AI. We've

Govern the Risk, Not the Buzzword

SPEAKER_00 2:38

used this phrase throughout the podcast, govern the risk, not the buzzword. Episode 5 is where I think it really matters. AI is not a single risk. It's a technology that can create, amplify or change many different kinds of risk. Suppose I use an approved AI tool to rewrite a paragraph of public marketing material. That's an AI use case. Now suppose an AI agent can modify production infrastructure. Also an AI use case. If both require exactly the same approval process, something has gone wrong. Governance should respond to consequence. What data is involved? What decision is being made? What action can the system take? Who could be affected? What happens if it gets something wrong? Can the outcome be reversed? Those questions tell us far more about risk than simply asking, does this use AI? Because eventually almost everything will. And if every AI-enabled capability requires special treatment forever, we're going to need a much bigger committee.

Not Everything Needs a Committee

SPEAKER_00 4:02

I should probably declare something at this point. I have nothing against committees. I've served on many of them. I've chaired some, I've probably created a few, for which I apologize. Committees can be useful when a decision genuinely requires multiple perspectives. AI certainly creates situations where security, privacy, legal, technology, data, and the business all need to contribute. But a committee should not become the place where every AI decision goes to live. Imagine someone wants to use an approved enterprise AI tool to summarize meeting notes. Do we really need 12 people discussing it? Hopefully not. Now imagine somebody wants to connect an autonomous agent to the financial system and allow it to approve transactions. I'll make time for that meeting. The trick is knowing the difference. Governance should escalate based on risk. Routine, low-risk uses should move quickly. Moderate risk users may need additional controls or specialist review. High-risk users should receive proper scrutiny and explicit approval, and some users may simply be unacceptable. That's not anti-innovation. That's how organizations make risk decisions about almost everything else. AI doesn't need an entirely separate universe. One of

Create a Fast Path to Yes

SPEAKER_00 5:19

the most important things an AI governance program can do is create a fast path to yes. Security teams sometimes focus heavily on defining what people can't do. Don't put confidential information into public AI. Don't connect unapproved tools, don't create unauthorized agents, don't upload customer data, don't do this, don't do that. All potentially sensible, but employees are trying to accomplish something. If the governance model only tells them what they can't use, they'll eventually find something they can. Usually about five minutes before an important deadline. So give them an approved route, approved enterprise AI platforms, approved development tools, approved models, approved data patterns, approved agent architectures, clear guidance on what information can be used, reusable security controls, pre-approved integrations, sandbox environments where people can experiment safely, make the secure option convenient, because convenience is one of the most underrated security controls in the enterprise. If the approved AI tool is easy to access, useful, and reasonably capable, people have much less incentive to find their own. The objective shouldn't be how do we stop people using AI? It should be how do we make the safe way of using AI the obvious way? I like

Risk Tiers Beat One Giant Process

SPEAKER_00 6:46

relatively simple risk tiers, not 37 categories, not a scoring model requiring a statistics degree, something people can actually understand. Maybe three or four levels. At the lowest level, perhaps the use case involves approved tools, non-sensitive information, no automated decisions, and no ability to take consequential actions. That should be easy. Perhaps even self-service. Then we have use cases involving internal information or more meaningful business processes. They may need additional review, then higher risk cases, sensitive personal data, regulated information, customer-facing decisions, employment decisions, security actions, financial transactions, production changes, autonomous agents. Those deserve more attention. And at the top, perhaps there are uses the organization simply won't permit. The exact tiers will vary. What matters is proportionality. Because if a 20-minute productivity experiment and a million-dollar autonomous business process go through the same governance process, people will rightly conclude that the process doesn't understand the technology. And once people stop respecting the governance model, you've got a much bigger problem than paperwork. One of the mistakes

Start With the Use Case

SPEAKER_00 8:06

I see in technology governance is starting with the technology instead of the business use case. Someone says, we want to use generative AI. For what? We're still figuring that out. That's going to make risk assessment rather difficult. Start with the outcome. What problem are we solving? Who is using the system? What information does it need? What output does it produce? Does it make a recommendation? Does it make a decision? Does it take an action? Who relies on the result? What happens if it's wrong? Suddenly the risk conversation becomes much more concrete. This also helps with something we discussed in episode two: ownership. A real use case has a business owner. Someone wants the capability because they expect it to create value. That's important because the business owner should remain involved in the risk decision. AI governance shouldn't become a mechanism for transferring every difficult decision to cyber security. I'm happy to identify security risk. I'm much less enthusiastic about becoming the owner of your account's payable process simply because somebody added AI to it. Another

Use the Governance You Already Have

SPEAKER_00 9:15

temptation is to create an entirely separate AI governance universe. New AI committee, new AI policy, new AI risk methodology, new AI intake process, new AI architecture board, new AI terminology, and obviously a new spreadsheet because apparently no governance framework is truly complete until somebody has created a spreadsheet. Some new governance will be necessary. AI does introduce genuinely different questions, but organizations already have a lot of useful machinery. Information security, privacy, data governance, enterprise architecture, third party risk, legal review, procurement, model risk, records management, compliance, internal audit, change management, use them. If an AI system processes personal information, privacy governance still matters. If a vendor provides the system, third party risk still matters. If it accesses sensitive data, information security controls still matter. If it changes production, change management principles still matter. AI governance should connect these disciplines, not duplicate all of them. Otherwise, we don't get governance. We get governance about governance. And I promise you nobody wants that meeting.

Policies People Might Actually Read

SPEAKER_00 10:35

We talked briefly about AI policies in episode one. I think they're worth revisiting. Because many organizations responded to generative AI by rapidly producing an acceptable use policy, which was sensible. The problem is that policies can become enormous, and if your AI policy is 28 pages long, employees may technically have access to it. That's not quite the same as having read it. The most important guidance should be simple. Which AI tools can I use? What information can I put into them? What can't I put into them? Can I use AI-generated content externally? When do I need human review? What users require additional approval? Who do I contact if I'm unsure? Give employees answers to the questions they're actually asking. You can have detailed standards behind that. Security requirements, architecture patterns, legal guidance, data handling requirements. But the everyday employee shouldn't need to become an AI governance specialist before asking a model to improve an email. A policy isn't successful because it exists. It's successful because people understand what they're expected to do. Innovation

Experimentation Needs Somewhere to Live

SPEAKER_00 11:49

requires experimentation, and experimentation is messy. People don't always know exactly what the use case will become when they start. That's normal. So organizations need somewhere safe for experimentation to happen. A sandbox. An approved AI environment. Synthetic or nonsensitive data, restricted integrations, limited permissions, clear boundaries, a place where teams can test ideas without immediately creating production risk. I think this is incredibly important, because if the only options are full production approval or don't do it, people will find option three. And option three usually involves a credit card. A good sandbox allows curiosity without pretending curiosity is risk-free. It also gives security teams visibility. We can see what people are trying to build. We can learn which controls are actually needed, and we can help successful experiments move toward production safely. Security becomes part of the innovation process rather than somebody invited in three days before launch to explain why something can't go live. I've attended a few of those meetings, they're rarely anyone's favorite. This is where I

Security Should Be an Accelerator

SPEAKER_00 13:05

think cybersecurity has a choice. We can position ourselves as the group that approves AI, or we can position ourselves as the group that helps the organization use AI safely. Those sound similar, they're not. If security becomes the gatekeeper for every AI decision, two things happen. First, we become a bottleneck. Second, we accidentally take ownership of decisions that belong somewhere else. I'd rather security provide expertise, secure architecture patterns, identity controls, data protection, threat modeling, agent security, logging requirements, testing, incident response, clear guidance, and then work with the business to understand the actual risk. The best security organizations I've worked with don't simply say no. They say, here's what would need to be true for us to do this safely. Sometimes the answer will still be no. There are risks an organization shouldn't accept, but no should be the result of understanding the risk. Not the default setting. Security is much more valuable when it helps the organization move faster with appropriate controls. That's what being a security multiplier actually means.

Governance Has to Move at AI Speed

SPEAKER_00 14:23

There's another challenge. AI is moving quickly, very quickly. A governance process designed around a six-month technology review cycle is going to struggle. Models change, capabilities change, vendors add features, agents gain new tools, enterprise platforms quietly introduce AI into products you already own? The system you reviewed six months ago may not be the system you're using today. So governance can't just approve something once and forget about it. We need lifecycle governance. What changed? Did the model change? Did the data sources change? Did permissions change? Did we add new tools? Did the agent gain new authority? Has the vendor changed how data is processed? Has the business use changed? A low-risk assistant can gradually become a high-risk agent without anybody deliberately deciding today we're going to dramatically increase the risk? It happens one useful feature at a time. That's why governance needs triggers for reassessment, not constant bureaucracy. Just recognition that material changes deserve another look. The risk follows the capability.

Measure Whether Governance Is Working

SPEAKER_00 15:36

Here's a question governance teams don't ask often enough. Is the governance actually working? Not how many AI assessments did we complete? That's activity. I'm interested in outcomes. How long does a low-risk AI use case take to approve? How many employees are using approved AI tools? How much unapproved AI usage are we finding? How many projects arrive at security immediately before launch? How many high-risk AI systems have clear owners? How many agents have known identities and documented permissions? How quickly can we identify AI systems across the enterprise? Are incidents increasing? Are exceptions increasing? Are people finding the process useful? If everything takes three months to approve, that's a governance metric, not necessarily a good one. If employees consistently bypass the process, that's also a governance metric. You can write the finest AI policy in the world. If everyone wrote around it, you don't have governance, you have literature.

The Board Doesn't Need to Approve the Chatbot

SPEAKER_00 16:39

Boards have an important role in AI governance, but that role is oversight. The board should understand how AI supports the organization's strategy, where material risks exist, what the organization's risk appetite looks like, whether management has appropriate governance in place, and whether significant incidents or exposures are being managed. The board probably doesn't need to approve the marketing team's AI writing assistant. That's management's job. One of the signs of mature governance is that decisions happen at the right level. Routine operational decisions remain operational. Higher risk decisions escalate appropriately. Material enterprise risks reach executive leadership, and where appropriate, the board. Everything shouldn't go upward. That's not governance, that's organizational anxiety. The goal is to give people enough authority to make the decisions they're accountable for. Which brings us back to something we've discussed throughout this series: ownership. Governance works when responsibility, authority, and accountability line up. When they don't, decisions either stall or disappear into the gaps. And risk rather likes gaps.

What Do You Do on Monday?

SPEAKER_00 17:53

So what do you do on Monday? First, look at your current AI approval process from the perspective of an employee trying to use it, not from the perspective of the person who designed it. How long does it take? Is it obvious where to start? Can employees understand what requires approval? Do low-risk uses move quickly? If the process feels painful to you, imagine how it feels to somebody who just wants to solve a business problem. Second, define your fast path. Which platforms are approved, which common use cases are already acceptable. What data can be used? What doesn't require additional review? Make that information easy to find. Third, define your escalation triggers, sensitive data, external decision making, high impact decisions, production changes, privileged access, autonomous actions, regulated processes, whatever matters in your organization. Fourth, connect AI governance to existing governance. Don't rebuild privacy, cybersecurity, procurement and data governance from scratch, and finally create somewhere safe to experiment. Because people are going to experiment with AI. The only question is whether they do it somewhere you can help them, or somewhere you discover six months later during an incident investigation. I know which one I'd prefer.

The Multiplier

SPEAKER_00 19:25

And now this week's multiplier. Take your organization's AI governance process and try to use it seriously. Pick a simple low-risk AI use case, something you would actually like an employee to be able to do, then walk through the process from beginning to end. How many forms? How many approvals? How many meetings? How many days? And most importantly, at any point would a reasonable employee decide it was easier just to go around you? If the answer is yes, don't immediately blame the employee. Fix the path. Because one of the best ways to reduce shadow AI isn't another monitoring tool. It's making approved AI easier to use than unapproved AI. Good governance doesn't just control risk, it creates a faster path to yes. That's this week's multiplier. Over these

Closing: A Faster Path to Yes

SPEAKER_00 20:14

first five episodes, we've covered quite a lot of ground. AI entering the enterprise, ownership, data, identity, agents, authority, and now governance. The common thread through all of it is that AI doesn't exist separately from the organization. It sits inside the systems, data, identities, processes, and decisions we already have. So AI governance shouldn't become a giant structure sitting above the organization shouting instructions downward. It should become part of how the organization works. The business owns outcomes. Data owners remain accountable for their data. Technology provides platforms and architecture. Security manages security risk. Privacy, legal and compliance bring their expertise. Leadership establishes risk appetite. And governance connects those responsibilities. The goal isn't maximum control. The goal is appropriate control. Because innovation without governance can create unacceptable risk. But governance without innovation creates something else. Irrelevance. People will find another way. Good AI governance should allow an organization to move quickly. Where the risk is low, carefully where the risk is higher, and deliberately where the consequences really matter. That's how I think we should approach AI. Not how do we control this technology, but how do we help the organization use this technology well? That's a much more interesting question, and probably a much more useful one. That's it for this episode of AI Security Multipliers. If you've enjoyed these first five episodes, please follow or subscribe to the show and share it with somebody who's trying to work through the same challenges. And I'd particularly like to hear how AI governance is working in your organization. What's working? What's painful? What have you tried? That absolutely didn't work. And if you've managed to create an AI governance process that employees genuinely enjoy using, I definitely want to hear from you because you may have achieved something considerably more difficult than artificial general intelligence. I'm Steve, and I'll see you next time.