--- title: "How Top GTM Teams Are Building the Operational Foundation AI Demands" id: "47993" type: "resources" slug: "closing-ai-readiness-gap-gtm-operations" published_at: "2026-09-24T15:42:16+00:00" modified_at: "2026-09-24T21:12:26+00:00" url: "https://www.leandata.com/resources/closing-ai-readiness-gap-gtm-operations/" markdown_url: "https://www.leandata.com/resources/closing-ai-readiness-gap-gtm-operations.md" excerpt: "See how top SDR teams use AI to route leads, draft outreach, and book meetings faster, without manual handoffs slowing down speed to lead." taxonomy_topic: - "AI" - "AI GTM" - "Intelligent Go-to-Market Orchestration" - "Orchestration" taxonomy_role: - "AI Transformation" - "Operations" - "Sales" taxonomy_content_type: - "Video" --- Video # How Top GTM Teams Are Building the Operational Foundation AI Demands AI GTMOperationsVideo **Summary** Most GTM teams have deployed AI, but few have the infrastructure to support it. In this session, Mike Madsen of LeanData talks with Jessica Kao of Adobe and Olga Traskova of Birdeye about closing that gap. The panel draws on LeanData’s 2026 AI readiness research for context. Their main takeaway is practical: agents perform only as well as the data, processes, and context behind them. **Meet the Speakers** **Mike Madsen**, Revenue Operations Leader, LeanData Mike leads LeanData’s revenue operations team and builds AI agents and workflows across the company’s GTM stack. **Jessica Kao**, Director of B2B Go-to-Market Transformation, Adobe A former B2B operations leader, Jessica now guides several companies each week through AI-driven operating model changes. **Olga Traskova**, VP of Revenue Operations, Birdeye Olga leads revenue operations across Birdeye’s full customer lifecycle and built a GTM engineering team to centralize AI development. ### What You’ll Learn - How Birdeye drives AI adoption by embedding apps directly in Salesforce, where sellers already work. - Why a single, named AI owner outperforms committees and open experimentation as a governance model. - Why data availability matters as much as data quality, as Birdeye learned when 30% of calls went unrecorded. - How to treat context as a strategy, and why giving an agent more context can produce worse results. - Where deterministic, rule-based workflows should stay in control and where probabilistic AI adds value. ### Why This Matters Early this year, many executives gave ops teams a simple mandate to go do AI. Teams built and deployed quickly. However, their data, business processes, and context stayed the same. That mismatch is the readiness gap. The risk compounds quickly. When a rep sees a flawed ICP definition, they question it. An agent simply executes on it, and it repeats the error at scale. As a result, AI success now depends on foundations that operations teams have always owned. Those foundations include data quality, clear process rules, governance, and a new discipline around shared context. ### Three Foundations for AI-Ready GTM Operations 1. **Data foundation.** Teams need data that is as clean as possible, with governance and visibility in place. The context behind that data matters just as much as the data itself. 2. **Business process truth.** Deterministic, documented rules define which tasks agents can perform and which fields they can change. 3. **Shared customer context.** Marketing, sales, and customer success work from the same business context. This shared understanding gives agents the judgment they need to act correctly. ## Webinar Transcript Click to Open **Mike Madsen** (00:14) Hello, everyone. Let me make sure I can start sharing my screen here. Alright, I really appreciate everyone joining today. My name is Mike Madsen, and I currently lead the revenue operations team here at LeanData. We’re really excited to talk today about a report we recently published, the 2026 AI readiness report on the state of AI in go-to-market. In today’s session, we’ll go through some of the findings, insights, and data from this report. We’ll also have two practitioners, whom I’ll introduce in a second. They come from two amazing companies, and they’ll share their experiences with AI readiness. That means getting your company ready for all the agents and infrastructure we’ll be deploying throughout the year. I’m really excited to get into this today. So let’s jump in. Why did we do this? Our starting thesis came from what we were hearing in the ops community. Everyone has been rolling out AI in their businesses. I think you all had a similar mandate to the one I had at the beginning of the year from executives: go do AI. Claude Code came out around February or March. It took the community by storm. Everyone started to build, everyone started to deploy, and everyone started getting AI into the hands of their end users as fast as they could. That was great. But one thing we all learned over the past six months is that our infrastructure never caught up to the deployment of our AI agents and use cases. Our data stayed the same. Our business processes stayed the same. The context stayed the same. That’s the gap we’ll talk about today. How do we go from quickly deploying agents to actually getting value into our business? There’s a huge amount of value in these agents and use cases. At the same time, we need to address the infrastructure gap we’re seeing across all of our customers. With that thesis, we surveyed roughly 150 people in the community and asked them some core questions. We’ll go through the results in a second. About half of the respondents work in revenue operations and marketing operations. So we’re confident many of these results reflect people like you who are listening today. These are the people living, breathing, and owning the GTM infrastructure, the data, the agents, and the tech stack. The core finding is that everyone has deployed AI. You can see the stat here. Roughly 80% of people have deployed AI. However, there’s a huge gap in having infrastructure ready for AI agents to consume your data, run your processes, and deliver ROI to your business. That’s what we’re excited to dive into today. Just a quick bit of housekeeping. Please put your questions and comments into the chat. At the end, we’ll pull out some questions to respond to. If you have a really interesting question, someone will highlight it for us, and we may address it during the presentation. The chat is the primary way we’ll capture questions to address at the end. Alright, let’s jump into the heart of the material. First, let’s introduce the two amazing practitioners I’m honored to have join me today. We have Jessica Kao, the Director of B2B Go-to-Market Transformation at Adobe. In her current role, she works with Adobe customers on their own go-to-market transformation. She has seen how many different industries and customers are going through this AI journey. We also have Olga Traskova, VP of Revenue Operations at Birdeye. She lives and breathes this every day within her business. She owns the go-to-market infrastructure and all the AI use cases across the go-to-market ops function. Olga, let’s start with you. Tell me a little bit about your go-to-market org at Birdeye. How have you been leading this AI transformation journey, and how has it unfolded at Birdeye? --- **Olga Traskova** (05:14) Thank you, Mike, and thank you for having me. I’m excited to be here with you and Jess. I lead revenue operations at Birdeye. We support the full customer lifecycle across marketing, sales, account management, and customer success. My team covers the traditional revenue operations foundation, from processes to systems to analytics, planning, and execution. Last year, we also added a go-to-market engineering capability. What pushed us to get serious about AI readiness was seeing how much experimentation was happening everywhere. As you mentioned, Claude Code went live at the beginning of the year, and suddenly everyone started using it for personal productivity. Teams used AI to research accounts, generate messaging, and summarize information, all to increase their personal productivity. Once it became obvious that everyone was eager to code and build, we ran a hackathon. In March of this year, we ran a go-to-market hackathon with sales, marketing, finance, and people teams. These were go-to-market teams, not business applications teams or engineers. They were building, coding, and generating apps in all kinds of AI environments, including Claude and Replit. That gave us solid ideas about where the pain was and what could actually help at scale. We then centralized this work on the go-to-market engineering team. The hackathon gave us a roadmap of the apps, automations, and AI workflows we were going to build. We realized this experimentation should be formalized under one central team. So we moved from asking where we can use AI to asking what has to be true for AI to operate safely and effectively at scale. --- **Mike Madsen** (07:56) I’m really excited to dig into ownership later in our discussion. I think it will be a very interesting conversation about who owns this strategy and how you build governance around it. Thank you, Olga. Jessica, over to you. What does your role look like at Adobe, and what are customers coming to you with right now related to AI? --- **Jessica Kao** (08:23) A lot. I was previously an operations leader at many B2B companies. Now at Adobe, I work with many different companies. This week alone, I’m seeing three. I work with them through people, processes, technology, and change management. It’s a really great opportunity to see trends across so many companies. I get to see the patterns of what’s working and the similar challenges different companies face. As we know in operations, it’s really the intersection of people and technology, especially with AI. We work through a lot of organization and operating model changes. What I hear consistently is that everyone understands the way we work has to change. But how? How do we have to work differently? Where do we need to pivot? I spend a great deal of time helping many companies through that transformation. As operations leaders, we’ve all helped companies through transformations. But now every company on the planet is trying to transform with AI. There are clear patterns. What does good look like? What do successful companies have in common? What are the common speed bumps companies face as they transition? I describe that transition as moving from single-player productivity use to an AI-infused workflow. I have the privilege of seeing and witnessing this across many companies, and the patterns are very similar. --- **Mike Madsen** (10:24) That’s great. I’m super interested to hear more about the trends in your customer base and the types of models people are deploying. Thanks, Jess. Now let’s jump to our first finding from the GTM readiness report, which shapes much of today’s conversation. The big idea is a concept we’ve coined as the readiness gap. The headline is that across our community of participants, deployment is outpacing infrastructure readiness. Pretty much every company has deployed an agent in their business. Going back to early in the year, there was a big push to get AI into businesses, and almost everyone followed through. The flip side is the gap we’re discussing. Only about a third of teams believe their infrastructure is ready to support these agents. We moved fast and got agents out the door. But as I mentioned at the start, the data, infrastructure, and foundation haven’t changed. It’s still a legacy model that isn’t structured to support today’s agents. That’s what we call the readiness gap. Olga, let’s start with you. Does this resonate with you? Does that gap feel real at Birdeye? And how have you built strong AI readiness infrastructure in your business? --- **Olga Traskova** (12:14) Absolutely. We started experimenting with AI through Claude-based skills and artifacts, and the outputs were valuable. But we quickly learned that even great AI-generated insights won’t drive adoption if they sit outside the seller’s or marketer’s workflow. That’s one of the gaps. If anyone, whether on the go-to-market engineering team, business applications, or engineering, builds something that isn’t positioned for adoption, it will die quickly. I believe it’s very important to meet people where they already work. Instead of asking sellers, marketers, and BDRs to navigate to an artifact, install a skill, or call a workflow, we built a go-to-market portal directly within Salesforce. I mentioned the go-to-market engineering team we established last year. That team’s resources and skill set allowed us to build a playground for AI apps and workflows directly within Salesforce. I agree 100% with what Jessica said. It all starts with people. As we navigate this AI world, having the right skills, expertise, and people is crucial. You need people who understand the business pains, processes, and challenges, as well as the technical capabilities. Having a solid team within revenue operations allowed us to build the go-to-market portal. Now we’re putting all kinds of applications there that help our teams with their day-to-day work. Account research is one example. It lives within the Salesforce account as one of the tabs on the page layout. It pulls internal and external signals, summarizes account research, and scores accounts. But I also realized that account research alone isn’t enough. Sellers and marketers don’t just need more information. They need to understand what to do next. So the next evolution of this app is moving from insights to action. We’ll start recommending next best actions: the best type of outreach, who to reach out to, and the next best step. That’s what we’re building now. Another lesson I learned is that apps and AI workflows require continuous evolution. If something you build lands well and gets adopted, keep improving it and building on top of it. The lesson is that producing an AI output is the easy part. The real work is embedding it into your day-to-day workflow, connecting it to the right moment, and turning that insight into a recommendation for what to do next. --- **Mike Madsen** (16:02) I couldn’t agree more. I love that you made the bet to build into the Salesforce ecosystem. If you’d asked people about that four months ago, they would have said SaaS was dead. Now the pendulum is swinging back. At Dreamforce, everyone was saying Agentforce is amazing. It’s fun to watch the trends change over time. Meeting end users, like sales and marketing teams, in the platforms they already use is a really interesting bet. I’m curious to dive into that more. Jessica, what about you? What are you seeing across your customer base around this gap? Are customers deploying agents quickly, then looking at their infrastructure and asking how to fix it? --- **Jessica Kao** (16:53) Absolutely. I think the whole world is seeing the same thing, especially ops people. As ops folks, we’re thinking, not that we told you so, but our job is even more important now. My hypothesis is that AI is a different beast in terms of technology. I look at it from a few angles. The first is what Olga said. You need to meet people where they are. We’ve all spent our careers getting sales and marketing teams to use technology. We know there are ways to get them to use it, and ways they simply won’t. We have decades of experience with that. AI adds another layer. Yes, we need to meet people in the interface they use. But AI really has an advantage when we point it like a spear at very specific places and workflows. Not every problem is an AI problem. Throwing AI on top of everything as a generic layer doesn’t work. It’s about being very specific. Where I’ve seen success is when ops people take the time to work closely with marketing and sales stakeholders. It’s not just asking about their challenges and how to fix them. It’s sitting down with them and watching them work. Literally, show me your screen and show me how you click through things. Understand where the handoffs happen from team to team. Getting to that level of detail and asking the right questions is a very specific skill, and ops folks are well suited for it. Training isn’t an 8 AM session and a 5 PM session for teams around the world. It’s being a strategic partner to go-to-market stakeholders, mapping their workflows, and asking the right questions. Then you act almost like a doctor, prescribing where AI should sit and building it into the actual workflow. That thorough understanding of how people work is what makes it successful. So is the resilience to keep going when something doesn’t work the first time. It doesn’t mean it was the wrong thing. It’s always the humans. The humans are always the variable. That’s one aspect. The second is that when we talk about the readiness gap, we’re applying an old framework for thinking about technology. Think about an integration. We’ve all done many of them. You connect the integration, check the box, and move on to the next thing. AI is different. It’s a continuous loop. Getting AI to spit out an output is easy. But is it the right output? Do we have a loop for constantly checking? That integration is a living, breathing thing. AI technology evolves so fast that you have to constantly maintain whatever you build. So we need dedicated people. Do we have the infrastructure, people, processes, and ways of working to support that? And how do you pivot as an organization? The third thing unifies everything. We have different models and interfaces, like Claude or ChatGPT. Pick your favorite. What stays tried and true for every company is context. More importantly, it’s shared context. Who owns it? Who’s allowed to update it? At what cadence? Who governs it? Context, long-term memory, governance, and traceability are all important. That’s how we address the readiness gap. All of that is net new. Companies typically don’t have context governance or a context graph with named owners. They don’t have a defined process for updating their ICP when a new product launches. --- **Mike Madsen** (21:58) Those are all brand new problems. If you apply an agent without governance or rules in place, it compounds the problem even faster. If you give flawed information to a human, they’ll flag it and say this ICP definition seems off. An agent doesn’t have an opinion. It just executes. You touched on an interesting topic that leads to our next finding, which is ownership. Who owns the AI strategy? And beyond that, who owns the context layer, the data layer, and all these new things we’re trying to figure out and build? We asked the community this core question. We found two primary ways people currently structure ownership and governance around AI. The biggest bucket, as you can see, is a committee. I think this is how most of us started. We didn’t know exactly who owned what or how it would be managed, and committees form easily around that. It’s the natural place people gravitated toward. The super interesting part of this finding is the next biggest ownership model: no one owns it. It’s just a free-for-all. Anecdotally, I think that reflects a move-fast mentality. People didn’t want governance or rules to get in the way of innovation and speed. It’s interesting to see these two structures in the results. Jessica, you’ve talked with many customers and likely seen many ownership models. What common trends are you seeing in how customers structure ownership of their AI strategy and agent management? --- **Jessica Kao** (23:59) We actually ran our own research, and it aligns exactly with what you’re seeing. More specifically, successful companies have a dedicated single AI owner. That doesn’t mean centers of excellence or steering committees go away. They’re necessary but not sufficient. The single owner is what matters. The companies that are ahead in maturity are the ones with an AI owner. Think about it. Without a captain of the ship, you’re just drifting. All of us live and breathe governance, data cleanliness, and data infrastructure. We understand how important that is and how to balance it with moving fast. Having a single owner correlates directly with how quickly companies achieve business outcomes and success with AI in go-to-market. --- **Mike Madsen** (25:06) I couldn’t agree more. If you look at typical ownership models for the tech stack or data, there are dedicated teams and owners. With AI, everyone has a seat at the table. There’s a lot of debate about where it should live and how to structure it. In my own practice and relationships, shifting toward a single owner gives teams far more ability to move and build quickly. In some cases, committees slow down the path forward. Olga, I know Birdeye went through a transformation around ownership. Talk a little about that transition and where you landed. --- **Olga Traskova** (26:00) That’s a compound question. Initially, ownership for us was distributed. Different teams experimented based on their own needs, which I think was healthy. It drove a lot of energy and helped us generate many use cases. But distributed experimentation creates duplicated tools, duplicated AI workflows, credit consumption, and inconsistent data practices. It’s about optimizing your own efficiency rather than building at scale. So we formalized go-to-market engineering under revenue operations. I truly believe revenue operations is well positioned to connect business priorities, data, technology, and frontline execution, and to drive adoption. We see the entire customer lifecycle. We understand the dependencies between teams and the customer journey. Having said that, we should also consider build versus buy. Just because you can build doesn’t mean you should. The barrier to entry for building with AI is extremely low. Everyone can build. But think about the ownership. I agree there should be one owner. What I learned is that when you build an AI product, you become a product manager. Now you’re responsible for administering, maintaining, and optimizing it, and for handling feedback and glitches. You’re all of it in one. I’ve built many apps. Now sales and marketing leaders come to me with requests and with things that aren’t working and need fixing. Do you really have the capacity to own that within your team? If you’re deciding between building an AI application or workflow and buying something already available, are you ready to maintain it? Who will own it 18 months from now? Three years from now? If that single owner changes roles or leaves the company, does the team have the capacity to take it on? That’s one of my key considerations around ownership. Every use case we build needs a clearly named business owner, a technical owner, and a measurable outcome. It also needs someone responsible for adoption. And it needs someone who continues to own it after it goes live, managing glitches, feedback, and improvements. --- **Mike Madsen** (29:08) I love that you touched on build versus buy. In the first half of the year, everyone was in build mode with Claude Code. People said, “We can build this. No one needs a CRM anymore. Let’s just build one.” I think the pendulum is swinging back, to your point. I’ve built a lot of apps and agents too, and they become a huge maintenance lift. It expands even further. You own the agent, then you own all the data feeding it. You also own the context Jessica mentioned that manages the agent. Your role compounds significantly. Without the resources and capacity, it isn’t a scalable solution. That brings us to our next topic. What’s the number one thing blocking teams from getting strong ROI or scaling AI use cases in the business? It’s data. As everyone has said many times, your agents are only as good as the data. More recently, you also hear the term context. Agents are only as good as the data and context you feed them. If you have a poor foundation and poor data, you pass bad outcomes to an agent that replicates them far faster than a human. The example I always use is the one I mentioned a second ago. If you pass a bad ICP or bad data point to a human, they’ll flag it. An AE will say, “This is wrong. I know better. I’m not acting on this.” An agent will just execute. Data quality is a huge challenge many people are experiencing. The next biggest challenge from the survey is stack integration. How do we move from a legacy model of point solutions, or a more consolidated solution, and start adding AI? As Jess mentioned, you can’t just throw AI on top of everything. It needs a more integrated strategy. Jessica, you’ve seen many customers’ points of view on data and how agents will execute against it. Are you hearing a lot about clean data? Is it more about context now? Or a blend of both? --- **Jessica Kao** (31:45) It’s absolutely both: data cleanliness and context. Let’s start with the pre-AI world. We had data, and it was never going to be squeaky clean. We operated in a world of imperfect data. That’s where human taste, judgment, and wisdom came in, all those squishy things. When I talk about operations, one of my best practices is to go on a date with your data. Know your data really well. Be best friends with your database, because then you know how to interpret it. If someone wants to cut a segment of customers, all of us immediately translate that in our heads. We know exactly how to do it. That’s wisdom and intent. In the pre-AI world, ops teams translated the intent of sales, marketing, and go-to-market through our wisdom and experience. That’s how we made things happen. Now AI goes directly into the data without that human intermediary. What fills that gap? Will our data ever be perfect? No. Can it be better? Absolutely. But what I often see is people swinging the other way. They say they have to wait until their data is clean. Your data will never be perfectly clean. We’ve always operated with imperfect data. It’s something we live with. So the question becomes how AI can enhance our workflows in a world of imperfect data. I think there are two aspects. First, it comes back to context. How do we take two decades of experience from myself, my team, and operations and give it to AI so it can operate with imperfect data? That shared business context is the glue that closes the gap. Second, we can use AI to improve data quality faster than we ever could from a resource perspective. So it’s two angles. Where I see many companies get stuck is thinking they have to wait until their data is perfect before doing anything. The successful companies understand you don’t know what to fix until you get in there. It’s like a chicken and egg. Business has never stopped because data was imperfect. Once you get in there, you learn where to fix things. It’s like learning to cook from a YouTube video. You’ll never really learn until you get your hands in it. You have to start, knowing you’ll operate in a world of imperfect data. All of us in operations have always done this. How do we take our knowledge and continually feed it context? That’s the role of humans. We talk about context engineering, but every role, whether marketing or sales, will need to do it continuously. It’s not one and done. Our role will be to continually give AI wisdom and context. --- **Mike Madsen** (35:29) I couldn’t agree more. Lately, I think of context as the lived experience you’re alluding to. When someone asks a certain question, I know which Salesforce fields to look at. I also know the business reason they care. Building that context into the agent will be the huge unlock we’re all working toward. It will be interesting to see how this develops over the next six months. Olga, what are you seeing at Birdeye? Talk about your journey, from rolling out agents in the early days to optimizing data cleanliness. What didn’t work, and what did? --- **Olga Traskova** (36:24) I agree with everything you’re both saying about data quality, context, and structured data. As you peel the onion or build foundational layers, data is the very first layer. I want to talk about data availability. Jess talked about integrations and the tech stack, and how we’ve historically worked with many tools to make sure data flows. At the beginning of the year, I was building a pipeline inspection agent, a forecasting accuracy agent, and an agent that highlights deal risks. I realized that if calls aren’t recorded, there’s no point building any AI solution. We’re still going back to Revenue Operations 101. You need the right tools, integrations, and data flows configured and working. We discovered that 30% of our calls weren’t being recorded. It was user error, a system issue, or an integration issue. Our pipeline inspection agent simply didn’t yield results. So to me, there’s data quality, and there’s data availability. You need to build that layer and ensure 100% of the data points you use for AI apps are there. Having trustworthy data at the point of decision is crucial. We spent time fixing that layer and all of the integrations. We made sure Salesloft was properly connected to Salesforce, and we worked with vendors to close the gaps. Only then could we build AI on top of it. --- **Mike Madsen** (38:22) Olga, I appreciate that background. It made me think of a question either of you can answer. Even with good data, there’s only so much data or context we can give an agent before it gets overwhelmed. I’ve seen cases where throwing everything at an agent makes it perform worse than giving it something narrower. Have you seen this personally, Jess, or in your customer base? We know some data is good and some is bad. What’s the right information to pass to an agent? How do we avoid throwing everything at it when much of it isn’t valuable? --- **Jessica Kao** (39:13) That’s what I’ve been talking about with customers for the past year: context is the new data. In the past, we focused on data cleanliness, activation, enrichment, and data strategy. Now context is a strategy, exactly to your point. Even in your own professional use of AI, have you ever run the same prompt on different models or at different effort levels? I’m sure you have. Sometimes more effort isn’t better. It’s actually worse. Sometimes I realize I gave it too much context, and it went off into left field. We’ve all experienced that individually. Now think about it at scale. It’s a combination of your ask, the model, the reasoning, and the context. Everyone is focused on building AI agents. But what’s the context strategy? We have data cleanliness, data hygiene, and data quality. We also have to strategically give AI the context it needs. That can be the difference between getting the outcome you want and veering off into left field. That brings us back to infrastructure, ways of working, and ownership. These are new human roles. When your UI becomes natural language, how do we translate probabilistic output into deterministic go-to-market actions? For example, say I want people to query their data and move from insights to action faster. If I want to report on campaign performance, a person might ask 20 different ways. But I want every version to trigger a very deterministic result. What guardrails do I need to make sure that happens? That’s not someone’s part-time job or side gig. It’s a full-time job for people and teams. It’s the adaptive layer that makes everything happen. I think it’s a net new focus area, and as AI grows, it’s a different type of maintenance. --- **Mike Madsen** (41:52) You’re speaking my language with deterministic and probabilistic scenarios. Personally, and as a business, we feel there’s definitely a hybrid model here. You don’t want to run probabilistic models across everything. Deterministic, rule-based workflows are still significantly better for many tasks than throwing an LLM at a problem or process step. It goes back to what you said about establishing business context and rules of engagement. What do agents do, and what stays a deterministic workflow? I talk a lot about process context. Your business has many process workflows, essentially from lead to cash. There are all these process steps, and in most cases they’re deterministic. So how do you capture that process context and pass it to an agent? And how do you blend it into probabilistic models? That’s where we’re heading as a business, and it’s something I’m personally interested in exploring. Olga, are you experiencing this at Birdeye? How are you thinking about context in your business? --- **Olga Traskova** (43:07) In simple language, to me it’s about nailing the prompt. You need to be very precise. Every AI app is powered by a prompt on the back end. You’re essentially asking questions and configuring prompts. You need to be very clear about how you want data returned, in what format, and what you want the AI to look at. From that perspective, I believe in human in the lead, not human in the loop. We still configure everything very precisely, and we spend a lot of time testing back and forth. Today, AI fills in more than 80% of the fields on the Salesforce opportunity object. As you can imagine, some fields are picklists or multi-select picklists, so you need to configure those prompts very precisely. Some back-and-forth testing is required. To me, it boils down to being very explicit in your prompt and connecting to the right data points. I think that’s what you’re calling deterministic, and I agree. --- **Mike Madsen** (44:25) I could geek out on context all day. But let’s jump to our fourth and final finding, which relates to trust. It starts from this idea. If agents run on mediocre or bad data without enough context, they’ll still produce outputs. Those outputs may not be very good. End users, like salespeople consuming a recommended next best action or email, may not trust them. As ops people, we may not be confident the agent is updating fields correctly, as in your example, Olga. It may be overwriting the right values with wrong ones. All of that leads to the trust issue the community reported back to us. Sixty percent of respondents said their top AI fear is an agent acting on the wrong record. In your example, Olga, that could mean updating the wrong opportunity, entering the wrong ARR amount, or entering the wrong next steps. That’s the top fear. What people are asking for is more visibility and more audit trails around what the agent is doing. Today, that’s a gap in the market and in these models. There isn’t a good way to track exactly why an agent performed a task. Why did it take this action? What was its rationale? What did it change? How do I track that, validate it, and QA it before I unleash it too far and it goes down a path I don’t want? Olga, you mentioned agents updating 80% of opportunities, which is amazing. How do you think about auditability, governance, and the rules that keep an agent from going rogue? --- **Olga Traskova** (46:25) Agents update 100% of opportunities and 80% of fields within those opportunities. That’s on purpose. We reserve some fields for humans to update. We want to distinguish what AI fills in from what humans fill in. It all comes back to what we just discussed: the underlying data, its availability, and its quality. When I build an AI workflow, we always start with data. Do we have the right data? Do we trust it? Then we ask more questions. What’s the impact? What’s the use case? How reversible is the decision? Who will own it? Can we maintain it long term, and what level of effort will that take? Should we build or buy? Who owns it on the business side and on the technical side? Some flows can be 100% AI-led, especially low-impact ones. But pipeline inspection, forecasting accuracy, and anything that goes to the C-suite, board, or executive level get a human in the lead. We inspect what AI produces. Trust requires visibility. So I make sure we can get into the back end of whatever AI is working on and spot-check it. --- **Mike Madsen** (48:05) So it sounds like you have some type of matrix. A low-risk task might go to an agent. A higher-risk task, like pipeline, might be human-led. Is there a matrix that determines the rules? --- **Olga Traskova** (48:20) That’s exactly it. Low risk could be account research, summarizing information, drafting emails, and flagging missing information. Medium risk would be account prioritization and deal scoring. High risk is where we involve humans and deterministic controls. That includes forecast commits, territory changes, customer commitments, deal quality, and so on. --- **Mike Madsen** (48:48) I love that. Jessica, maybe speak to this quickly, since we’re running a little behind. What are you seeing with customers and industries around auditing, visibility, and the fear that AI will go rogue? --- **Jessica Kao** (49:05) At Adobe, our philosophy is very big on governance and traceability. I think it should be 100%. You want to know. Think about the stack you typically manage as an operations person. You have users, roles, and permissions. Things are locked down, like which objects you can see in your CRM and who can read or write fields. That’s table stakes. That’s 101. Why would that be any different with AI, or with anyone changing data? We want a full trail and traceability. That is governance. Governance isn’t the sexy thing, but it’s so important. Here’s an example. I was at a conference chatting with a peer group, and people were very excited about AI. They wanted to put AI on top of all their data so everyone could ask it questions. I used to own analytics for marketing QBRs. My team owned all the dashboards. They were fixed and locked down, and those were the ones you used. In a world where anyone can query the data, guess what happened at the QBR? Everyone brought their own dashboards. They spent the entire time showing conflicting data and debating which number was right. They never got to insights, let alone action. Everyone had asked their own questions and come with their own dashboards. So it comes back to everything that’s tried and true around governance. Having a strategy for everything we’ve discussed today is essential to making AI successful. --- **Mike Madsen** (51:26) I love it. That’s a great segue to our final commentary on where businesses should make their bets. To your point, it’s about having a strategy, and we see it falling into three pillars. Our point of view is that a successful AI strategy needs all three of the things we’ve discussed today. By the way, this conversation has been great, and I really appreciate the dialogue. The first pillar is data foundation. That means structured, clean, or as clean as possible data. It means governance models, visibility, the right tables, and all the unsexy work behind a strong data foundation. More importantly, it means having the context behind that data. The second is business process truth. We talked about deterministic workflows, documentation, and rules, like the ones Olga described. The agent can do these tasks against these fields because those are the rules we set. You need strong business process truth that is deterministic in nature so you can control and manage it. The third is the huge unlock we’ll see in the short term: shared customer context. Marketing, sales, and customer experience all work from the same business context for their processes, data, and agents. Before, people worked purely from data. Jess, I really like how you framed context as lived experience. How we pass that information to an agent gives it, for lack of a better term, a brain to do the right thing. That context is what unlocks these capabilities. We see these three pillars as the foundation for a strong AI strategy going forward. Before Q&A, I’ll pass it to Jess and Olga for final words. We covered many topics and concepts. Where should people start to get the ball rolling? --- **Jessica Kao** (53:38) We’re all being asked to show success. So what does that look like when there are infinite things you could do? Where do you start? I’m a big believer in choosing something, showing success, and building momentum. Don’t try to boil the ocean. Triangulate what you control and own, your skill sets, the technology available to you, and the data available to you. There’s a lot you could do. But often, accessing certain data would take much more alignment and work with other teams. So ask what you control in your domain. What can you do, and what can you show? What ties directly to ROI, impact, and revenue? Everything we do in operations should tie directly to revenue. It’s about translating that and showing success. Start strategically by looking at the workflow and dissecting it. Point AI precisely where it can augment the work. Then execute, show success, build momentum, and keep moving forward. --- **Mike Madsen** (54:54) Love it. Olga, any final words for the group? --- **Olga Traskova** (54:59) Exactly what you both said. You summarized it very neatly. Don’t start with the tool. Start with the business problem and the outcome that matters. Map the full workflow. Check your data quality and data availability. Do you believe the data? Are your data and processes strong enough to support the workflow you’re trying to build? So start now. Build the foundation your workflow will rely on, and then earn the right to automate more. --- **Mike Madsen** (55:32) Amazing. Thank you again, Jessica and Olga, for this great conversation. I really appreciate the dialogue. Let’s open it up for Q&A. If you have questions, please put them in the chat. We’ll give it a minute or two before we close out. Let me look at the questions here. I’ll paraphrase one. We mentioned that sometimes you need to sit with end users to understand their workflows. Jessica, I think you talked about this. How do you approach deploying AI across very large teams where you can’t see every person’s workflow? How do you understand team members’ workflows and needs at a large company before deploying agents? --- **Jessica Kao** (56:46) It starts with the use case or scenario. You’re not trying to map what everyone does all at once across their roles and remits. You pick a particular business process. For example, you might map the production of paid media campaigns and how it moves across different teams. Whether you call it a use case or a scenario, it only touches specific people. Then you focus on a representative or core group from each part of that process. You interview them, ask the right questions, and map it out. Or it could be something in sales, which Olga may have mapped. You start with that business process or scenario. Again, we’re not trying to boil the ocean or improve AI for an entire team at once. AI has the greatest impact across teams, at the handoffs. --- **Mike Madsen** (57:57) Olga, anything to add about understanding end users’ processes and deploying AI to solve their problems? --- **Olga Traskova** (58:04) We have a very prescriptive sales process. It’s fully configured, and there’s very little autonomy within it. We don’t ask sellers how they like to do things. We tell them how they should do things. That solves it. Then it’s back to meeting sellers where they are. Build a tool or workflow for them wherever they work, whether that’s Salesforce, HubSpot, or another platform. It can be guidance, automation, or recommended next best actions. It can be deal scoring, lead scoring, or account scoring with insights. Map it to their process, whether that’s following up on leads, creating opportunities, or following your sales methodology. As Jessica said, map it to outcomes and revenue, to what actually drives the bottom line, and start there. --- **Mike Madsen** (59:05) Amazing. We’re coming up on time. To close out, if you liked today’s conversation with Jessica and Olga, thank you both, we’re having more of these in a couple of weeks. On October 5 and 6, LeanData is hosting OpsStars at the InterContinental in downtown San Francisco. It’s our tenth year, and we’re super excited. Jessica will speak at the event on the future of ops and turning AI pilots into business outcomes. She’ll cover why pilots never make it into production and how ops people can change that. If you’d like to join us, please sign up using this QR code. Finally, thank you all for joining. To learn more about the report, use the QR code here to read the full version offline. If you’d like to reach out to any of us, feel free to connect with us directly on LinkedIn. We’re always available to anyone who wants to learn more. Jessica and Olga, I really appreciate your time and the conversation. I look forward to more of these discussions with you both. Cheers. --- **Olga Traskova** (01:00:13) Thank you, Mike. ## Frequently Asked Questions ### How does Birdeye decide which AI tasks need human oversight? Olga uses a risk matrix. Low-risk tasks, such as account research and email drafts, can run fully AI-led. Medium-risk tasks include account prioritization and deal scoring. High-risk work keeps a human in the lead with deterministic controls. Examples include forecast commits, territory changes, and customer commitments. ### What should teams confirm before they build an AI workflow instead of buying one? Olga suggests asking who will maintain the tool in 18 months. Building an AI app effectively makes you its product manager. Each use case also needs a named business owner, a technical owner, and a measurable outcome. Someone must own adoption and ongoing improvements after launch. ### How do you map workflows for AI across a large organization? Jessica recommends starting with one business process rather than an entire team. Next, interview a representative group from each team that touches that process. Watch them work on screen and map every handoff. In her experience, AI delivers the greatest impact at those cross-team handoffs.