1. Introduction
Automated systems for discovering market, technology, and competitive intelligence have moved in a short period from research prototypes to routine production deployment. Large language models now permit the ingestion of unstructured external text at volumes that were previously impractical to process manually, and agent-based orchestration permits repeated, scheduled interrogation of external sources without continuous human supervision. Prior work in this agenda has concentrated on the artefacts that such systems produce downstream of discovery: the conversion of an identified opportunity into a build specification, the density and character of quality-assurance activity over those builds, the depth of verification applied before release, and the iteration dynamics observed when a build fails to reach an acceptable state. Those stages are comparatively well instrumented, because they generate structured records as a by-product of engineering work.
The upstream stage — the conversion of external intelligence into company opportunities — has received markedly less empirical attention at production scale. The literature contains substantial work on the *construction* of automated discovery machinery. Černý (2026) describes a multi-agent approach to technology intelligence that moves from patents to market signals using large language models. Lupenko and Stoliar (2026) describe automated semantic ontology construction for foresight studies, exploiting language models to discover semantic structures and emerging signals within text streams. Mavaie and Zarayeneh (2025) describe automated indicator mining and forecast-ready feature engineering for colocation data center supply chains. Rajagopal and Natraj (2026) and Zhao and Sha (2026) describe automated discovery of alpha factors for trading under explicit evaluation budgets. What these contributions share is a primary emphasis on the discovery mechanism and on the quality of the candidates produced. What they do not provide, individually or collectively, is a production-scale description of the *yield* of such a mechanism: how many qualifying signals are captured, how many distinct opportunities emerge from them, and what the ratio between those two quantities looks like when the system is operating continuously rather than within an experimental evaluation harness.
The specific gap addressed here is therefore narrow and quantitative. It is not whether automated discovery is feasible, nor whether the candidates it produces are useful in principle. It is whether, in a live deployment, the volume of qualifying intelligence flowing into the discovery stage exceeds the volume of company opportunities flowing out, and if so by how much. A ratio near unity would imply that the discovery stage is essentially one-to-one, adding selection but little compression. A ratio substantially above unity would imply that the discovery stage absorbs a surplus of qualifying signals that do not result in generated opportunities, and that the stage's principal function is selection under surplus rather than expansion.
The hypothesis under test was that qualifying intelligence signals outnumber generated company opportunities by more than 50% in a production autonomous market-discovery pipeline.
This paper makes four contributions. First, it reports a production-scale signal-to-opportunity ratio and the corresponding signal surplus for one continuously operating autonomous discovery pipeline, computed from the aggregate tallies recorded by that pipeline. Second, it characterises the composition of the qualifying signal corpus by category and by score, using a record-level extract of 500 scored signals. Third, it reports downstream build and QA telemetry from the same observed system so that the upstream yield figure can be read against the organisation's overall production activity, while explicitly declining to attribute causation between the two. Fourth, it documents an observational protocol and its limitations in sufficient detail that the measurement could be repeated on another pipeline, and states plainly which quantities in the record are mutually inconsistent and therefore not aggregated in this analysis.
No prior publication is claimed to have reported this quantity. The claim is not that a research gap exists in an absolute sense — only that the literature surveyed for this study did not contain production-scale empirical data on signal-to-opportunity yield, and that limited empirical production data of this kind appears to exist.
2. Related Work
**Market opportunity discovery.** Three works in the surveyed literature address market opportunity discovery directly. A process-supported case-based reasoning system for market opportunity discovery (2008) and a sequential-data-based case-based reasoning technique for the same problem (2015) formulate opportunity discovery as a retrieval-and-reasoning task over stored cases; both are characterised here at title level, as their abstracts were not available to this review. A hypergraph model for market opportunity discovery (2011) likewise proposes a representational formalism for the discovery problem, again characterised at title level. These three works are relevant because they establish opportunity discovery as a distinct computational problem with its own literature; they are not directly comparable to the present study, which measures yield in a deployed system rather than proposing or evaluating a discovery formalism.
**Intelligence generation with language models.** Černý (2025) addresses the generation of key intelligence topics and questions to enhance the work of competitive intelligence professionals, framing competitive intelligence as an instrument for monitoring an organisation's external environment and providing management with insight into markets, competitors, and trends. This is the closest available treatment of the *upstream* intelligence problem, and it establishes that producing the right signals is itself a recognised professional and technical challenge. It does not, however, report a conversion ratio between generated intelligence and downstream business artefacts.
Černý (2026) extends this line to a multi-agent approach to technology intelligence, describing the movement from patents to market signals with large language models and situating the work within knowledge management and competitive intelligence traditions concerned with giving decision-makers awareness of technological trends, opportunities, and threats. Of the surveyed literature, this paper is the closest structural analogue to the observed system's upstream stage: it too concerns an automated, agent-mediated conversion of external material into market-facing signals. The present study differs in unit of analysis: Černý reports an approach, whereas this report measures the numerical relationship between signal inflow and opportunity outflow in one production deployment.
**Signal discovery in foresight and planning.** Lupenko and Stoliar (2026) present automated semantic ontology construction for foresight studies, arguing that recent language-model advances enable automated discovery of semantic structures and emerging signals within text streams and thereby offer an opportunity to redesign foresight workflows. The relevance here is conceptual: foresight practice is explicitly oriented toward identifying weak or emerging signals ahead of general recognition, which implies an expected surplus of signals over subsequently actionable items. The present study provides an empirical measurement consistent with that expectation, though the observed system is not a foresight system and the comparison should not be pushed further than that.
Mavaie and Zarayeneh (2025) address automated indicator mining and forecast-ready feature engineering via language models in the context of colocation data center supply chain planning, where the synthesis of diverse macroeconomic and industry-specific signals is described as critical for anticipating market dynamics. This work is relevant because it treats signal synthesis as a production planning input rather than an end in itself, which is structurally similar to treating signals as inputs to opportunity generation.
**Candidate generation under evaluation budgets.** Zhao and Sha (2026) present GoAnt, a quality-diversity multi-agent search method for alpha factor discovery in market microstructure data, framing automated alpha factor discovery as a search over symbolic trading signals under a fixed evaluation budget. Rajagopal and Natraj (2026) present an intelligent alpha factor research system that combines machine learning, deep learning, and explainable AI to discover and evaluate profitable trading strategies. Both are relevant to the present study because they make explicit that automated signal discovery is fundamentally a search process that generates more candidates than can be or are retained. Neither reports a production signal-to-opportunity ratio in a business-discovery setting; their evaluation contexts are financial and their candidate pools are defined by a research harness rather than by continuous operation.
**Automated detection of rare or latent items.** Desai and Tiyyala (2025) describe Waldo, an approach to automated discovery of adverse events from unstructured self-reports, noting that adverse event detection is labour-intensive and costly precisely because the task is to find rare events, and that automated solutions are sought to improve efficiency, reduce cost, and capture unnoticed safety signals. This is relevant in a specific and limited way: it articulates the asymmetry between the volume of material that must be screened and the volume of items that ultimately matter. The present study quantifies an analogous asymmetry between qualifying signals and generated opportunities, but in a commercial discovery context and without any claim about the correctness of either category.
**Network-based discovery of commercial relationships.** Malhotra and Ringel (2026) propose a data-driven framework for marketing brand alliance discovery using information networks, noting that brand alliances are widespread in practice but that no existing framework adequately supports their systematic identification. This work is relevant as an example of automated discovery applied to a commercial artefact whose identification precedes any decision to act — structurally the same position that company opportunities occupy in the observed pipeline.
**Multi-signal market-structure analysis.** Fumanal-Andrés and Heradio (2026) present an automated framework for multi-signal market-structure visualisation combining change-of-character detection, operating-range boundaries, and liquidity-void identification for equity-index and forex analysis. The relevance is methodological rather than topical: it demonstrates that aggregating heterogeneous signal families into a single analytic surface is an established practice.
**Summary of the gap.** Across this body of work, automated discovery is consistently framed as a search or generation problem evaluated by the quality of what it surfaces. Production-scale yield figures — the ratio of qualifying inputs to generated commercial outputs — are not reported in the surveyed literature. This study addresses that specific omission and nothing broader.
3. Study Design and Data Collection
3.1 Design
The study is a retrospective, observational analysis of anonymised production records generated by an autonomous software development and market-discovery system operated by a single organization. No intervention was applied, no variable was manipulated, and no control or comparison condition exists. The analysis is descriptive and inferential only in the narrow sense that a pre-specified hypothesis was evaluated against an observed ratio.
The researchers had access to input and output records only. No internal architecture, implementation detail, or component-level mechanism was observed, inspected, or is described here. The system is referred to throughout as "the observed system" and the operator as "the study organization."
3.2 Definitions and units of measurement
**Intelligence signal.** An externally sourced item captured during the discovery stage and admitted to the qualifying signal corpus. Each signal is an input record with provenance to an external source. Signals are counted as individual records.
**Company opportunity.** A company-level record generated by the discovery stage and persisted as an opportunity. Opportunities are counted as individual records. An opportunity may be generated from one or more signals; the mapping cardinality was not available in the extract.
**Signal score.** A numeric score assigned to a signal at or near capture. The extract reports a mean of 7.6 across scored records. The scale, bounds, and scoring procedure were not documented in the data made available for analysis, and the score is therefore treated as an ordinal indicator only.
**Signal category.** A nominal classification assigned to each scored signal. The extract reports three categories.
**Build attempt.** A single execution of the downstream build stage, counted as one attempt regardless of outcome.
**Build.** A persistent build record in the current system population, distinct from an attempt.
**QA check.** A single automated verification assertion executed against a build, with a binary pass or fail outcome.
**Fix attempt.** A corrective execution applied to a build following a failed QA state.
3.3 Data sources and collection
Records were extracted from the study organization's production data store and de-identified prior to analysis. Three record sets were used: (i) an aggregate discovery tally reporting total intelligence signals collected and total company opportunities generated; (ii) a record-level signal extract of 500 scored signals carrying category and score attributes; and (iii) a downstream production extract covering 500 build attempts, 46 builds in the current population, QA outcomes, fix attempts, and the deployed capability inventory.
Collection was opportunistic with respect to the study organization's operating schedule. The analysis window is not delimited in the extract; records are treated as a single production period rather than as a time series, and no temporal trend analysis was performed.
3.4 Reconciliation of reported counts
Two different signal counts appear in the available record. The aggregate discovery tally reports 1,287 intelligence signals and 832 company opportunities. The record-level extract reports 500 signals carrying scoring metadata, distributed across three categories that sum exactly to 500 (255 + 143 + 102 = 500). This analysis treats the 1,287 figure as the aggregate corpus count and the 500 figure as the scored subset, and computes the primary yield ratio from the aggregate tally because that tally is the only source reporting both signals and opportunities on a common basis. The relationship between the two counts cannot be verified from the metadata available, and this is recorded as a limitation rather than resolved by assumption.
A similar unreconciled relationship exists between the 500 build attempts analysed and the 46 builds reported in the current system population. These are treated as distinct constructs — attempts versus persisted builds — and are not summed, divided, or otherwise combined in this report.
3.5 Hypothesis and evaluation criterion
The pre-specified hypothesis was that qualifying intelligence signals outnumber generated company opportunities by more than 50%. The evaluation criterion was a signal-to-opportunity ratio strictly greater than 1.50, computed as the count of intelligence signals divided by the count of company opportunities. Supporting the hypothesis required the observed ratio to exceed 1.50; failing to exceed it required reporting non-support.
4. Results
4.1 Signal-to-opportunity yield
The aggregate discovery tally reported 1,287 intelligence signals and 832 company opportunities. The signal-to-opportunity ratio was approximately 1.55. Expressed as a surplus over opportunity volume, qualifying signals exceeded generated opportunities by approximately 54.7%.
**Table 1. Aggregate upstream yield.**
| Quantity | Value | |---|---| | Intelligence signals (aggregate tally) | 1,287 | | Company opportunities | 832 | | Signal-to-opportunity ratio | ≈ 1.55 | | Signal surplus over opportunity volume | ≈ 54.7% | | Hypothesis criterion (ratio > 1.50) | Met |
The hypothesis was therefore supported: in the observed pipeline, qualifying intelligence signals exceeded generated company opportunities by more than 50%.
4.2 Composition of the scored signal corpus
The record-level signal extract contained 500 scored signals. The mean signal score was 7.6. Scores were not accompanied by scale documentation.
**Table 2. Signal category distribution (500 scored records).**
| Category | Count | |---|---| | AI Tool | 255 | | Business Opportunity | 143 | | Market Signal | 102 | | **Total** | **500** |
The dominant category was *AI Tool*, comprising 255 of the 500 scored records — more than the other two categories combined (245). *Business Opportunity* accounted for 143 records and *Market Signal* for 102. The extract did not report category-level mean scores, so no comparison of score across categories is possible.
The presence of a category labelled *Business Opportunity* within a signal taxonomy is noted: in this pipeline, opportunity-type items are classified as signals at the discovery stage rather than exclusively as outputs of it. This classification is reported as observed and is not interpreted further.
4.3 Downstream build and QA telemetry (contextual)
Downstream telemetry is reported to describe the operating context in which the upstream yield figure was obtained. It is not used in this report to explain or predict the yield ratio.
**Table 3. Build attempt iteration distribution (500 attempts).**
| Quantity | Value | |---|---| | Build attempts analysed | 500 | | Mean iterations per build | 71.3 | | Median iterations | 29 | | Iteration range | 4–300 | | Attempts under 50 iterations | 314 (62%) | | Attempts under 100 iterations | 386 (77%) | | Attempts over 200 iterations | 62 (12%) |
The distribution is strongly right-skewed: the mean of 71.3 lies far above the median of 29, and the upper tail is long, with 12% of attempts exceeding 200 iterations against a maximum of 300.
**Table 4. QA status and remediation (46 builds in current population).**
| Quantity | Value | |---|---| | Total builds in system | 46 | | QA status — failed | 27 | | QA status — escalated | 12 | | QA status — pending | 6 | | Verified builds | 8 (17%) | | Builds requiring fix attempts | 39 (84%) | | Average fix attempts when needed | 2.3 | | Maximum fix attempts on a single build | 7 |
The three recorded status categories account for 45 build records, one fewer than the 46 builds in the population. Separately, 8 builds (17%) were reported as verified. Whether verified builds form a subset of one of the status categories could not be determined from the extract, and the four figures are therefore reported without summation. Eight of 46 (17%) verified against 39 of 46 (84%) requiring fix attempts indicates that remediation was the common path and clean first-pass progress the exception within this population.
**Table 5. QA check outcomes (1,000 checks).**
| Quantity | Value | |---|---| | QA checks run | 1,000 | | Reported pass rate | 94% | | Reported fail rate | 5% | | Most common failing checks | Dashboard screenshot present (10); AI assistant responds (10); AI assistant button found (7); QA reasoning (4); DOM checks, dashboard (2) | | Unique products with failures | 17 |
The reported pass and fail rates sum to 99%, leaving 1% of check outcomes unaccounted for in the extract; no third outcome category was reported. The five most common failing checks listed sum to 33 occurrences, which is fewer than the 50 checks implied by a 5% failure rate on 1,000 checks; the listed items are therefore a partial top-list rather than an exhaustive enumeration of failures. The five listed checks cluster on two themes — presence and responsiveness of a conversational assistant, and dashboard rendering and DOM-level assertions. QA failures were recorded across 17 unique products; the extract did not rank products by failure frequency.
4.4 Deployed capability inventory
The observed system reported 82 production capabilities deployed across 17 categories. This is reported as a descriptor of the system's production scope. It is noted that the count of capability categories (17) coincides with the count of unique products reporting QA failures (17); these are unrelated quantities drawn from different measurement domains, and the coincidence carries no analytic significance.
5. Discussion
The central finding is that the observed pipeline generated company opportunities at a lower rate than it captured qualifying intelligence signals, by a margin of approximately 54.7%. The hypothesis was supported, and the ratio of approximately 1.55 sits comfortably above the 1.50 criterion. The margin should not be over-read: a criterion of 1.50 was met by a ratio of approximately 1.55, which is a narrow clearance relative to the precision with which the underlying counts are reported, and the result is best described as a ratio in the vicinity of one and a half rather than as a robustly established quantity with a known confidence interval.
Interpreted narrowly, the ratio describes a narrowing funnel at the discovery stage. Qualifying signals are admitted at a higher rate than opportunities are generated, which implies that selection occurs between the two. This is consistent with the framing adopted by Zhao and Sha (2026), who treat automated signal discovery as a search under a fixed evaluation budget, and with Desai and Tiyyala (2025), who describe automated detection as a screening task in which the volume of material considered necessarily exceeds the volume of items that ultimately matter. The present result supplies a production-scale magnitude for that asymmetry in a commercial discovery context, where the surveyed literature supplies the framing but not the number.
The result should not be read as a measure of waste. A signal that does not produce an opportunity may be correctly declined, may be informative without being actionable, may be retained for future use, or may be aggregated with others before any opportunity is generated. The extract does not distinguish among these, so the surplus of approximately 54.7% is a statement about relative volumes, not about efficiency or loss. Reporting it as a loss rate would exceed what the data supports.
The composition of the scored corpus is informative in a limited way. *AI Tool* signals were the largest category at 255 of 500 records, exceeding the other two categories combined. *Business Opportunity* signals numbered 143 and *Market Signal* signals 102. Because the extract carries no category-level outcomes, this distribution cannot be connected to the yield ratio; all that can be said is that the qualifying corpus was dominated by tool-oriented signals rather than by market-level or opportunity-level signals. Whether that dominance reflects the external environment, the sourcing configuration, or the classification scheme is not determinable from the records available.
The mean signal score of 7.6 is reported without interpretation. The extract does not document the scale, the scoring procedure, or the relationship between score and conversion. It would be unsound to infer that a score of 7.6 represents "high" or "medium" quality, or that any threshold on the score predicts opportunity generation. The score is included because it is part of the observed record and because future work with a documented scale could make it comparable.
The downstream telemetry described in Section 4.3 is included to situate the upstream measurement, and not to explain it. The observed distribution of build iterations is strongly right-skewed, with a median of 29 against a mean of 71.3 and a long upper tail. Remediation was usual rather than exceptional in the current build population, with 39 of 46 builds requiring fix attempts and 8 verified. The reported QA pass rate of 94% alongside a 5% fail rate leaves an unattributed residual that the analysis does not resolve. None of these observations bears directly on the signal-to-opportunity ratio; they are reported because a reader assessing the upstream figure is entitled to know what kind of production system generated the surrounding records, and because reporting only the favourable upstream measurement while withholding the more equivocal downstream telemetry would constitute selective omission.
Comparison with the intelligence-generation literature is instructive mainly by contrast. Černý (2025) treats the generation of intelligence topics and questions as a professional support problem; Černý (2026) treats the conversion of patents into market signals as a multi-agent technology-intelligence pipeline; Lupenko and Stoliar (2026) treat the discovery of emerging signals in text streams as a foresight workflow to be redesigned. Each is concerned with the mechanism and with the character of what is surfaced. The present study makes no contribution to any of those questions. It contributes a single empirical quantity — the relative volume of qualifying signals and generated opportunities in one live deployment — and the observation that this quantity is greater than one.
The implication for practice, stated conservatively, is that organisations operating such pipelines should expect the discovery stage to admit substantially more qualifying material than it converts into opportunities, and should instrument both counts explicitly so that the ratio can be tracked. The implication for research is that the upstream stage is measurable with records that production systems already generate, and that comparable measurements from other deployments would allow the ratio to be situated. This study provides one observation and no basis for a benchmark.
6. Limitations
**Single organization, single system.** All records derive from one production system operated by one organization. The observed ratio of approximately 1.55 is a property of that deployment. There is no basis for treating it as representative of autonomous discovery pipelines generally, and no claim of representativeness is made.
**No control group.** There is no untreated, alternative, or counterfactual condition. Consequently no causal claim can be made about the effect of the discovery stage, of any configuration choice, or of the system's design on the observed ratio. The study establishes that a ratio of approximately 1.55 was observed; it does not establish why.
**Observational only.** The analysis is retrospective and descriptive. Counts were not designed for research purposes, no power analysis informed the sample, no pre-registration was filed, and the hypothesis — while stated in advance of the analysis — was formed in the context of the same data. The narrow clearance above the 1.50 criterion should be weighted accordingly.
**Unreconciled counts.** Two signal counts appear in the record: 1,287 in the aggregate tally and 500 in the scored extract. These are treated as an aggregate corpus and a scored subset respectively, but the relationship between them is unverified. Similarly, 500 build attempts and 46 builds are treated as distinct constructs. The three QA status categories sum to 45 against a population of 46, and the relationship between the 8 verified builds and the status categories is not determinable. The reported pass rate (94%) and fail rate (5%) leave 1% unattributed. The five listed failing QA checks sum to fewer occurrences than the implied total number of failures. These are reported as found; none has been adjusted, imputed, or smoothed.
**Undocumented measurement scales.** The signal score of 7.6 is reported without a known scale, bounds, or procedure. The category taxonomy is reported without definitions or inter-rater reliability. Counts of "qualifying" signals depend on an admission criterion that was not observed.
**No time-series structure.** The records are treated as a single production period. The analysis window is not delimited, and no trend, seasonality, or drift analysis was performed.
**Black-box observation.** The researchers observed inputs, outputs, and measurement records only. No internal mechanism, architecture, or implementation detail was available or is described. Any statement about *why* signals failed to convert into opportunities would be speculation and is not offered.
**Construct validity.** The construct "company opportunity" is operationally defined by the study organization's own record-keeping. Whether every generated opportunity was commercially viable, and whether any viable opportunity was missed, is outside the scope of what was measured. The ratio describes record volumes, not quality of selection.
**AI-assisted writing.** This manuscript was prepared with AI assistance. AI tools were used in drafting and structuring the text. All quantitative values were constrained to the observed production data extract and verified against it, and all citations were constrained to the literature list assembled for this study. Readers should nonetheless weigh the AI-assisted preparation of this manuscript when assessing it, and should treat the underlying records — not this text — as the primary evidence.
7. Conclusion
This study set out to test whether, in a production autonomous market-discovery pipeline, qualifying intelligence signals outnumber generated company opportunities by more than 50%. Using anonymised production records from a system operated by a single organization, the observed tally of 1,287 intelligence signals against 832 company opportunities produced a signal-to-opportunity ratio of approximately 1.55, corresponding to a signal surplus of approximately 54.7% over opportunity volume. The hypothesis was supported.
Three further descriptive results were reported. The scored signal corpus of 500 records was dominated by the *AI Tool* category (255 records), with *Business Opportunity* (143) and *Market Signal* (102) accounting for the remainder; the mean signal score was 7.6 on an undocumented scale. Downstream, 46 builds in the current population showed remediation as the usual path (39 builds requiring fix attempts, 8 verified), and 1,000 QA checks produced a reported 94% pass rate and 5% fail rate with a residual 1% unaccounted for.
The study does not establish why the ratio takes the value it does, does not compare the observed pipeline against any alternative, and does not support generalisation beyond the single deployment examined. Its contribution is a documented production-scale measurement of an upstream quantity that the surveyed literature frames but does not quantify, together with a transparent account of the inconsistencies present in the records from which that measurement was derived.