"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?

The honest answer? I'm wrestling with the signal-to-noise problem in a pretty acute way right now. Every observability vendor — Datadog, Honeycomb, New Relic, take your pick — is slapping "AI-powered" on everything, but when I dig into what that actually means under the hood, it's usually just anomaly detection with a chatbot wrapper and a 40% price increase. What I actually need is something that can correlate across my traces, logs, and metrics and surface *why* something is breaking, not just *that* something is breaking — and I haven't seen a vendor nail that in a way that doesn't require me to babysit the model constantly. The deeper tension for me is build vs. buy: I've got engineers who could wire something up on top of our existing OpenTelemetry pipeline using Claude or GPT-4, but then I own that forever, and my security team will have opinions about what telemetry data we're shipping to an external LLM endpoint.

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

The core problem for me is signal-to-noise ratio at scale. We're generating terabytes of observability data — logs, traces, metrics — and the gap between "we have all this data" and "we actually understand what's happening in production" is enormous. AI feels like the obvious wedge to close that gap, but right now most of what's being sold to me is either glorified alerting with an LLM bolted on, or it's a black box that I can't audit, can't integrate cleanly into our existing Datadog and Grafana stack, and definitely can't trust with anything touching customer data. The real question I'm trying to answer is: can AI actually reduce mean time to detection and resolution in a way I can measure, or are we just paying for a better-looking dashboard that my SREs will stop trusting within six months?

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

Good looks like AI that's actually embedded in the signal layer — not bolted on as a chat interface over my logs. I want the system to understand the *context* of an anomaly: is this a deploy artifact, is this a cascading dependency failure, is this actually customer-impacting versus just noisy infrastructure churn? Right now we're maybe 30% of the way there — we've got some decent alerting through Datadog with their Watchdog stuff, but half the time it's still a human doing the correlation work that should be automated. The gap isn't the data, honestly, it's that none of these vendors have figured out how to make the AI *opinionated* in a way that maps to my specific system topology rather than generic patterns.

What would change your perspective on this entirely?

Honestly? Show me a system that can actually close the loop autonomously — not just surface an anomaly and hand it back to a human, but reason through the blast radius, correlate it to a deployment event, and either remediate or escalate with full audit trail. Right now everything I see is "AI-assisted" which is just a fancier dashboard with a chatbot bolted on. If someone demoed me genuine agentic observability that I could trust in production without a babysitter, where the model's decisions are explainable and the security posture is defensible to my compliance team, I'd completely reassign my priors. The other thing that would move me is seeing the data gravity problem solved — we're talking petabytes of telemetry living in different silos, and until there's a coherent answer for how AI reasons across that without me rebuilding my entire data pipeline, it's still a science project.

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

The question nobody asks me is: "What does your observability data architecture actually look like, and is it even *ready* for AI to operate on?" Everyone wants to talk about which AI-powered observability tool to buy, but they skip the foundational question of whether your telemetry is clean, consistent, and semantically meaningful enough for a model to do anything useful with it. I'd love for someone to ask me that because the honest answer at most orgs — including mine six months ago — is "no, it's a dumpster fire of inconsistent labels, missing context, and data that lives in four different systems that don't talk to each other." You can't just point an LLM at Datadog and call it AI-driven observability; garbage in, garbage out, and now you've paid a premium for confident-sounding garbage. That's the maturity gap nobody wants to name directly.

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

The thing I keep coming back to is that we've got maybe six or seven observability tools in our stack right now — Datadog, some Grafana, a little Honeycomb experimentation — and every single vendor is slapping "AI-powered insights" on their marketing, but when you actually kick the tires, it's glorified anomaly detection we could've built ourselves three years ago. What I'm genuinely wrestling with is whether to wait for the tooling to mature or start building opinionated internal abstractions on top of these APIs before my engineers drown in alert noise. The security implications of piping production telemetry through some vendor's LLM backend also keeps me up at night — nobody's having that conversation clearly enough. And honestly, the vendor fatigue is real; I don't want to sign another three-year contract for something that gets commoditized in eighteen months.

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 petabytes of telemetry data across our microservices architecture, and the AI tooling that vendors are selling us is basically just "we'll summarize your logs better" — that's not solving anything meaningful. What I actually need is causality, not correlation — I need a system that can tell me *why* something is degrading, not just pattern-match on what degraded last time and surface a pretty dashboard about it. The maturity gap I keep running into is that most of these "AI-powered observability" tools are doing glorified anomaly detection dressed up with GPT wrappers, and my team sees through that immediately.

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

Good looks like AI that's actually embedded in the signal path — not bolted on as a chatbot wrapper over my existing dashboards. I want anomaly detection that reasons across my entire telemetry graph, correlates a latency spike in service A with a memory leak in service B three hops away, and surfaces that *before* my on-call engineer gets paged at 2am. Where am I today? Honestly we're at maybe 30% of that vision. We've got some decent threshold-based alerting with a thin ML layer on top in Datadog, and we're experimenting with their AI features, but it still feels like I'm paying for a sports car and only using it to idle in a parking lot. The models don't have enough context about *my* system's topology to be genuinely useful — they're pattern-matching on generic signals, not reasoning about my specific service dependencies and SLOs.

What would change your perspective on this entirely?

Honestly? If I saw a vendor actually demonstrate — not pitch, *demonstrate* — that their AI caught a real incident faster than my on-call engineer would have, with full explainability on *why* it fired and zero false positives in a 30-day window, I'd be genuinely moved. Right now everyone's showing me demos with cherry-picked data sets and calling it production-ready, which is the same song we heard from every APM vendor for the last decade. The thing that would really flip my thinking is if the security posture around these AI observability tools was airtight — right now I'm being asked to pipe my most sensitive telemetry data into someone's cloud model and just trust them, which is a non-starter for me at a Series C where we're actively pursuing enterprise deals with compliance requirements. Show me a SOC 2 Type II, show me on-prem or VPC deployment options, show me the data residency story — *then* we can talk about whether the AI layer is actually worth evaluating.

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

The question nobody asks me is: "What does your observability data *actually cost you* to store and query, and is the AI layer making that better or worse?" Everyone wants to talk about the glamorous stuff — anomaly detection, AIOps, natural language querying of your logs — but nobody's doing the math on the fact that you're now running LLM inference *on top of* already expensive telemetry pipelines, and your Datadog bill just went from painful to genuinely absurd. I've watched AI-enhanced observability features triple our data retention costs because suddenly everything looks "potentially relevant" to the model so nothing gets filtered aggressively at the edge anymore. The vendors are selling you insight, but they're not showing you the unit economics of intelligence at scale — and that's the conversation that actually determines whether any of this survives a board-level cost review.

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

Honestly, the thing I keep banging my head against is the gap between what AI observability tools *promise* and what actually ships value to our on-call engineers. Like, we've been trialing a couple of tools — we looked at Datadog's AI-assisted anomaly detection and some of the newer stuff from Honeycomb — and the demo is always incredible, but then you get into actual usage and it's like... the signal-to-noise problem just moved one layer up. Instead of alert fatigue, now you have *AI recommendation fatigue*. The other thing I'm genuinely wrestling with is the organizational buy-in piece. Our CTO is bullish on AI everything right now, which creates this pressure to ship "AI-powered" features into our observability stack before we've actually validated whether engineers trust the outputs enough to act on them. I'm a research-first person — I want to talk to the on-call engineers, understand their actual mental model during an incident, and *then* figure out where AI fits — but there's this top-down pressure that's basically skipping that whole discovery phase. And honestly the maturity question is real — I don't think most fintech engineering orgs, including ours, have the data hygiene and instrumentation baseline that makes AI observability actually useful. You can't AI your way out of bad telemetry fundamentals.

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

The biggest gap I keep running into is that everyone's treating AI in observability like it's purely a tooling problem — just drop in some AIOps platform and suddenly your on-call rotation is magically smarter. But the actual blocker is way more people-and-process than it is technology. Like, I've seen orgs spin up Datadog's AI features or New Relic's anomaly detection and then wonder why adoption is basically zero six months later — it's because nobody changed how the team actually responds to alerts or investigates incidents. What I really want to understand is where that maturity gap actually lives — is it in the signal quality upstream, is it in trust gaps where engineers just don't act on AI-generated recommendations, or is it organizational stuff like nobody owns the observability practice at all? From a PM lens, if I can't identify the specific job-to-be-done that AI is solving in that workflow, I can't build or buy toward it intelligently.

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

Honestly, "good" for me in observability with AI looks like: the system surfaces the *so what* before I even have to ask. Like, not just "here's an anomaly spike at 2am" but "here's why it happened, here's the blast radius, here's what your on-call engineer should do first." Right now we're basically still in dashboard hell — beautiful Grafana boards that tell you numbers but make you do all the cognitive work yourself. We're probably a solid 18 months away from that north star, and the gap isn't really the tooling — it's that our observability data is messy enough that any AI layer sitting on top of it is just going to confidently hallucinate correlations. The "garbage in, garbage out" problem is very real and nobody wants to say it out loud because it's unglamorous work. We've got instrumentation gaps, inconsistent naming conventions across services, and honestly the team has tribal knowledge that never made it into structured logs, so the AI has no shot at the full picture yet.

What would change your perspective on this entirely?

Honestly, if I saw a vendor actually demo AI observability that could close the loop autonomously — like, it detects an anomaly, traces it to a root cause, proposes a fix, and then validates that fix didn't break anything else — without a human in the middle rubber-stamping every step, that would shift my thinking pretty dramatically. Right now everything I've seen is glorified alerting with a ChatGPT wrapper on top, and the sales pitch is always way ahead of the actual product. The other thing that would move me is seeing real adoption data from engineering orgs that aren't just the vendor's cherry-picked case studies — like, talk to me about the teams that tried this and hit walls, what were the actual blockers, how did they get past them. The honest failure stories are way more useful to me than another "we reduced MTTR by 40%" slide, because that's where the real maturity picture lives. Show me the messy middle, not just the success theater.

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

Honestly, the question nobody's asking is: **"Who actually owns the alert noise problem before you layer AI on top of it?"** Like, everyone's racing to put AI on their observability stack, but if your underlying data quality is garbage and your on-call runbooks are nonexistent, you're just automating chaos. I've watched vendors come in and pitch "AI-powered anomaly detection" and I'm sitting there thinking — cool, but your model is training on a signal-to-noise ratio that's basically 10% signal. The more interesting conversation to me is about the organizational accountability gap: is this an SRE problem, a platform team problem, a PM problem? Nobody wants to own it, so AI becomes this convenient scapegoat where you can say you're "doing something" without actually fixing the process underneath.

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

Honestly, the thing that's keeping me up right now is this gap between what AI observability tools *promise* and what actually happens when you put them in front of an engineering team. We've been piloting a couple of AI-assisted alerting and anomaly detection layers on top of our existing stack — we're using Datadog pretty heavily — and the signal-to-noise problem hasn't gone away, it's just gotten fancier. Like, the AI is confidently surfacing "insights" but my engineers are still doing the same triage dance they were doing before, just with an extra step of validating whether the AI recommendation is even remotely grounded in our system context. What I'm really wrestling with is the adoption question from a product lens — because I think about this the same way I'd think about any new feature rollout. Who's the actual user, what's their workflow, and are we solving a real pain point or are we just bolting AI onto a process that's fundamentally broken at a different layer? Right now I feel like we're in that hype-driven phase where leadership sees "AI observability" and thinks it's a checkbox, but the engineering team's trust in the tool is close to zero because the false positive rate is still brutal. That trust deficit is the real blocker, not the technology itself.

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

Honestly, the thing I keep coming back to is the signal-to-noise problem — our observability tooling is generating *so much* data that even with AI layered on top, we're not actually getting faster to root cause, we're just getting faster summaries of noise. The real question for me is whether AI in observability is genuinely reducing mean time to resolution or just creating a prettier dashboard that engineers still have to manually interpret. From a PM lens, I want to understand what the actual workflow looks like before and after AI gets introduced, because I've seen too many teams buy into a tool, bolt it on, and call it an AI-powered observability practice without any real before/after measurement. It's the same trap as agile theater — you're going through the motions without validating the outcome. The maturity gap isn't just technical, it's that most orgs don't even have a baseline to measure improvement against.

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

Good, to me, looks like AI that's embedded in the observability loop in a way that actually closes the feedback cycle — you get an anomaly surfaced, you get probable root cause with context, and an engineer can make a decision in under five minutes without digging through five different dashboards. The tooling should fade into the background and just accelerate the human judgment call, not replace it. Where we are today? Honestly, we're at like a 3 out of 10. We've got Datadog with some of their AI features turned on, but it's mostly noise reduction and pattern flagging — it's not actually giving engineers actionable synthesis, it's still dumping a wall of information and calling that "AI-powered." The gap between the marketing and the actual workflow improvement is genuinely frustrating when you're trying to build a lean, fast-moving eng team.

What would change your perspective on this entirely?

Honestly, the thing that would flip my thinking the most is seeing actual before-and-after data on mean time to resolution that's *attributable* to AI observability tooling specifically — not just "we also hired three more SREs and retooled our on-call rotation" mixed in there. Right now every vendor deck I see is correlation masquerading as causation, and as someone who's spent years trying to separate signal from noise in user research, that pattern is really obvious to me. The other thing — and this is more structural — if I saw an engineering org that had genuinely figured out the people-and-process side *before* layering in the AI tools, and could show me that the tooling then actually scaled the gains, I'd be a real believer. Every case study I've dug into so far, the orgs that "succeeded" with AI observability already had strong incident culture, clean telemetry pipelines, and psychological safety around postmortems. So was it the AI, or did they just already have their house in order? That's the question I can't get a straight answer on from anyone.

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

Honestly, the question nobody's asking is: **"What happens to institutional knowledge when AI starts summarizing your incidents?"** Like, everyone's so focused on "can the AI detect the anomaly faster" — which, fine, yes, sometimes — but nobody's interrogating what gets *lost* when engineers stop writing postmortems themselves and just let a model synthesize it. That's where the real nuance lives, the stuff that doesn't exist in any training data — "oh yeah, that alert fires every time payments does a deploy on a Tuesday because of this legacy quirk" — and I worry we're optimizing for speed at the expense of that organizational memory. At my company we're already seeing junior engineers who just trust the AI summary without digging in, and that's a maturity problem nobody's tracking.

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

Honestly? Attribution is eating my brain right now, and it has nothing to do with observability in the engineering sense — but when I heard this study was about AI in observability workflows, my first thought was "God, I wish we had that figured out on the GTM side too." From what I see bleeding over from our eng team, they're drowning in alert noise and they've got these AI-assisted monitoring tools that are supposed to surface what matters, but nobody's actually confident the signal is real — it feels a lot like our lead scoring problem where the AI says "this is a hot lead" and sales ignores it because they don't trust the model. The maturity gap I keep hearing about from our engineers is less about the tooling existing and more about nobody having built the feedback loops to make the AI actually learn what "bad" looks like in production. That's the same thing I see in martech — we deploy six tools, they all have "AI baked in," and none of them talk to each other so the intelligence is basically siloed and useless.

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

Honestly? Attribution and pipeline efficiency — those are my north stars for basically everything. Like, I'm constantly trying to figure out which channels are actually moving the needle versus which ones just *look* good in a last-touch model, and that problem bleeds into every tool decision I make. So when I'm evaluating anything new — AI, observability, whatever — I'm asking "does this help me spend smarter and prove it?" because at Series A you don't get the luxury of vague brand metrics, the board wants CAC payback and they want it now. The biggest thing I need to solve is closing the feedback loop between what I'm spending and what's actually converting to closed-won, and most tools I've seen are still pretty garbage at that handoff part.

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

Good looks like me being able to walk into a pipeline review and say "here's exactly where we're leaking, here's the channel mix that's actually driving qualified opps, and here's the CAC trend by segment" — all without spending three hours in spreadsheets beforehand. Like, the data tells the story automatically and I can just act on it. How far am I? Honestly, pretty far. We've got HubSpot, we've got some intent data from Bombora, we've got Clearbit enrichment — and I still can't get a clean answer on which touchpoints are actually closing deals because sales never feeds the closed-won data back properly, so nothing learns. It's the classic feedback loop problem that I think most B2B teams just accept as a fact of life but shouldn't.

What would change your perspective on this entirely?

Honestly? If someone showed me undeniable ROI tied directly to pipeline impact, I'd flip immediately. Like, not "we reduced MTTR by 23%" — I don't care about that metric, my engineering counterparts do — but show me that faster incident resolution meant we didn't churn that $200k ARR account, *that* changes the conversation for me. The other thing that would move the needle is if the attribution actually closed the loop. Right now when I talk to our engineering team about observability tooling, it's a completely separate world from the revenue impact — nobody's connecting "AI caught this anomaly" to "here's what that was worth to the business." If a vendor could actually demonstrate that feedback loop end-to-end, with real customer data, not a deck full of logos and vague efficiency claims, I'd be a genuine believer instead of a skeptic sitting in budget meetings.

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

Honestly? Nobody ever asks me "what does good AI actually look like when it's working?" Everyone's obsessed with the adoption question — are you using it, how much, what's your maturity score — but nobody asks what the outcome looks like when it's firing on all cylinders. For me it's not about the tool, it's about whether my pipeline metrics are cleaner and my CAC is defensible at board time — that's the real north star. I'd also love someone to ask about the feedback loop problem, because that's where everything dies: AI tools are only as smart as the data you feed back into them, and most orgs, including mine honestly, are terrible at closing that loop between what converted and what didn't. That's not a technology gap, that's a process and culture gap, and nobody wants to have that uncomfortable conversation.

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

Honestly, the thing that keeps me up at night right now has nothing to do with observability directly — it's that my eng team is drowning in alert noise and I can't get clean data on what's actually causing pipeline drop-off versus what's a system issue versus what's a campaign issue. Like, we're spending real CAC dollars driving traffic and I can't tell if conversion problems are a product bug, a latency thing, or just bad targeting. The attribution headache bleeds into everything — if observability tooling had better AI that could surface "hey, this degradation event correlated with a 23% drop in trial completions last Tuesday," that would actually change how I allocate budget. Right now I'm getting dashboards with a thousand metrics and nobody's telling me which ones connect to revenue outcomes. It's the same problem I see in martech honestly — everyone's baked AI into their product but it's mostly AI-washed noise, not signal.

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

Honestly, my brain doesn't live in observability land day-to-day — that's more of an eng problem — but from where I sit, the thing that keeps me up at night is whether the tools my engineering team uses actually translate into pipeline impact. Like, if observability AI is catching incidents faster, does that mean our uptime SLAs look better in sales decks? Does it reduce churn signals we can actually market against? The disconnect I see is that eng buys these tools in a silo, marketing never hears about the wins, and we can't close the loop on whether the investment is actually moving revenue metrics we can point to. That attribution gap between "our stack is more reliable because of AI" and "here's how that converts to retention or expansion ARR" — that's the unsolved thing for me.

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

Honestly, "good" for me is full-funnel visibility where I can see exactly which channels are driving pipeline that actually closes — not just MQLs that evaporate. Like, I want to open a dashboard on Monday morning and know within five minutes where to double down and where to cut, without spending two hours in spreadsheets trying to reconcile Salesforce data with HubSpot data with LinkedIn campaign manager. How far am I from that today? Pretty damn far. The feedback loop between what marketing touches and what sales actually closes is still mostly broken — sales reps don't update the CRM consistently, so any "AI-powered" attribution I'm theoretically getting from my tools is garbage-in-garbage-out. I've got like four tools that all claim to do attribution and they all tell me something different, which means I basically trust none of them.

What would change your perspective on this entirely?

Honestly? If I saw a vendor actually close the feedback loop end-to-end without me having to duct-tape five tools together. Like, right now my mental model of AI in observability is basically "fancy alerting with a confidence score slapped on it" — if someone showed me a system that actually learned from incident resolutions and fed that back into pipeline attribution or customer health scoring in a way that moved CAC, I'd be all-in immediately. The thing that would really flip me is seeing a real case study — not a Gartner magic quadrant slide, an actual company our size — where engineering and marketing are sharing the same signal layer and it's visibly compressing time-to-pipeline. Show me that and I'll go to bat for the budget internally tomorrow.

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

Honestly? Nobody ever asks me "what does good observability data actually do for your pipeline model?" Like, I sit on the demand gen side and I'm constantly fighting for budget based on what's converting, but if the engineering team's observability stack is throwing false alerts or missing incidents, that directly tanks our SLAs, which tanks our customer retention, which tanks the expansion revenue I'm counting on in my attribution models. The question I want someone to ask is: "how does observability maturity actually connect to revenue outcomes?" Because right now it feels like eng teams are optimizing their alerting dashboards in a vacuum while I'm over here watching churn spike in accounts where we had three incidents last quarter, and nobody's drawing that line explicitly. Someone needs to build that bridge between the ops data and the GTM data, and I genuinely think that's where the biggest unlock is — not just "is AI finding anomalies faster" but "is finding anomalies faster actually protecting the revenue base?"

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

Honestly, what's keeping me up at night is the gap between what vendors are promising around AI observability and what I'm actually seeing deployed at scale in real engineering orgs. I sit in a lot of cross-functional meetings — we're a Series B, so I'm close to the product and eng teams — and I keep hearing "we're piloting AI for anomaly detection" or "we're experimenting with AIOps" but nobody can show me a before/after on MTTR or incident costs that actually holds up to scrutiny. The thing I wrestle with most is whether this is a genuine maturity problem — like, orgs just aren't ready operationally — or whether the tooling itself is still mostly marketing-layer AI that's been sprinkled on top of existing observability stacks to justify a price increase. Because from where I sit, there's a massive difference between those two diagnoses, and they lead to completely different GTM and adoption strategies. I've been burned enough times by "AI-powered" dashboards that are just prettier regex to be deeply skeptical.

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

Honestly, the thing I keep coming back to is the ROI attribution problem. Everyone on the engineering side is excited about AI-powered observability — anomaly detection, automated root cause analysis, whatever — but nobody can cleanly show me the dollar value of a faster MTTR or fewer false positives waking up an on-call engineer at 2am. Until someone builds a credible business case framework around this, it's going to stay in the "cool demo" bucket for most organizations. I've seen this exact pattern with every wave of tooling — big promise, murky value, and then the procurement cycle stalls because the CFO asks one hard question and nobody has an answer.

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

"Good" to me is when your observability stack is actually closing the loop autonomously — you're not just getting alerted that something's on fire, the system is correlating signals, surfacing root cause, and ideally suggesting or executing remediation without someone having to page through dashboards at 2am. Think of it like the difference between a smoke alarm and a sprinkler system. Right now, most of what I'm seeing in the market — even from vendors slapping "AI-powered" on everything — is still basically a fancier smoke alarm. We're probably at a 3 out of 10 on that maturity scale, and the gap isn't really a technology problem anymore, it's a data quality and organizational trust problem — engineers don't trust the AI recommendations enough to let them act, and honestly given how noisy most alerting environments are, I don't blame them.

What would change your perspective on this entirely?

Honestly? Show me a P&L impact. Not "we reduced MTTR by 23%" in a vendor case study that conveniently never names the company or shows the baseline — I mean a real CFO-signed-off number that says AI in observability saved us X dollars or accelerated revenue by Y. The thing that would flip me is seeing an engineering org actually attribute cost savings from reduced incident response to a business outcome that shows up in the financials, not just an engineering KPI dashboard. Right now it feels like we're measuring the wrong things — everyone's optimizing for "faster alerts" when the actual question is "did we retain that enterprise customer because we caught the issue before they noticed," and nobody's connecting those dots.

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

Honestly? Nobody ever asks "what does success actually look like in 18 months, and how will you know you got there?" Everyone's so focused on the adoption metrics — seats activated, features used, dashboards built — but nobody's tying it back to whether the engineering org is actually shipping faster or catching incidents before customers do. I'd love someone to push me on the accountability loop: if we invest in AI observability tooling, who owns proving the ROI, and what happens when nobody can answer that question in Q4 review? Because right now that accountability gap is exactly why half these implementations stall out — you get the pilot, you get the excitement, and then nobody can defend the renewal because the success criteria were never defined upfront.

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

Honestly, what's keeping me up at night is the gap between what vendors are promising and what's actually moving the needle. I sit in on enough cross-functional meetings with our engineering and DevOps teams to know that everyone's got some AI "baked into" their observability stack — Datadog, New Relic, whoever — but when I ask "show me the ROI," I get a lot of hand-waving about reduced MTTR and faster incident response that nobody's actually baseline-measured. The real wrestling match for me is figuring out whether AI in observability is genuinely transformative or whether it's just the new version of slapping "machine learning" on your alerting thresholds and calling it innovation. I've been burned enough times in my agency days by vendors selling magic fairy dust to be deeply skeptical, and I need to see concrete before-and-after metrics, not a demo environment where everything works perfectly.

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

Honestly, the thing that keeps me up at night — and I know this is a research interview so I'll be direct — is the signal-to-noise problem. We're drowning in observability data, alerts firing constantly, and the promise of AI is that it can actually triage that noise into something actionable. But what I keep seeing is vendors slapping "AI-powered" on dashboards that are basically just glorified regex rules, and my engineering partners are rightfully skeptical. The real question I need answered is: where does AI in observability actually move the needle on MTTR and incident cost versus where is it just feature marketing to justify a higher price tier?

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

Honestly, "good" to me in observability with AI is when the system closes the loop without me having to babysit it — anomaly detected, root cause surfaced, relevant context pulled, and the right team paged, all before a customer even files a ticket. That's the dream state. Where we actually are today? We've got Datadog with some ML-based alerting baked in, and it catches stuff, but it's still generating noise I have to train my eng team to filter through — so we're maybe a third of the way there. The gap isn't really the detection piece anymore, it's that the AI can flag *something's wrong* but the "so what do I do about it" layer is still largely human, and that's where the time drain lives. It's like having a smoke detector but no sprinkler system.

What would change your perspective on this entirely?

Honestly? Show me the P&L impact. Not "we reduced MTTR by 23%" in a case study PDF that some vendor's customer success team wrote — I want to see what that actually cost the business before, and what it costs after, fully loaded with the implementation overhead, the engineering time to tune it, the ongoing licensing. If someone put a real before-and-after in front of me that wasn't AI-washed marketing copy — like actual incident cost reduction tied to revenue impact or customer churn prevention — I'd move from skeptic to advocate pretty fast. The problem is every vendor is selling me fairy dust right now, slapping "AI-powered" on dashboards that are basically just slightly smarter alerts, and I've been burned enough times in my agency days to know the difference between a tool that moves metrics and one that just moves slides.

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

Honestly, the question nobody asks is: "What does good actually look like two years from now, and how do you measure the delta from today?" Everyone's so focused on the adoption story — "are you using AI in observability, yes or no" — but nobody's building a rigorous baseline. I came from agency world where we had to justify every dollar with attribution models, and I look at how engineering orgs are evaluating AI observability tools and it's basically vibes and vendor-supplied case studies. You can't optimize what you don't measure, and right now most of these conversations are just theater — a lot of "we deployed Copilot and engineers seem happier" with zero connection to incident reduction rates or MTTR. The real question is whether anyone is actually treating this like a business transformation with KPIs attached, or whether we're just collectively convincing ourselves the ROI is there because the alternative — admitting we spent a ton on AI tooling and can't prove it worked — is too uncomfortable to say out loud.