29 Jul 2026 · 29m · Shahir Daya · 20 quoted facts
Shahir Daya of Zafin argues that banks are deploying AI agents faster than they can govern them. Teams run hundreds of agents across different models and data, yet most institutions cannot say who authorized an action, what it touched, or what it cost. His answer is a central control tower covering orchestration, access guardrails, proof of work and cost tracking, with oversight that varies by workflow instead of one rule for everything.
The central tension is speed versus defensibility. Early agent wins come quickly, and Daya describes pilots that multiply in months. But regulators hold the institution accountable and token costs can jump sharply, so missing controls surface around a year in. Retrofitting them takes far longer than building them in. He urges starting with small, well-understood, boring workflows and governing the workflow, not each individual agent.
Daya says one agent can work across many workflows and one workflow can use many agents, so agent-level controls do not scale. Regulators ask about risk, human judgment and evidence at each workflow step.
Daya cites Safe Software going from about $20,000 to $100,000 a month on AI tools in six months. He also cites Uber exhausting its 2026 AI coding budget by April, which shows why per-task cost tracking matters.
Daya says existing operating models were built for human-centric workflows and cannot answer who authorized an agent, what data it used, or where a human stepped in. That leaves a material governance gap.
Zafin treats agents with zero trust and least privilege, like staff, and logs prompts, plans and every step as unchangeable proof of work. That evidence is what lets an institution defend agent actions to auditors.
Daya describes deploying fifty agents in ninety days, then discovering around month twelve that cost, risk and accountability cannot be explained. Bolting controls on afterward takes roughly eighteen months.
Agents differ in how long they run, what data they touch and which tools they call, so a single rule either blocks useful work or leaves risky agents under-watched. Daya compares the fix to an air traffic control tower that gives different flights different clearance.
It is a central layer that tracks every agent's identity, authority, data, tools, budget and evidence, much as an airport tower tracks every aircraft. Daya says it combines orchestration, guardrails, proof of work and cost governance.
It is a permanent record of what an agent did, including its prompt, its phase-by-phase plan and a full transcript of every step. Daya says this is what lets a regulator see what an agent did and why.
Token usage is a variable compute cost, not a fixed software fee, and it can grow fast. Daya also says AI providers are moving to usage-based billing, with analysts projecting enterprise AI bills 30 to 50% higher.
Daya says to start with a boring, well-understood workflow where the fix is localized and the outcome is measurable. He warns against leading with a flashy transformation project that looks impressive but is hard to prove.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Zafin is described as an AI platform company that provides governed agent orchestration and control capabilities for regulated institutions.
“Zafin is an AI platform company that provides governed agent orchestration and control capabilities for regulated institutions.”
Shahir Daya says that in cybersecurity, institutions that acted early were better positioned than those that waited for a crisis.
“the institutions that acted early were positioned much, much better than the ones that waited for a crisis to happen”
Shahir Daya says Uber's president publicly said the company used up its entire 2026 AI coding budget by April and could not link the token surge to better customer outcomes.
“saying the company burned through its entire 2026 A I coding budget by April in four months. And they could not link the surge in token consumption to better outcomes for customers.”
Shahir Daya says Safe Software went from about $20,000 to $100,000 a month on AI tools in six months.
“six months ago, they were spending about 20,000 a month on A I tools. Today, they're spending $100,000 a month.”
Shahir Daya says Anthropic moved its enterprise customers from flat-rate plans to usage-based billing earlier in the year.
“You saw Anthropic, they moved from they moved their enterprise customers from a flat rate plans to usage based billing”
Shahir Daya says analysts project AI pricing will normalize to real infrastructure costs, raising enterprise AI bills another 30 to 50% above current levels.
“Analysts are actually projecting that the pricing normalizes, you know, the pricing normalizes to reflect real infrastructure costs. Enterprise A I bills will go another thirty to 50% above current levels.”
Shahir Daya says Lloyds Banking Group recently put an AI agent in its boardroom to help directors analyze materials and prepare for meetings.
“I think Lloyd's Banking Group recently put an A I agent in its own boardroom to help directors actually analyze the materials and prepare for a meeting.”
Shahir Daya says North American regulators have warned that AI summarization tools may cause directors to miss critical details by not reading materials in full.
“I believe the North American regulators have already raised concerns about that, warning that A I summarization tools may actually cause directors to miss critical details by not reading the materials in full.”
Shahir Daya says Zafin AOS defect-fixing agents reproduce a defect, fix it, prove both steps and then create a pull request with the fix.
“They know the defect, they have to fix, they reproduce the defect, they fix it, they prove that they could reproduce it, they prove that they can fix it, and they create a pull request with the fix.”
Shahir Daya says a scheduled agent in Zafin AOS runs every Saturday at 5 a.m. to do a vulnerability scan and penetration test of a production application.
“I have an agent in Zafin a o s again that actually comes up every uh Saturday at five a m. And it does a vulnerability security scan and a penetration test of an application I have in production.”
Shahir Daya says Zafin's orchestration moves each case through defined states, including to do, in progress, awaiting a human, review and done.
“Here's a to do, then it's in progress, then it's awaiting a human, maybe, and then finally, it asks the human a question, the human answers, then it goes into a review, and then it's done.”
Shahir Daya says Zafin's agents operate under the same identity system, role-based access control, zero trust and least-privilege principles as employees.
“Agents need to live inside the same governance envelope that our employees do, the same identity system, the same role based access control system, same principles of zero trust and least privilege.”
Shahir Daya defines agent sprawl as several teams quietly building near-identical agents with slightly different prompts, running against the same data without anyone knowing.
“The situation where three different teams have quietly composed three near identical agents with slightly different system prompts, and no one knows they're running against the same data and so on.”
Shahir Daya says Zafin's agents document their work in far more detail than human engineers do, including recording videos of what they are doing.
“But the agent is so verbose and they put so much detail, they record videos of what they're doing.”
Shahir Daya says Zafin AOS captures each agent's phase-by-phase execution plan and full step transcript in a record that cannot be changed.
“And the execution plan, phase by phase, is captured as a plan file, the full transcript of what the agent is doing, every single step is all documented and can't be changed.”
Shahir Daya says Zafin AOS runs evaluations and experiments against different models using different rubrics to pick the right model for each task.
“So we in in our A O S platform, we have evaluations, we run experiments against different models with different rubrics.”
Shahir Daya says institutions that skip controls early may spend about eighteen months retrofitting them, which is harder than designing them in from the start.
“And then you spend the next eighteen months trying to bolt on those capabilities back into what whatever you've built, which is far harder than designing them correctly from the start.”
Shahir Daya says he spent twenty-seven years at IBM before joining Zafin and saw that regulated organizations must lock down their taxonomy early or struggle for a decade.
“I spent twenty seven years at IBM before joining Zafin, and the pattern I saw over and over again in regulated environments is that the taxonomy is the thing you either lock down early or you fight for the next decade.”
Shahir Daya argues that institutions will use multiple large language models and mix internally built and purchased agents across cloud and on-premises workloads.
“You will use multiple large language models, you will build some agents internally, you'll buy others, some workloads will run in the cloud, other workloads you'll be forced to put on prem”
Shahir Daya says Zafin was the first customer of its own Zafin AOS, which now handles most of its defect fixing and builds enhancements and new features.
“We ourselves at Zafin, we were customer zero for our own Zafin AOS operating system. Now it's doing most of our fixing of our defects, building new enhancement requests, new features.”
Daya frames AI agent oversight as following the path cybersecurity took, where early movers fared far better than those who waited for a crisis.
“cybersecurity was not a standing item on the agenda of board meetings”
He sets out the pressures facing banks: scattered agents, a governance gap, cost and regulatory scrutiny.
“I see four key risk inflection points that are very urgent right now.”
Real examples show AI spend rising sharply while providers move toward usage-based billing.
“burned through its entire 2026 A I coding budget by April”
Daya explains why a central tower, not the individual planes, lets many different agents operate safely at scale.
“is not the airplanes, it's the control tower”
Each unit of work moves through a small set of clean states, so every transition can be explained.
“every case moves through a very defined, deterministic workflow”
Agents get employee-style access limits, and their detailed records give auditors evidence of what happened.
“Evidence is so important, and proof of work is so important.”
Tracking spend per task and testing which model fits each job keeps cost and quality in balance.
“In this control tower, you need to be able to know what things are costing you.”
Fast starters can stall later if they cannot explain cost, risk or accountability across their agents.
“not always the ones that scale fastest in the end”
Regulators care about workflows and their controls, so a clear taxonomy separating workflows from agents holds up.
“Regulators do not want to talk about your agents”
A shared operating layer works across mixed models and environments, and Daya advises starting small with measurable work.
“establish the control plane as infrastructure, not another governance forum”
Welcome everyone to the Emerge AI in Business podcast. Today's guest is Shahir Daya, Chief Product and Technology Officer at Zafin. Zafin is an AI platform company that provides governed agent orchestration and control capabilities for regulated institutions. Shahir discusses why regulated institutions need different levels of oversight for different AI agents, rather than applying one policy across the board, and how a centralized control layer can track what each agent is authorized to do, what it acted on, and what it cost.
He explains what separates an agent that's ready for live production from one still running as a pilot, and why institutions that skip these controls early on face a much harder path adding them back in later. Today's episode is sponsored by Zafin. If you offer AI products or services into the enterprise, you need to find enterprise dealers with the relevance. That means the right title, at the right type of organization, and, of course, readiness. Emerge attracts V P and higher ranking enterprise audiences who are already convinced that they need to move beyond traditional I T.
To learn the exact strategies we use to help leading A I brands and startups connect with their ideal enterprise A I buyers, visit emerj.com/AD1. That's E M E R J dot com slash A D one. Now the conversation with Shahir. Shahir, welcome to the Emerge A I in Business podcast.
Thank you so much, Yolandi, thank you for having me.
Yeah, great conversation waiting for us, and I'm excited to pick your brain on something that I think needs more attention. It feels like everyone's racing to get agents to do the work, right, uh, approving things and moving money and touching records and doing all of these things. But this is happening before anyone's fully worked out who's going to be watching those agents once they're in it, once they're doing it. And that is going to be the big part of our conversation today is figuring out how that is supposed to work.
So, where are you seeing the biggest gap between how fast banks want to put A I agents into real workflows and how ready their governance is to handle all of that?
So that's a really good starting question. Look, let's start with an analogy, okay? Because I think it will help frame the whole conversation around regulated organizations, implementing agents and risk, because there is some significant risk inflection points that are emerging. Okay, and I think there's maybe three, maybe four risk inflection points that I see, right? If you think more than a decade ago, right, cybersecurity was not a standing item on the agenda of board meetings. It is now, right?
But it took high profile failures, very public high profile failures. It took significant pressures from the regulators, sustained pressure from the regulators, and the institutions that acted early were positioned much, much better than the ones that waited for a crisis to happen. And I think that decisioning governance, the ability for an institution, a bank, to explain the decisions it makes is on the same kind of trajectory as cybersecurity. And to give you an example, you know, two similar customers get a different offer.
Or a rate that was negotiated at the deal table doesn't match the actual rate that gets billed. Or a compliance officer asking to demonstrate pricing consistency has a lot of manual reconstruction of the decisions that were made, and these processes are not designed to be reconstructed. That's the baseline governance risk, okay? This already exists at the banks. A I didn't really create it, but A I is accelerating it much faster than the governance frameworks we have today can actually respond.
It's it's really amplifying it. And from where I sit as someone who's responsible for product and technology at an organization that works with banks, and we have a product that's with the banks. I see four key risk inflection points that are very urgent right now. The first is distributed work. Every team is running agents. Engineering has coding agents, the operations guys have workflow agents, they have S R E agents, legal has agents around documents and understanding documents, the customer service organization has service agents.
Everyone is delivering agents. Now, these agents are really producing defensible productivity gains. I'm not saying that they aren't. They are producing real gains, but it's on their own turf, right? So if you zoom out, and as we zoomed out, you're looking at a very different picture. You've got hundreds of these agents running across the enterprise using different models. They're coming from different providers, they're acting on different data, they're producing different kinds of output, and no single view of who authorized what, what actually got done, and what did it cost?
That's not a governance model, right? That ambient automation, and I'd argue we've seen this movie before, right? Digital channels moved faster than the operating models around them. Cloud adoption moved faster than the control frameworks around cloud. A P I's the same thing. So agentic A I is in the same place, except it's compressed because of how fast all of this is actually moving. And I think that the second inflection point is this governance gap, okay? When you insert an agent in a workflow, see, we've all got these workflows, but they're human centric workflows, and we're trying to insert agents into the workflows.
You inherit a new set of problems, a new set of questions that that old operating model can't answer, right? Who authorized the agent to act? Like, what data did it act on? Where was the human judgment in that workflow? And what evidence remains that we can point to when someone asks? So those questions all sit inside the daily work of financial institutions, right? When there's a pricing change or a disclosure gets updated, you or a fraud review happens, or or it's a lending workflow.
All of those questions, you know, who authorized the agent to act? What data did it act on? Where was the human? All of those questions apply to all of the things that we do in financial services. Now, can we explain these things? Can we explain a pricing change? Can we explain an offer outcome? Can we explain if a regulator asks us a question? All of this is extremely, extremely important. And if management cannot answer these kinds of questions with confidence, then there is a material gap in governance.
And what I've observed with a lot of our client base is that most banks cannot answer a lot of these questions today. It's not because they're not trying, it's because their operating models were never built to answer these kinds of questions. Where an agent is doing work and an agent has made a decision and is being used. Then the third inflection point, which is also very important, is cost pressures. This is going to surprise a lot of people over the next eight, nine, twelve months.
Agentic A I changes the economics of work in a way that most institutions have not fully internalized. I'll give you a concrete example because these are on the record today. You've all probably seen Uber's president go public very, I think, early this year, saying the company burned through its entire 2026 A I coding budget by April in four months. And they could not link the surge in token consumption to better outcomes for customers.
Right? So, and this is not an Uber problem. This is not. This is a signal about the whole category, and it's not just giants like Uber. The other day I read in the Globe and Mail, I think it was last week. And the Globe and Mail ran a piece on safe software, a Canadian company in British Columbia somewhere. And six months ago, they were spending about 20,000 a month on A I tools. Today, they're spending $100,000 a month.
Their C E O publicly said that that's over $1 million a year, and they are they really want to get this under control because it's a fivefold increase in six months. In a well-run company that is genuinely trying to be disciplined about it. And we've seen the market itself repricing, too. You saw Anthropic, they moved from they moved their enterprise customers from a flat rate plans to usage based billing, tried to actual compute, and they did that earlier this year.
And GitHub copilot followed. Analysts are actually projecting that the pricing normalizes, you know, the pricing normalizes to reflect real infrastructure costs. Enterprise A I bills will go another thirty to 50% above current levels. This is not the future, this is now. And speaking as someone who watches token budget every day for my teams, I'll tell you this shift really does matter because it's no longer a per s line item. It's a variable compute cost and you have to actively manage it.
And then the fourth inflection point is obviously regulatory scrutiny. Regulators, auditors, the board, management teams, they still look to the institution for accountability. You can't delegate work to A I. Like, you're never gonna be able to say, well, the A I did it, right, or the agent did it. You cannot delegate that responsibility for the outcome. That was true when we digitized channels, it was true when we moved to the cloud, and it's more true now because agentic A I can make decisions and take actions on your behalf that used to require a human being in the room.
And A I is now entering governance itself. I think Lloyd's Banking Group recently put an A I agent in its own boardroom to help directors actually analyze the materials and prepare for a meeting. And I believe the North American regulators have already raised concerns about that, warning that A I summarization tools may actually cause directors to miss critical details by not reading the materials in full. So the signal is very clear. A I is being embedded into governance functions, and the governance frameworks needed to manage it are still catching up, right?
And the core question is, can your institutions explain itself? Not eventually, not after two weeks with manual reconstruction, but immediately. We have regulators asking questions. We need an answer. You should be able to give it to them in the next five, ten minutes. Sure, it's actually scary listening to all of this, and it sounds like it's not just a single problem or a single gap. There are multiple things to take into consideration. And when we're thinking governance, I keep on coming back to this idea that lots of institutions currently are treating governance as kind of like a light switch.
Like, uh, either everything is locked down tight, or it's all just running loose. So the lights are either on or off. Um, but you've talked about this more like air traffic control, if I remember correctly, where different flights get different levels of clearance depending on what they're carrying or where they're heading and things like that. So can you talk us through why that distinction matters? Yeah, so it all comes down to the capabilities that are required to manage these AI agents in a very responsible way as these agents interact with sensitive systems, sensitive data, and there's, you know, human decisions points that are injected into all of these workflows.
So we think about airspace over a major airport. There's hundreds of aircraft, different sizes, different destinations, they've got different fuel loads, different levels of urgency. What lets the system actually function safely and at scale is not the airplanes, it's the control tower. So when you look at ZAF in A O S, we have a control plane and there is a control tower. And the tower knows every aircraft that is in the airspace, its identity, its clearance, its route, priority, uh, what altitude it's at, whether it's ahead or behind schedule.
It sees the whole picture and it coordinates the traffic. So no two planes, you know, ever try to occupy the same space. That would end horribly. So if you translate this to agents, right? In an enterprise, you're going to have agents that are finite, one time agents, you know, they have a start and an end, and they run, and they get the job done. And this is, for example, agents in ZAF in A O S that fix defects in our product, right?
They know the defect, they have to fix, they reproduce the defect, they fix it, they prove that they could reproduce it, they prove that they can fix it, and they create a pull request with the fix. That's a one time kind of agent. There are agents that might be running on a schedule. You know, I have an agent in Zafin a o s again that actually comes up every uh Saturday at five a m. And it does a vulnerability security scan and a penetration test of an application I have in production.
And whatever vulnerabilities it finds, it actually creates work for the other agents to go fix those vulnerabilities. And that might happen on a on a schedule. And then I have agents that are continuously running. There's agents that are monitoring my logs, looking for errors and acting on those errors, right? So uh, you're going to have many types of agents that are running. And when are they running? Under whose authority are they running? What data are they using?
What tools are they calling? These agents need access to tools, and what are they spending? What's their budget? And what evidence are they producing? You need that control tower, and so there's key components in the control tower. The orchestration layer is what handles the orchestrating of these agents. So when my team designed ours, we made very specific architectural choices. It sits between the intent from the human and the execution by the agent. And every unit of work becomes a case, and every case moves through a very defined, deterministic workflow with a small number of very clean status transitions.
Here's a to do, then it's in progress, then it's awaiting a human, maybe, and then finally, it asks the human a question, the human answers, then it goes into a review, and then it's done. And you have to break it down into these small states and then have the orchestration engine walk it through these states. The point of the of a state machine like that is not the states themselves. It's that every transition is deterministic, there's no hidden state, no side effects.
And when a case moves to a waiting human back to in progress, you know exactly what unblocked it, who went and answered it, and and what unblocked that agent from actually continuing. The whole orchestration is one capability, but enterprise guardrails on this workflow is also extremely important. Agents need to live inside the same governance envelope that our employees do, the same identity system, the same role based access control system, same principles of zero trust and least privilege.
Our agents, when they're instantiated, we do not trust them at all. And then when we do trust them to do a piece of work, we give them only the access and the authorization to do that piece of work. I don't want them veering off and trying to, hey, I have access to this other thing, maybe I'll go look at this other thing. We don't want them to do that. So zero trust and list privilege, just like we do with employees, is extremely, extremely important.
And all of their behavior is logged, all of their provisions are logged, and and and so on. And this is not a philosophical preference. It's how you get to a place where you can defend the model to your risk and compliance teams, to your auditors, and to the regulators. It's also about how you avoid what I call agent sprawl, right? The situation where three different teams have quietly composed three near identical agents with slightly different system prompts, and no one knows they're running against the same data and so on.
It's in our control tower, everyone sees the agent registry and they know what agents exist and they don't create their own. And the guardrails was one, the next one is proof of work. Evidence is so important, and proof of work is so important. What we learned very quickly was we used to use a tool called Jira, which is what our engineers use to put their work that they do, the tickets and so on. Humans, they could write a description of a defect, for example, that there's a bug in our software.
And the agent would take that and would rewrite it. And you look at the significantly more detail and thoroughness that an agent does, it's very significant. And when a human does the work, they'll go back and document very briefly where they are. Most of the time, they don't even go back. The last comment will be, Yeah, I'm done, it's fixed. But the agent is so verbose and they put so much detail, they record videos of what they're doing.
And so this is this proof of work is extremely important because when a regulator asks, which agent, what did they do, what were they thinking, what is the prompt that was sent to them? All of this is documented in in Zafin AOS as proof of work. And the execution plan, phase by phase, is captured as a plan file, the full transcript of what the agent is doing, every single step is all documented and can't be changed.
So the proof of work is an extremely, extremely important piece. You know, that's the difference between using AI and actually trusting AI in an enterprise. And the fourth, you know, capability is cost governance. In this control tower, you need to be able to know what things are costing you. And it's not just token cost. Token cost, there's a lot of focus because the minute the subsidies come off and the token providers start actually charging you for the actual compute that's being used, a lot of people are going to go bankrupt.
They're not going to be able to afford all of this. It is extremely expensive and we've learned those lessons the hard way as well. But you want to know every single task an agent is doing, where is it spending money, which model is being used, could it have used a different model? So we in in our A O S platform, we have evaluations, we run experiments against different models with different rubrics. Sometimes cost is important to us, sometimes correctness is more important than cost.
And based on that rubric and based on the experiments, we figure out for a particular task, what model should I use so that, you know, I get the characteristics I want from its answer. So all of our traces have all of the cost, all of the models that are used, and we have, you know, obviously boundaries in terms of this is the max cost a particular agent can use for a particular kind of task and so on.
So governance of cost is extremely, extremely important as well. And that's part and parcel of a control tower, right? So you want to have the observability, you want to have the uh guardrails, you want to have cost governance, and you want to have the proof of work. That is the control tower, and that analogy with the uh control tower of an uh airport. A lot to think about there, and I think there's a lot of value in uh what you just said and great way of explaining it, great picture that you're painting.
It's it's really, it makes a lot of sense. So the last part of this conversation is normally where we leave our listeners with kind of uh a game plan or something like that. But we're gonna phrase it a bit differently this time around because I think that this doesn't get talked about enough. Everyone shows a great pilot, right? Um, and then it quietly just disappears, or it goes live, and the cost, like you just mentioned, turns out to be a mess.
Because nobody was actually watching what model was doing, what work and why, and and how the puzzle pieces fit together. So what's genuinely different about an AI agent that's ready for production in a regulated institution versus one that's still just a promising pilot? Yeah, see, the the institutions that move fastest at the start of this whole agentic A I wave are not always the ones that scale fastest in the end. Okay, there is this. There's a version of this where you deploy fifty agents in like ninety days.
You see real early wins. And then somewhere around, you know, month twelve, you realize you cannot explain what is happening across. You cannot attribute cost, you can't see risk, you can't prove accountability. And then you spend the next eighteen months trying to bolt on those capabilities back into what whatever you've built, which is far harder than designing them correctly from the start. So if you're going to build enterprise grade agents that are ready for a regulated industry, you've got to get it done right from the start.
And so the first move is is really, you have to define the controls at the workflow level. You know, everybody thinks it's at the agent level. I know it's. It's a bit counterintuitive, and it's it is the one banking customers press us hardest on, right? It's very counterintuitive, and these customers keep pressing us hard. Regulators do not want to talk about your agents, right? They want to talk about the workflow. They want to talk about the workflow.
They want to know what the universe of risk is, which workflows are in scope, what controls apply to each step in the workflow, where does human judgment actually fit into the workflow? Like, what evidence is retained at every step of this workflow? So here's why this framing really matters. If you build your controls around individual agents, they will not scale because one agent can work across many workflows over its lifetime, and one workflow can be run by many agents.
You need to build your controls around the workflows, and the taxonomy holds up. That taxonomy with workflow principles and agent principles cleanly separated is what lets you explain yourself to regulators, to the auditors, to like internal risk teams, and to anyone who walks into the conversation. You got to get that right, and everything downstream is easier. You know, I spent twenty seven years at IBM before joining Zafin, and the pattern I saw over and over again in regulated environments is that the taxonomy is the thing you either lock down early or you fight for the next decade.
The second thing that's important is establish the control plane as infrastructure, not another governance forum. You know, the control plane, this control tower and the control plane, it's. It's not a committee that meets on Wednesdays, it's not like a monthly reporting layer. It is the operating infrastructure that lets agentic work move through the institution with consistent authority, with routing, with observability, with cost visibility, with proof of work and evidence. The reason this matters is that most institutions are not going to standardize on one model, on one framework, on one vendor, or one environment.
That debate is over. You will use multiple large language models, you will build some agents internally, you'll buy others, some workloads will run in the cloud, other workloads you'll be forced to put on prem, depending on how sensitive the data is, and they should run on prem. But some and some workloads will be modern, others will be dependent on some legacy systems that you have. The mistake is assuming this can be solved by picking the right model and the right framework.
Models are changing, frameworks will change, the use cases will change. The question is how institutions manage work across all of them. So that is what the control plane does. It creates a common operating layer across very different environments. So you can govern how work is assigned, how work is executed, how it's reviewed, how it's evidenced, without forcing every part of your enterprise onto the same stack. So, and then what is important also is when you start, don't try to start with the flashiest use.
Don't lead with a major transformation project of some kind. Start where the fix and the scope is localized, where the outcome is something you can measure, and where the workflow is well understood, right? Full stack transformations are complicated and they take a long time to prove out. So small slivers of value with a very tight scope delegated to agents inside a well defined workflow. That is how you're going to build the pattern and the proof point.
I think there's a line that I really like, I'm trying to remember it. Uh, from um Gassin Karian, a UK consultancy chair, and um, you know, he said, running before walking is how many institutions end up with expensive tools solving trivial problems or powerful tools bolted onto broken processes. And that is that is the exact trap that we are trying to avoid. So start with a boring problem. Where is the decisions bottleneck? Where do people spend time on low value work?
Where do customers hit friction? You know, those are the candidates. We ourselves at Zafin, we were customer zero for our own Zafin AOS operating system. Now it's doing most of our fixing of our defects, building new enhancement requests, new features. It's handling it all, and the humans are on the loop outside, governing the agents that are actually uh doing all of this. Shaker, what a real non sugar coating candid answer. And I think that there's no hiding it, and a lot of what you said makes so much sense that you would think that our leaders already think about it.
But I think that, just like you mentioned, that last that last point, that there's a lot of hype around, oh, we just need to do something with AI because everyone else is, and they're not bolting it onto the right workflow to start with, not with the boring things. We want to start with the extravagant things because that's going to look impressive to the board. But it's not going to look so impressive when it's not working, or when it costs way more than you anticipated.
Exactly. Great conversation. Thank you so much for sharing your insights with us. There's a lot to take from this and a lot to think about, um, very valuable insights. Thank you so much for sharing that with us today. Thank you so much, Lani. The pleasure was all mine. Wrapping up today's episode, let's look at the three key takeaways from our conversation with Shahir. First, applying the same governance rules to every AI agent doesn't reduce risk, it just hides where the real risk is sitting.
Second, control spots around individual agents don't scale. The workflow, not the agent, has to be the unit institutions govern. And finally, the difference between a pilot and a production ready deployment comes down to whether cost, accountability, and evidence were designed in from the start, not added afterward. If you offer AI products or services into the enterprise, you need to find enterprise dealers with the relevance. That means the right title, at the right type of organization, and readiness.
Emerge attracts V P and higher ranking enterprise audiences who are already convinced that they need to move beyond traditional I T. To learn the exact strategies we use to help leading A I brands and startups connect with their ideal enterprise A I buyers, visit emerj.com/AD1. That's E M E R J dot com slash A D one. On behalf of the team at Emerge, we'll see you on the next episode.