![[Un]Churned by Gainsight](https://substackcdn.com/image/fetch/$s_!7AoO!,w_120,h_120,c_fill,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffe2167ac-0bcf-4575-9712-8d5ef3588851_300x300.png)
![[Un]Churned by Gainsight](https://substackcdn.com/image/fetch/$s_!hKlf!,e_trim:10:white/e_trim:10:transparent/h_72,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe14b36dd-52b9-48a3-9f93-3f6a459d55ff_1344x256.png)
Above the Model: Where AI Value Is Really Built
What building autonomous agents taught me about where AI value is actually moving.
![[Un]Churned's avatar](https://substackcdn.com/image/fetch/$s_!vkJ0!,w_108,h_108,c_fill,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0464ad30-26c2-4f32-b429-ae4283dd5586_200x200.png)
Models keep getting better. Inference keeps getting cheaper. And yet organizations are hitting token cost ceilings with little attributable ROI to show for it.
The bottle neck is not intelligence. Most organizations just haven’t built the layer where the value actually lives. There are two common problems that stand between AI investment and the value it delivers:
The first is that raw intelligence doesn’t automatically translate into economic value. The system that can handle real-world scenarios reliably, decides how much reasoning a task deserves, and when to involve a human, that’s the bottleneck. Not the model.
The second is that better models increase in demand. They unlock new uses. More workflows become viable. More decisions become worth augmenting. More customers become worth reaching. The demand for intelligence is effectively infinite.
These two problems point to the same answer. Think of the foundation model as the brain—general purpose, powerful, expensive to run at scale. Economic value comes not from the brain itself, but from the neural pathways that encode specific skills. Those pathways are built through messy iterations in the real world. They improve through repetition. They get extremely efficient at the task.
Models are now becoming an input. The system and harness around it is becoming the product. That shift has significant implications for Customer Success leaders trying to make sense of their AI roadmaps, and for understanding why some organizations are compounding advantage right now while others are still waiting for the technology to deliver.
Acquiring the Skill
I started as a data scientist. Eight years and several AI paradigm shifts later, I’m Director of AI Product at Gainsight, building Atlas, our autonomous agent platform that manages entire customer lifecycles. This journey has shaped the way I think about the gap between AI capability and AI value.
Building production agents is much harder than building good demos. The value capture lies in the last mile, and the last mile is where most organizations underestimate the work.
Traditional software engineering is about expressing logic. You define the rules upfront and the system runs on it. Agent engineering is about governing judgment. The rules don’t cover everything and the edge cases only show up in production. LLMs have jagged capability profiles. They can perform brilliantly on one task and fail unpredictably on an adjacent one.
My team experienced this firsthand while implementing intelligent outreach for an Atlas customer. A “Do Not Contact” designation had been entered into a middle name field as a workaround. It had always been handled implicitly by the humans who knew to look for it. When translating emails into gendered languages, a similar problem surfaced. Templates needed to be gender-neutral in ways that had never been formally specified.
These are edge cases in our data that nobody anticipated and weren’t visible until the agent hit them.
The second problem is different in kind: institutional knowledge that was never written down because it never needed to be. The full set of reply categories—what the system should do when a customer says they’ve already renewed, or that the wrong person was contacted, or that they want to escalate, existed almost entirely in people’s heads and in years of email threads. Humans handled it implicitly. The agent couldn’t.
Your team carries an enormous amount of undocumented operational knowledge. Deploying AI forces that knowledge to the surface in a process that’s neither fast nor clean. What looks like an AI problem is often a knowledge problem. What looks like a failed agent is often an organization meeting itself for the first time.
Codifying the Pattern
A key limitation of LLMs is that they cannot learn on the job. They don’t fail, correct, and improve the way a practitioner does. The neural pathway doesn’t strengthen on its own. That’s what the system above the model has to do.
The “Do Not Contact” case I mentioned earlier is a good example of how this works in practice. The fix wasn’t a single patch. We separated contact name cleaning into its own agent, reviewed every eval failure in that category, and tightened prompt instructions with targeted examples.
Issues caught during human review went into judge prompts and validation datasets. That specific case required the cleaning agent to flag contacts as unusable rather than attempt to clean them. It was a decision that required rethinking task decomposition entirely and only became clear after seeing the failure in production.
If you want a closer look at how we actually build and stress-test the evaluation layer behind Atlas agents, we wrote about it here: Before the Agent Ships: The Human Layer Behind Gainsight’s Agent Evaluation Process. For more weekly insights on AI, retention, and Customer Success, subscribe to the [Un]Churned Substack.
Currently the learning process requires careful distillation of observed issues and fixing them across the system. Our engineers use AI every step of the way, but it’s inefficient and requires supervision. Learning harnesses will improve over time, but I expect there will remain a system above the model responsible for turning experience into capability.
Neural pathways don’t form from a single exposure. They strengthen through repetition, through failure, through correction. You can’t just read a book on riding a bike, hop on, and compete in a triathlon. You fall, correct, fall again, and eventually your body internalizes something your brain couldn’t have calculated upfront. Once it clicks, it’s automatic and nearly impossible to unlearn.
This process produces the raw material for operational improvement today and training substrate for more specialized, efficient models tomorrow. Organizations deploying agents now and instrumenting them well are accumulating a moat most haven’t fully recognized yet.
Executing Efficiently
Once the skills are acquired and the learning system is in place, the optimization problem shifts. The goal is no longer capability. It’s efficiency, which is another defining characteristic of neural pathways.
The demand for intelligence is effectively infinite. The answer is spending it deliberately. Not every task deserves frontier reasoning. Not every task needs a language model at all. As domain-specific systems mature, fine-tuning smaller models on your own deployment data becomes standard practice. The encoded knowledge accumulated through deployment is what makes that specialization possible, and adds another layer of defensibility to this layer.
The Pathway Is the Product
That’s the real opportunity above the model. Not building a better brain, but building a better system for putting it to work.
The current gap between AI investment and attributable ROI is not evidence that the technology is overhyped. Intelligence isn’t the bottleneck for a majority of economic tasks anymore. The system that makes it useful is still being built, and the value it unlocks will be larger than what came before.
The assertions in this article are meant for the next year or two. Or months. Or days. Beyond that, we’re all guessing. We might end up with a Technocore negotiating in the Datasphere while humans go on pilgrimages. IYKYK. Until then, there’s plenty of interesting work to be done above the model.
Shantan Reddy is Director of Product, AI at Gainsight, where he leads Atlas, Gainsight's autonomous agent platform for Customer Success. Outside work, he plays racquet sports and reads too much sci-fi.