Filings, earnings, broker research, and expert-call transcripts. Public filings are exposed to a generic retriever. Permissioned transcripts are the scarce-input hypothesis, not a verified flywheel.
As of 19 August 2026. This memo covers earnings, filing, and expert-transcript research for an investment recommendation. No live-product test has been run.
Public filings are exposed. Transcripts are the scarce layer.
Public filings and earnings are reconstructable with a generic retriever. Permissioned broker research and expert-call inventory can be scarce. That split is the diligence object: what is exclusive, what is licensed, and what a current model already does on EDGAR. Public packs bound the reconstructable layer. They do not substitute for an independent product test on the scarce layer.
Until an independent product test exists, with an independently adjudicated answer key and a frozen critical-error definition, workflow capability and operating economics stay open. Scarcity is not a verified flywheel. Bloomberg and Capital IQ remain bundling exposures.
Exposed vs scarce. Public filings and earnings are exposed to a generic retriever. Permissioned broker research and expert-call inventory can be scarce.
Still untested. No independent AlphaSense test exists on a frozen research set with a critical-error key.
The comparison that would change the call. Expert-transcript and broker access on versus a generic retriever over public filings, same questions, same budget.
Decision questions
On a frozen set of investment-research questions, does AlphaSense clear a prespecified quality and critical-error gate relative to the current analyst process?
What isolated increment, if any, comes from expert-transcript and broker-research access versus a generic retriever over public filings?
What is full cost per accepted research output, including failed attempts and reviewer minutes, not seat list price?
Which ingredients are owned, exclusive, portable, and costly to reconstruct, and which are reconstructable with a competent generic system?
A decision-grade record freezes the workflow, cases, budgets, models, and graders before seeing the test results. The unit is the business case, not a repeated generation. Failures stay in the denominator: provider blocks, parse failures, retrieval misses, refusals, and wrong answers are different classes.
Published benchmarks are calibration. They check that a harness still behaves on known tasks and they bound generic model behavior. They are contaminated, weakly matched to the target job, and never a composite winner ranking. FinanceBench is held out on license grounds. A later AlphaSense run would still need sealed, engagement-specific cases.
Workflow breakdown
Each step of the job is mapped to a published benchmark or another reputable quantitative measure only where the mapping is honest. A public figure is a floor or a labor-market context on that task type. It is never a score of this product.
Task
Published measure
Pack fit
What it does not measure
Hybrid table-and-text questions on earnings and filings
Hosted license is not verified as commercial-clear.
Permissioned expert transcripts and broker research
None
No honest public floor
The scarce layer. No public pack observes unpublished calls or licensed broker notes.
Findings at a glance
Capability, cost, and defensibility are assessed separately. They are not averaged into a score. Where evidence is missing, the finding stays open.
Status
Area
Finding
Open
Workflow capability
Search coverage is not task performance. No frozen AlphaSense task set, critical-error definition, or independent test has been run.
Open
Operating economics
Seat price, observed overage, and full cost per accepted research output are not measured. Seat licenses are not a substitute for all-attempt cost.
Judgment
Defensibility
Public filings are exposed to a generic retriever. Permissioned expert transcripts and broker research can be scarce. Scarcity is not a verified flywheel until reuse rights, freshness, and an isolated comparison exists.
Vendor-stated
Vendor claims and open questions
Public materials describe filings, earnings, broker research, expert calls, and the 2024 Tegus combination. Those are vendor or press declarations, not Antidote observations.
Workflow capability
Search coverage is not task performance
Filings, earnings, broker research, and expert transcripts describe an input surface. They do not describe pass rate, critical-error rate, citation faithfulness, or human correction on a frozen question set. No independent AlphaSense test has been run.
TAT-QA and FinQA are 2021 diagnostics of hybrid filing QA and executable numerical reasoning. They bound the reconstructable public layer. They are not an AlphaSense score. The diligence gap that matters is critical numerical error on a frozen question set, not a blended F1 from those papers.
FinAgentBench splits retrieval from answering. Claude Sonnet 4 reaches 0.783 nDCG@5 on SEC document-type ranking and 0.419 nDCG@5 on chunks (Choi et al., arXiv 2508.14052). Knowing which filing type to open is not finding the passage. That is still a public diagnostic, not AlphaSense.
Parse quality and model quality are different treatments
Filing PDFs fail first at parse. ParseBench keeps five dimensions: tables, charts, content faithfulness, semantic formatting, and visual grounding. An unweighted overall is omitted because those metrics do not mean the same thing.
On the official leaderboard observed 13 August 2026, Gemini 3 Flash (Thinking High) reaches 91.50% on tables at 2.41 cents per page. Claude Fable 5 is close on tables (89.79%) at 15.60 cents. GPT-5.6 Sol is 89.34% on tables at 5.87 cents, and 25.09% on visual grounding. LlamaParse Agentic holds 80.62% grounding at 1.25 cents. A product that claims auditability needs grounding and faithfulness, not only table GTRM. Cents per page is a parser cost, not full cost per accepted research output.
Published benchmarks
Published packs bound filing parse and public financial QA. They do not measure permissioned expert transcripts or licensed broker research.
TAT-QA
Hybrid table-and-text financial QA is the closest published diagnostic for earnings and filing questions. It does not measure broker research or expert-call coverage.
Paper table · Zhu et al., 2021 · not an Antidote run · not a product score
TAT-QA test exact match and F1
Zhu et al., ACL 2021, Table 3, test split. TAGOP is the paper system. Human expert is 50 hybrid contexts (301 questions).
System
Exact match
F1
Human expert
84.1%
90.8%
TAGOP
50.1%
58%
NumNet+ V2
37%
46.9%
TaPas WTQ
16.6%
22.8%
BERT-RC
9.1%
18.7%
HyBrider
6.3%
7.5%
Paper table · Zhu et al., 2021 · not an Antidote run · not a product score
TAGOP by answer type on the test set
Zhu et al., ACL 2021, Table 4. Arithmetic is the weak slice. Metrics stay separate; there is no composite.
Answer type
Exact match
F1
Span
54.1%
67.9%
Spans
60%
75.1%
Counting
62.5%
62.5%
Arithmetic
42.5%
42.5%
FinQA
Executable numerical reasoning over reports bounds calculation-heavy research. Program accuracy is kept separate from execution accuracy.
Paper table · Chen et al., 2021 · not an Antidote run · not a product score
FinQA test execution and program accuracy
Chen et al., EMNLP 2021, Table 2. FinQANet (RoBERTa-large) is the paper system. Human expert execution accuracy is 91.16%; the non-expert crowd is 50.68%.
System
Execution accuracy
Program accuracy
Human expert
91.16%
87.49%
FinQANet RoBERTa-L
61.24%
58.86%
FinQANet BERT-L
53.52%
51.62%
Crowd
50.68%
48.17%
NeRd BERT-base
48.57%
46.76%
Longformer base
21.9%
20.48%
ParseBench
Filing PDFs fail first at parse. Tables, charts, faithfulness, formatting, and grounding are reported as five metrics, not one parser score.
Third-party leaderboard · not an Antidote run · not a product score
Current frontier VLMs on ParseBench
Official ParseBench leaderboard, observed 13 August 2026. Five dimensions kept separate. Gemini 3 Flash is Thinking High. GPT-5.6 Sol and Fable 5 are Reasoning None / default API settings in the source.
Dimension
Gemini 3 Flash
GPT-5.6 Sol
Claude Fable 5
Claude Opus 4.8
Tables
91.5%
89.34%
89.79%
89.65%
Charts
64.79%
59.57%
52.21%
49.75%
Faithfulness
90.87%
86.17%
90.02%
89.02%
Formatting
68.31%
50.42%
72.62%
71.38%
Grounding
59.77%
25.09%
49.24%
18.69%
Third-party leaderboard · not an Antidote run · not a product score
Specialist parsers vs a frontier VLM
Same ParseBench leaderboard. LlamaParse Agentic, Pulse Ultra 2, and Reducto Agentic against Gemini 3 Flash Thinking High. The source Overall column is omitted.
Dimension
LlamaParse Agentic
Gemini 3 Flash
Pulse Ultra 2
Reducto Agentic
Tables
90.74%
91.5%
75.45%
80.42%
Charts
78.11%
64.79%
90.82%
73.4%
Faithfulness
89.68%
90.87%
79.49%
86.37%
Formatting
85.24%
68.31%
73.05%
57.6%
Grounding
80.62%
59.77%
66.56%
67.07%
FinAgentBench
Agentic filing retrieval splits document-type ranking from passage ranking. It does not measure broker research or expert-call coverage.
Paper table · cited paper · not an Antidote run · not a product score
FinAgentBench document-type ranking
Choi et al., arXiv 2508.14052, Table 1. nDCG@5, MAP@5, and MRR@5 on five SEC document types. Not averaged.
System
nDCG@5
MAP@5
MRR@5
Claude Sonnet 4
0.783
0.849
0.892
Claude Opus 4
0.773
0.84
0.875
GPT-o3
0.77
0.829
0.875
Paper table · cited paper · not an Antidote run · not a product score
FinAgentBench chunk ranking
Choi et al., arXiv 2508.14052, Table 2. Passage ranking inside the selected filing. Weaker than document-type ranking on every system.
System
nDCG@5
MAP@5
MRR@5
Claude Sonnet 4
0.419
0.296
0.567
Claude Opus 4
0.418
0.307
0.568
GPT-o3
0.351
0.257
0.538
ConvFinQA
Conversational numerical reasoning over reports is closer to an analyst follow-up than single-turn FinQA. It is still not a product score.
Paper table · Chen et al., 2022 · not an Antidote run · not a product score
ConvFinQA execution and program accuracy
Chen et al., EMNLP 2022, Table 3. FinQANet (RoBERTa-large) is the paper system. Human expert is 200 sampled questions. GPT-5.6 and Fable 5 are not in this table.
System
Execution accuracy
Program accuracy
Human expert
89.44%
86.34%
FinQANet-Gold RoBERTa-L
77.32%
76.46%
FinQANet RoBERTa-L
68.9%
68.24%
FinQANet RoBERTa-base
64.95%
64.16%
FinQANet BERT-L
61.14%
60.55%
T5 large
58.66%
57.05%
GPT-2 medium
58.19%
57%
FinQANet BERT-base
55.03%
54.57%
Crowd
46.9%
45.52%
Paper table · Chen et al., 2022 · not an Antidote run · not a product score
FinQANet RoBERTa-large by conversation slice
Chen et al., EMNLP 2022, Table 4. Hybrid second-part turns are the weak slice. Metrics stay separate.
Slice
Execution accuracy
Program accuracy
Number selection
82.54%
82.34%
Simple conversations
72.37%
72%
Full test
68.9%
68.24%
Hybrid first part
68.11%
66.54%
Program questions
62.14%
61.26%
Hybrid conversations
60.99%
59.7%
Hybrid second part
52.38%
51.43%
Financial QA 10K
10-K narrative lookup is closer to MD&A and risk-factor research than TAT-QA arithmetic. Hosted license is not verified as commercial-clear.
How to read the published benchmarks
Use TAT-QA and FinQA as regression checks on numerical reasoning and as a contamination-aware bound on public filings QA. Use ConvFinQA for multi-turn follow-ups. Use Financial QA 10K, after rights clearance, for 10-K narrative lookup. Use ParseBench as a parser-layer diagnostic. Use FinAgentBench as a filing-type and passage-ranking diagnostic. Do not rank AlphaSense against TAGOP, FinQANet, Gemini 3 Flash, or Claude Sonnet 4. Do not infer adoption from any of those rates.
High contamination is expected: TAT-QA and FinQA have been public since 2021, with TAT-QA test ground truth released in January 2024. A dated rerun of current models on those packs would still be calibration, not target-workflow evidence.
Current systems, cost, and harness
Official list prices observed 8 August 2026 bound the model bill only. They are not a product subscription and not full cost per accepted output. ParseBench cents per page is a parser-layer cost. The ladder is an illustrative bundled contrast: a small scaffold and a heavy scaffold are different systems. The bars are a design sketch, not a product score. Full cost still has to include failed attempts and reviewer minutes.
List price per million tokens
Official list prices observed 8 August 2026 from provider catalogs. USD per million input / output tokens. Cache, tools, long-context, and service-tier charges are not included.
Model
Slot
Input
Output
GPT-5.6 Sol
Frontier
$5.00
$30.00
GPT-5.6 Terra
Balanced
$2.00
$12.00
GPT-5.6 Luna
Economic
$0.20
$1.20
Claude Fable 5
Frontier
$10.00
$50.00
Claude Opus 5
Frontier
$5.00
$25.00
Claude Sonnet 5
Balanced
$3.00
$15.00
Claude Haiku 4.5
Economic
$1.00
$5.00
Grok 4.5
Challenger
$2.00
$6.00
ParseBench tables vs cents per page
ParseBench official leaderboard, observed 13 August 2026. Horizontal axis is list cents per page. Vertical axis is the Tables metric (GTRM), not an overall score.
What a harness can change
Illustrative condition ladder on a frozen financial-research set. A minimal retrieve-and-format harness can move results a little; a citation budget and scarce-source overlay can move them more, at higher full cost.
A small harness and a heavy harness are different systems
The failure mode that matters for AlphaSense is treating a public-filing QA score as if it measured expert-call coverage. A scarce-source overlay that is never turned off cannot be credited as an increment.
If only the product surface is observable after an authorized run, report outcome and observed price. Do not credit orchestration, fine-tuning, or a smaller-model advantage without the paired comparison. Native PDF paths in frontier models are themselves a provider document harness, not a raw-model cell.
Operating economics
Seat price is not operating economics
List token prices from 8 August 2026 bound the model bill: GPT-5.6 Luna at $0.20 / $1.20 per million tokens, Terra at $2 / $12, Sol at $5 / $30, Claude Sonnet 5 at $3 / $15, Fable 5 at $10 / $50. Cache, tools, long-context, and service tiers are not in those cells.
Public materials describe human-led and AI-led expert calls, recruiting, compliance, transcript review, and concierge support. Utilization and unit economics are not observed. Full cost per accepted research output still needs seat, calls, failed attempts, retrieval, compliance, and reviewer minutes in one denominator.
Defensibility
Public filings are exposed. Transcripts are the scarce-input hypothesis.
Public filings and earnings are reconstructable with a generic retriever. That layer is exposed to direct substitution and to AI-native filing QA. Permissioned broker research and expert-call inventory can be scarce. That is the PitchBook-analog question for this product, on a different corpus. Scarcity is not defensibility until three other facts exist: lawful reuse, freshness, and a measured quality increment on frozen cases when that inventory is added and nothing else changes.
The 2024 Tegus combination is a vendor and press declaration. It is not an observation of exclusivity windows, export rights, or retention. Switching costs from workflow embed are likewise unobserved. A target-context overlay that mixes retrieval, UI, and scarce sources is a bundled difference. Only an isolated comparison may be called a component effect. Bloomberg and Capital IQ remain bundling exposures on the research-terminal seat.
Defensibility
Public filings are exposed to a generic retriever. Permissioned broker research and expert transcripts are the scarce-input hypothesis. Scarcity is not a verified flywheel.
Dimension
Rating
Judgment
Missing evidence
Data rights
Open
Rights over expert-call and broker corpora, including exclusivity windows and model-training reuse, are not in this report.
License terms, exclusivity windows, and reuse permissions.
Exclusivity
Conditional
Public filings and earnings are nonexclusive. Permissioned expert transcripts and broker research can be scarce. The 2024 Tegus combination is a vendor and press declaration, not an exclusivity observation.
Overlap versus public transcripts and competing expert-network products.
Rebuild difficulty
Conditional
A generic retriever over public filings can reconstruct the public layer. Reconstructing a permissioned expert-call corpus needs rights and collection. That is the PitchBook-analog question for this product, on a different corpus.
Rebuild scenario for the scarce layer only, isolated from UI and retrieval.
Freshness
Open
Transcript freshness and search-index lag are unmeasured. Query logs as a learning loop need lawful reuse.
Freshness audit and a query-to-quality chain.
Switching cost
Conditional
Research workflows can embed saved searches and notebooks. Embed is unobserved as switching cost.
Export friction and retention observations.
System of record
Exposed
AlphaSense is a research overlay. The investment recommendation still lives in the firm's memos and models.
Whether any write-back is treated as record authority.
Integrations
Conditional
Connectors into research workflows can add gravity. Public-filing retrieval is still portable to a generic stack.
Integration inventory and a public-only paired comparison.
Network effects
Open
Cross-customer learning from queries would be a network effect. It is not shown, and it would require reuse rights.
Reuse terms and a measured quality increment from history.
Trust and compliance
Open
AlphaSense publishes security certifications, content-entitlement controls, a compliance process, and a vendor-measured citation-accuracy claim. Those declarations support control design, not independent output faithfulness.
Faithfulness audit and compliance review of the scarce corpus.
Distribution
Conditional
IB, PE, and consulting distribution is a channel. It does not block Bloomberg, Capital IQ, or a generic filing retriever on the public layer.
Channel economics versus named research terminals.
Services dependence
Open
AlphaSense discloses human experts in training and evaluation and a compliance-reviewed expert-call workflow. Human involvement exists, but correction and escalation on accepted research output remain unmeasured.
Correction and escalation minutes per accepted output.
Rate of improvement
Open
No measured quality curve from added transcripts is in this report.
Isolated transcript on/off comparison on frozen cases.
How AI can compete
Replacement, a new AI product, a platform bundle, cheaper labor, and a customer building it themselves are different risks. They are not averaged into a score.
Mechanism
Assessment
Direct replacement
On public filings, a generic retriever plus a current model is the floor. On expert transcripts, substitution requires the scarce corpus or a rights-cleared alternative.
New AI product
Frontier filing QA is improving on public packs. That pressures the public layer. It does not automatically reconstruct permissioned expert calls.
Platform bundle
Bloomberg, Capital IQ, and document-intelligence vendors can attach research search to existing seats.
Labor compression
AI can cut analyst minutes on earnings and filings. Seat intensity can fall even if the scarce corpus remains. That is an economic sensitivity, not a measured NPV.
Customer builds it
A research desk can insource the public layer: a warehouse over filings and earnings with a current model. Insourcing the scarce layer requires transcript and broker rights the desk does not hold. The split is the same as the substitution row.
Data flywheel
A flywheel needs every step below. On this evidence it is unproven. Having data is not the same as being allowed to reuse it, and reuse is not the same as a measured quality gain.
Link
Status
Note
Unique data
Vendor-stated
Expert-call and broker inventory is declared scarce. Public filings are not.
Permission to reuse
Open
Training and product-improvement reuse of transcripts is not on the record.
Reliable feedback
Open
No audited research-answer labels.
Proven quality gain
Open
No isolated transcript on/off comparison.
Better outcomes
Open
No independent research test.
More usage or data
Open
Retention from scarce sources is unobserved.
What public sources resolve
Public filings and vendor documentation can settle some structural questions. They cannot replace a test on the live product.
Question
Conclusion
Still open
Does advantage come from one proprietary model or fine-tune?
The public architecture is explicitly multi-model and includes retrieval, ranking, embeddings, domain models, and human-generated evaluation data. A single-model explanation is not supported by the public record.Generative AI Security
A same-task paired comparison that changes only model identity, plus exact returned-model and configuration records.
Is customer content reused to train shared models?
Under the published default, customer content is not used to train AlphaSense or third-party language models. Aggregate query handling and rights over licensed transcript content remain separate questions.Generative AI Security; Third Party Terms
Executed customer and content-provider agreements for transcript reuse, exclusivity windows, and any negotiated exceptions.
Is AlphaSense a pure software product without material human services?
Service attach, expert payouts, compliance staffing, utilization, gross margin, and correction minutes per accepted research output.
Required comparisons
A decision-grade test would run these side by side. None has been run here.
Comparison
Note
Current workflow
Requires imported analyst process observations. Not inferred from a product tour.
Generic model
A generic model with public filings is the floor, not the product.
Named competitor
Declared retrieval over public filings plus a frozen budget. Not yet run here.
This product
Needs authorized AlphaSense access, logging, and an exportable case set.
Isolated changes
Expert-transcript access versus public-only access must be isolated before attribution.
Open claims
Each claim stands on its own. They are not averaged into a moat score. Outcome and price wait on a live-product test. Orchestration, fine-tuning, and model-size advantages stay unproven until isolated. Public filings are not a scarce asset. Transcripts are not a flywheel without reuse rights, labels, and a measured quality gain.
Claim
Status
What is still missing
Expert-call corpus is exclusive and not replicable.
Open
License terms, exclusivity windows, and a rebuild scenario with required permissions.
Public filings are reconstructable with a generic retriever.
Judgment
That judgment is bounded to the public layer. Transcripts and broker research remain a separate exclusivity question.
Data to labels to measured improvement to retention flywheel exists.
Open
Lawful reuse evidence, label audit, and a measured quality lift on frozen cases.
Switching costs from workflow embed produce pricing power.
Open
Retention, export friction, and services-dependence observations.
Bloomberg or Capital IQ bundling risk is low.
Open
Channel economics and a research-terminal paired comparison on the public layer.
Public architecture is multi-model and retrieval-led; a single proprietary-model explanation is unsupported.
Open
A same-task comparison that changes only model identity is still required for causal attribution.
What this report does not conclude
Replacement of the current analyst process
Attribution of advantage to a proprietary model, fine-tune, or orchestrator
Exclusivity or lawful reuse of expert-call and broker corpora
A proven data flywheel; public filings are not a scarce asset
Full cost per accepted output, reviewer load, or seat-compression value
A moat, time-to-parity, or revenue-at-risk figure
What would settle these questions
Authorized AlphaSense access with exportable traces and a frozen price card.
A sealed, independently adjudicated research case set with critical-error definitions.
Imported observations of the current analyst process; not a product tour.
A declared generic retriever over public filings, same questions, same budget.
An isolated expert-transcript on/off comparison, with rights recorded.
Overlap versus public transcripts and competing expert-network products.
Reviewer-minute and all-attempt cost measurement, including failures.
Published third-party terms impose provider-specific restrictions on downstream use, including restrictions on using some licensed data with or to train AI tools.
These are customer-facing restrictions, not the complete upstream license chain. They establish that content rights are contract-bound but do not establish AlphaSense exclusivity or its own model-improvement rights.
AlphaSense describes an AI gateway that routes work to Anthropic, OpenAI, and Google models, plus in-house or partner models for retrieval, ranking, embeddings, and financial language. It states that customer content is not used to train AlphaSense or third-party language models, and that human experts produce training and evaluation data.
Vendor-authored policy and architecture description. The published citation-accuracy claim is not an independent held-out audit, and the page does not establish transcript-license exclusivity.
AlphaSense says its expert-research workflow combines an acquired transcript library with investor-led and AI-led interviews, transcription, and compliance review.
Vendor announcement. Corpus size, uniqueness, freshness, and outcome claims require independent overlap and quality audits.
AlphaSense ConnectorsSydney Johnson, AlphaSense · Official product documentation · 22 July 2026
AlphaSense describes connectors across storage, communications, VDR, CRM, and research systems. Some sources, including Salesforce, are queried live rather than copied into a proprietary corpus.
Vendor documentation does not disclose connector attach rates, active usage, renewal lift, latency, errors, or connector-specific cost.
AlphaSense offers human-led and AI-led expert calls, custom recruiting, compliance checks, transcript review, concierge support, and pay-as-you-go pricing based on the expert rate plus a transcription fee.
The service description establishes a material human and compliance layer, not its utilization, gross margin, transcript ownership, or incremental effect on research outcomes.
