Research note · July 2026

Earnings Call Interrogator

When is a small, locally-run AI model good enough for a narrow, repeatable task, and when is the premium for paid cloud models worth paying? I use earnings-call Q&A extraction as the test: same transcripts, scoring, and PDF template; only the model changes.

Models tested3
Calls per model10
Local model cost$0
TrainingQwen3-8B · GCP L4

The problem

When is a small, locally-run AI good enough for a narrow, repeatable task, and when is the premium for paid cloud models worth paying?

Public companies hold quarterly earnings calls. The interesting content is the Q&A: what was asked, how management answered, where answers were partial, and how much of the call was scripted versus live. I built an end-to-end extractor that produces a structured research PDF, then used that product as the measurement harness.

How it works

Ingest, preprocess, extract, score, report.

Preprocessing handles FactSet and Motley Fool transcript formats, splits management prepared remarks from analyst Q&A, and tags who is speaking. Extraction returns a six-field structured output. Scoring adds a credibility rating, how much of the call was live Q&A (versus scripted), analyst attribution, and evasion flags (answers that seem to dodge the question). PDFs are watermarked by model.

Training the AI

Base model: Qwen3-8B, an 8-billion-parameter open-weight model, adapted via 4-bit QLoRA (a memory-efficient training method) on a GCP NVIDIA L4. Version 1 used 4,448 pattern-matched labels and was unusable (training loss 3.67). Version 2 used 499 structured labels from a GPT-4o review and became the production path (loss 0.016, ~9.4 hours). Local serving via Ollama (a tool for running AI models locally) using a compressed model file on a mini PC for measured CPU runs.

Results

Cost vs. speed trade-off

Publication figure used in the research brief. Bubble size and labels include evasion-flag density, the percentage of questions where the model surfaced something worth reviewing.

Figure 2 · Cost vs. speed trade-off Static chart
Cost vs. speed trade-off
X = time per call (seconds) · Y = cost per call in cents. Solid = measured. Hollow × = estimated Qwen-on-GPU time (~6× faster than CPU; not re-benchmarked). Labels: time · cost · evasion-flag density.
Figure 3 · Time per call distribution
Latency dispersion by engine
Ten calls per model; diamond = average.
Figure 4 · Evasion-flag density
Evasion capture density
Evasion flags per question. These are review candidates, not confirmed evasions.

Qwen3-8B v2

$0
Locally-trained specialist
Time per call (CPU / GPU est.)
~205s / ~34s
Evasion flags surfaced
~11%
First-pass
8/10 (10/10 after retry)

GPT-5.5

~11¢
Best value paid option
Time per call
~67s
Evasion flags surfaced
~39%
First-pass
10/10

Claude Fable 5

~31¢
Fastest, most flags, highest cost
Time per call
~55s
Evasion flags surfaced
~33%
First-pass
10/10
Scale

Quarterly volume → annual cloud cost

Projection from measured per-call costs. Drag the slider to explore volume scenarios.

Figure 5 · Cost curve
X = earnings calls screened per quarter. Y = projected annual API spend.
Volume explorer

Calls screened per quarter

1,000 / quarter

ModelProjected annual API cost
Qwen (local)$0
GPT-5.5$0
Fable 5$0

Annual cost = quarterly calls × 4 × cost per call. Local hardware and power excluded.

Interpretation

Decision framework

How I read the results: solid reference tables, not interactive widgets.

When to lean which way

SituationLean toward
Low volume, vendor data sharing acceptableGPT-5.5
Need max speed/flags and QA budget existsFable 5
High volume, privacy, or unit-cost controlLocal Qwen (GPU if speed matters)
Unclear volume / early pilotStart with a paid cloud model; reassess at ~500–1,000 calls per quarter

Limits of this note

ConstraintImplication
10 companies testedDirectional, not large-sample proof
Credibility / evasion scoresRule-of-thumb scores for review, not independently verified
GPU speed estimateEstimated from throughput ratios, not re-run
ScopeResearch note. Not a product claim.

Main lesson from training: label quality beat label quantity. The extraction and scoring workflow matters more than which model you use.

More

PDFs and related notes