"How are engineering organizations adopting AI within observability workflows, what is the maturity gap, and what is blocking scale?"

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

Honestly, the thing I keep coming back to is the signal-to-noise problem. We've got observability tooling that's generating enormous amounts of data — traces, logs, metrics — and the promise of AI is that it helps us make sense of that faster. But right now we're at this awkward middle stage where vendors are slapping "AI-powered" on everything without being able to clearly explain what's actually different under the hood. I've had more than one conversation that felt like that loop where someone just keeps saying "built with AI at its core" without being able to tell me what that means for my on-call engineer at 2am. The other thing I'm wrestling with is the integration layer. We're API-heavy, we care a lot about how things connect, and a lot of the AI observability features I'm seeing are fairly locked into their own ecosystem. That creates real tension for us because we've already got consolidation fatigue from too many point solutions. So the question isn't just "does this AI feature work," it's "does it work *within* the way we've already wired things together." Those are pretty different questions and most sales conversations don't get there.

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

The core problem for us right now is signal-to-noise. We're generating enormous amounts of observability data — traces, logs, metrics — and the AI tooling that's been bolted on top of it mostly just surfaces more of the same noise faster. What I actually need is something that can tell me *why* something happened, not just that it happened. The secondary piece is integration. I've got five or six observability and monitoring tools already in the stack, and every vendor is now claiming AI is "built at their core" — and when you push them on what that actually means technically, it usually means they added an LLM wrapper to their existing query interface. That's not solving the problem. So the real question for me is: can AI meaningfully reduce the cognitive load on my on-call engineers, or are we just moving the alert fatigue from a dashboard to a chat interface? That's what I'm trying to figure out.

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

Good, to me, looks like a closed-loop system where anomaly detection surfaces something, there's enough context automatically assembled — traces, logs, recent deploys, dependency state — that an engineer doesn't have to spend the first 45 minutes just figuring out what they're even looking at. And ideally the system can at least draft a hypothesis about root cause, not just alert. Where are we today? Probably 40% of the way there. We've got decent instrumentation with OpenTelemetry, we're using Datadog for aggregation, and we've wired in some of their AI-adjacent features. But the "automatically assembled context" part is still pretty manual in practice — someone still has to go pull the deployment history, check who merged what, cross-reference the runbook. The correlation is better than it was two years ago, but it's not good enough that I'd trust it without a human in the loop for anything customer-impacting. The bigger gap for us honestly isn't the detection side, it's the signal-to-noise problem upstream. Our alert volume is still too high, which means even when something smart surfaces, it's competing with a lot of noise and people have trained themselves to be skeptical. Fixing that is partly a tooling problem and partly a discipline problem with how we've configured things over time.

What would change your perspective on this entirely?

That's a good question to sit with for a second. Honestly, what would move the needle most for me is seeing a well-documented case where AI-driven anomaly detection actually caught something that a traditional threshold-based alert *wouldn't* have caught — and where the signal-to-noise ratio was genuinely better, not just differently distributed. Right now a lot of what I see is vendors replacing one kind of alert fatigue with another. The other thing that would shift my thinking is better answers around data residency and model auditability. We're in a regulated-adjacent space and when I ask vendors "where does my telemetry go when you're training or fine-tuning your models," I get hand-wavy responses. If that gets cleaner — real contractual clarity, not just a trust-us blog post — that removes a meaningful blocker for us. And frankly, if I saw a credible path to integrating these AI observability features deeply through APIs rather than forcing us into another monolithic platform, I'd be a lot more open to the conversation. The "built with AI at its core" pitch I keep hearing doesn't tell me anything concrete about how it actually fits into what we've already built.

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

Honestly, I don't have a burning "hidden question" I'm dying to answer. But if I were to flag something that comes up in my day-to-day that doesn't get enough airtime in these conversations — it's probably around the data pipeline problem underneath AI observability. Everyone wants to talk about the AI layer, the inference, the anomaly detection. Nobody asks about whether your telemetry data is actually clean and structured enough for any of that to work. In my experience that's where a lot of these initiatives quietly stall — not because the AI is bad, but because the underlying data is a mess. That feels like the more honest conversation to have.

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

Honestly, the thing I keep coming back to is signal-to-noise. We've got observability tooling that generates enormous amounts of data — traces, logs, metrics — and the promise of AI is that it collapses that into something actionable. But what I'm actually seeing in practice is that we've layered AI features on top of already noisy systems, and now I have to evaluate whether the AI summarization is trustworthy or whether it's just laundering bad data with a confident tone. The second thing is vendor consolidation fatigue. Every tool in our stack — Datadog, whatever else — has suddenly "baked in AI" and I'm getting the same circular answers about what that actually means technically. Someone told me their product was "purpose-built for AI" and when I pushed on the specifics, it was basically a wrapper on an LLM with access to my telemetry. That's not nothing, but it's also not the transformation they're pitching. So practically, I'm trying to figure out where the actual leverage is — is it anomaly detection, is it incident summarization, is it something closer to root cause correlation — and whether I should be building lightweight internal tooling against our own APIs or buying something that genuinely does the hard work. I don't have a clear answer yet, which is part of why I find this space interesting right now.

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

The core problem for me right now is signal-to-noise at scale. We're generating an enormous volume of observability data — logs, traces, metrics, spans — and the question is whether AI can actually help my team cut through that to find what matters, versus just adding another layer of tooling that my engineers have to babysit. The secondary thing I'm trying to figure out is build versus buy. There are a lot of vendors right now slapping "AI-powered" on their observability products, and when I push them on what that actually means technically, I often get circular answers — it's AI at its core, it's purpose-built for AI. That doesn't tell me anything about whether it's a fine-tuned model on my telemetry data, a wrapper on an LLM, or just anomaly detection they've been doing for five years with a new label. So the real problem is: I need enough clarity to make a defensible architectural decision, and right now the market is noisy enough that I'm skeptical of most of what I'm hearing.

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

Good, to me, looks like AI that's actually embedded in the signal-to-noise problem — where my on-call engineers aren't drowning in alert storms at 2am, and the system has enough context to say "here are the three things that actually matter right now and here's why." Not just anomaly detection, but correlated causality across logs, traces, and metrics with enough explanation that a junior engineer can act on it without escalating. How far are we from that today? Honestly, further than the vendors want us to believe. We have pieces — we're using some AI-assisted alerting, we've got Datadog doing some of the correlation work — but it still feels like we're stitching together four different products that each have their own AI layer, and none of them are talking to each other in a meaningful way. The integration tax is real. The thing that's missing most is context persistence. The AI sees a spike, flags it, but it doesn't know that we just shipped a feature to 10% of users or that there's a known flakiness issue in one of our upstream dependencies. Until we solve the "AI needs to know what we know" problem, it's going to keep feeling like a smart tool that doesn't understand our system. That gap is probably 18 to 24 months away from being genuinely closed, and that's optimistic.

What would change your perspective on this entirely?

Honestly, the thing that would move the needle most for me is seeing reliable, consistent signal reduction that I can actually verify myself — not a vendor demo, not a cherry-picked case study. If an AI layer in our observability stack could demonstrably show me "here are the 200 alerts we suppressed this week, here's why, and here's the audit trail proving none of them were false negatives" — that would shift my posture significantly. The other thing is data residency and model transparency. Right now a lot of these tools are essentially black boxes sitting on top of my telemetry data, and I don't have great visibility into what's being sent where. If a vendor could genuinely solve the trust and auditability problem — not just claim they do — I'd be a lot more willing to go deeper. Beyond that, I don't have a strong view on what else would fundamentally change my perspective. Those two things are the blockers I keep coming back to.

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

Honestly, the question I'd want someone to ask is: "How do you actually trust the output?" Everyone's focused on adoption rates and feature sets, but nobody's drilling into the verification layer. When an AI surface in your observability stack flags an anomaly or suggests a root cause, what's your process for validating that before an on-call engineer acts on it? That's where I see the real maturity gap — not in whether teams have deployed these tools, but in whether they've built any systematic way to audit the AI's reasoning against ground truth. The security-first piece of my brain keeps coming back to this. You're essentially letting a model make inferences about your production environment, and if you can't interrogate that chain of reasoning, you've just added another black box on top of an already complex system.

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

Yeah, so the thing I keep coming back to is the gap between what AI tooling promises in the observability space and what actually gets operationalized day-to-day. Like, we hear a lot about anomaly detection, intelligent alerting, that kind of thing — but when I talk to our engineering team, the reality is most of them are still triaging alerts manually and just using AI as a fancier search layer on top of logs. What I'm genuinely wrestling with right now is the people and process side of it more than the tooling. We could probably adopt better AI-assisted observability tomorrow if we had clearer ownership of who acts on what the AI surfaces. The technology isn't really the bottleneck — it's that the workflow around it hasn't been redesigned to account for it. And then there's a trust dimension. Engineers are pretty skeptical by nature, and if an AI-generated insight turns out to be wrong a few times, adoption stalls. So I'm thinking a lot about how you build confidence incrementally — like, what's the MVP of AI in your observability workflow that actually demonstrates value before you ask people to rely on it more heavily.

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

From my vantage point as a PM, the most pressing question is really about the gap between what AI tooling *promises* in observability and what engineering teams can actually operationalize day-to-day. Like, there's a lot of vendor noise around "AI-powered" everything, but when you dig in, teams are still doing a lot of manual triage and the AI features are more bolt-on than genuinely embedded in the workflow. The thing I want to understand better is where the real blockers are — is it data quality, is it trust in the outputs, is it just that the tooling isn't mature enough yet? Because those have pretty different solutions. In my experience, the people and process side tends to be underestimated relative to the tooling side, and I suspect that's true here too.

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

Good, to me, looks like AI being embedded in the observability workflow in a way where it's actually closing the loop — not just surfacing anomalies but helping the team understand root cause faster and feeding that learning back into how we build. Like, the system gets smarter over time because of how it's being used, not just because someone retrained a model. Where we are today is pretty far from that. Right now what I see is more like AI-assisted alerting with a lot of human babysitting on top. Engineers still have to do a ton of manual correlation work, and honestly the tooling is fragmented enough that you end up with noise in one place and signal buried somewhere else. The people side is also real — even when tools are good, getting the team to actually trust and act on AI recommendations requires more change management than anyone budgets for. The gap I'd name is less about the technology ceiling and more about process maturity and data quality. The AI is only as useful as the observability data you're feeding it, and a lot of orgs, including mine at times, haven't done the unglamorous work of cleaning that up first. So the "AI observability" story can sound really compelling in a demo and then land pretty flat in practice.

What would change your perspective on this entirely?

Honestly, the thing that would shift my perspective the most is seeing a clear before-and-after on alert fatigue reduction that's actually attributable to AI — not just a vendor's case study, but something from a team I trust where they can walk me through the methodology. Right now a lot of what I hear is "we deployed this AI-powered observability tool and things got better," but there's so much confounding in that — team maturity, better on-call practices, infra changes happening simultaneously. The other thing would be evidence that the people side got solved, not just the tooling side. Most of what I've seen is organizations spending the majority of their energy on the tech stack and then 12-18 months in realizing the adoption is shallow because the processes and communication norms didn't change around it. If someone showed me a genuine culture change story alongside the tooling story, that would move me.

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

Honestly, I don't have a burning "question nobody asks" moment here. But if I were to flag something that comes up less than it should... I'd say the question around *who owns the output* when AI surfaces an anomaly or an alert in an observability context. Like, we talk a lot about whether the AI is accurate, whether it reduces noise — but in my experience working cross-functionally, the harder problem is organizational. Who acts on it? What's the escalation path? That's a people and process question, not a tooling question. In our world, I've seen teams get pretty excited about a new tool and then 12-18 months in realize the workflow around it was never actually defined. The tool works fine but the humans didn't change their behavior to match it. That feels relevant to observability with AI layered on top — the detection might get better, but if the incident response process is still ad hoc, you haven't actually moved the needle much. I don't have a sharp take on how to solve it, but I think that accountability and workflow design question gets underweighted compared to "is the AI good enough."

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

Yeah, so the thing I keep coming back to is the gap between what AI in observability promises versus what actually gets adopted at scale. Like, we have some tooling in place — we use a couple of AI-assisted alerting and anomaly detection features — but honestly the utilization is pretty uneven across the team. What I'm wrestling with right now is less about the technology itself and more about the workflow integration piece. It's kind of like what I see in product work generally — you can have the right tool, but if it doesn't fit into how people actually work day to day, adoption stalls. Our engineers are skeptical of AI-generated insights that they can't trace back to something concrete. There's a trust gap there. And then from my PM perspective, I'm also dealing with the tension between leadership wanting us to lean into AI more aggressively — like, "why aren't we using AI everywhere" — and the engineering team rightfully pushing back on questions around reliability and signal quality. The noise-to-signal ratio in alerting is already a pain point without AI making it harder to understand why something fired. So I don't have a clean answer yet. It's more like we're in this middle stage where we've got some AI features turned on but we haven't really closed the loop on whether they're actually making our engineers faster or more effective. That measurement piece is still pretty underdeveloped on our end.

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

From a PM perspective, the thing I keep coming back to is: where does AI actually reduce friction in the observability workflow versus where is it just adding a layer of complexity? Like, the tooling conversation tends to dominate, but in my experience the harder problem is almost always the process and the people side. Specifically for observability — the question I'd want answered is whether teams are actually changing how they respond to incidents and alerts because of AI assistance, or whether they've just bolted an AI feature onto an existing workflow and called it done. Because those are really different maturity levels, and I suspect most orgs are closer to the latter.

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

Good, to me, looks like AI that's actually integrated into the signal-to-noise problem in observability — like, you've got so much telemetry data coming in that the real issue is surfacing what actually matters before an engineer even has to look at it. So "good" would be anomaly detection that's contextual, alert fatigue reduction that learns from how your team actually responds, and maybe some root cause suggestions that are grounded in your specific system topology, not just generic patterns. Where we are today? Honestly, pretty far from that in most orgs I've seen or talked to. A lot of what gets called "AI in observability" right now is still pretty shallow — it's a copilot on top of dashboards, or it's basic threshold-based alerting dressed up with ML language in the marketing copy. There's a gap between the demo and what's actually in production. From my side specifically, we're a fintech startup, so we're dealing with the usual constraints — the engineering team is lean, and getting them to invest time in tuning an AI layer on top of observability is a hard sell when there are product features competing for that same bandwidth. So even if the tooling were further along, the organizational readiness question is real. It's people and process as much as it is the technology.

What would change your perspective on this entirely?

That's a good question. I think the thing that would shift my view the most is seeing reliable, auditable signal come out of AI-driven observability at scale — not just in a demo environment or a controlled pilot, but in production at a company with real compliance constraints. Right now a lot of what I see is "it works great in the sandbox," and then you put it in front of a security or compliance review and things get complicated fast. The other thing — and this is more of a process concern than a technology concern — is if I saw organizations actually close the loop between AI-surfaced insights and team behavior. The tooling side is getting better, but most places I talk to still have the humans ignoring or overriding the AI recommendations because trust hasn't been established. If someone showed me a case where that trust was genuinely earned over time, with clear feedback mechanisms, that would change how optimistic I am about the adoption curve. Right now I'd say we're better at building the tools than we are at building the workflows around them.

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

Honestly, I don't have a strong "secret question nobody's asking" take here. But one thing I do think gets underweighted in these conversations is the people and process side versus the tooling side. Like, everyone wants to talk about which AI observability vendor you're evaluating, but the harder question is whether your engineering org actually has the cultural alignment and workflows to absorb a new capability — because you can buy a great tool and 12-18 months later realize adoption is basically zero because nobody changed how they work. For us specifically, I'd love if someone asked more about how you validate that AI-generated insights in observability are actually trustworthy before engineers start acting on them automatically. We're a fintech, so there's a real compliance and risk dimension to that. It's not a blocker per se, but it's a question we haven't fully answered.

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

Honestly, the observability piece isn't my core domain — I'm on the demand gen side, so I'm not the one in the weeds with engineering tooling day to day. But where it does intersect with my world is attribution and pipeline data quality. Like, I'm constantly trying to understand what's actually happening across our funnel, and a lot of the underlying data infrastructure questions that engineering deals with in observability feel analogous to what I deal with on the marketing analytics side. What I'm wrestling with more broadly is that we have a lot of tools that all claim to have AI baked in now, and the question is always whether that's actually moving the needle or just feature bloat. My instinct is that most teams — including ours — are pretty early in terms of using AI in a systematic way versus just having it as a checkbox somewhere in the product. So I'd be curious what you're seeing on the engineering side, because I suspect the maturity gap you're describing is real.

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

Honestly, from where I sit — and I know this interview is focused on engineering and observability specifically — the thread that connects to my world is attribution and signal quality. Like, if observability tooling is generating AI-driven alerts or anomaly detection, someone downstream has to act on that signal. And if the signal-to-noise ratio is bad, you lose trust in the whole system fast. The parallel I'd draw from marketing ops is that we've deployed probably five or six AI-enabled tools in the last year, and the ones that actually stuck were the ones solving a specific, well-defined bottleneck — not the ones promising to transform everything. I'd imagine engineering teams are hitting the same wall with AI in observability: the generic "AI-powered" label gets slapped on everything, but the question of whether it's actually reducing mean time to resolution or just adding another dashboard nobody trusts — that's the real thing to understand. So I guess the core question I'd want answered is: where does AI in observability actually close a loop versus where does it just surface more noise that humans still have to sort through manually?

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

Good, in my world, is when I can actually trust the numbers I'm looking at — like, I know what's driving pipeline, I know which channels are working, and I'm not spending half my week reconciling data between platforms or chasing the engineering team to fix a broken event. Where we are today? Honestly, further than I'd like. We've got decent tooling — things are stitched together reasonably well — but the feedback loops are still pretty manual. Sales isn't reliably closing the loop back to marketing on what actually converted, so the scoring models we're using are kind of flying partially blind. That's the piece that frustrates me most. I don't have a strong view on the AI observability side specifically since that's a bit outside my direct scope, but from a general "is the data trustworthy and actionable" standpoint, I'd say we're maybe at a 5 or 6 out of 10. Good enough to operate, not good enough to scale confidently.

What would change your perspective on this entirely?

Honestly, probably seeing a really clean proof point from a company I actually respect — not a vendor case study, but like a peer at a similar-stage company saying "here's exactly what we did, here's what changed in our on-call workflow, here's the before and after." The other thing that would shift me is if the attribution and signal quality actually got better as a result of AI in observability. Right now I hear a lot of "AI-powered" everything, and half the time it's just a dashboard with a slightly smarter filter. If I saw it genuinely reducing noise to the point where eng teams were faster to root cause *and* that fed back into product reliability in ways marketing could actually point to — that would matter to me. But I'll be honest, I don't have a strong view on the deep technical side of this. My exposure is mostly where observability intersects with pipeline and uptime affecting customer experience. So my perspective is probably already pretty narrow.

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

Honestly, I don't have some profound "nobody's asking this" insight to offer here. I guess the thing that comes up a lot in my actual day-to-day is the feedback loop problem — like, how does AI-assisted observability or any AI tooling actually get smarter over time if the data going back into it is garbage? We talk a lot about deploying AI, but not enough about what it takes to maintain quality inputs so the system learns what "good" actually looks like. That feels relevant whether you're in engineering ops or demand gen. But I'll be honest, I'm not deep enough in the observability space specifically to claim that's a gap in *this* conversation. It might already be a well-worn topic for engineering teams.

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

Honestly, the thing I keep coming back to is attribution — but not in the way I usually think about it. In my role I'm constantly trying to figure out which channels are actually driving pipeline, and we use a handful of tools that are starting to bake AI into their reporting and scoring. The problem is I don't fully trust what they're telling me. Like, every piece of software I use has some version of "AI-powered insights" now, but when I dig into what's actually happening under the hood, it's not always clear. So the question I'm wrestling with is — how much of this is genuinely useful signal versus just a feature flag someone added to justify a price increase? The other thing is that we're a Series A company, so CAC is a real constraint. Adding another tool that promises to improve our observability into pipeline health has a high bar to clear — it needs to obviously beat what we're already doing, not just marginally. And right now I'd say most of what I've seen is marginal. I don't have deep visibility into how our engineering team thinks about observability specifically, so I can speak more to the go-to-market side of this — but I'm curious where you're finding the real gaps.

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

Honestly, my biggest problem is pipeline visibility and attribution — knowing which channels are actually driving revenue, not just leads. That's not directly about observability in the engineering sense, but when I think about how AI is being applied in our own workflows, the same core problem shows up: we're generating a lot of signal but the tooling to make sense of it is still pretty immature. So I guess if I'm being honest, the question I keep coming back to is whether the AI layer on top of these observability tools actually reduces the time to insight, or whether it just adds another dashboard nobody trusts. Because we've been burned by that pattern before — you buy the tool, it promises to surface the right thing at the right time, and then you're still doing manual triage anyway.

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

From my seat in demand gen, "good" is when I can actually trust the data I'm making decisions on — like when a campaign fires, I know within a reasonable timeframe whether it's actually moving pipeline or just inflating vanity metrics. That feedback loop being clean and fast is the dream. How far am I from that today? Pretty far, honestly. The attribution problem alone keeps me up at night. We've got data sitting in five different places and the "source of truth" conversation never really resolves. We've added tools that are supposed to help but every piece of software we use has AI baked in now and half of it is solving for its own attribution, not mine. So it's more noise than signal in a lot of cases. The feedback loop between what converts to closed-won versus what I'm optimizing toward is still pretty broken. Sales doesn't consistently feed that back, so whatever scoring we have doesn't actually learn what good looks like. That's the gap I'd most want to close.

What would change your perspective on this entirely?

Honestly, the thing that would change my perspective most is if I started seeing attribution actually work end-to-end in these AI observability tools. Right now my skepticism is partly borrowed from my own world — we can't get clean closed-loop data between marketing and sales, and from what I hear, engineering teams have similar gaps where the AI flags something but nobody closes the loop on whether the resolution was actually tied to that alert. If someone showed me a real example — not a demo environment, like a production case — where an AI observability tool caught an anomaly, traced it to a root cause, and you could verify downstream that it actually saved X hours or prevented Y incident, I'd take the maturity claims more seriously. The ROI narratives exist but they feel vendor-generated to me right now. The other thing would be simpler tooling. Every piece of software my team uses already has AI baked in, and the ask to learn something net new has to clear a pretty high bar. If these observability platforms got to the point where the AI layer was just invisible and obviously useful — not a separate feature you have to configure and champion internally — that would shift how I think about adoption friction for engineering teams too.

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

Honestly, I'm not sure I have some big reveal sitting there waiting to be asked. But if I think about it... maybe something like "what does good actually look like before you add AI to it?" Because I see a lot of conversations about AI for observability or AI for whatever, and the underlying data hygiene and process discipline just isn't there. You can't bolt intelligence onto chaos and expect it to work. That's not a dramatic insight, it's just something I keep running into — the feedback loops are broken, the data's a mess, and then people are surprised the AI isn't learning anything useful.

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

Honestly, the thing I keep coming back to is the signal-to-noise problem. Like, every observability vendor right now is telling me they have AI baked in — anomaly detection, AIOps, whatever — but when I actually dig in with our engineering counterparts, it's hard to figure out what's genuinely useful versus what's just the product team slapping "AI" on a feature that already existed. I sit on the marketing side, so I'm not the one configuring these tools day-to-day, but I'm close enough to the buying conversation to see where deals slow down. And what I'm noticing is that even when engineering teams are interested, there's a real gap between piloting something and actually scaling it into how they operate. The "good enough" threshold is surprisingly high — teams already have workflows, they have on-call rotations, they have dashboards they trust, and a new AI layer has to clear that bar before anyone's willing to change behavior. That maturity gap is what I'd say is genuinely interesting to me right now. Not whether AI in observability works in theory, but why adoption stalls after the pilot.

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

From where I sit — and I'll caveat that I'm marketing, not engineering — the thing I keep bumping into is the gap between "we have AI in our observability stack" and "we actually trust it enough to act on it." Like, vendors are very good at showing you dashboards with anomaly detection and predictive alerts, but the teams I talk to internally and through peers are still running parallel processes. They're not replacing human judgment, they're layering AI on top of existing workflows and hoping it adds value. So the core question I'd want answered is: what does it actually take for an engineering team to shift from AI as a supplemental layer to AI as the primary signal? Because until that trust gap closes, I don't think you're really talking about adoption — you're talking about evaluation mode that never ends.

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

Good, to me, looks like a closed loop — where your observability data is actually informing decisions without someone having to manually dig through dashboards and write a summary for a Slack post. Like, the system surfaces the right signal, to the right person, at the right time, with enough context that they can act on it. That's the north star. How far are we from that? Honestly, pretty far in most orgs I'm aware of. What I see more commonly is teams that have layered AI features onto existing tooling — some anomaly detection here, some log summarization there — but it's not really connected. You still have an SRE manually triaging alerts and correlating across three different tools. The "AI" part is kind of bolted on rather than load-bearing. I don't have a strong view on the exact maturity numbers across the industry, but from what I observe in conversations with engineering counterparts and vendors pitching us, most teams are somewhere in the middle — they've got the ingredients but not the recipe.

What would change your perspective on this entirely?

Honestly, the thing that would shift my view the most is seeing a really clean, documented case study where AI in observability meaningfully reduced MTTR or prevented an outage — not a vendor-produced one, but something from an engineering team that's willing to show the before and after with actual numbers. Right now a lot of what I hear feels like the "sprinkle AI on it" phenomenon — every monitoring tool is saying they've got AI-powered anomaly detection, AI-powered root cause analysis, and it's hard to separate signal from noise when everyone's using the same language to describe very different levels of capability. If I saw peer-validated evidence that a team at similar scale actually reduced on-call burden or alert fatigue in a measurable way, that would make me take the category more seriously. Right now I don't have a strong enough view on the engineering side specifically to be more precise than that.

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

Honestly, I don't have some burning "secret question" that nobody's asking. But if I had to pick something... I think the conversation around AI in observability tends to focus heavily on the engineering side — which makes sense, it's an engineering workflow topic — but nobody's really asking how the signal from these systems actually flows back into business decisions. Like, if your AI-assisted observability tool catches an incident faster, what's the downstream effect on customer retention metrics, on NRR? That translation layer between engineering outcomes and business outcomes is still pretty murky in most orgs I've seen, and I think that's where the real adoption friction lives. Engineers can see the value, but getting budget and executive buy-in requires that business case to be airtight, and right now I don't think the vendors are helping make that connection clearly enough.

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

Honestly, the thing I keep coming back to is the gap between what vendors are promising with AI in observability and what I'm actually seeing work in practice. I'm not an engineering leader — I sit on the go-to-market side — but I work closely enough with our eng and DevOps teams to see where the friction is. The noise level is just extraordinary right now. Every tool we already use has "AI" baked into it somewhere, and then we're getting pitched on standalone AI observability solutions on top of that. At some point you have to ask whether any of this is actually moving the needle on incident response or anomaly detection, or whether it's mostly feature theater. The specific tension I'm wrestling with is around adoption — like, even when the technology is genuinely useful, getting engineers to change their workflows is hard. People default to what's "good enough," and that inertia is a real blocking factor that I don't think vendors account for enough in how they position this stuff. I don't have a strong view on the maturity curve specifics since that's not my direct domain, but from a GTM and organizational perspective, the question of "who owns AI in observability" inside a company seems genuinely unsettled, and that ambiguity slows everything down.

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

From my vantage point in marketing, I'm adjacent to this rather than living it daily, so take that for what it's worth. But the pattern I see across our engineering and product counterparts is that the tooling has gotten ahead of the workflow. Everyone's got some AI baked into their observability stack now — it's almost table stakes — but there's a real gap between "we have this capability" and "we actually trust it enough to act on it without a human in the loop." The thing I'd want to understand is where that trust breaks down. Is it a data quality problem? A model explainability problem? Because those have pretty different solutions, and I've watched too many teams buy a tool, get burned once when it surfaced a false positive or missed something obvious, and then basically revert to the old workflow. At that point you've paid for something that nobody uses at full capacity. I don't have a strong view on the specific technical blockers, but from a go-to-market and adoption standpoint, that "good enough" inertia is real. Getting people to change an on-call workflow they've relied on for three years requires the new thing to be dramatically better, not just marginally better.

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

Good, to me, is when the signal-to-noise ratio actually flips — where your observability tooling is surfacing the right alerts with enough context that an engineer doesn't have to go on a 45-minute investigation just to understand what's happening. Like, the system tells you not just *that* something broke, but *why*, and ideally *what to do about it*. From where we sit today, I'd say we're maybe halfway there at best. We've got tools with AI features baked in, but a lot of it still feels like the vendors sprinkled some AI on top to hit a talking point rather than fundamentally rethinking the workflow. The alert fatigue problem is still very real on our engineering side — I hear about it constantly. The gap isn't really the technology at this point, it's the data quality and process discipline underneath it. If your telemetry data is messy or your tagging is inconsistent, no AI layer is going to save you. So "good" also requires getting the hygiene right first, and that's honestly the harder problem for most teams.

What would change your perspective on this entirely?

Honestly, the thing that would move the needle for me is seeing a clear, documented case where AI in observability actually reduced incident response time in a meaningful way — not a vendor case study with cherry-picked numbers, but something I could validate. Like, a peer at a comparable company saying "we went from 45 minutes to 12 minutes mean time to resolution and here's the before-and-after data." Right now a lot of what I see around AI and observability feels like the same pattern I've watched in martech — someone slaps "AI-powered" on an alerting dashboard and calls it transformational. I've been burned enough times by that framing that I need to see the actual workflow change, not just the feature announcement. If the engineering counterparts I work with started coming to me saying "this is genuinely reducing toil and we can redeploy that engineering capacity toward product work," that would change my view pretty quickly. That's a concrete, measurable outcome I can work with.

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

Honestly, I don't have some big contrarian insight that nobody's surfacing. I think the questions I'd want more of are just... more grounded versions of what people are already asking. Like, less "what's your AI strategy" and more "what's actually working in production today versus what's still basically a demo." That gap between the pitch and reality is where I spend most of my time trying to calibrate, and most conversations I'm in skip over it pretty fast. The other thing I'd push on more is the feedback loop question — not glamorous, but in any system where AI is supposed to get smarter over time, who's actually closing the loop and making sure the model learns from what happened? In my experience that accountability tends to live nowhere.