Enterprise spending on generative AI climbed to between $30 billion and $40 billion through mid-2025, with 95% of pilots producing no measurable impact on profit and loss. The broader AI agents market sat at $7.6 billion in 2025, on track to reach $47.1 billion by 2030. Most of the gap between investment and return surfaces in production, where workflows that succeeded in a demo collapse under the weight of unit economics. Every token consumed and every retry attempted lands on the bill.
Cornelius Renken leads AI Apply at Kombo, an HR tech platform that turns fragmented job submissions into one-click applications across enterprise systems like Workday, iCIMS, and SuccessFactors. In addition to his technical leadership, Renken is a member of the Fast Company Executive Board, a highly selective, vetted community of business leaders shaping the future of business through technology and design. At Kombo, he recently founded the company’s first applied-AI initiative, scaled the engineering team, and now leads a platform that operates across thousands of employer configurations.
The Pilot-to-Production Cost Cliff
Gartner forecasts that 40% of agentic AI projects will be scrapped by 2027 from cost overruns, weak risk controls, and unclear business value. Carnegie Mellon’s evaluations of leading agents put their multi-step task completion rate at 30 to 35%, a number that drops further as task chains lengthen. The pattern is consistent across the data. Demos prove capability; production exposes economics. A model that solves a problem 30% of the time is not a 30% solution at scale, because the failures compound into retry costs and manual recovery, and they cost something harder to price: trust.
AI Apply was framed from its first commit as a unit-economics problem. The product automates job submissions into enterprise applicant tracking systems, and the customers paying for it are job boards and recruitment agencies whose own pricing models depend on a tight cost per application. Renken set the bar early. The agent had to decrease the CPA, the cost a job agency pays to push an application from a board to an employer, down to a level that left margin for everyone in the chain. The work pulled the cost down up to 8x. That number, more than any single accuracy benchmark, unlocked the platform’s adoption by leading recruitment players, which have collectively pushed hundreds of thousands of submissions through the system and now report 3 to 5 times more completed applications and a 99.9% delivery success rate.
“Cost is a design constraint, not a finance problem you solve later,” Renken says. “If the math does not work at the unit level, no amount of accuracy improvement saves the product. That gets decided in the architecture, not in the demo.”
Model Selection as Engineering Discipline
LLM inference costs have fallen by roughly a factor of 1,000 in three years, with GPT-4-equivalent performance now available at around $0.40 per million tokens compared with about $20 in late 2022. Inside that decline sits a wide spread. Switching from a frontier model to a smaller one for the right subset of tasks can cut per-task spend by 70 to 90%, and in extreme cases by 190x for high-volume routine workloads. The economics push toward routing rather than monoliths: send each step to the cheapest model that can finish it correctly, and reserve the expensive ones for the steps that actually need them.
Inside the platform, model selection runs as a live decision per step in the workflow. A semantic form-understanding step that would burn through reasoning tokens on a frontier model is routed to a smaller, cheaper one when the surface complexity is low. A high-stakes interpretation step, like resolving an ambiguous validation rule on a Workday form, escalates to the larger model only when the cheaper one signals uncertainty. The architecture also caches and trims context aggressively, because every additional token in the prompt compounds across thousands of forms a day. The result is a workflow that holds margin even as employer configurations multiply.
Renken puts it bluntly. “The cheap model is the right tool most of the time. The mistake is reaching for the expensive one because you trust it more, when what you actually need is to know exactly when the cheap one is going to be wrong.”
The Caching Layer That Keeps the Lights On
Operational research on production LLM systems finds that disciplined prompt caching, semantic caching, and exact-match caching can deliver 40 to 60% cost reductions without measurable quality loss. The applicant tracking system market the platform integrates into reached $17.22 billion in 2025 and is projected to climb to $34.83 billion by 2034, with much of that growth flowing into AI-driven features. As the agentic layer on top of those systems grows, the cost question stops being whether to use AI and becomes which results need to come from a model at all, and which can be served from memory.
The system implements caching at multiple layers. Form structures, once interpreted, are cached so identical fields on the same employer system never get re-interpreted. Validation rules, once learned, are stored against the system that produced them. This is the design pattern Renken laid out in his DZone article on tools for building deterministic LLM systems, where the argument is that a deterministic LLM workflow is largely a discipline of remembering what you have already paid to find out. The caching layer is also what makes telemetry useful, because it gives the system a stable reference for what should and should not trigger a new model call.
“Caching is the unglamorous part of LLM engineering and also the part that decides whether you have a business,” Renken observes. “Each cache hit is a model call you did not have to pay for. Multiply that across thousands of submissions a day and the difference between viable and unviable shows up in the gross margin line.”
When Workflows Run for Two Hours
Long-horizon agent workflows are where the cost question becomes a survival question. A two-hour run that fails at minute 110 has spent compute, tokens, and human attention with nothing to show for it, and that loss has to be priced into every successful run. Carnegie Mellon benchmarks put leading agents’ multi-step success rate at 30 to 35%, which means the median uncontrolled long-horizon agent loses money on most attempts. The discipline that closes the gap is not better models; it is recovery.
AI Apply embeds self-healing into the workflow itself. The agent emits structured signals at each decision point, validates the state of the form against expectations, and chooses a recovery path when the surface deviates from what it expected. If a CAPTCHA appears mid-flow, the system has a defined response. If a session expires, the system re-authenticates and resumes from the last validated state rather than from scratch. Renken has written about this orientation toward designed obsolescence in his HackerNoon piece on writing code that will be obsolete in two years, where he argues that the engineering decisions worth making are the ones that make the system cheap to repair when reality moves.
“Self-healing is a cost-control discipline before it is a reliability one,” Renken explains. “If a workflow can recover from minute 110 instead of restarting at minute zero, you have just saved 109 minutes of compute and the human investigation that goes with it. That is what makes a long workflow shippable.”
Beyond Hiring: A Template for LLM Economics
Gartner projects that 33% of enterprise software will include agentic AI by 2028, and the AI agents market is on track to grow from $7.6 billion in 2025 to $47.1 billion by 2030. 61% of organizations are already exploring agent development, and roughly 70% of corporate AI experimentation is happening inside HR, where hiring volumes and integration complexity stress-test the same engineering questions other industries will face once their workflows reach comparable scale. The disciplines that decide the platform’s gross margin (model routing, intelligent caching, designed-in recovery) generalize wherever LLM agents have to run autonomously across systems they do not control.
The implications travel well beyond hiring. A claims-processing agent stitching together insurance backends, a procurement bot navigating supplier portals, and a compliance checker traversing government filing systems each face the same economic equation: the cost of running an agent has to stay below the cost of the human work it replaces, and that constraint has to hold across thousands of variations of the underlying systems. The standard the field is converging on, from correctness-as-goal to repeatability-as-standard, is at root an economic standard. A workflow that runs the same way every time can be priced. A workflow that improvises cannot.
“The teams that figure this out first will not be the ones with the most advanced models,” Renken concludes. “They will be the ones who treated the boring parts as the actual product. Caching, routing, recovery. That is where the economics live, and that is where the next decade of enterprise AI gets built.”
Leave a Reply