Methodology · White paper

Stop underestimating
LLM projects.

Most AI project estimates are wrong by 3–10×. DRIFT is a 7-dimension framework that quantifies the factors classical methods miss.

By Brain Breaking LLC · 2025

The DRIFT Formula


T_total = 800·Cx·A + 0.75·C + O·(18 + 3·Rd + 4.8·V·A) + 150·Tc

Cx — Context ComplexityA — Agentic DepthC — Corpus SizeO — Orchestration StepsRd — Retrieval DepthV — Validation RigorTc — Tool Call Count

Get the full white paper

Worked examples, dimension scoring guides, and a comparison to classical CCV risk theory. Free — enter your email below.

No spam. We'll send the PDF to your inbox.

The framework

7 dimensions of LLM complexity

Each dimension captures a distinct source of effort that classical estimation methods ignore entirely.

Cx

Context Complexity

How large and structured is the context the model must reason over? Unstructured docs, multi-file codebases, and interleaved tool outputs all increase Cx.

A

Agentic Depth

Does the system make decisions autonomously across multiple steps? Single-turn Q&A scores low; multi-step agents with branching logic score high.

C

Corpus Size

The volume of documents, code, or data the system must index, retrieve from, or reason over at runtime.

O

Orchestration Steps

How many discrete model calls, tool invocations, or sub-agent handoffs does a single task require? Each step compounds latency and error rates.

Rd

Retrieval Depth

How many retrieval hops does the system perform? Single vector lookup vs. multi-hop reasoning over knowledge graphs have very different costs.

V

Validation Rigor

How strictly must outputs be verified? Structured output driving downstream systems needs schema checks, evals, and human review loops.

Tc

Tool Call Count

The number of distinct external tools the model can invoke. Each tool requires integration work, error handling, and permission design.

Try it

DRIFT Calculator

Adjust each dimension to see your estimated token budget and cost in real time.

Cx Context Complexity 2

1 — Trivial3 — Moderate5 — Cross-cutting

A Agentic Depth 1.5

1 — Single-turn3 — Multi-step5 — Autonomous

C Corpus Size 2,000 lines

015,000 lines30,000 lines

O Orchestration Steps 5 steps

030 steps60 steps

Rd Retrieval Depth 1

1 — None3 — Multi-hop5 — Graph

V Validation Rigor 2

1 — None3 — Schema checks5 — Human review

Tc Tool Call Count 10 calls

075 calls150 calls

Get AI estimate →

Answer a few questions and we'll score your project's DRIFT dimensions.

T_total

2,400 tokens

Formula breakdown

800·Cx·A 2,400

0.75·C 1,500

O·(18+3·Rd+4.8·V·A) 320

150·Tc 1,500

Story points

≈ 3.0 story points

Cost estimate (output tokens)

Haiku 4.5 $1/1M in · $5/1M out

$0.01

Sonnet 4.6 $3/1M in · $15/1M out

$0.04

Opus 4.7 $5/1M in · $25/1M out

$0.06

Risk level

Low — straightforward

Stop underestimatingLLM projects.