"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?

Yeah, so the thing I'm actively wrestling with right now is the gap between what the observability vendors are promising with their AI features and what actually works in production. Every vendor we talk to — Datadog, Honeycomb, whoever — has some version of "AI-powered anomaly detection" or "intelligent alerting" baked in now. And honestly, a lot of it feels like the AI got slapped on as a checkbox rather than built into the core workflow in a meaningful way. The specific tension for me is around signal versus noise. We already have an alert fatigue problem. If the AI layer isn't materially better at prioritization, it's just adding another surface area I have to trust — and trust is where I get stuck. Like, what's the model actually doing? What data is it trained on? Is it my telemetry data, their aggregate data across customers? Those questions matter a lot from a security standpoint and I don't always get clean answers. The other piece is the build-versus-buy question, which I'm never fully done asking. We've got engineers who could wire something together using LLM APIs on top of our own telemetry pipeline, and I genuinely don't know if that's smarter than paying a premium for a vendor's native AI features that I can't fully inspect. That's the open question I don't have a good answer to 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 in our observability data. We're generating enormous volumes of telemetry — traces, logs, metrics — and the honest answer is that most of it goes unread. The question I keep coming back to is: can AI actually help us surface the stuff that matters before an engineer has to go dig for it, or is it just going to add another layer of dashboards nobody looks at? The secondary piece is integration. We've accumulated a fair number of tools in our stack over the years — that vendor fatigue is real — and I'm skeptical of any AI observability capability that requires me to bolt on yet another platform with its own API contract and data model. So the build-versus-buy calculus here isn't simple. If someone's pitching me an AI layer on top of observability, I need to understand exactly what APIs it exposes, what data it needs access to, and whether it can sit on top of what we already have rather than replace it.

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

Good, to me, is when AI is actually embedded in the signal layer — not bolted on top. So when an anomaly surfaces, the system has already correlated it across services, suppressed the noise, and given my on-call engineer enough context that they're not starting from zero at 2am. The AI is doing the triage work, not just reformatting the same alert into a slightly different UI. Where are we from that today? Honestly, we're closer on the detection side than the response side. We've got some reasonable anomaly detection working in our metrics pipeline, but the correlation across logs, traces, and events is still pretty manual. My engineers are still doing a lot of the "okay, this fired, now let me go figure out why" work that I'd want automated. I'd say we're maybe halfway to what I'd call good, and the gap is less about the AI capability itself and more about the data quality and normalization work that has to happen upstream before any model can do something useful with it. The other piece I'd flag — and this is a real frustration — is that every vendor is telling me their thing is "built with AI at its core," and when I push on what that actually means in terms of my workflow, the answers get vague fast. So part of what's slowing us down is just the evaluation overhead of sorting signal from noise in the vendor landscape itself.

What would change your perspective on this entirely?

Honestly? If someone showed me a system that could close the loop autonomously — not just surface an anomaly and say "hey, look at this," but actually trace it back through distributed traces, correlate it with a recent deployment, and either remediate or escalate with enough context that my on-call engineer isn't starting from zero at 2am — that would move the needle significantly. Right now most of what I see is pattern matching dressed up as intelligence. The "AI" label gets slapped on threshold-based alerting with a slightly fancier UI, and vendors can't actually explain the mechanism when you push them on it. I've had those conversations where you ask "what does AI at its core mean for your product" and you just get the phrase repeated back at you. What would genuinely change my view is longitudinal evidence — show me a org that's been running this for 18 months, show me how alert fatigue metrics moved, show me the false positive rate over time as the model learned their environment. Not a demo environment, a real production system with real cardinality. That kind of proof is almost never what gets put in front of me.

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

Honestly, the question I'd want someone to ask is: "What happens to your observability data quality when you layer AI on top of a fragmented tool stack?" Because that's where I actually spend mental energy. Everyone's asking about which AI features vendors are shipping or what the ROI looks like, but nobody's asking whether the underlying signal is clean enough for AI to do anything useful with. If your telemetry is inconsistent, if your tagging conventions are a mess, if half your services aren't even instrumented properly — the AI layer just surfaces noise faster. Garbage in, garbage out, but now it's automated garbage. That's the maturity gap nobody wants to talk about because it doesn't make for a good demo.

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 signal-to-noise problem. We have more observability data than we've ever had — traces, logs, metrics, all of it — and the promise of AI is that it helps you actually make sense of that at scale. But what I'm finding in practice is that we're getting a lot of vendor noise around "AI-powered" this and "intelligent" that, and when you actually dig in and ask what that means technically, it's often pretty thin. The specific thing I'm wrestling with right now is: do we build lightweight AI layers on top of our existing observability stack — we're running a mix of OpenTelemetry, some Datadog — or do we go with one of these vendors who's promising AI-native observability? And every time I evaluate one of those vendors, I end up in this circular conversation where "AI at the core" doesn't actually translate into a clear technical answer about what's different. That's frustrating. The other piece is the data quality problem. AI-assisted anomaly detection or root cause analysis is only as good as the telemetry underneath it, and honestly a lot of our telemetry is still inconsistent. So we're debating whether we should be investing in cleaning that up first before we layer AI on top of it, or whether some of these tools are actually robust enough to handle messy inputs. I don't have a clean answer on that yet.

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

The core problem for me right now is signal-to-noise ratio at scale. We're generating an enormous amount of observability data — traces, logs, metrics — and the question is whether AI can actually help us surface what matters before a human has to go digging. That's the genuine value proposition I'm trying to evaluate. The secondary thing underneath that is trust. If an AI system tells my on-call engineer "this is the root cause," I need to understand the confidence level and the reasoning, not just the output. Because if it's wrong at 2am and someone acts on it, that's a bad outcome. So explainability isn't a nice-to-have for me, it's a prerequisite.

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

Good, in this context, means AI that's actually embedded in the signal layer — where anomaly detection isn't just alerting me that something broke after the fact, but correlating across services, understanding blast radius, and surfacing a ranked hypothesis about root cause before my on-call engineer has even opened their laptop. That's the north star. Where are we today? Honestly, we're using a couple of the AI features in our observability tooling — we're on Datadog — and they're fine. The anomaly detection catches some things. The natural language query stuff is useful for engineers who aren't fluent in their query language. But it still feels like features bolted on rather than a coherent intelligence layer. The tool is doing pattern matching; it's not doing reasoning. I'd say we're maybe a third of the way there. The gap isn't really the vendor's fault — part of it is that our telemetry isn't clean enough yet for AI to do much more than we're asking it to do. Garbage in, garbage out. So we're spending cycles on data hygiene and instrumentation consistency before we can even unlock the next tier of what these tools claim to offer. That's the unsexy part nobody talks about in the pitch cycle.

What would change your perspective on this entirely?

Honestly, the thing that would move the needle most for me is seeing genuine, durable signal separation from the noise layer. Right now when I evaluate AI observability tools, I can't tell if I'm looking at a real capability or just a wrapper around an LLM that's been dressed up with some dashboards. If a vendor could demonstrate — in my actual environment, with my data — that their AI is surfacing anomalies I genuinely would have missed, and doing it consistently over months not just in a demo, that changes the conversation. The other thing would be better integration story at the API level. Most of what I'm seeing right now is point solutions that want to be yet another pane of glass. If the AI layer could slot cleanly into the data pipeline we already have — talking to our existing telemetry stack without forcing us into their proprietary data store — I'd take it a lot more seriously. Right now the build-vs-buy calculus keeps tipping toward build precisely because the integration story is so weak.

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

Honestly, the question I'd want someone to ask is: "How are you managing the organizational trust layer between your engineers and AI-generated observability outputs?" Everyone asks about tooling and integration. Nobody asks whether your team actually acts on what the AI surfaces, or whether they've learned to quietly ignore it because it's been wrong too many times. That trust erosion is silent and it compounds — and once engineers start treating AI alerts the same way they treat alert fatigue from traditional monitoring, you've basically lost the value proposition entirely. The maturity conversation tends to stay at the tooling level, but the harder problem is behavioral and cultural.

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* and what's actually operationally useful day-to-day. Like, we're a fintech, so we have compliance and reliability concerns that don't go away just because something is labeled AI-powered. And I see a lot of vendor pitches that lead with "AI" as if that's sufficient — but honestly that framing is almost meaningless at this point given how many different things it covers. What I'm actually wrestling with is the people and process side more than the tooling side. We could adopt a better AI-assisted observability tool tomorrow, but if the on-call rotation doesn't have shared context about what the model is flagging and why, we haven't really solved anything. It's less a technology problem and more a workflow and trust problem — do engineers actually trust the anomaly detection enough to act on it without digging into the raw data themselves? And then there's the organizational piece — who owns the outcomes when an AI-assisted alert fires and turns out to be noise? That accountability question hasn't been cleanly answered in most places I've seen.

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

From where I sit as a PM, the thing I keep coming back to is the gap between what AI tooling *promises* in observability and what engineering teams can actually operationalize day-to-day. Like, we've got tools throwing anomaly detection and log summarization at teams, but if the underlying data quality is poor or the alert taxonomy is a mess, the AI just amplifies noise instead of reducing it. The deeper question for me is really about process readiness — are teams actually set up to act on what AI surfaces, or are we bolting AI onto a workflow that was already broken? I've seen that pattern in other contexts where you invest in the shiny tooling but 12-18 months in you realize the people and process side wasn't there to support it. So the most important thing I'd want to understand is: what's the actual blocker — is it the technology maturity, is it team workflow and culture, or is it something like data access and governance? Because the solution looks pretty different depending on which of those is actually the constraint.

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

Good, to me, is when AI in observability actually shortens the loop between "something's wrong" to "we understand why and we're fixing it." Right now we have alerts firing, someone has to context-switch, dig through dashboards, correlate logs manually — that whole triage cycle is really where the time bleeds out. So the ideal state is: AI surfaces the right signal, gives you a plausible starting hypothesis, and your on-call engineer is confirming rather than discovering. How far are we from that? Honestly, we're partway there in some narrow areas. Anomaly detection has gotten genuinely useful. But the gap I see most is that the AI recommendations still lack enough context about *our* business logic — what's actually a critical flow versus noise. It's a similar problem to what you see with generic AI tools broadly: trained on public patterns, not on the nuance of your specific system. So there's a lot of alert tuning and prompt engineering and customization work that falls back on the team. I'd say we're maybe at a 40-50% confidence level in the AI-surfaced insights on any given incident. That's not nothing, but it's not good enough to reduce cognitive load meaningfully at 2am. The trust threshold is still the blocker.

What would change your perspective on this entirely?

That's a good question to sit with for a second. Honestly, the thing that would shift my thinking the most is seeing a clear, documented feedback loop where AI-generated observability insights actually changed a decision — and that decision had a measurable outcome. Right now a lot of what I see or hear about is AI surfacing patterns or anomalies, but then it's kind of a dead end. Someone still has to interpret it, act on it, and there's no clean line back to "the AI recommendation caused this outcome." If I saw that working at scale — not just in a demo environment or a single-team pilot — that would genuinely change how I think about the maturity question. Because right now my mental model is that most orgs are still in the "AI as a fancy alert filter" stage, not "AI as something integrated into how we actually make engineering decisions." The other thing, honestly, is if the people and process side got solved alongside the tooling. I've seen enough situations where the tech is fine but the org adoption isn't there. If someone showed me a credible model for how teams actually change their workflows — not just deploy a new tool — that would move me.

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

Honestly, I think the question that gets skipped is something like: "Who actually owns the signal quality problem?" Because a lot of the AI observability conversation is focused on the tooling layer — like, what model are you running anomaly detection with, which vendor are you using — and not enough on the upstream data quality issues that make those tools underperform. In my experience, if your telemetry is noisy or inconsistently instrumented across services, no amount of AI on top of it is going to save you. You're just automating garbage. And from a PM perspective, that's actually a people and process problem before it's a tooling problem. Someone has to own instrumentation standards, someone has to care about alert hygiene. But in most orgs I've seen or talked to, that ownership is fuzzy. So I'd love for researchers to dig into that more — who's accountable, and how does that change as you try to scale AI-assisted observability.

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 AI tooling promises for observability and what actually gets operationalized at scale. Like, we're a fintech startup, so we're not a massive engineering org, but we still have enough complexity in our payment flows and compliance monitoring that observability matters a lot. What I'm wrestling with right now is really a people-and-process question more than a tooling question. We've got engineers who are genuinely interested in using AI-assisted alerting and anomaly detection, but the workflow around it — who owns the signal, how do you act on it, how does it fit into on-call rotations — that's still pretty undefined for us. The technology is kind of outpacing our process maturity. And then there's the trust question. When an AI surfaces something in your observability stack, how confident are your engineers actually in that signal versus tuned, hand-crafted alerts they've built over time? I don't have a clean answer there yet. We're still in that early phase where people are experimenting but nobody's fully committed to letting AI-driven insights drive incident response decisions. So yeah — less of a vendor evaluation problem for us right now, more of a "how do we build the organizational habits around this" problem.

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

From a PM perspective, the thing I'm most focused on understanding is where AI in observability actually moves the needle versus where it's just... noise. Because right now there's a lot of "we added AI to this dashboard" and it's hard to tell if that's genuinely improving how engineers detect and respond to issues, or if it's just a feature checkbox. The maturity gap question is interesting to me specifically because in my experience — and we're a fintech so reliability and compliance matter a lot — the tooling conversation tends to outpace the process and people conversation. Teams adopt a new AI-powered observability tool and then six to twelve months in they realize the alerts are still noisy, the on-call engineer still has to do a ton of manual correlation, and they haven't actually changed how they triage. So the tech moved but the workflow didn't. What I'd really want to understand from this research is what the actual blockers to workflow change look like — is it trust in the AI outputs, is it data quality upstream, is it org structure? I don't have a strong view yet on which of those dominates, which is partly why I find this topic worth digging into.

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

Good, to me, looks like a world where AI in observability is actually closing the loop — not just surfacing anomalies but helping the on-call engineer understand *why* something is happening and what the likely remediation path is, without them having to go dig through five different dashboards. Basically compressing that time from alert to context to action. Where we are today feels pretty far from that, honestly. Most of what I see — even in tools that market themselves as "AI-powered" — is smarter alerting or anomaly detection, which is useful, but it's still pretty shallow. The narrative layer, the "here's what this means for your service and your users right now," that's still largely manual. Engineers are still doing a lot of the connective tissue work themselves. And from a PM perspective, the blocker I keep running into is less about the models and more about data quality and integration depth. The AI is only as good as the telemetry you're feeding it, and a lot of orgs — including ours at points — have fragmented or inconsistent instrumentation. So you end up with an AI layer sitting on top of a shaky foundation, and the outputs aren't trustworthy enough for engineers to actually act on confidently. I'd say we're probably in early-to-mid innings on this, which is fine, but I think there's a risk that orgs invest in the AI surface without fixing the underlying data problems first.

What would change your perspective on this entirely?

Honestly, the thing that would shift my view the most is seeing real evidence of AI in observability that actually closes the loop autonomously — not just surfaces an anomaly or drafts a summary, but meaningfully reduces time-to-resolution in a way that's measurable and reproducible across different team contexts, not just one well-resourced team at a big tech company. Right now a lot of what I see feels like the tooling side is getting a lot of attention but the process and people side isn't keeping pace. Like, if I saw organizations where the AI recommendations were actually trusted enough to act on without a human double-checking every single one, and where that trust was earned through demonstrated accuracy over time — that would tell me we've crossed a real maturity threshold. The other thing that would change my perspective is if the data access and governance problems got meaningfully easier. A lot of the blocking I observe isn't really about the AI capability itself, it's about fragmented data pipelines and compliance constraints that make it hard to even feed the right context into these tools. If that got solved in a practical way, I think adoption would accelerate pretty quickly. But I'll be honest — I don't have super deep visibility into the observability space specifically from an engineering org standpoint. That's more adjacent to my role than central to it, so take my framing with that caveat.

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

Honestly, I don't have some dramatic "nobody asks this" hot take. But one thing that comes up in my actual work that I don't hear framed well in these conversations is: who owns the AI output when it's wrong? Like, in observability specifically — if an AI system surfaces a false positive or misses an incident, and my engineers have started to trust that signal... what's the accountability model? That feels adjacent to risk and reliability concerns, which tend to get deprioritized when everyone's excited about the capability side. It's not a glamorous question, but it matters for whether teams actually commit to these tools at scale versus just piloting them indefinitely.

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

Honestly, the thing I keep bumping into is attribution — not even in the traditional marketing sense, but more like... when we're looking at observability data and trying to understand what's actually driving a pipeline issue or a system degradation, there's a lot of noise. Engineering teams are starting to layer AI tooling on top of that noise, but I'm not sure it's making the signal clearer yet. From where I sit in demand gen, I care about system reliability insofar as it affects customer experience and retention, which feeds back into our CAC and expansion numbers. So when engineering is adopting AI for observability, I want to know: is it actually reducing mean time to resolution, or is it just another tool in a stack that's already pretty bloated? We've got tool sprawl problems in marketing, and from what I can tell, engineering teams aren't immune to that either. The maturity gap piece is real though. I don't have a strong view on the technical specifics, but I do see the pattern — teams deploy something, it works okay in narrow conditions, and then scaling it is where things fall apart. That feedback loop problem, where the AI never quite learns what "good" looks like because nobody closes the loop systematically, that shows up in my world constantly and I'd guess observability is similar.

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

Honestly, my honest answer is that this is a little outside my day-to-day — I'm on the demand gen side, not in engineering or DevOps. So I'll give you what I can. From where I sit, the thing that seems to matter most in any AI adoption story is whether the tooling actually fits into how people already work. I see this on the marketing side constantly — every piece of software we use already has AI baked in somewhere, and the question is never "does it have AI" anymore, it's "does this actually change my workflow in a meaningful way or is it just another layer I have to manage." I'd assume that's true for observability teams too. The other thing I'd flag, just from watching how tooling decisions get made at our stage of company, is the buy-in problem. Even when the technology works, you've got people who are skeptical, teams that have existing processes, and limited bandwidth to learn something new. That friction is real and it's usually underestimated. But I want to be straight with you — I don't have a strong view on the specifics of AI in observability workflows. That's not a space I operate in directly.

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

Good, in my world, is pretty simple to define even if it's hard to execute: I want clean attribution from first touch to closed-won, and I want to know which channels are actually driving pipeline that converts — not just MQLs that sales ignores. That's the north star. How far am I from it? Honestly, pretty far on the attribution side. We've got data coming from multiple sources and the feedback loop between marketing and sales is still broken in the ways you'd expect — sales isn't consistently logging what actually closed and why, so any scoring model we try to build is working off incomplete signal. It's not a tool problem at this point, it's a process and alignment problem. The AI layer on top of that feels premature for us right now. Every tool I get pitched claims it'll fix lead scoring or automate the handoff, but if the underlying data hygiene isn't there, you're just getting garbage out faster. So "good" for me starts with getting the basics right before I layer in anything more sophisticated.

What would change your perspective on this entirely?

Honestly, the thing that would flip my view most is if someone showed me clean, validated evidence that AI-assisted observability actually reduced incident response time in a way that traced back to pipeline impact or revenue protection — not just "we caught the bug faster" but something that connects to real business outcomes. Right now a lot of what I see is "AI" getting slapped on existing monitoring tools, and it's hard to tell what's genuinely net new capability versus rebadging. If a vendor could demonstrate that the AI component specifically — not just better dashboards or more alerts — is what drove the outcome, that would shift how I think about the buy-in problem on the go-to-market side. I don't have a strong view on the engineering side specifically, but from where I sit, the tools that win are the ones that can articulate a tight before-and-after. The feedback loop problem is real across most AI tooling right now — if observability platforms solved that cleanly, I'd take them a lot more seriously.

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

Honestly, I'm not sure I have a burning one for this particular topic — observability and engineering workflows aren't really my world day-to-day. If I'm stretching into adjacent territory that actually does affect me, I'd probably say something like: "How does AI-assisted monitoring and alerting actually feed back into product decisions or GTM prioritization?" Because on the marketing side, I'm often flying blind on product reliability signals that could inform how we're positioning or which segments we're targeting. But I don't have a strong view on the engineering observability side specifically — that's a few walls away from where I sit.

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

Honestly, my day-to-day is pretty removed from observability workflows directly — that's more of an engineering and DevOps concern. But where it touches me is on the attribution and pipeline side. Like, when our engineering team is trying to figure out what's broken in our product or our data infrastructure, that has downstream effects on my reporting and my ability to trust the numbers I'm making decisions off of. So what I'm wrestling with more broadly is just... how much of the AI tooling we're buying actually delivers signal versus noise. We've got AI baked into half our martech stack already, and half the time I'm not sure if it's genuinely surfacing insights or just repackaging what I already knew in a nicer dashboard. That skepticism probably applies to observability tooling too — I'd want to know if the AI anomaly detection is actually catching things earlier or just adding another alert I have to triage. The other thing is tool sprawl. Every vendor says they have AI now. It needs to add real value beyond whatever we already have that's "good enough," otherwise it's just noise on top of noise.

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

Honestly, this is a bit outside my day-to-day — I'm on the demand gen side, not engineering or DevOps. So I don't have a strong view on observability workflows specifically. But if I'm thinking about it from a general "AI adoption in technical teams" angle, the pattern I see everywhere is the same: people buy or build AI-adjacent tools and then struggle to get the feedback loops right. Like, the system never learns what "good" actually looks like because the data going back in is messy or incomplete. That feels like it would apply to observability just as much as it does to lead scoring or any other workflow. The other thing I'd guess is relevant — and I see this in my own stack — is that every tool already has AI baked in now, so the question isn't really "do we have AI," it's "is the AI we have actually doing something useful versus just being a feature checkbox." I'd want to understand whether teams are actually getting signal reduction or faster resolution, or if it's just noise with a nicer UI on top.

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

Good, to me, is when I can close the loop between a campaign touch and actual pipeline influence without spending three days in spreadsheets trying to reconcile data sources. Like, I want attribution that's reasonably trustworthy, CAC by channel that I actually believe, and enough signal to make a defensible budget decision in real time rather than retrospectively. How far am I? Pretty far, honestly. We've got data sitting in HubSpot, our ad platforms, and a few other tools, and stitching that together is still mostly manual or relies on workflows that break. The feedback loop from sales back into marketing on which leads actually closed is weak — sales isn't consistently logging what I need them to log, so the scoring models are learning from incomplete data. It's a solvable problem, but it requires buy-in across teams and someone to actually own the data hygiene piece, and that's where it stalls out.

What would change your perspective on this entirely?

Honestly, the thing that would shift my view the most is seeing clean attribution data come out of these AI observability tools and actually connect back to business outcomes — not just "we caught an incident 12 minutes faster." Like, if someone could show me a workflow where the AI flagged an anomaly, that anomaly was traced to a pipeline degradation, and that degradation was costing X in conversion or revenue — and the tool surfaced all of that automatically — I'd take the whole category a lot more seriously. Right now most of what I hear is engineering-centric metrics that don't translate well across the org. I don't have a strong view on the technical maturity side specifically, but from a buyer psychology standpoint, the framing is still pretty siloed. If vendors started speaking in terms of business impact rather than just MTTR or alert volume, that would move the needle for me in terms of how I think about it.

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

Honestly, I'm not sure I have some profound unanswered question here. But if I'm thinking about what's actually relevant to my world — even though I'm on the demand gen side, not engineering — it's probably something like: "How does the attribution and data feedback problem get solved before you layer AI on top of it?" Because from what I see, a lot of teams are excited to add AI to their observability or analytics workflows, but the underlying data hygiene is still a mess. Sales never tells marketing which leads closed, engineers aren't tagging incidents consistently — and then you're training models or building AI workflows on top of noisy inputs. The AI doesn't learn what "good" looks like because nobody agreed on what good looks like in the first place. That feels like the more honest conversation to have before talking about AI adoption maturity.

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 and what's actually operationally useful. We get pitched constantly — I'd say our eng and DevOps teams are fielding AI-in-observability pitches almost weekly — and a lot of it feels like "we sprinkled AI on our dashboards" without a clear answer to what problem that actually solves faster or cheaper than what we already have. The more specific tension for me is around signal versus noise. Observability tooling already generates an overwhelming amount of data. The pitch is usually that AI helps you cut through that — anomaly detection, root cause surfacing, whatever. But when I push on it, the underlying data hygiene question rarely has a good answer. If the telemetry pipelines are messy, adding an AI layer on top just gives you confident-sounding noise instead of regular noise. The third thing, and this is more of a slow burn, is that the teams doing the most interesting work in this space seem to be cobbling things together themselves — not buying a turnkey solution. Which makes the vendor landscape feel a little disconnected from where practitioners actually are.

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

Honestly, from where I sit — and I'll caveat that I'm on the marketing side, not engineering — the thing I keep bumping into is the signal-to-noise problem. We generate enormous amounts of observability data, and the question is whether AI is actually helping teams cut through that to find what matters, or whether it's just adding another layer of complexity on top of an already noisy stack. The tooling we already use has AI baked in to varying degrees, and a lot of it still feels like it's solving for "impressive demo" rather than "actually reduces alert fatigue or MTTR in practice." So the core question for me is whether AI in observability is genuinely changing how engineering teams triage and respond, or whether it's mostly a positioning story vendors are telling right now.

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

Good, in this context, means the system surfaces the right signal before I have to go looking for it. Like, if something's degrading in a way that's going to affect customers or pipeline, I want that flagged automatically with enough context that the team can act — not just a noisy alert that says "hey, something's off, good luck." Where we are today is... honestly pretty reactive still. We've got tooling in place, some of it has AI features baked in, but a lot of those features feel more like checkbox items than genuinely useful capabilities. The gap between "the tool has AI" and "the AI is actually reducing investigation time" is real. I'd say we're maybe a third of the way to what I'd consider good. The infrastructure is there but the workflows around it — who acts on what, how alerts get triaged, how findings connect back to business impact — that's still pretty manual and inconsistent.

What would change your perspective on this entirely?

Honestly, the thing that would shift my view the most is seeing a clear, documented feedback loop where the AI isn't just surfacing alerts but actually getting smarter over time based on what engineers do with those alerts. Right now a lot of what I see feels like it's a one-way street — the system flags something, a human resolves it, and that context just disappears. If someone showed me a production case where the model demonstrably reduced mean time to resolution over six months because it was learning from closed incidents, with actual before-and-after data, that would move me. Not a demo environment, not a cherry-picked benchmark — real operational data from a team at scale. The other thing, honestly, is if the tooling got good enough that it didn't require a dedicated person just to babysit it. A lot of the AI fatigue I'm aware of comes from the overhead of tuning and maintaining these systems, and until that cost drops significantly, the ROI math is hard to close for most teams.

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

Honestly, I don't have a burning "question nobody's asking" framing for you. But if I were to point at something that doesn't come up enough in these conversations — it's probably around the feedback loop problem. A lot of the AI in observability discussion is focused on the detection side, the surface layer. But nobody's really asking whether the signal that trains or informs these AI systems is actually clean. In my world on the marketing side, I see this constantly — AI scoring tools that never learn what a good outcome looks like because the data loop between sales and marketing is broken. I'd imagine the same thing happens in engineering orgs with observability: the AI flags things, engineers resolve them, but whether that resolution feedback actually gets back into the system to improve it over time is an open question. That's where I'd poke if I were evaluating one of these vendors.

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's actually delivering value day-to-day for engineering teams. I sit close enough to our engineering org to see the tool evaluations happening, and there's a real pattern of — every monitoring or observability platform has "AI" baked in now, but half the time it's just a chat interface bolted on top of the same dashboards we already had. The more specific frustration is around signal-to-noise. The promise is that AI helps you cut through alert fatigue, surfaces the right anomalies faster — but I'm not seeing strong evidence internally that we've meaningfully reduced the time engineers spend triaging noise. It's still a lot of manual context-switching. And then from a marketing lens, I'm wrestling with how we talk about this category without falling into the same AI-washing trap that makes buyers skeptical before the conversation even starts. Everyone's burned enough times that saying "AI-powered" almost works against you now unless you can immediately ground it in a specific workflow outcome.

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

From my vantage point in marketing, the observability and engineering workflow side isn't my direct domain — so I want to be upfront about that. But what I do understand really well is the adoption problem, and I think that's actually the core question here. The pattern I see over and over is that you've got teams who've deployed AI tooling but can't clearly articulate what it's actually doing for them. Like, six tools in the stack and half of them have unclear ROI. That's not an observability-specific problem, that's a maturity problem. And I'd want to understand whether the blocker is the tooling itself, the workflow integration, or just that nobody's defined what "good" looks like in measurable terms before they bought in. The maturity gap question is the one I'd push hardest on.

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 workflow — not a separate tool someone has to remember to go open. Like, you get an anomaly in your observability stack, and the system doesn't just surface it, it gives you context on probable cause, relevant runbooks, maybe even a confidence score on the diagnosis. That's the dream. Where we are today? Honestly, most of what I see is still pretty surface-level — it's alert triage, maybe some log summarization. Useful, but not transformative. The gap between "AI is in the product" and "AI is actually changing how my on-call engineer operates" is still pretty wide for most orgs I'm aware of. We're closer to the former than the latter.

What would change your perspective on this entirely?

Honestly, the thing that would shift my view most is seeing rigorous before-and-after data from engineering orgs that have actually scaled AI in observability — not a vendor case study, but something closer to a controlled comparison. Like, here's MTTD before, here's MTTD after, here's the noise reduction rate, here's the alert fatigue metric. Right now most of what I see is either early pilots with cherry-picked wins or vendor marketing dressed up as outcomes. The other thing — and this is maybe more fundamental — is if I started seeing teams report that the AI actually *learned* over time in their specific environment. A lot of the frustration I hear is that these tools are good out of the box but the feedback loop breaks down. Nobody closes the loop on which alerts actually mattered, so the model never improves. If someone showed me a credible example of that flywheel actually working at scale, I'd take the whole space a lot more seriously.

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

Honestly, I don't have a burning "question nobody's asking" that I've been sitting on. That framing can lead to manufactured insight. If anything, what I'd say is — most conversations I'm in focus on the technology itself, and fewer people ask about the organizational change management side. Like, who owns the AI observability workflow once it's in place? Is it engineering, is it DevOps, is it a platform team? In my experience watching tools get adopted and then quietly shelved, the governance question matters as much as the capability question. But that's a pretty generic observation, not a dramatic revelation.