"How are security organizations adopting AI within security workflows, what are the governance blockers, and how ready are they for agentic security?"

Tell me what's top of mind for you on this topic right now — what are you wrestling with?

Right now the thing I keep coming back to is the data boundary question. We're a B2B SaaS company, so we're handling customer data, and every time someone on my team wants to wire up an AI capability — whether that's for our security workflows internally or something product-facing — the first conversation is always: where does the prompt actually go, and who sees it? We've got some teams leaning toward third-party hosted models, and that's fine for certain use cases, but for anything touching security telemetry or customer configuration data, I'm not comfortable with that yet. We've been exploring more of a private deployment model, which adds operational overhead, but it's the trade-off I'm willing to make. The other thing I'm wrestling with is governance velocity. The security use cases where AI actually seems useful — log analysis, anomaly correlation, some of the triage work — those are also the cases where the data sensitivity is highest. So the approval cycles to actually get something into production are slow, and by the time we've gone through the review process, half the vendors we were evaluating have already pivoted their roadmap. That's a real friction point for us right now.

What's the most important thing you need to understand or solve here?

Right now, the thing I'm most focused on is figuring out where we can actually let AI touch security workflows without creating a bigger exposure than the one we're trying to close. That's not a philosophical question for us — it's very operational. The prompt data problem is real. When your security tooling is processing incident data, vulnerability details, internal config information — and that's going into a third-party model — you have to have a clear answer for where that data lives and who can see it. We don't have a clean answer on that yet for all our use cases, and that's the primary blocker right now.

What does 'good' look like to you — and how far are you from that today?

For us, "good" looks like AI-augmented security workflows where the human stays in the decision loop for anything consequential, but the routine triage and enrichment work is largely automated. Good also means I can audit every action an AI component took — full explainability, full logging — and I'm not depending on a vendor's black box telling me "trust us, it worked." How far are we from that? Closer on the detection and triage side than on the response side. We've got some reasonable tooling for enriching alerts and correlating events. Where we're still pretty far is agentic behavior — anything where the AI is taking action rather than just generating a recommendation. The governance and audit trail requirements for that aren't mature in the products I've evaluated, and frankly our own internal approval workflows aren't designed for AI-initiated actions yet. The other gap is data residency. A lot of the compelling AI security capabilities are cloud-delivered, and I'm not fully comfortable with what's leaving our environment to power those inferences. That's not a blocker today, but it's a constraint on how aggressively we adopt.

What would change your perspective on this entirely?

That's a fair question. If I saw a vendor actually demonstrate — not just claim — that their agentic system could operate within our existing access control boundaries without requiring us to over-provision permissions, that would shift my thinking considerably. Right now the ask is usually "give the agent broad access and trust us," and that's a hard no in our environment. The other thing would be better auditability. If I could see a clean, verifiable log of exactly what an agent did, what it queried, what it touched, and why — structured enough that my compliance team could actually review it — that would move the needle. Right now most of what I've seen is more like "here's a summary of what happened," which isn't the same thing.

What question are you not being asked that you wish someone would ask?

That's a tough one to answer on the spot. Let me think for a second. I think the question I don't get asked enough is: "What does your AI governance actually look like in practice, not in policy?" Everyone wants to talk about what AI tools you're evaluating or what use cases you're excited about. But nobody asks how you're actually operationalizing the guardrails — who owns the decision when an AI-assisted workflow touches customer data, what your review process looks like for prompt content leaving your environment, how you've structured your acceptable use boundaries. We spent a lot of time getting that right internally before we deployed anything meaningful. And when I talk to peers, most of them have a policy document somewhere but the actual enforcement mechanisms are pretty thin. That gap between the written policy and the runtime behavior is where I think a lot of organizations are going to have problems.

Tell me what's top of mind for you on this topic right now — what are you wrestling with?

Right now the thing I keep coming back to is data residency and what actually leaves our environment when we use any of these AI-assisted security tools. Every vendor we talk to says "oh, we're secure, your data stays isolated" — but when you actually dig into the architecture, a lot of them are just wrapping OpenAI or another third-party LLM behind their UI. So we're essentially asking: are we comfortable with our threat telemetry, our vulnerability data, our incident context flowing through someone else's infrastructure? The second thing is more of an organizational challenge. We've got pressure from leadership to show AI progress, but my security team is asking reasonable questions about governance — who approves what these tools can do, what decisions can they make autonomously versus flag for a human. We don't have a clean answer to that yet, and I don't want to just deploy something because it demos well.

What's the most important thing you need to understand or solve here?

For us, the core problem right now is figuring out where AI actually sits in the trust boundary. We're a SaaS company, so we're handling customer data, and the moment you start routing any of that through a third-party model — even with enterprise agreements and data processing addendums — you've introduced a dependency you can't fully audit. That's the thing I keep coming back to. It's not that I don't believe the vendors when they say they're secure, it's that I can't verify it at the depth I need to. So the question we're trying to answer is: where do we run models ourselves, where do we accept the vendor risk, and how do we draw that line consistently across teams — because right now every team is making that call independently.

What does 'good' look like to you — and how far are you from that today?

For us, "good" looks like AI that's actually integrated into the security workflow at the API level — not bolted-on chat interfaces where analysts are copy-pasting findings into a separate tool. It means the system can pull context from our SIEM, correlate against our asset inventory, and surface something actionable without a human having to stitch it all together manually. And good also means we know exactly where the data goes. Right now I have real concerns about prompts containing environment-specific information — IP ranges, internal tooling names, config details — leaving our perimeter and hitting a third-party model endpoint. We don't have a clean answer for that yet. We're evaluating private deployment options but the compute cost and operational overhead aren't trivial. So where are we today? Probably 40% of the way there on the workflow integration side, and further back than that on the data governance piece. The tooling is maturing, but the controls around it aren't keeping pace.

What would change your perspective on this entirely?

For us, the thing that would probably shift my thinking the most is seeing a real production deployment — not a demo, not a case study written by the vendor's marketing team — where an agentic security workflow ran autonomously for six-plus months and the audit trail held up under scrutiny. Not "it caught X threats," but "here's how it behaved when it got something wrong, here's how it escalated, here's the rollback." Right now the governance story around autonomous action in security is pretty thin. If someone could show me a credible framework where the blast radius of a bad agentic decision is actually contained — not just described as contained — that would move me. The data tenancy and prompt routing questions also need cleaner answers before I'd let something with real permissions operate in our environment.

What question are you not being asked that you wish someone would ask?

That's an interesting framing. Let me think for a second. I'd probably want someone to ask: "How do you actually govern AI tool sprawl across your security org when every vendor is claiming their product is AI-native?" Because right now, we have vendors coming in saying their product is "built with AI at its core" — and when you push them on what that actually means technically, it often just means they've wired an API to an LLM and slapped a chat interface on top. My team has to sort through that, and there's no clean framework for it yet. The governance question underneath that is: how do I know which of these tools is actually sending my security telemetry or incident data to a third-party model endpoint, and what are the data retention policies on that? That's the conversation I want to have, and most interviewers are still asking me about use cases rather than the control layer around those use cases.

Tell me what's top of mind for you on this topic right now — what are you wrestling with?

Right now the thing I keep coming back to is governance lag. We're moving fast on the product side — piloting AI tools for fraud detection workflows, some anomaly flagging — but the policy and oversight layer is just not keeping pace. Like, engineering wants to ship, leadership sees the competitive pressure, and I'm sitting in the middle asking "okay, but who owns the decision when the model flags something incorrectly and we act on it?" The data sensitivity piece is also real for us. We're fintech, so we're dealing with transaction data, user financial behavior — and every time someone wants to integrate a new AI tool, there's this legitimate question of what's actually leaving our environment and where it's going. We haven't fully resolved the self-hosted versus third-party API question yet. Those two things — accountability gaps and data boundary clarity — are what I'm actively wrestling with right now.

What's the most important thing you need to understand or solve here?

For us right now, the biggest question is really around where AI can actually be trusted to act versus where it still needs a human in the loop. We're starting to explore some AI-assisted tooling on the security and compliance side — things like anomaly detection, automated policy checks — and the governance question comes up immediately. It's less about whether the technology works and more about: who owns the decision when an AI flags something or, eventually, takes an action? That accountability gap is what we keep circling back to internally.

What does 'good' look like to you — and how far are you from that today?

For us, "good" looks like AI that's embedded in actual workflows rather than sitting as a separate tool people have to remember to use. So concretely — a security analyst getting a relevant, contextualized recommendation at the moment they're triaging something, not going off to a chat interface and prompting manually. The other piece is auditability. In fintech, we can't just say "the AI flagged it." We need to show regulators and our compliance team why a decision was made and what data touched it along the way. That's the bar. How far are we? Somewhere in the middle. We've deployed a few AI-enabled tools — some are process-specific, some are more general-purpose — and the process-specific ones are actually delivering. But we're still stitching things together, and the governance layer is honestly behind the tooling. We know what we want the workflow to look like; we haven't fully solved for who owns the output when something goes wrong.

What would change your perspective on this entirely?

That's a good question to sit with. I think if I saw a well-documented case — not a vendor case study, but something peer-reviewed or at least written up by the security team themselves — where an agentic system made a consequential decision autonomously and it went well repeatedly, not just once, that would shift my thinking on readiness timelines. Right now the gap for me is the lack of institutional memory around failures. We're all building these governance frameworks somewhat blind because there isn't a lot of honest post-mortem sharing about where AI in security workflows actually broke down. If that started circulating more openly, I'd have better signal to calibrate against.

What question are you not being asked that you wish someone would ask?

The one I'd actually want someone to dig into is: how do you build trust incrementally with engineering teams around AI security tooling, rather than just mandating adoption from the top down? A lot of the conversation I see is about governance frameworks and compliance checklists, but the real friction in my experience is that engineers have legitimate concerns about how these tools work, what data they're touching, whether the outputs are reliable — and if you don't address that at the team level, you end up with shadow usage or surface-level compliance. Neither is great. I don't see enough research asking how security orgs are actually managing that internal trust-building process.

Tell me what's top of mind for you on this topic right now — what are you wrestling with?

Right now the thing I keep coming back to is the gap between what our exec team *thinks* AI can do for security workflows and what our security and engineering teams actually feel comfortable deploying. The pressure from leadership is very real — everything needs to have an AI story — but when you start asking the harder questions about data handling, model access controls, and what happens when an automated decision is wrong, the conversations get complicated fast. For us specifically, we're in fintech, so the compliance surface is already significant. Any time we talk about routing sensitive data through a third-party AI service, someone raises a legitimate concern about whether that data is being used for training, how it's stored, who can see it. We haven't fully resolved that. We're somewhere in the middle — trying to find a configuration that gives us actual capability without just hoping the vendor's terms of service are airtight. The other thing I'm wrestling with is ownership. Like, if an AI-assisted tool flags something or takes an action in a security context, who's accountable for that? That question doesn't have a clean answer in our org right now, and I think we need to sort it out before we go much further.

What's the most important thing you need to understand or solve here?

For us, the biggest unsolved question right now is where human judgment stays in the loop versus where we can actually let automation run. We're a fintech, so there's always a compliance angle — but even setting that aside, I genuinely don't know yet which security decisions are safe to hand off to an AI agent and which ones need a person signing off. That boundary isn't defined for us. The second piece is data boundaries. We're already having internal debates about what's okay to send to third-party AI services — supplier data, customer PII, transaction patterns — and the guidance from leadership is inconsistent. We need a clearer framework there before we can scale any of this responsibly.

What does 'good' look like to you — and how far are you from that today?

For us, "good" looks like AI being embedded in actual workflows rather than sitting as a separate tool people have to remember to open. Like, the insight surfaces where the decision is being made — not in some dashboard that requires a separate login and context switch. How far are we? Honestly somewhere in the middle. We've got a few AI-assisted features in production that are genuinely useful, but the governance layer is still pretty manual. We're still having conversations like "okay, what data can actually flow into this model, and who signed off on that?" every time we want to expand scope. That slows iteration down more than the engineering work does. The other gap is that our security workflows in particular are still pretty human-in-the-loop heavy — which isn't necessarily wrong, but the criteria for when to trust an automated output versus flag for review aren't consistently defined yet. We've got it in some places and it's ad hoc in others. That's probably the thing I'd fix first if I could.

What would change your perspective on this entirely?

That's a good question to sit with. I think the thing that would shift my view most is seeing a real track record of agentic AI operating in a high-stakes security context without requiring constant human intervention to catch mistakes. Right now a lot of what I see is demos and pilots — and the gap between a controlled demo and production behavior at scale is where I get skeptical. If I saw two or three years of documented outcomes — not just efficiency claims, but actual error rates, escalation rates, what happened when the agent got it wrong — that would move me. The governance piece would follow from that evidence naturally.

What question are you not being asked that you wish someone would ask?

That's a hard one to answer off the top of my head. Let me think for a second. Maybe... "How does your governance process actually slow down adoption, and is that always a bad thing?" Because in fintech, we have compliance and regulatory overhead that forces us to be deliberate before we deploy anything. And sometimes that friction is genuinely useful — it surfaces data handling questions we'd otherwise skip past. Most conversations I'm in frame governance as purely a blocker. But sometimes the compliance review catches something real, like a vendor's data retention policy that doesn't align with what we've committed to customers. I'd rather have that conversation before we've already embedded a tool into a workflow. I don't have a strong broader view on how other security orgs are navigating this, but from where I sit, the governance process isn't just overhead — it's occasionally doing actual work.

Tell me what's top of mind for you on this topic right now — what are you wrestling with?

Right now the biggest thing I'm wrestling with is that every tool in my stack has some form of AI baked in — HubSpot, 6sense, whatever — but I genuinely can't tell you which of those AI features are actually moving the needle versus just being a checkbox the vendor added to justify a price increase. That's the attribution problem compounded. It's not just "which channel drove the pipeline," it's "which AI-assisted action within that channel did anything useful." The other thing is that every piece of software I evaluate now leads with AI in the pitch, and most of them can't explain the actual business impact in concrete terms. I'll ask a rep "what does this do differently because of AI" and I get circular answers. So I'm skeptical by default now, which probably means I'm also slower to adopt things that might actually matter. On the security side — which I know is more your focus — I'm a buyer, not a practitioner, so I see it from the vendor relationship angle. We're being asked to send more data through more AI-powered tools and our legal and IT teams are starting to push back on what's going in, where it's stored, who owns it. That conversation is happening more frequently and it's starting to slow down procurement cycles for us.

What's the most important thing you need to understand or solve here?

Right now my biggest problem is pipeline efficiency — specifically, knowing which channels and campaigns are actually driving revenue, not just MQLs. Attribution is a mess when you've got six or seven touchpoints before a deal closes, and a lot of the AI-assisted tools I'm using either don't connect back to closed-won data or require sales to actually log things consistently, which doesn't happen. The feedback loop between marketing and sales is where everything breaks down for us. If I don't know which leads converted, I can't train anything — AI-powered or otherwise — to get smarter about what good looks like. So before I can even think about layering in more sophisticated tooling, I need cleaner data hygiene and a reliable handoff process.

What does 'good' look like to you — and how far are you from that today?

For us, "good" is having a clear, defensible picture of what's actually driving pipeline — not just first-touch or last-touch, but something I can stand behind in a board conversation. And then having the ability to act on that quickly enough that it actually matters. Right now, we're a decent ways off. The feedback loop between marketing and sales is probably our biggest gap — we're generating leads, scoring them, doing the enrichment stuff, but the closed-won data coming back from sales is inconsistent. So whatever scoring model we're running, it's not learning what "good" actually looks like in practice. That's frustrating because it's not really a tool problem, it's a process and alignment problem. The AI tooling we have is embedded in the platforms we already use — HubSpot, some ad channel stuff — so it's not like we're missing access. It's more that the underlying data hygiene has to be right first, and we're still cleaning that up.

What would change your perspective on this entirely?

If I actually saw a clean, end-to-end example of an agentic workflow in a security context that didn't require a dedicated engineer to babysit it — that would move me. Right now every demo I've seen has a person somewhere in the loop doing the thing that supposedly makes it "autonomous." The other thing would be better attribution on outcomes. I'm wired to think in pipeline and CAC terms — so I apply that same scrutiny to tooling. If someone could show me: here's the threat we caught, here's what it would have cost if it landed, here's what the tool cost — that kind of closed-loop accounting would genuinely shift how I think about the investment case for agentic security tools.

What question are you not being asked that you wish someone would ask?

That's a good one to end on. I'd say... nobody really asks about what happens to your demand gen motion when your ICP starts using AI agents to do their own research and evaluation. Like, we're still optimizing for the human buyer journey — content touches, intent signals, nurture sequences — but if the person I'm trying to reach is delegating initial vendor research to an AI assistant, my whole attribution model breaks. My SEO strategy breaks. My nurture cadence becomes irrelevant. I don't have a clean answer to it. But it feels like the thing that could fundamentally change how demand gen works, and I'm not seeing a lot of people in my position talking about it seriously yet.

Tell me what's top of mind for you on this topic right now — what are you wrestling with?

Right now the thing taking up the most mental space for me is attribution, honestly in a different context than security — but it bleeds into everything. We're a lean team, I'm running demand gen, and every tool we add claims to have AI baked in. The problem is half of them can't actually tell me what the AI is doing differently than what existed before. On the security side specifically — which I know is the framing here — we've got tools in our stack that flag things, generate summaries, surface recommendations, and I'm genuinely unsure how much of our security posture is actually improved versus how much we've just added another dashboard. That's the real tension for me: governance around these tools is pretty immature internally. Nobody's formally signed off on what data goes where, what prompts are being sent to third-party services, what's in scope for AI to act on versus just surface. The agentic piece is interesting but honestly feels a step ahead of where we are. We haven't fully figured out how to govern the AI tools we already have, so the idea of autonomous action in a security context is a little abstract for us right now.

What's the most important thing you need to understand or solve here?

Right now my biggest problem is understanding which channels are actually driving pipeline — not just first-touch or last-touch, but the full picture. Attribution is genuinely messy for us. A prospect might see a LinkedIn ad, come to a webinar, get a sequence from SDR, and then convert — and depending on which model I look at, a completely different channel gets credit. That bleeds directly into CAC conversations, which bleed into budget decisions. If I can't confidently say "this channel is working at this efficiency," I'm basically making bets with incomplete data.

What does 'good' look like to you — and how far are you from that today?

For us, "good" means knowing with reasonable confidence which channels and motions are actually driving pipeline — not just first-touch or last-touch attribution, but something closer to the full picture. And then being able to act on that fast enough that it actually matters within a given quarter. Right now we're pretty far from that. We've got data living in HubSpot, some ad platforms, a few enrichment tools, and they don't talk to each other cleanly. The feedback loop from sales back to marketing on which leads actually closed is basically nonexistent — sales isn't logging it consistently, so I'm optimizing off incomplete signal. That's probably the single biggest gap between where we are and where "good" would be. The AI layer on top of all that feels premature until the data hygiene problem is solved. I can bolt on a scoring model or intent tool, but if the underlying data is messy, I'm just getting confident wrong answers faster.

What would change your perspective on this entirely?

If I actually saw a clear, auditable trail from an AI-generated action back to a business outcome — not just "the model flagged this" but here's what it caught, here's what it prevented, here's what it would have cost us — that would move the needle for me. Right now a lot of what I see is directionally interesting but hard to tie to anything concrete. The other thing would be better feedback loops. In my world, the classic failure mode is sales never telling marketing which leads actually closed, so the scoring model never learns. I suspect security AI has a version of that same problem — the system acts on something, but there's no clean signal about whether that action was right or wrong, so you're not actually improving over time. If someone showed me a workflow where that loop was genuinely closed, I'd take the governance and autonomy conversation a lot more seriously.

What question are you not being asked that you wish someone would ask?

That's a good one to end on. I'd probably say: nobody asks about the feedback loop between sales and marketing when it comes to AI-generated pipeline signals. We're using AI tools to score leads, predict intent, surface accounts — but if sales isn't closing the loop on which of those actually converted, the model never learns what "good" looks like. The AI just keeps optimizing toward the wrong thing. We've run into this ourselves. The tooling is fine, but the process around it is broken, and that's not a technology problem — it's a coordination problem that AI can't fix on its own.

Tell me what's top of mind for you on this topic right now — what are you wrestling with?

Right now the thing I keep coming back to is the gap between what vendors are promising around AI-driven security and what we can actually verify in terms of outcomes. We're a Series B company, so every dollar matters, and I'm getting pitched constantly on tools where the "AI" differentiation is basically circular — it's built with AI, what does that mean, well it's AI-native. That loop gets old fast. The more practical tension for me is around data governance. We handle customer data, and anytime we're evaluating a tool that's sending prompts or telemetry to a third-party model, I want to know exactly where that data goes and who else might be training on it. Our security team is cautious about that, reasonably so, and it creates real friction in the adoption process even when the tool itself looks promising. So we're caught between pressure from leadership to move faster on AI adoption and a procurement and security review process that — again, reasonably — wants answers that vendors aren't always prepared to give clearly.

What's the most important thing you need to understand or solve here?

Right now, for us, the biggest question is really about where AI actually moves the needle in security workflows versus where it's just layered on top of existing processes without changing much. We've deployed several tools that have AI components, and the honest challenge is figuring out which ones are actually compressing time or reducing risk, and which ones are just checking a box. The "built with AI at its core" pitches we get from vendors are almost impossible to evaluate because nobody can articulate what that means operationally. The governance piece is the other side of that. Our security team is reasonably sophisticated, but the moment you start talking about AI systems making or influencing decisions autonomously, the risk and compliance conversation gets complicated fast — and that's not a blocker I can just push through as a marketing leader.

What does 'good' look like to you — and how far are you from that today?

For us, "good" means AI that's actually embedded in workflow rather than sitting as a separate tool someone has to consciously decide to use. The difference between a capability people adopt because it's genuinely faster and one that requires a behavior change every time — that gap matters a lot to adoption rates. Right now we're probably somewhere in the middle. We've got AI baked into several tools we already use, which is the easier win — it's just there, people use it. The harder part is the more deliberate use cases where someone has to think "okay, now I'm going to use AI for this" — that's where we see inconsistency. The thing I'd add is that "good" also means I can actually measure what it's changing. I don't have a strong view yet on what the right metrics are for some of these workflows — that's honestly one of the more unsettled parts for us.

What would change your perspective on this entirely?

That's a fair question. I think the thing that would genuinely shift my view is seeing a well-documented case study where an agentic security workflow ran autonomously — made decisions, took remediation actions — and the outcomes were auditable and defensible after the fact. Not a vendor deck, but like, a real post-mortem where you can trace exactly what the AI did and why, and it holds up to scrutiny. Right now the governance piece is what I keep coming back to. If someone showed me that the accountability chain actually works — that when something goes wrong there's a clear record and a clear owner — I'd be a lot more open to pushing further into agentic territory. The compliance and risk management side of this isn't glamorous, but it's what actually determines whether we can deploy something broadly or keep it siloed as a pilot.

What question are you not being asked that you wish someone would ask?

That's a good one to end on. I think the question that doesn't come up enough is: "How do you actually get security and marketing — or security and any business function — to agree on what 'good enough' governance looks like before you deploy something?" Everyone's asking about adoption rates or which AI tools teams are using, but the harder problem is the internal alignment piece. In my experience, security wants air-tight controls, finance wants ROI justification, and the business unit just wants to ship. Nobody's asking how those three groups actually reach a workable decision together — and that bottleneck is probably slowing AI adoption more than any technical limitation.

Tell me what's top of mind for you on this topic right now — what are you wrestling with?

Right now the thing I keep coming back to is the gap between what our security vendors are claiming AI does for them versus what I can actually verify as a buyer. Every pitch deck says "AI-powered" or "AI at its core," and when you push on what that means mechanically — like what's the actual decision being automated, what's the error rate, what happens when it's wrong — the answers get vague fast. From a marketing perspective, I'm also dealing with the internal governance question more than I expected to. We've got teams wanting to pipe customer data and competitive intel into third-party AI tools, and the security org is rightly pushing back on that. There's no clean policy yet, and the business pressure to just move faster keeps running into "wait, what data is actually leaving our environment?" That tension is real and unresolved for us.

What's the most important thing you need to understand or solve here?

For us right now, the core question is really about where AI actually moves the needle in security workflows versus where it's just adding another layer of complexity we have to manage. We've got tools across the stack that all claim AI capabilities, and half the time it's hard to tell what's genuinely differentiated versus what's just a feature someone bolted on to update the pitch deck. The governance piece is probably where I'd focus most — who owns decisions about what data flows into these systems, what the approval process looks like, and how you avoid the situation where a well-meaning team deploys something that creates exposure you didn't sign off on. That's the unsexy part that doesn't get enough attention relative to the capability conversations.

What does 'good' look like to you — and how far are you from that today?

For us, "good" means AI is actually embedded in the workflow, not sitting as a separate tool that someone has to remember to open. Like, the insights surface where the work is happening — in the platforms we're already using — and there's a clear line from AI output to a decision or action we can measure. How far are we from that? Closer on some things than others. The tools we've layered AI into that are already part of our stack — that's working reasonably well. Where we fall short is the more fragmented stuff, where someone's running a prompt in a separate window and copy-pasting results back. That's not really integration, that's just a fancier search engine. Getting from that to something that's actually embedded and auditable is where most of the remaining work sits.

What would change your perspective on this entirely?

That's a fair question. I think the thing that would move me most is seeing a clear, traceable line from an AI security capability to a business outcome that I can verify — not a case study written by the vendor, but something like a peer at a comparable company saying "here's what we measured before and after, here's the methodology." Right now a lot of what I see is directionally compelling but not rigorous enough for me to stake budget on. The other thing that would shift my view on agentic security specifically is seeing governance frameworks that actually match the autonomy level of the tool. If someone showed me a credible model for how you scope what an agent can and can't do, with real audit trails, I'd take the category a lot more seriously. Because right now my concern isn't whether the AI is capable — it's whether we can actually oversee it in a way that doesn't create more exposure than it solves.

What question are you not being asked that you wish someone would ask?

The governance gap between what security teams are allowed to do with AI and what they're actually doing in practice. There's a real shadow AI problem brewing where individual analysts are running sensitive data through whatever tool is fastest, and the official policy hasn't caught up. That's the conversation I think gets skipped because everyone wants to talk about capabilities and use cases, not the messy compliance reality underneath.