Episode 1 · Oct 7, 2026 · 25 min
Why Your AI Gets Worse The Longer You Use It (with Jack Taubl from Ocean.io)
Prefer audio? Listen here
Chapters13 chapters
TranscriptThe full conversation · 5,321 words · tap a time to play from there
if you're in marketing or sales or you're really involved in GTM in any way, then you're almost definitely using AI to some degree. We're all playing around with different tools, but the challenge is if you're playing around with these tools, you've likely been working on something, and you start to get it to work really well, and then all of a sudden it stops working, or the quality gets worse, or, you know, if you're using it for prospects, then you end up surfacing the same people over and over again. And it can be really challenging and, uh, hard to figure out how you get over that. Welcome to Pipeline. I'm your host, Joseph Lewin, and I'm super excited to have today's guest, Jack Tobol, on from Ocean.io, 'cause he's been working on this same stuff for a while, and he's figured out a few things about how to use AI effectively, both for pipeline generation and how to close deals. So Jack, welcome to the show
Joe, thanks for having me. Glad to be
Yeah, so we've had a couple chats, and normally I start off with, uh, a different question here, but, um, you were able to explain really well why that's happening with AI, 'cause I've definitely had that happen where I've worked on it, get something working super well, and then the quality just degrades over time, especially again, going back to the list building side. So why is that happening?
There's a couple of layers to it. So if you think about it foundationally, the margin that AI providers or LLMs specifically, whether you're Anthropic or cl- or ChatGPT, Gemini, whoever, their margin comes from the efficiency of their models and the amount of work that they do. So these models are naturally geared to become as efficient as possible, and if you think about the length of time in one single Claude chat thread, for example, you'll get the line that says, "Compacting our conversation so that we can keep talking." That is Claude's way of, you know, managing costs and managing the, uh, you know, the amount of storage and the amount of memory that they have to worry about, the amount of context that they have to retain. So while to your point, you know, I'm a, a huge proponent of AI, that is-- it's fundamentally designed to erode the longer you use it if you're staying in a single chat.
Now, the other reason behind it is that a lot of times when people are going into a chat thed- thread, they have a very ambiguous thing that they want the AI to do. So if you go in and say, "What does a good XYZ look like?" You are going to… It's going to drift around a lot, and you're not going to come to a conclusion faster because you're not being specific. So I think the two biggest reasons why it happens is, one, it's designed to happen, and you need to be, you need to be proactive about how you're working with it, and two, there's a lack of specificity and a lot of ambiguity that goes into what, uh, into what we ask AI, and if you ask ambiguous questions, you're gonna get ambiguous results.
Yeah. So let's, let's dig into each of those. So on the, on the layer of, um, staying in the same chat, I mean, the, the reason you would stay in the same chat or co-working session or whatever would be so that it does keep contact so that you don't have to start over from scratch every time. And, uh, but then to your point, the longer you stay in that chat, then it erodes over time. So you get to a point where it's working super well and you're like, yes, you're finding… Let's just say you're using it for finding the right prospects and, and you're getting in and, and you finally get it to where it's really juiced up and you're like, "This is awesome." And then suddenly it resurfaces the same group of people again, or goes back over and, you know, finds some of the same, same info that it surfaced last time and you're like, "No, that's not right." And then it doesn't actually, it can't actually go back and, and reference very easily the stuff that is already compacted.
So how do you, how do you deal with that? How do you start solving that problem?
So at a super tactical level, uh, like there's a more strategic element to it, but for like, just for a takeaway for your audience that is easy to implement, my recommendation is whatever, if you're working on something that is broad and doesn't necessarily have a specific end in mind, do that work within the confines of a Claude project. Because each chat that you create within a Claude project has the context from other chats. And so the minute when I have, when I start to get an answer that I'm skeptical of in a chat within a project and I'm confused as to whether or I don't trust it, I will immediately switch to a separate chat and potentially switch the reasoning model that I'm using as well, right? Because the context from that first chat is still there, but now I'm getting fresh eyes on it and not dealing with the same, you know, uh, drift that the first chat is essentially. So that's one really tactical way of going about it.
But at a more strategic level, this is why I joined Ocean.io where I am today. You really, across all your GTM tech, you need contextual grounding that is going to make sure that everything is oper- all of your systems are operating off the same information. So for us, this is contextual intelligence around the accounts that you're selling into. We have a deep understanding of virtually every B2B organization that there is in a way that others don't, and that context is what allows us to act as the underlying data foundation that all of your other systems run on. So the way that I describe this is you have NetSuite over here where you have customers, you have Salesforce which has accounts, you have HubSpot that has companies, and then you have, uh, let's say Zendesk that has custom- or, uh, customers. I forget what their naming convention is. All three of th- those three records all have different naming conventions.
They're part of a different schema and for that reason you're gonna have a different data set potentially in each one of those systems. What Ocean's doing is providing you with that underlying company record contextual layer that, that exists as the source of truth in all of those different systems. So now rather than operating on three different records and three dis- different systems for Joe, now Joe sits in the middle and is persisted into all of those systems. So that's at a high level, in order for AI to work, it needs some kind of contextual grounding and that's what Ocean offers for go-to-market.
So I wanna take this from two angles because I actually think what you guys are doing is really interesting, and, um, I've have a lot of use cases that I've been working on, and I think what you're, you guys are doing could potentially solve. So I wanna dig deeper i-into that. But first, let's take it back a step and go like, okay, a-agnostic from Ocean, what you're talking about is you, you basically need some- something outside of Claude that Claude is going to go back to, to pull information from to make sure that it's accurate and that all of that information is staying up to date. So, like, a different example of that might be, um, you get a Supabase project set up to store some kind of information. I've been messing around with that for building lists and stuff like that. So then that way, every time that I'm going to search for a new list for a new client, it's going back, and instead of starting from scratch or me trying to use the same project forever, I could start a new, uh, or a new, new session, not project, but I could start a new session in the same project using a skill that then, you know, layers these different pieces on top of each other, and one of those is gonna be going to Supabase and figuring out who are the current clients who've been on the podcast in the past, um, and, and then who have we already said isn't a good fit for the podcast from previous research, and then it's pulling all of that context into the initial search, and that way we're not starting from scratch every time.
So is that kind of, uh, uh, the idea that you're getting to? You need something that it's going back to to ground, uh, so you're not just starting from scratch every time?
Absolutely. And I, like, I don't want my LLM of choice to be a storage layer, right? And I, think this is where the world is heading. Like Cl- all the-- whether it's Claude or, uh, ChatGPT, take your pick, the whole world is moving towards this new motion of, uh, the LLM is the execution layer, and it's about getting your secret sauce to be the grounding behind that LLM. There's a lot of things about Cla- I have a lot of thoughts about Claude Force and everything that's happening with it, but to be honest, I think it's the right move. Even just this week, they said very publicly, like, "Hey, we are going to be the structured, categorized dataset that you run all of your work Claude workflows on top of." So I think it's-- I wouldn't tru- like I frankly wouldn't trust Claude to maintain a database for me that lived within Claude. Would I trust it to update it when I tell it to? Absolutely, but it's gotta be external and it's gotta be a, a, a hyper objective and specific to you because if you start, if you're without that, again, same context window deterioration problem, the results are going to deteriorate over
Yeah, I don't wanna go too far down a rabbit trail, but I think that something that I've been thinking for the last little while, and then you're one of the first people outside of me living in my bubble that I've heard voice the same thing, is platforms are gonna start to go away. Not necessarily completely, but where we used to go into five different plat- I mean, I, I'm gonna just lay this out because it's, it's what I used to do, is like, okay, I go to build a list, I go to Sales Navigator, I do a bunch of searches with all these crazy… You know, I mean, this is like a year ago, showing people how to do this . So it's like, okay, you go in, you, you try to find people, then you have to go to their individual profiles and check and make sure it's right and whatever. Then maybe you go to Apollo or Seamless or whatever tool you're using to pull contact information from. Um, and then, you know, you're basically jumping from LinkedIn to Apollo, and then maybe I have my CRM and I'm jumping into there and, you know, going back and forth between all of them.
Whereas the way that I'm working now is, does this tool have an MCP? And if it doesn't, mm, it's gotta have a really good reason for me to, to use it at this point and, and get it. If it doesn't and it has an API that I can make my own MCP, I'll think about it. If it doesn't have an API at all and I have to go into their platform to build it, I'm, I'm gonna be very, very hesitant to do that at all. So then moving towards Claude being the thing that I'm using on a regular basis, it connects with Apollo and lightfield, the CRM that we're using, and all the other platforms I'm using. And then Claude is basically like the, the, the piece that I'm interacting with. It knows when to go to all the other tools, either on its own or through a skill that I create, and then it goes and pulls all the right information from the right places. And then what you're talking about is, okay, yes, that's the right approach.
That's how you should do it. And in that mix, you need something that's essentially the single source of truth for, for the accounts that you're going after, and if you trust your CRM, it's not a database and, uh, like it's not updated off of current information. It's not gonna be as rich as something like Ocean, uh, .io or, or, you know, even what I'm building in Supabase. On a small scale, maybe that works. Maybe if you're an individual seller or you're working by yourself, maybe you build something out like that. But then where you run into trouble custom building something like that is gonna be, um, it- at scale with a whole team of people who are constantly updating that across all of those different tools, and now it's not just sellers, it's, you know, potentially pulling that data in for your, your, your other teams to be able to use as well. And now it's that same single source of truth issue.
Like if you ha- if, if you have data coming in from all these different tools, then something's gonna get screwed up and, and dropped along the way. So am I, am I thinking about that right?
y- you articulated it, you, you, you articulated it beautifully. In fact, I'm gonna steal some of the stuff that you just said because it was actually really good. Uh, thank you for that. And, and you're exactly right. I started off my career at a, at a consulting firm called Serious Decisions, and that firm was specifically focused on how do we get sales, marketing, and product management aligned. Because if you're aligned, you're gonna be more efficient in your go-to-market, you're going to generate more revenue. And there were four pillars to what alignment is. It's what do we sell? Who do we sell it to? How do we define a lead? And how do we measure results? I think most companies m- are misaligned largely because of the data that they're working off of. So for example, if you're talking about account-based marketing, which is marketing and sales working together strategically, if you have a different data set in your HubSpot instance than your Salesforce instance, then it's really difficult to do account da- based marketing because everybody's working off of a different foundation.
So with Ocean, what we're doing is making sure that if I am running an ABM strategy against Adobe, I know that Adobe is Adobe and all the little child, child, children that are underneath it from a hierarchy perspe- standpoint, all the duplicates, I don't have to worry about those. I trust this account record. So you're exactly right, and having all of your systems in sync, once you, once you've done it, there's no benefit - there's no telling how much you're gonna benefit in terms of efficiency gain and, and productivity from your team. So think about how much time we've spent in our careers cleaning up Salesforce data or CRM data or something
it's the worst. It's literally the worst. Uh, and then you can never really do it because, um, I used to work in an engineering software company, and it wasn't, it wasn't CRM data, it was all of these parts that their engineers spent hundreds or thousands of hours building that get lost in their PLM system, which would be like their, you know, version of the CRM. Uh, and no engineer can ever find it again, even though they built something or used that same part. And I mean, the amount of cost in that situation was, like, astronomical. It's insane. So
And it compounds too because everybody, like when a new person comes in after the previous person over-engineered it, they've got to figure out a solution which is inherently going to be over-engineered because they can't figure out what this person did. And then the next person has to come up with an even more complicated and over-engineered solution because they can't figure out what the previous two did. So it's actually a really compounding problem that it's one of those hidden costs that don't really show up on a balance sheet
Yeah. Uh, uh, totally agree. Okay. So what I would love for you to do, and, you know, normally we don't focus quite so much on the company that you work at, but I think with Ocean, it's just interesting because, um, it really fits super well with some of the things that we were talking about before, just as far as creating that data layer, and it's something that I know that a lot of teams are really struggling to figure out how to do. Um, and you've got super expensive tools out there and, um, you know, I know we talked about Clay and some of the challenges with that, where you end up potentially pl- plo- proliferating, uh, bad data indefinitely, and then you're spending all this money maybe on autopilot co- that can cost you an incredible amount of money for a contact that maybe nobody will even ever look at. So you can run into a bunch of issues there. So what I'd love for you to do is just lay out an example, um, either of how you've been, been using, uh, Ocean or something that you've seen with a client that kinda illustrates the value of creating that single source of truth.
Absolutely. So I'll, I'll take this from two lenses because they're w- contextual contextual intelligence is applied both at a very strategic level and a tactical level. So I'll s- first start with the strategic level, and then I'll talk about how I use it on a day-to-day basis, which is the more tactical side of things. So when you couple the level of depth Ocean has in terms of our contextual intelligence around all the companies that are in your total addressable market with your deal data, your actual closed won, closed lost opportunity data, that produces, that unlocks what we call segmentation, meaning I understand where I should direct my go-to-market efforts and my go-to-market spend in order to hit my budget. So if I know that I win more in healthcare than I do in bi- or, like, let's say, with payers and providers versus biosciences, I know I win more with biosciences, that's where I'm going to focus.
That's how I'm going to do my head count planning. That's how I'm going to do my territory assignments. That's how I'm going to understand what percentage of my pipeline needs to come from sales source versus marketing source. This is really the foundation of how I'm gonna make a plan to hit my number. So that's where we help you set the strategy. Then at the execution layer, now I know that my reps are focused on the right accounts, and they're not wa- wasting their time on companies that are never gonna turn into customers, prospects that are never gonna turn into customers. And on top of that, I'm running automations that make life easier for my team. So if, Joe, you close a deal with Acme Corporation, you're gonna get a Slack message that says, "Hey, congrats on closing Acme Corporation. Here's five other prospects that look just like them. You should focus on these next." So what Ocean provides is that context layer that I was talking about, uh, more broadly from a tech standpoint, but realistically, we're talking about helping you set the strategy for how you're gonna hit your number as a CRO, and then providing the intelligence that your team needs to execute at a tactical level.
So you're focused on the right things, your team is executing against the right t- things, and then you can monitor performance over time. Am I making sense? I know, I know I
So then one, one layer beyond that, so then we've talked a lot about your LLMs. Let's just say then you're a sales rep, you're using Claude. How does that then interact with, with like your daily tool that you'd be using versus going into the platform since that's kind of the, the thing we set up front?
The theme of this conversation. Yeah, yeah, absolutely. So the, we're not just focused on companies within your TAM, we're also focused on the people within those companies, right? So that's the type of logic that we can apply. So if I have, um, let's say Salesforce assigned to me as a prospect, I need to know who to talk to within Salesforce. I need to know who else is using Salesforce. I need to know about partner adjacency, et cetera, et cetera. So I, when I ask that question, my Claude instance is trained to go to Ocean first for the information I need and the context I need. When I ask for an account briefing, if I get a new, you know, something new assigned to me, Ocean is one of the first resources that gets pulled into what the com- the composition of that briefing looks like. So the idea here is that we're bringing that contextual intelligence that we have around your book of business into wherever you're doing the work.
So hopefully that, uh, that helps a little bit. It's, it's mainly to make me smarter and tell me what I don't know about the prospects and customers that I need to be
Yeah, then you guys have basically a database with tons of information. So like could I, in theory, build a tool using a bunch of search tools that are gonna go do some search? Yeah, but it's only what I can access on the, on the internet, only what's visible from the outside, and you don't necessarily get all of that data. That's a lot harder to just go scrape
Right. So we, we crawl about 67 million websites and our refresh rate is very, very fast. I think it's somewhere like a million a day or something like that. So it's real… It's, it-- The refresh rate, it's always constant. Here's the difference. Like technically there are companies that crawl X million websites per day and, and, you know, do more of that. What's different about Ocean is the fact that we've been doing this since 2015, which means that we've had to painstakingly go through the process of figuring out the math that allows us to vectorize this context in a way that's useful. Because you can go crawl whatever you like, you can go capture whatever you like off a landing page, but until you have figured out the calculations that allow you to say, "This vector looks like this vector and therefore we should cluster them," you're not going to be really getting intelligence. You're gonna be getting educated guesses at best.
And so the data volume is, in my opinion, it's very important, but what's more important is the amount of time that Michael and team have spent really, really fine-tuning the algorithm in a way that all- associates the vectors accord- appropriately. So I don't know. I know that's super, super high level, um, but hopefully that makes sense in terms of the difference in approach, because it's more about the time we've put in and the math behind it than it is about what you can and can't do
Yeah, I mean, and something that you touched on is the vector piece. And so, like, you can go out and get all kinds of information, but it's not about how much you have, but being able to pull in the right information at the right time, which is what a vector is gonna help you to, to do. Um, so that,
and you know, it's funny, it's funny you put it that way because that's really our approach to the market. Like, the way that companies have historically tried to solve the problems that we're tackling right now is let's buy more data and somewhere in there are the companies we should be targeting. It's looking for a needle in a haystack. The analogy I always use is a phone book. Like, the other data providers of the world, they'll tell you who you could call. Ocean's taking a very different approach. We wanna tell you who you specifically should call. So it's less about how much data you have and more about do you have the right
Yeah, that makes sense. Okay, so I, I want you to-- I mean, obviously, uh, you know, what could somebody do on Monday morning? They could, uh, you can reach out to Jack and check out Ocean, but, uh, but, but short of doing that, like, okay, so we've talked about a lot of different things, and some of it's very specific to Ocean, which I think is really interesting, and I, I do think is a tool that people should check out. Um, but just taking it back to, like, a seller on Monday morning, they're sitting down, they need to do some, some research, and maybe they're not-- they can't just go out and get whatever tool they want and, and, you know, with, with what they have available. How can they start using some of what you shared about today to, to create better contacts and just get, get more out of, um, Claude or whatever AI tool they're using?
Yep. So there's three things super easy to implement. The first one, like I talked about earlier, is projects, and the second thing is within the confines of projects. So if you are looking at your Claude view on the right-hand side of the screen, there is an actual context window where you can upload files that serve as a knowledge base for that project. So an example would be, I am working on a proposal for a company and I need to ground this proposal in something real. I'm gonna download a copy of their most recent 10K report and add that to the project context. That way the whole project is grounded in where they're trying to go, what their revenue growth strategy is, what their challenges have been, what their financials look like, et cetera, et cetera. So that provides the grounding that is then going to maintain more of the context within the confines of that, uh, that, uh, project, which already has shared context across the different chats anyway.
So that's a super quick and dirty way to get better results out of this. The other thing that I can't recommend enough is, uh, scheduled actions. So I have one that I run every, uh, every day at like 4:00 a.m. so I don't go through any more credits that says-- It puts together my daily briefing. Like, what do I need to do today? What are the-- What do I have outstanding? Where should I be? You know, what do I need to tackle? Where should I be spending my time? Doing that in a, uh, in a scheduled manner is basically a constant refresh of the context. So, um, it- w- would you work with it the way that you would work with a chat thread? No. But it, the frequency of the refresh is what makes it so that I can trust that, that what I need to do today list is actually correct. So the three, three easy things: use projects, add context to your projects, and use scheduled acce-
I love it. And I'll add one more that I know we talked about before, which is start connecting MCPs to the tools that you're using and see how Claude's able to pull them in at exactly the right time and make way better use of those tools than what you can if you're trying to log into the platform and download an Excel sheet and upload it.
Claude want, Claude wants help. Claude likes talking to those things. So Claude wants to help. Help him out.
awesome. I love it. Uh, okay, so I'm gonna ask you a, a, um, totally unb- non-business related, uh, question, and I'm gonna throw this one out to you on the fly. So if you had a, if you have a, a day off this weekend where you're by yourself and, and you can go, uh, y- you have the day to yourself, would you sit and read a book or would you go, go for a hike or go for a run?
I guess it depends on the day. In gener- generally speaking, probably more of a stay-at-home read-a-book guy. Um, yeah, and I feel like there's so much noise i- just in life in general today that sometimes when you have the quiet, you really gotta e- embrace that and celebrate it. So yeah, I'd say I'm more of a stay-at-home
Love it. Are there any, uh, interesting books you've read in the last couple months or year?
Let me see. The, um, I just finished, uh, "My Inventions" by Nikola Tesla, which is a great read, really interesting eye into that guy's, uh, y- you know, the just the way he thought and the way he visualized things. I thought that was really interesting. Um, I'm reading a book by Henry Dick Thompson about Vietnam specifically, uh, you know, the behind the lines vision, uh, missions in Laos and Cambodia that we weren't supposed to be on. I think that period of history is really fascinating. So big history buff. Um, and then there's probably three or four others that are, that are-- I'm drawing a blank on right now, but
I put you on the spot, so awesome. Well, Jack, thanks so much. Uh, how can people find out more about you and what you're up to?
Yeah, absolutely. If my, uh, I'm available on LinkedIn, uh, my email is jta@ocean.io, and that's the website, ocean.io. Would love to connect. I'm always open to having conversations and learning more about what people are working on. It's a fun and exciting time in the world, so maybe there's an opportunity for us to work together, maybe there's not, but either way, just love hearing from folks like Joe who are doing innovative things with this new tech paradigm. So
Yeah, with that, thanks so much for tuning in. Don't forget to subscribe wherever you listen to your favorite podcasts, and we'll see you on the next episode
Your AI was working great, and then it started handing back the same twenty companies it gave you last month.
On this episode of Pipeline, Joseph Lewin talks with Jack Taubl, Director of Sales at Ocean.io, about why AI output decays the longer you work inside one chat, and what to do about it. Jack's explanation is structural. Providers earn their margin on the efficiency of their models, so a long thread gets compacted to control cost, and quality erodes by design rather than by accident. His fix takes three steps a seller can run Monday morning: work inside a project instead of one endless thread, load that project with real grounding documents like a prospect's most recent 10-K, and set a scheduled action that rebuilds your daily briefing on a fixed refresh. He runs his at 4:00 a.m. to keep credits down.
Chapters:
Intro: why AI stops working for GTM
Why AI output degrades: compaction and vague prompts
Fixing context drift with Claude Projects
Contextual grounding: one account record across every system
The LLM is the execution layer, not the database
Why MCPs are replacing platform-hopping
Bad data, misaligned teams, and the compounding hidden cost
Turning account data into segmentation and territory strategy
Pulling account context into Claude for rep briefings
Why more data isn’t the differentiator
Monday-morning playbook: Projects, context files, scheduled tasks, MCPs
Rapid fire: books over hikes
Where to find Jack
What you can use by Monday:
- Move broad, open-ended work into a project instead of one long chat, so every new chat inherits the context from the others.
- Upload real grounding documents into the project knowledge before you ask for anything, like a prospect's most recent 10-K.
- When an answer feels off, open a fresh chat inside the same project and switch the reasoning model instead of arguing with the drift.
- Set a scheduled action that rebuilds your daily briefing on a fixed refresh, so the list you work from is current.
- Ask specific questions with an end in mind. Ambiguous questions return ambiguous results every time.
Pipeline is short conversations with GTM peers who are actually filling pipeline right now, so you leave with something you can use by Monday. New episodes weekly on Apple Podcasts, Spotify, and wherever you listen. Produced by Podcast 2 Pipeline.