In June, we ran a hands-on web scraping workshop at New York Tech Week. One attendee asked a question that stuck with us as it was then echoed by others and has cropped up in client meetings too: “Why can’t we just use Claude (or ChatGPT, or the latest LLM) to build our scraping pipeline?” The answer deserves a deeper explanation than we had time for in a workshop. So we published a white paper. Here’s why.

LLMs can turn a blank page into a working proof of concept in minutes. That’s real. But we’ve watched teams hit a wall moving those prototypes into production. The question they ask, “Why can’t we just keep using the LLM?”, sounds reasonable until you consider what happens on run two, run two hundred, and run two thousand.

The Problem

Large language models are non-deterministic by construction. The same prompt can produce different output from one run to the next. For a chatbot, that variability is a curiosity. For a system extracting prices, financial disclosures, or inventory levels that feed trading models, compliance filings, or repricing engines, it’s disqualifying.

Research on ChatGPT’s code generation shows the scale. Sending the identical prompt five times produced completely different test outcomes on 75.8% of one benchmark’s problems. Even at temperature=0, the setting most associated with removing randomness, 43.6% of problems still diverged across five identical runs.

What This Breaks

Non-deterministic extraction creates four operational problems:

Silent field drift: Same page yields different values across runs, with no alert. Wrong data flows silently into your pipeline.

Fabrication under ambiguity: When uncertain, the model produces plausible-looking hallucinations instead of failing clearly.

No baseline for change detection: When output changes, you can’t tell if the website changed or the model just answered differently this time.

Unpredictable cost: Model API calls compound across catalogs. Reconciliation strategies multiply costs and latency.

Why Institutions Can’t Accept This

Investment banks need to reproduce any historical data point on demand per SEC Rule 17a-4. Live LLM calls can’t guarantee that.

Retailers feed pricing data into automated repricing engines. A hallucinated price hits production immediately.

Data resellers sell feeds under SLAs guaranteeing schema stability and field accuracy. Non-determinism makes those commitments indefensible.

The Architecture That Works

Use an LLM to turn plain English into a working prototype. Engineers refine it, add validation, lock it down. The deterministic runtime executes the same extraction logic every time, with no model inference involved.

That prototype becomes a versioned, auditable artifact. It can be tested, code-reviewed, and rolled back. Every run writes a full audit trail. You get the speed of AI-assisted development without sacrificing the reliability institutional data programs require.

This is what guides Sequentum Agent Builder. Users describe the task. Agent Builder builds a prototype in minutes. Engineers refine it. The deterministic runtime executes it identically every time, with full audit trails and zero AI inference in the extraction step.

Read the Full Argument

We published a white paper that walks through the evidence, the ways non-determinism breaks institutional use cases, and the architecture that works.

Download “Why Not Just Use Claude?”

Reach out to us at info@sequentum.com. We’d love to help you solve your data challenges.