Go back to Market Research
COMET provides software and services for incident reporting, investigation, root cause analysis, audits, and supplier assurance. Its Supply Chain module manages supplier information, assessments, documents, and corrective actions. COMET Signals analyzes health, safety, environmental, and quality records; its AI Assistant helps investigators prepare and analyze investigations. Customers include operators in energy, rail, utilities, construction, and other industries with significant operational hazards. (1, 2, 4, 5)
Overall supply chain score: 4.9/10.
| Dimension | Score |
|---|---|
| Supply chain depth | 4.2/10 |
| Decision and optimization substance | 4.0/10 |
| Product and architecture integrity | 6.0/10 |
| Technical transparency | 5.4/10 |
| Vendor seriousness | 5.0/10 |
COMET’s substance lies in organizing investigations and improving how organizations learn from failures. Its supplier module extends that approach to qualification and risk assessment. The claim to optimize the entire supply chain is much broader than the mechanism: assessors configure scores, review suppliers, and assign actions. There is no disclosed calculation selecting purchases, suppliers, or inventory quantities through their expected economic consequences. (2, 3)
Signals’ principal weakness is the jump from patterns in records to claims about causes and effective interventions. A query can be correctly grounded in reports and still support the wrong operational conclusion. Tracking an indicator before and after an action does not, by itself, establish that the action caused the change. (4)
COMET versus Lokad
Suppose a supplier repeatedly delivers defective parts. COMET can organize the reports, connect them to audits, examine failed controls, update the supplier’s risk assessment, and assign corrective work. That addresses an important organizational problem: failures often recur because information remains scattered and nobody follows the action through. (2, 3, 4)
The associated supply-chain choice is different: how much to order from that supplier, whether to qualify another, and how much inventory to hold while corrective work takes effect. These choices depend on uncertain defect rates, receipt dates, demand, replacement costs, capacity, and cash. Lokad’s inventory-optimization approach maintains numerical decision logic to evaluate such commitments economically. Its Supply Chain Scientists develop and revise that logic, with recurring executable decisions as the output.
COMET maintains an investigation and assurance process; Lokad maintains a decision process for the flow of goods. Neither a risk category nor a root-cause label supplies the probabilities and monetary trade-offs needed for the latter. The perspective in Introduction to Supply Chain is useful here: additional information earns its cost when it improves a decision. A thorough investigation can inform that calculation, but the calculation remains to be built.
Supplier assurance and the limits of risk scores
The Supply Chain page describes supplier onboarding, certification and document review, assessments, and action management. Its custom scoring considers expenditure, environmental impact, safety, security, and other company-defined criteria. Screens show risk tiers and a likelihood/severity matrix. Assessors produce an overall supplier score using the customer’s configuration. (2)
These are workflow and classification mechanisms. A supplier ranked as high risk is not automatically the supplier from which less should be purchased. A sole source for a critical spare, a replaceable commodity vendor, and a low-volume subcontractor can have similar risk labels but require very different responses. The economic decision needs the cost and feasibility of alternatives, the quantity exposed, the duration of a disruption, and the likelihood that mitigation changes the outcome.
An ordinal matrix does not supply that calculation. Labels such as likely or catastrophic need an operational definition, a time horizon, and a relationship to exposure. Combining them with expenditure and other scores can hide rather than resolve trade-offs. COMET publishes the assessment workflow, but no calibration of supplier scores against subsequent losses or a decision rule allocating a mitigation budget across suppliers.
This makes the page’s claims of an all-in-one supply-chain solution and optimization of the entire process excessive. The software standardizes supplier assurance. It does not document the stochastic projection of future demand, lead times, defects, and interruptions that would support risk-adjusted purchasing or inventory optimization.
Investigation mechanisms and causal claims
COMET Investigate organizes evidence in a Significant Investigation Data grid, examines failed or absent barriers, analyzes human factors, and codes findings through five Root Maps: Communication, Operating Environment, Management, Equipment, and Training. Power BI dashboards expose patterns, investigator performance, closure times, and the consistency of root-cause selection. (3)
This methodology can make investigations comparable and prevent a superficial conclusion such as operator error from ending the inquiry. Its practical demands are visible in a December 2020 Lloyd’s Register case: a drill-pipe failure had caused approximately 24 days of nonproductive time; the investigation combined drilling records, procedures, independent metallurgy expertise, and a seconded senior investigator. The six-week review produced four COMET factors, nine associated root causes, and ten recommendations. Those are documented investigative outputs, rather than measured reductions in future drilling losses. (17)
Signals 2.0 classifies records and answers queries. It describes source verification, confidence assessments, abstention, versioned prompts, and historical replay after taxonomy changes. This makes analysis more inspectable and easier to revise. (4)
However, consistency with a record is not validation of its causal interpretation. A report can omit a precursor, repeat an investigator’s mistaken explanation, or reflect pressure to select an acceptable category. A verifier may confirm that the explanation appears in the evidence without establishing that it explains the failure. Predicting a future incident also requires evaluating alerts on records available before that incident, including missed events and the workload caused by false alarms. The reviewed material does not publish that evaluation.
Measuring an intervention requires more than a changed indicator
Signals attaches evidence, ownership, dates, costs, and tracked measures to actions. It presents before-and-after indicators as proof of effectiveness: a useful measurement trail, but a weak causal test. (4)
For example, incident counts may fall after a shutdown simply because fewer hours were worked. A new reporting campaign may increase near-miss reports while improving safety. An intervention introduced after an unusual spike may appear successful as the count returns to its normal range. The analyst must distinguish those effects from the action being evaluated. Operational context helps, but displaying it beside a chart does not resolve attribution.
COMET understands this problem in its own domain. Its October 2026 article on underreporting explains how targets, bonuses, and contractor scorecards can reward silence. It also describes Signals monitoring falling reporting rates and declining narrative quality. This is a substantive defense against misleading indicators. It should constrain the product’s much stronger claim that a before-and-after number proves an intervention’s effect. (15)
The same standard applies to the vendor’s results. The McDermott case page reports a consistent global investigation process and use of historical HSE data. A separate testimonial on that page attributes a 29% reduction in first-aid cases in 2022 to COMET at Corning. The latter is a customer-reported outcome, but lacks exposure-adjusted rates, a comparison group, and a breakdown of software versus other changes. It predates the 2026 AI Assistant. (16)
Technology history and roadmap integrity
The company emerged from the 2021 combination of STC Global and INSISO and adopted the COMET trading brand in May 2024. BGF records a growth investment in STC INSISO in 2023. The lineage combines investigation methods, training, software development, and business-process work. (10, 18)
The pre-2023 AI claims deserve particular scrutiny. In September 2022, COMET announced a Network Rail Southern-region pilot using natural-language analysis of tens of thousands of HSE records, with a launch planned after the pilot ended that year. In October, STC INSISO acquired the remaining intellectual property and software of Signals from its collaboration with Lloyd’s Register. The announcement described commercial pilots at Network Rail, McDermott, and Coca-Cola, and claimed the ability to predict latent root causes that could produce future incidents. (8, 9)
Text classification and detection of recurring themes were plausible applications of machine learning. The stronger promise of predicting causes required prospective evidence that these announcements did not provide. Later McDermott material supports continued use of the analytics, but does not quantify predictive performance. The original ambition should therefore be judged by the outcomes delivered, rather than by the continued use of the AI label. (16)
The newer products are more specific. The legacy Companion page identifies an OpenAI-powered chatbot for product guidance, configuration help, and training reinforcement. The April 15, 2026 AI Assistant launch describes drafting within investigations and portfolio analysis through Companion Plus, with human approval of every change. It says this newer capability does not use OpenAI or a public large language model, and excludes legally privileged investigations from portfolio analysis. These are different named capabilities; the supplier needs to make their data paths and version boundaries clear. (6, 7)
The roadmap has continuity. Its weak point remains the escalation from finding patterns to claiming effective prevention, without corresponding outcome validation.
Bill of software materials and architecture
The strongest materials evidence comes from the public service definition, product documentation, and a dated employer vacancy.
| Layer | Documented materials | Scope |
|---|---|---|
| Core hosting | Microsoft Azure in regions appropriate to the customer | 2024 G-Cloud service definition |
| Investigation analytics | Microsoft Power BI | COMET Investigate’s described dashboards |
| Web development | .NET Core/C#, Angular, React, SQL Server/T-SQL, REST APIs | August 2025 developer responsibilities in COMET’s technology branch |
| Signals 2.0 | Azure; tenant resources detailed below | Current product architecture description |
| Legacy Companion | OpenAI-powered conversational support | Product-guidance chatbot, distinct from the 2026 Assistant |
The vacancy covers COMET’s custom systems, platforms, and integrations. It establishes actual development responsibilities, while leaving the precise allocation of those components across products unspecified. The published materials show a largely Microsoft-oriented application business. Model families and analytic compute requirements remain unspecified.
Signals analyzes records from existing HSEQ systems. The 2024 service definition also offers API development to connect existing EHS systems and distinguishes configuration from bespoke changes. That boundary creates an ongoing obligation to maintain field meanings, identities, taxonomies, and integrations as the source systems change. (4, 12)
Security by mechanism
Signals describes personal-data redaction, separate tenant databases, indexes and queues, and processing within the customer’s Azure environment without an outside model. The Assistant launch further states that client data does not contribute to a shared model and privileged investigations are excluded from portfolio analysis. These are meaningful design claims about exposure and containment. (4, 6)
The 2024 service definition provides point-in-time recovery and encryption, while assigning customers responsibility for their own archives. Neither that document nor the current product account specifies recovery copies protected from a compromised administrator, a tested restoration objective, or the precise privileges of analytic agents. Those are practical remaining gaps. ISO alignment and compliance language cannot fill them. (12)
Changes to classification, queries, and agent behavior require permissions, review, reproducible evaluation, and rollback. Revision history helps diagnose a mistake after it happens; restricting an agent’s tools limits the damage it can cause. (4)
Productivity, maintenance, and total cost
COMET’s AI Assistant claims to reduce investigation documentation time by more than 30%. It drafts material for investigators to accept, change, or reject, and preserves human judgment. This is a scoped productivity claim: documentation can be a substantial burden, but saving time on that activity does not establish the same reduction in total investigation work. Interviews, evidence collection, technical analysis, review, and corrective-action follow-up remain. The product page supplies no workload study or measured staffing denominator. (5)
Human review has a clear purpose in contested or consequential investigations. The July 2025 account describes interviewing witnesses, assessing credibility, and evaluating operational context as central tasks. The relevant post-2022 test is how much preparation, coding, searching, and routine follow-up can be removed while preserving those responsibilities. Approval of substantive conclusions does not require manually redoing every intermediate operation. (19)
Correcting a taxonomy can improve many subsequent analyses. The expensive work is testing whether the revised definitions improve decisions, preserve comparability, and reduce false alarms. An administrator can quickly change a field; deciding what the field should mean is the expensive part. (4)
This connects to Holimization: the categories, operational assumptions, and available corrective actions must themselves be open to revision. A system that consistently sorts reports into an inadequate taxonomy can become efficiently wrong. The economic validity of the resulting priorities still requires evaluation.
The published 2024 G-Cloud pricing makes several ownership costs visible. One module in the UK/Europe region cost £41,400 for the first year; the schedule included standard 5% annual increases and additional regional charges. It separately listed £7,500 for API implementation and £6,500 for a two-day practitioner course for up to twelve delegates. These are historical framework prices, not a current customer quote. They show why software fees alone are an incomplete measure of the investment. (13)
The service definition adds configuration support, super-user training, process development, communications, and change management. Recurrent costs also include refreshing supplier records, maintaining risk definitions, reviewing classifications, and following up actions. An AI guide may reduce assistance requests, but it does not eliminate this work. A productivity assessment should count vendor and customer effort together. (7, 12)
Accountability and inspectability
COMET’s public 2024 G-Cloud terms promise performance substantially in accordance with documentation and reasonable skill and care, with correction of nonconformance at the vendor’s expense. The same document assigns customers sole responsibility for results and conclusions, excludes specified economic losses, and generally limits aggregate liability to six months of fees or a lower agreed amount, subject to stated exceptions. This is responsibility for contracted software behavior, with much less responsibility for the business outcome. (14)
Two other provisions affect evaluation. Performance-test results are classified as confidential, which can obstruct public comparison. The exit clause requires a request for the most recent data backup within ten days of termination and allows reasonable return expenses. That is a more restrictive migration arrangement than merely saying customers own their data. (14)
The stronger commercial test would connect each intervention to its expected benefit, actual cost, operational exposure, and subsequent evidence. Calling an indicator movement proof of effectiveness, while leaving results entirely with the customer, sets a much lower standard than the claims of prevention and risk intelligence suggest.
Supply chain score
The overall 4.9/10 is the rounded arithmetic mean of 25 equally weighted sub-scores. The assessment credits COMET’s investigation mechanisms and evaluates the supply-chain decisions its supplier and risk claims imply.
Supply chain depth: 4.2/10
| Criterion | Score |
|---|---|
| Economic framing | 4/10 |
| Decision end-state | 3/10 |
| Conceptual sharpness on supply chain | 4/10 |
| Freedom from obsolete doctrinal centerpieces | 4/10 |
| Robustness against KPI theater | 6/10 |
Average: 4.2/10.
Economic framing. The missing calculation is the value of competing mitigation choices: which expenditure is most likely to prevent a larger loss? (2, 4)
Decision end-state. The outputs are assessments, investigations, recommendations, and owned actions. There is no documented calculation producing economically selected purchase quantities or sourcing allocations. (2, 5)
Conceptual sharpness on supply chain. Supplier assurance and operational investigation have clear mechanisms. Claims to optimize the entire supply chain outrun that scope. (2, 3)
Freedom from obsolete doctrinal centerpieces. The offer avoids conventional safety-stock and S&OP machinery, but configurable risk matrices can still replace an economic problem with subjective scores. (2)
Robustness against KPI theater. COMET’s analysis of suppressed reporting is strong. Its test of intervention effectiveness needs the same skepticism about indicator movement. (4, 15)
Decision and optimization substance: 4.0/10
| Criterion | Score |
|---|---|
| Probabilistic modeling depth | 3/10 |
| Distinctive optimization or ML substance | 5/10 |
| Real-world constraint handling | 3/10 |
| Decision automation and labor productivity | 5/10 |
| Resilience under real operational complexity | 4/10 |
Average: 4.0/10.
Probabilistic modeling depth. Confidence in a classification cannot price future operational exposure. The disclosed decision mechanism lacks that probability-to-loss calculation. (2, 4)
Distinctive optimization or ML substance. The analytic architecture supports inspection and correction. The missing prospective evaluation prevents a stronger judgment about predictive value. (4, 8)
Real-world constraint handling. Investigations preserve technical and organizational context, as the drilling case shows. The software does not disclose optimization under purchasing, capacity, or budget constraints. (2, 17)
Decision automation and labor productivity. Preparation work can be reduced substantially. The claimed documentation saving leaves the complete investigation and follow-up workload unmeasured. (4, 5)
Resilience under real operational complexity. Rare-event prediction and causal interpretation remain vulnerable to incomplete and selectively reported evidence. The treatment of underreporting is a useful partial response. (4, 15)
Product and architecture integrity: 6.0/10
| Criterion | Score |
|---|---|
| Architectural coherence | 7/10 |
| System-boundary clarity | 7/10 |
| Security seriousness | 6/10 |
| Software parsimony and ownership cost | 5/10 |
| Compatibility with programmatic and agent-assisted operations | 5/10 |
Average: 6.0/10.
Architectural coherence. Incident records, structured investigation, supplier assurance, and analytic overlays share a recognizable operational purpose. The Signals acquisition has a traceable technical lineage. (1, 8)
System-boundary clarity. The assurance and investigation responsibilities are clear. Buyers still need an explicit map of data movement across Companion, Companion Plus, and Signals. (4, 6, 7)
Security seriousness. The stated containment mechanisms earn credit. Recovery independence and agent permissions require more specific disclosure. (4, 6, 12)
Software parsimony and ownership cost. Microsoft-oriented development and shared configuration suggest a coherent technical base. The cost of integrations, training, taxonomies, and analytic operation remains material. (11, 12, 13)
Compatibility with programmatic and agent-assisted operations. Integration and repeatable evaluation are supported. Recurring action execution has a less explicit public interface. (4, 12)
Technical transparency: 5.4/10
| Criterion | Score |
|---|---|
| Public technical documentation | 6/10 |
| Inspectability without vendor mediation | 5/10 |
| Portability and lock-in visibility | 5/10 |
| Implementation-method transparency | 6/10 |
| Evidence density behind technical claims | 5/10 |
Average: 5.4/10.
Public technical documentation. The architecture and service definition allow substantive scrutiny. Reproducible model evaluation and detailed integration contracts remain less accessible. (4, 12)
Inspectability without vendor mediation. Product workflows, employer engineering evidence, and public framework documents can be examined. There is no public reproducible evaluation of incident prediction or causal recommendations. (8, 11, 14)
Portability and lock-in visibility. Customer ownership and export are explicit, as are a short exit-request deadline and possible return expenses. Trained taxonomy use and historical classifications add migration work beyond exporting records. (12, 14)
Implementation-method transparency. Configuration, integration, training, and organizational embedding are documented and partly priced. This makes the human component more visible than a claim of instant AI deployment would. (12, 13)
Evidence density behind technical claims. Named deployments establish operational use. They leave the quality of prediction and intervention selection largely unmeasured. (4, 9, 16)
Vendor seriousness: 5.0/10
| Criterion | Score |
|---|---|
| Technical seriousness of public communication | 5/10 |
| Technology roadmap integrity | 6/10 |
| Conceptual sharpness | 6/10 |
| Economic accountability | 3/10 |
| Defensibility in an agentic-software world | 5/10 |
Average: 5.0/10.
Technical seriousness of public communication. Operational examples improve the account. Broad guarantees of complete investigations demand much stronger validation. (4, 5, 17)
Technology roadmap integrity. The record connects investigation methods, 2022 language analytics, and newer drafting and verification. Continuity earns credit; the early predictive promise still lacks published outcome validation. (6, 8, 9)
Conceptual sharpness. The treatment of failed barriers, organizational context, and reporting incentives is substantive. Risk scores and before-and-after indicators are weaker substitutes for economic and causal reasoning. (2, 3, 15)
Economic accountability. Contract terms place results with customers, while public outcome claims lack a complete costed comparison. (4, 14, 16)
Defensibility in an agentic-software world. Operational methodology, controlled evidence, and maintained classification can retain value. Chat-based assistance and document drafting are less durable advantages; sustained value depends on improving intervention quality and reducing total work. (3, 4, 5)
Sources
Undated pages were accessed on November 4, 2026. The G-Cloud documents are the vendor’s published 2024 framework materials; their prices and terms are not represented as universal current customer arrangements.
COMET, Platform overview. Modules, methodology, operating sectors, and product boundaries.
COMET, Supply Chain. Supplier onboarding, configurable risk assessment, documents, actions, and suite-wide claims.
COMET, Investigate. Root Maps, evidence organization, human factors, and integrated Power BI analysis.
COMET, AI Data Analytics. Technical architecture and operational claims.
COMET, AI Assistant. Drafting, portfolio analysis, documentation-time claim, and approval workflow.
[6] April 2026 Assistant launch
COMET, Introducing COMET AI Assistant, April 15, 2026. Product split, approval, model/data-boundary claims, and privileged-case exclusion.
COMET, Companion. OpenAI-powered product assistance, configuration guidance, and training reinforcement.
COMET, Acquisition of the AI product developed with Lloyd’s Register, October 18, 2022. IP ownership, pilot scope, language analysis, and predictive claims.
COMET, Signals pilot at Network Rail, September 23, 2022. Intended records, workflow, other pilots, and launch timing.
[10] Corporate lineage and rebrand
COMET, STC INSISO transitions to COMET, May 23, 2024. Earlier merger and adoption of the trading brand.
COMET, Full Stack Developer, August 11, 2025. Employer-defined responsibilities and scope of the technology branch.
COMET/STC INSISO, G-Cloud 14 service definition, May 2024, ODT. Hosting, recovery, support, configuration, integration, training, and exit arrangements.
COMET/STC INSISO, G-Cloud 14 pricing, May 2024, ODS. Module/region fees, annual increases, API implementation, and courses.
COMET/STC INSISO, G-Cloud 14 terms and conditions, May 2024, ODT. Conformity obligations, customer responsibility, performance-test confidentiality, liability, and export conditions.
[15] Reporting incentives and data quality
COMET, Near-miss underreporting, October 26, 2026. Incentive distortion, reporting quality, and Signals monitoring.
[16] Deployment and outcome claims
COMET, McDermott case study. Public case summary and separately identified Corning and other customer testimonials.
COMET, Lloyd’s Register drilling-performance case, December 18, 2020. Incident scope, investigative work, and delivered recommendations.
BGF, STC INSISO portfolio entry. Records a 2023 growth investment and the software/services business.
COMET, AI and human judgment in high-risk decisions, July 24, 2025. Intended investigative responsibilities and limits of automation.