Agentic AI
New framework tackles cost estimation for AI-driven software development
Researchers propose ACEM, a model designed to predict expenses in projects where autonomous agents handle coding while humans oversee and validate.
A fundamental shift is underway in software development as autonomous AI agents increasingly take on coding and implementation tasks. But the tools used to estimate project costs for decades—models like COCOMO II and Story Points—were built for human-led teams and no longer capture the true expenses of this new paradigm.
Researchers have introduced ACEM (Agentic Cost Estimation Model), a framework that rethinks how to forecast expenses when large language models and autonomous agents do the heavy lifting while humans focus on planning, oversight, and quality assurance.
Why existing models fall short#
Traditional cost estimation assumes that software development cost flows largely from human effort in design, implementation, and testing. But when AI agents drive implementation, the cost structure changes fundamentally. Development now incurs LLM token consumption as agents process information and generate code, human-in-the-loop oversight as people validate and correct agent outputs, and infrastructure expenses for running the agent platform itself.
These costs are also unpredictable in ways human effort is not. The same task may consume wildly different numbers of tokens across multiple agent attempts, follow different reasoning paths, and require varying amounts of human correction.
How ACEM works#
The model breaks total agentic development cost into three components:
- LLM cost: Driven by token consumption across agent actions
- Human oversight cost: Effort required for validation and correction
- Infrastructure cost: Resources needed for agent orchestration and tooling
To account for variability, ACEM introduces three key constructs:
- Revision Factor (RF): Captures additional token consumption from rejected outputs and retry loops
- Context Factor (CF): Reflects rising token usage as context windows grow during development
- HITL Intensity Score (HIS): A four-level classification system for human oversight requirements, ranging from minimal to intensive
The framework also maps established project-sizing metrics—Use Case Points, Story Points, and Function Points—to estimated token counts, allowing teams to leverage existing data for agentic forecasting.
