Vlad: Here's the thing nobody wants to say out loud this week. A company doubles its AI spend, doubles its revenue in the same stretch, and the CEO still won't tell you one caused the other. That's not a failure. That's the most honest thing I've read from a tech executive in months.

Vlad: Because everyone else is out there claiming the connection. This guy looked at the timeline and said, I can't prove it. That's today's episode in one sentence: production AI is running ahead of anyone's ability to actually govern it, or even measure what it's doing for you.

Kate: And it's not just the ROI question, Vlad. The infrastructure underneath it is being built the same way, fast, ungoverned, patched together after something breaks.

Zoltan: Which is the entire history of enterprise IT, so, welcome to the club.

Vlad: Right. Let's get into it.

Vlad: Welcome to *Hack the Stack*, the weekly show for senior IT and AI leadership who actually have to deliver, govern, and maintain AI in production. I'm your host, Vladimir Cvijanovic, joined by Kate, our technology and data architecture lead, and Zoltan, our old-school IT skeptic. Every Tuesday, we skip the vendor marketing pitch and break down what's actually happening across the stack. Let me get right into this week's first topic.

Kate: So let's start with Temporal, because this is the case study everyone in our audience should actually read start to finish. Temporal, workflow orchestration company, about 500 employees, 200-plus engineers. AI spend is up roughly 5x this year. Revenue's roughly doubled since January. And the CEO, Samar Abbas, goes on record saying he can't prove the two are connected.

Vlad: Which is rare.

Kate: Extremely rare. Most CEOs would take that timeline and run a victory lap. He didn't. And what I actually care about is what's underneath that headline, because they didn't just buy tools and hope. They built real governance around this. Hard dollar caps on AI spend, reviewed every two weeks. Not a soft target, an actual cap. A company-wide ban on open-weight models for security and privacy reasons. A central AI team that has to vet any new model before it touches a real workflow.

Zoltan: Every two weeks. So for up to thirteen days, the spend can run wherever it wants, and then somebody looks at a spreadsheet.

Kate: It's a review cadence, Zoltan, not a blank check. The point is it exists at all. Most companies at this stage have nothing.

Zoltan: I'm not saying it's nothing. I'm saying "biweekly review" is doing a lot of work in that sentence. Somewhere there's an engineer who found that gap on day two.

Vlad: Fair, but let's not lose the more interesting part. Kate, what happened to code review at Temporal?

Kate: This is the part I think our audience will actually feel in their gut. Code generation got so fast that code review became the bottleneck, not writing the code. And that changed the culture. It's now considered completely normal to bounce back a colleague's AI-generated pull request immediately, no review comment softening, nothing. Before this, that would've read as rude.

Vlad: That's a real cultural shift, not a tooling one.

Kate: Right, and it had to be paired with a hard rule: anything touching critical infrastructure gets a human reviewer, full stop, no matter how the code was produced. They didn't let velocity override the one place where a mistake actually costs you.

Zoltan: And still. Five times the spend. The CEO's own quote, they've got "AI slop" showing up in code, in slide decks, and on their own public website. Their public website. That's not a hypothetical risk I'm inventing, that's the CEO of the company saying it out loud.

Vlad: Which is honestly the most useful line in this whole piece for our audience. Every Finance Director listening right now is being told by somebody that AI investment equals guaranteed return. Temporal is the counter-example. Right size, right governance mechanisms, actual controls in place, and the CEO still can't cleanly claim the win.

Kate: And I'd push back gently on reading that as a failure. Engineering shipped 20 to 30 percent faster with the same headcount. That's real. It's just not the same thing as "AI caused our revenue to double," and Abbas is smart enough to know the difference.

Zoltan: He's also smart enough to know a $4 million AWS bill from ungoverned agent spend, somewhere else, is the actual reason his own caps got set as hard as they are. That's the part that should scare people. He built his governance model off someone else's disaster, not off best practice.

Vlad: That's actually the through-line for a lot of what we're covering today. Nobody's building this stuff from a playbook. They're building it from a scar.

Kate: And I think there's a nuance worth sitting on. Temporal's headcount is right in the range our audience actually operates at, roughly 500 employees, 200-plus engineers. This isn't a hyperscaler with infinite budget experimenting for the sake of it. This is a company the exact size of a lot of the CEE mid-market firms we talk about every week, making the exact same bet, and still landing on "I can't prove it."

Vlad: Which should recalibrate expectations for anyone in that position right now.

Kate: Recalibrate, not discourage. The 20 to 30 percent delivery speed gain, same headcount, that's real and it's measurable on its own terms. The mistake would be assuming that number automatically rolls up into revenue. It doesn't, not cleanly, not on a quarter-by-quarter basis, and Abbas is honest enough to say so instead of drawing a straight line that isn't there.

Zoltan: Which is worth repeating for anyone skimming this episode for a quote to put in a slide deck. The CEO said he can't prove it. Not "we're still measuring it," not "the data's still coming in." Can't prove it. That's a different, more honest sentence, and people are going to quietly edit it into something more convenient before this week is out.

Vlad: That's why we're reading the actual piece and not the LinkedIn summary of it. Anything else on Temporal before we move on?

Kate: Just one thing. The model-vetting requirement, a central team has to approve any new model before it touches a live workflow, that's the piece I'd tell any IT director to copy first. It's cheap to implement, it doesn't slow anyone down meaningfully, and it's the single control that would have prevented most of the messes we're about to get into with the next story.

Zoltan: Cheap and effective. Write that down, it doesn't happen often in this industry.

Vlad: Which brings us to the second piece, and this one's got real numbers behind it. VentureBeat, reporting out of VB Transform, three companies, Replit, Kilo Code, and Symbotic, all separately dealing with the same problem: AI coding agents blowing straight through budgets.

Kate: And the numbers are wild once you actually sit with them. At Kilo Code, engineers write or read code themselves about one percent of the time now. One percent. Everything else is agents.

Zoltan: One percent. So what exactly are they paying these engineers to do.

Kate: Reviewing, directing, architecting, the actual judgment calls, Zoltan, which is the point. Their co-founder said it directly, cost per pull request is the metric she watches now, not lines shipped.

Vlad: Give me the number that made you stop reading.

Kate: One individual engineer at Kilo Code ran up a $600-a-day bill on agentic tooling before anyone flagged it. And at Symbotic, it gets worse, because it wasn't even an engineer. A support employee, outside engineering entirely, ran an automation on GPT-5.5 Pro Max and burned through what the article calls, quote, an insane amount of money, before anyone noticed.

Vlad: Wait, genuinely, outside engineering.

Kate: Outside engineering. That's the sentence I want every IT director listening to sit with. This is moving out of IT's visibility. Support, sales, ops, anyone with API access to an agentic tool can now generate a bill that used to require a whole cloud migration to rack up.

Zoltan: So the good news is, we've democratized the ability to accidentally spend a fortune.

Vlad: That's exactly it. So how are these three companies actually fixing it, because that's the useful part.

Kate: Three completely different approaches, and none of them off the shelf. Replit scores every AI-generated pull request with an automated risk score. Low risk merges itself. Higher risk routes to a human. They call it "human on the loop" instead of "human in the loop," which I actually think is a smart distinction. They also run the whole agent fleet inside isolated cloud VMs behind token proxies, so access is controlled at the infrastructure layer, not just at the code review stage.

Zoltan: Infrastructure layer control. Finally, someone remembered that's where control is supposed to live.

Kate: Don't get used to agreeing with me, it happens rarely enough as it is.

Vlad: Kate, keep going. Kilo Code and Symbotic.

Kate: Kilo Code decoupled the tooling from the model entirely, they've got a routing gateway that supports over 500 different models. Expensive frontier models get sent to architecture and planning work, cheap open-weight models handle routine execution. That's genuinely elegant engineering, you route spend to where the intelligence is actually needed.

Zoltan: That part, I'll give them. Cost-aware routing at that scale isn't a hack, that's real infrastructure thinking.

Vlad: Noted for the record, Zoltan liked something.

Zoltan: Don't make it a pattern.

Kate: And then Symbotic went the blunt route. Per-employee monthly cost tiers, on a manager-facing dashboard, so someone gets moved up or down a spending tier based on actual usage and role. And here's the part that tells you the real story, that system only got built after Cursor changed its pricing terms and killed a legacy flat-rate discount Symbotic had been running on. They didn't build governance proactively. They built it the week the free lunch ended.

Zoltan: Which is my entire point about all of this. Nobody builds the fence until the first person walks off the cliff. And the article itself admits there's no standard here. Three companies, three bespoke fixes, no shared playbook. Every company adopting this today is re-solving a problem three other companies each solved differently.

Vlad: That's a real gap. If you're a mid-market IT director in Bucharest or Warsaw right now, there's no vendor you can call and buy this from. You're building it yourself, from scratch, probably after your own version of the $600-a-day surprise.

Kate: Which is exactly why this matters more for our audience than a Silicon Valley trend piece usually does. This isn't "adopt AI coding tools." It's "adopt AI coding tools and build the cost governance layer at the same time, because nobody's going to sell it to you pre-built."

Vlad: Let's stay on the Symbotic detail for a second, because I think it's the most instructive part for anyone running finance and IT together. The governance didn't come from a risk assessment. It came from a vendor changing pricing terms.

Kate: Right, and that's a pattern our audience should actively plan for, not just react to. Legacy discounts, promotional pricing, free tiers, these all eventually normalize. If your entire cost model depends on a vendor's current pricing staying exactly where it is, you don't have a cost model, you have a countdown.

Zoltan: This is the technofeudal point again, whether anyone wants to hear it that way or not. You get comfortable on favorable terms, you build habits and headcount around those terms, and then the landlord raises the rent once you're dependent enough that leaving costs more than paying it.

Vlad: Zoltan, keep it concrete.

Zoltan: It is concrete. Symbotic is the concrete example. They didn't build governance because it was good practice. They built it the week the free lunch ended. That's not a hypothetical, that's literally what happened, per the article.

Kate: Which is exactly why I'd tell anyone listening, don't wait for your version of that week. Build the cost-tiering, the per-role budget visibility, the model-routing logic, before the pricing changes, not after. It's cheaper every time you do it early.

Vlad: And the one percent number from Kilo Code, does that hold up under scrutiny, or is that a flashy stat that falls apart if you push on it?

Kate: I think it holds up, with a caveat. One percent of an engineer's time actually writing or reading code by hand doesn't mean engineers stopped mattering, it means the job moved up a layer, toward direction, review, and architecture. That's consistent with what Temporal described too, review became the bottleneck, not writing. Two completely separate companies landing on the same structural shift independently is the kind of signal I actually trust.

Zoltan: Two independent data points agreeing is worth more than either headline alone. I'll concede that much.

Vlad: Okay, and this next one is, I think, the real spine underneath both of those stories. VentureBeat again, also out of VB Transform, five startups all racing to fix what the headline calls agents that can't talk to each other, can't be trusted with permissions, and can't be audited.

Zoltan: Five startups. Five different vendors, all telling you their piece solves the trust problem. This is starting to sound familiar.

Vlad: Say more.

Zoltan: It's the exact same shape as the last story. No standard, five companies patching five different holes in the same boat. Except this time the boat is "can an autonomous agent be trusted with access to your systems at all."

Kate: Which, to be fair to the piece, it doesn't hide that. One of the CTOs quoted, Ben Hylak at Raindrop AI, says it straight, "this problem is getting a lot worse as models and agents improve, and I think there's good reason to believe it will continue to get worse." That's not vendor spin, that's someone telling you the problem scales faster than the fix.

Vlad: Walk me through what each of these five is actually building, because "agent governance" is vague enough to mean nothing if we don't get specific.

Kate: Fair. BAND builds a coordination layer, compatible with both A2A and MCP protocols, that lets a human step directly into an agent-to-agent conversation mid-task, not just review the outcome after the fact. Their CTO put it plainly, the transport layer has to be solved first, how the agents actually talk to each other in real time.

Zoltan: Which they still haven't solved, they've built a way to eavesdrop on the conversation while it's happening. That's not the same as solving it.

Kate: It's a meaningful step, Zoltan, being able to intervene mid-task instead of only auditing afterward is a real capability gap being closed.

Vlad: Next.

Kate: Raindrop AI, the one Hylak's quote came from, builds an audit log that captures every message, tool call, retry, and error an agent produces, all in one unified log. And this is the part I actually find clever, they can simulate a proposed fix against real past user behavior before it ever goes live, specifically to catch side effects before they hit production.

Vlad: That's a genuinely useful idea, a staging environment for agent behavior changes, not just for code changes.

Kate: Exactly. Then Arcade.dev, they're building least-privilege, time-scoped authorization for agents, so every action is attributable to an exact moment with the smallest possible scope of access. It can run on-prem inside what they call a clean room, gated by your existing identity provider and role-based access controls.

Zoltan: On-prem, gated by existing RBAC. Now that's a sentence I can respect. Their CTO said the abstraction has been wrong, and honestly, he's not wrong. Most of security tooling for agents has been bolted onto human-identity assumptions that don't hold.

Vlad: Second one today.

Zoltan: Don't get excited, it's still five vendors solving five slices of one unsolved problem. Being right about the diagnosis doesn't mean the industry has the cure.

Kate: And Conifers applies all of this specifically to cybersecurity, breaking down the silos between threat intelligence, threat hunting, detection, and response, so one agent can move across all four without a human relaying information between separate tools. Their claimed numbers are dramatic, incident containment time down from seven hours to twelve minutes, complex investigations down to under four minutes.

Zoltan: Those are vendor numbers. I'd want to see that audited by someone who isn't selling the product.

Vlad: Reasonable. And the fifth?

Kate: Omilia, customer experience agents, over three billion calls a year, claims 30 to 45 percent faster time to resolution and 21 times more upsell revenue compared to human agents. But here's the honest part of their pitch, even their most mature deployments only run at 80 to 90 percent automation. Their own CPO says human in the loop is still, quote, very fundamental.

Vlad: So even the vendor with the biggest scale and the best numbers is telling you not to go fully autonomous.

Kate: Right. Which I think is the single most credible data point in the whole piece. If the company with three billion calls a year and the strongest ROI story still won't cross into full automation, that's not caution for caution's sake. That's them telling you where the actual ceiling is right now.

Zoltan: The technofeudal read on this is simple. Five well-funded startups building five incompatible moats around the same unsolved problem isn't progress, it's five landlords racing to own the fence before anyone agrees on where the property line goes.

Vlad: Zoltan, don't turn this into a lecture, give me the practical version.

Zoltan: The practical version is, whichever of these five you pick, you're locked into their protocol, their audit format, their idea of least privilege. There's no portability yet. Choose carefully, because switching later won't be free.

Vlad: That's a fair, concrete warning, actually. Kate, does that land, or are you going to push back?

Kate: No, that one lands. I'd frame it less darkly, but the underlying point is right. This is greenfield. Whatever you adopt now, you're making an architectural bet, not just a purchasing decision.

Vlad: Let's talk about the protocols for a second, A2A and MCP, because I think a lot of listeners hear those acronyms and tune out, and that's a mistake.

Kate: It is, because this is genuinely foundational. MCP, Model Context Protocol, is how an agent gets structured access to real data and tools instead of guessing from a prompt. A2A is the emerging standard for how one agent hands work to another agent. Neither one is fully mature, but BAND explicitly building compatibility with both, instead of picking one horse, tells you where the smart money thinks this is heading, toward interoperability, not a single winner-take-all standard.

Zoltan: Or it tells you BAND doesn't want to bet wrong and get stranded on the losing protocol. Hedging isn't the same as insight.

Kate: It's also not the same as being wrong, Zoltan. Sometimes hedging is the correct engineering decision when the standard genuinely hasn't settled yet.

Vlad: Let's come back to Omilia for a second, because that stat, three billion calls a year, is enormous, and I want to make sure we're not letting the scale distract from the actual lesson.

Kate: The lesson isn't the scale, it's the ceiling. Even with three billion calls of production data behind them, they're capping automation at 80 to 90 percent on purpose. That's not a technology limitation they haven't solved yet, their own CPO frames it as a deliberate choice, human in the loop is fundamental. That's the most senior, most experienced voice in this entire piece telling you where the actual limit is right now, not where the marketing wants you to believe it is.

Vlad: So if the most mature deployment in the industry still holds the line at 90 percent, anyone telling a CEE mid-market client they can run fully autonomous agents in production today is either ahead of the entire industry, or lying.

Zoltan: I know which one I'd bet on.

Vlad: Last one, and this is the one I think ties everything else together. Forget agents for a second, this is about the data underneath them. Google Cloud Blog, a piece on mainframe modernization, and the argument is basically, everyone's framing this wrong.

Kate: This is the one I was most excited to get into. The author's core point is that treating mainframe modernization as "convert COBOL to Java, line by line" is the wrong model entirely, because the real complexity isn't the code, it's decades of tightly coupled dependencies. Proprietary data formats like VSAM, flat files, IMS hierarchical structures, none of which an AI agent can just read directly. Transaction monitors like CICS and IMS TM, where a single transaction scenario can span millions of lines. And then integration protocols binding the whole thing together at the edges, CTG, IMS Connect, MQ, LU 6.2 Sockets.

Zoltan: Finally, someone at a hyperscaler admits the mainframe isn't a legacy embarrassment, it's genuinely hard engineering that most of the industry has forgotten how to respect.

Vlad: That's the second grudging compliment out of you today, Zoltan. Are you feeling alright?

Zoltan: I've spent thirty years watching people who've never touched a mainframe explain how simple it would be to replace one. It's satisfying to see it acknowledged in writing for once. Don't read too much into it.

Kate: So their approach, four pillars, assessment, modernization, de-risking, and data migration. What I find genuinely well-designed is the assessment tool, it reverse-engineers the existing codebase before anyone touches anything, producing dependency maps, extracting business rules into both plain-language requirements and visual decision trees, and generating technical documentation automatically. All of that feeds into the actual modernization agents through MCP, so they're working from real application context instead of guessing at what forty-year-old COBOL is actually supposed to do.

Vlad: And then there are two different paths depending on the workload, is that right?

Kate: Right, this is the part I think is the most mature idea in the whole piece. If the business logic itself is a competitive differentiator, worth improving, you take a "rewrite and reimagine" path. But if it's something like a core general ledger, where regulatory compliance depends on the output matching exactly, you take a stricter "deterministic, like-for-like" path. Byte-for-byte identical output for any given input, only the underlying data structure gets modernized, the compliance logic stays untouched.

Zoltan: And how exactly do you prove "byte for byte identical" before you bet the company on it.

Kate: That's their de-risking product, Dual Run. Before you retire the legacy system, it runs live production transactions through both the old mainframe and the new cloud environment simultaneously, and compares the outputs side by side, protocols, messages, data changes, all of it, until the two systems prove complete equivalence. Only then does the old system actually get switched off.

Zoltan: That's a real de-risking mechanism, not a marketing slide. Running old and new in parallel until they provably agree, that's how you're actually supposed to migrate something you can't afford to get wrong.

Vlad: Three times, Zoltan.

Zoltan: I said it's a real mechanism, I didn't say I trust the company selling it to me for free. This is still Google Cloud's own blog, promoting Google Cloud's own product suite. There's no cost, no timeline, anywhere in this piece. Try pricing this out and see how quickly the tone changes.

Vlad: That's a fair caveat to put on the record, this is vendor content, and it should be read that way. But the architecture itself, the idea that you assess deeply before you touch anything, and pick a rewrite-versus-preserve path based on whether the logic is actually a competitive asset, that's sound thinking regardless of who's selling it.

Kate: And it directly ties to why the first three stories matter at all. None of the spend governance, none of the agent trust infrastructure, matters if the data your agents are supposed to reason over is still locked inside a format from 1985 that nothing modern can read. This is the foundation the other three stories are quietly assuming already exists.

Vlad: Walk me through the financial-services example they used, because I think concrete beats abstract every time on this topic.

Kate: It's a composite example, not a named client, but it's specific enough to be useful. Stable, high-volume back-office batch jobs, nightly statement processing, get the like-for-like path, cut mainframe processing costs, minimal risk, because the logic itself isn't a differentiator worth touching. The core general ledger gets the strict deterministic path, moves to Cloud SQL for better analytics, but the compliance logic stays byte-for-byte identical. And the customer-facing loan origination platform, which they consider a genuine competitive differentiator, gets the full rewrite path so it can support real-time approvals.

Zoltan: Which is the correct hierarchy, for once. Touch the thing that makes you money last, or not at all, and touch the thing that's pure operational cost first.

Vlad: That's a rare full sentence of agreement from you today.

Zoltan: Don't misread it as optimism. I still want to know what happens when Dual Run finds a mismatch between the old and new system on transaction ten thousand, after nine thousand nine hundred ninety-nine matched perfectly. Does the whole migration stop, or does someone quietly wave it through because the project's already six months behind schedule.

Kate: That's a completely fair operational question, and to be clear, the article doesn't answer it. It describes the comparison mechanism, not the governance process around what happens when the comparison fails. That's a real gap.

Vlad: Note that as an open question for our audience to ask directly if they're evaluating this. Anything else on the mainframe piece before we tie it together?

Kate: Just that I'd flag this as required reading for any CEE mid-market company in banking, insurance, or manufacturing still running core systems on a mainframe, which is a lot of them. The four-pillar structure, assess deeply, then choose your modernization path deliberately, then de-risk with parallel running, then migrate the data, that's a sound sequence regardless of which vendor's tools you actually use to execute it.

Vlad: So let's pull the thread together, because I think it's a real one. Temporal shows you can build governance and still not be able to prove your ROI. Replit, Kilo Code, and Symbotic show you what happens when governance lags adoption, six-hundred-dollar-a-day surprises and a support employee accidentally burning the budget. The five agent-trust startups show that the infrastructure to even make governance possible at the protocol level doesn't exist yet as a standard, everyone's building their own incompatible version. And Google's mainframe piece shows the whole conversation is moot if your underlying data isn't even in a shape your agents can use.

Kate: Which, put together, tells you exactly where a CEE mid-market company or a PE portfolio company needs to actually start. Not with the flashiest agent demo. With an honest assessment of your data foundation, and hard spend caps before you scale anything, not after.

Zoltan: And a very healthy suspicion of anyone telling you this part is already solved. It isn't. Four different stories this week, four different companies, and not one of them has a clean, repeatable answer. That should tell you something.

Vlad: It tells me the companies actually worth learning from this week weren't the ones with the best AI story. They were the ones honest enough to say what's still broken.

Kate: And there's a sequencing lesson buried in here too, one I don't think gets said out loud often enough. Every company we covered today learned its lesson in almost the same order: adopt fast, discover the gap the hard way, then build the governance after the fact. Not one of them designed the controls first and the adoption second.

Vlad: Which is the exact order our audience has the chance to reverse, if they're paying attention this week instead of next year.

Zoltan: That would be a first. Usually the scar comes before the fence, every single time, in every organization I've ever worked in. I'll believe it when I see a client actually do it the other way around.

Kate: Then let's make sure at least one of them does, Zoltan. That's kind of the point of the show.

Vlad: So here's the takeaway, and it's a specific one, not a platitude. If you're running an AI program right now and nobody in your organization can answer "what does an hour of unsupervised agent activity actually cost us, and who's watching it," you don't have an AI program, you have an open tab, and you don't know the total yet.

Kate: Fix that before you scale anything else. Genuinely, that's the cheapest problem to solve on this list, and the most expensive one to ignore.

Zoltan: And check what format your data's actually sitting in before anyone promises you an agent that understands it.

Vlad: That wraps up this week's episode of *Hack the Stack*. If you're building or scaling AI infrastructure in your organization and want to discuss how to solve these delivery or governance bottlenecks, let's talk. You can book a call directly via the link in the show notes, or send me a direct message on LinkedIn.

Kate: We'll be back next Tuesday with another breakdown of what's real and what's hype across the stack.

Zoltan: Try not to deploy anything into production before then.

Vlad: Until next week.