Challenges

3

• 3E

  • Challenge number 47

    How can we use multi-modal AI to automatically extract technical information from circuit diagrams and OEM product manuals to build the digital electrical architecture of a grid-scale battery?

    When onboarding a new Battery Energy Storage System (BESS) to our analytics platform, a major bottleneck is manually mapping the physical and electrical architecture. This data is trapped in static PDFs, schematic diagrams, and unstructured text manuals. We want to explore computer vision and natural language processing to automate this mapping. We are looking for academic input on the best machine learning approaches to parse complex engineering schematics and correlate them with text into a structured digital format. We can provide sample schematics and product manuals to facilitate the discussion.

A

• Amazon

  • Challenge number 18

    Can we trust LLM as a judge in production, and know when it has quietly stopped being trustworthy?

    Once LLM application is deployed widely, the bottleneck moves from building systems to trusting them, and a common practice is to have one model grade another's output. Two harder problems remain open. First, a judge can look well-calibrated on average severity while getting the rank order of systems wrong so a team can ship the wrong system while every headline metric looks healthy. Second, the judge is itself a moving target: a revised judging prompt or a model version change acts as a brand-new measuring instrument, breaking comparability with all historical result. A further complication is that the expert labels we calibrate against are noisy and contested, which violates the unbiased-ground-truth assumption the newer statistical methods rely on. We would like to discuss principled ways to (a) guarantee rank-order fidelity (b) monitor and version automated judges so drift is detectable; and (c) do both when the human ground truth is itself unreliable.

B

• Barclays

  • Challenge number 19

    How can organisations best design and implement effective AI guardrails and governance mechanisms that promote fair and responsible AI usage, mitigate undesirable model behaviour, and protect against malicious or unintended user interactions, when introducing AI at scale?

    As AI adoption grows across financial services, organisations are increasingly using AI to enhance information discovery, automate processes, and improve customer and colleague experiences. This challenge focuses on the key considerations for delivering trustworthy AI at scale, focussing on systematic evaluation, effective governance and guardrails to promote responsible usage. Guardrails provide the governance, safety, and compliance controls required to mitigate risk, protect sensitive information, and ensure AI systems operate within the policies and values of Barclays.

• Blue Energy

  • Challenge number 23

    Blue Energy wants to use AI to support nuclear power design and construction projects. However, they will be subject to commercial and legal security constraints, and we require very high integrity of AI output. How should we structure our approach to achieve these?

    Blue Energy is a nuclear power plant development company deploying small modular reactor (SMR) fleets that are financeable, scalable, and cost-competitive. We are currently seeking ways to leverage state-of-the-art AI technology to innovate in nuclear project development.

    The underlying pattern across our use cases is the same: matching a structured set of requirements, including technical and legal aspects, against a large body of unstructured documents, with every match cited to its source and reviewed by an engineer. Three specific use cases:

    1. Site selection. We screen candidate sites for SMR deployment. This means extracting hundreds of siting requirements from regulations and guidance documents, then matching them against site evidence such as geotechnical reports, flooding studies, and grid interconnection assessments. In addition to that, Blue Energy has its own site screening characteristics. The output is a screening matrix per site showing which requirements are satisfied, which have gaps, and which lack data, with each conclusion cited to a specific document section.

    2. Construction and commissioning. During construction, defects are recorded as free-text nonconformance reports and punch lists. We want to automatically link each defect to the specification requirement it violates and the schedule activities it blocks, then draft the engineering disposition with citations to the governing code and to precedent dispositions from the same project. Engineers review and sign every disposition; the tool never closes anything on its own.

    3. Supply chain quality. Components arrive with certification documents such as material test reports. We want to extract the reported values (material properties, heat numbers, procedure references) and verify them automatically against the purchase order acceptance criteria, flagging discrepancies with the cited values from both documents.

    The research questions where we need academic expertise:

    • Extraction completeness and nuclear QA compliance. Standard evaluation tells us when the model is wrong, but in our setting the dangerous failures are omission and hallucination: a requirement the model failed to extract, evidence it failed to link, or a statement it fabricated. How do we measure and bound what the model missed or invented?

    • Grounding guarantees. What techniques give us confidence that every generated statement is genuinely supported by its cited source, beyond simple quote matching? How can such qualitative measures be compared within the AI framework?

    • Secure deployment. Our documents include export-controlled and supplier-proprietary material. What architectures let a small company use frontier LLMs on such data while meeting nuclear quality assurance and cyber security requirements?

E

• ENODA Ltd

  • Challenge number 36

    Bottom-up demand and flexibility forecasting for distribution networks. Distribution networks are becoming harder to forecast as demand, embedded generation, electric vehicles, heat pumps, batteries, and flexible loads all interact across different locations and time horizons. Operators increasingly need forecasts not only at grid-supply-point or substation level, but also at more granular points across the distribution system, including smart meters, feeders, substations, SCADA nodes, and controllable assets. How can AI methods produce reliable bottom-up forecasts of demand and flexibility availability across a distribution network, using heterogeneous data sources with different resolutions, quality levels, ownership boundaries, and uncertainty characteristics?

    Short- and medium-term demand forecasts support balancing, scheduling, voltage optimisation, DER integration, outage planning, and flexibility procurement. However, the data available for forecasting is highly heterogeneous: smart meter data may be delayed or incomplete, SCADA measurements may be real-time but spatially sparse, topology models may be imperfect, and weather, behavioural, market, and asset-availability data may exist at different resolutions.

    A practical forecasting approach would need to predict not only expected demand, but also the uncertainty around that demand and the flexibility available from EVs, heat, storage, commercial loads, and demand response. The challenge is to create models that are accurate, uncertainty-aware, geographically meaningful, and operationally useful from feeder level through to wider system-level decision-making.

    • Attachment

• Euryan

  • Challenge number 42

    When a neuro-symbolic system encounters an entity or failure mechanism outside its formal ontology, what should it do, and how can that behaviour be made more auditable? Can a system admit provisional, uncertain knowledge without forfeiting the determinism that makes it certifiable in the first place?

    Euryan's platform pairs a deterministic engineering knowledge graph with a schema-constrained transformer, so that every output is traceable to an asserted fact rather than generated. This gives auditability for regulated design work, but the guarantee holds only for entities the ontology already contains.

H

• HFD Datavita Limited

  • Challenge number 33

    What combination of technical evidence would allow a technical platform operator to prove that a customer's data and models never left its sovereign, secure infrastructure?

    We are developing a platform with Jisc to provide Trusted Research Environments to UK Universities. This ability to provided audit, evidence which is admissible as fact in a legal context is very relevant to the current requirements for Sovereignty and Security guarantees.

    Sovereignty claims such as "your data stays in the UK" and "your AI never leaves our infrastructure" currently rest on contractual promises rather than demonstrated facts. A credible technical proof would likely combine hardware attestation, confidential computing, cryptographically signed supply chains and tamper-evident logs — but any such proof raises questions about forgery resistance, circumvention, and whether it would carry evidential weight with regulators, auditors or courts, including in what form, against what standard, and where liability sits if the proof later fails.

    Why we are asking it

    - a) It is the assurance and evidence engine — a key build item for the TRE component of our platform: A signed supply chain, tenancy isolation and continuously generated evidence pack are the raw inputs to this proof.

    - b) It is a hyperscaler critique with academic authority: Their sovereignty is contractual, ours would bestructural. A DataVita-initiated research agenda on provable sovereignty

    - c) UK law reform is converging: post-Horizon, the direction of travel is from presuming computer evidence is reliable toward requiring operators to demonstrate control of their systems. That requirement is what we want to develop in the form of an evidence engine to automate that data collection.

• Historic Environment Scotland

  • Challenge number 46

    In an (historic) environment where accuracy, evidence and source material are foundational is there a role for AI in our records?

    The National Record of the Historic Environment (NRHE) represents over a century of fieldwork and recording of Scotland's Historic Environment. Linked to a paper and digital archive, the NRHE contains location and descriptive data about more than 350,000 places in Scotland, from find spots of Roman coins, to sunken ships and skate parks. All of these places make up the historic environment we experience in our daily lives.

    The records within the NRHE are varied; some contain only basic locational information and a line or two of text, while others list decades of archaeological investigation and links to significant archive material. All of these records are used by planners, academics and the public to make decisions impacting our futures and build up our story of Scotland's past.

    We've used machine learning to identify monuments prior to fieldwork and to help link records held in different datasets, but all with a high degree of human validation. While tools such as Co-Pilot are being rolled out, organisations such as ours are grappling with the longer-term implications. How do we record use of AI so that users of our records in another century understand what we did, and how do we leverage this tool in a trustworthy way and one which attempts to replicate the value-driven thought processes of professionals who have worked with these records for decades?

    We've had a long-term ambition to be able to create user-friendly site descriptions, to enable anyone to understand what a place record may represent. How can we do this, but by not simply mashing text together? For example, how could systems make a value judgement on 'authority' in order to weigh what to include or where to add emphasis. How do we move beyond mashing up data and towards value-driven synthesis. This could include questions about how to include domain or knowledge led parameters in summarising, with some underlying principles that include the value judgements instinctively made by those who have worked with the record for many years. As an example, the UK national mapping agency, the Ordnance Survey will have recorded accurate locations with sufficient descriptions, but classifications have become refined over time. Some descriptions may indicate less committed fieldworkers in past decades, while interpretations made in the 1960s will not have had access to the Ordnance Survey first edition maps we now all have instant views of.

    As a public sector organisation, we are unlikely to access funding for the creation of significant digital infrastructure, so solutions need to be “home grown” and able to be maintained within existing capacity, yet in the face of increasing generic pressure to “use AI.”

• HSBC Bank PLC

  • Challenge number 14

    Create a balance sheet optimisation engine that drives the reallocation of interest-earning assets to business units and financial instruments that generate the highest return (optimising Net Interest Income levels)

    Banks manage large, complex balance sheets where capital is deployed across multiple business units, products, and geographies. While some portfolios consistently generate stronger net interest income (NII) per unit of interest earning assets, others underperform due to legacy product mix, pricing, or inefficient deployment. The challenge is that balance sheet resources can’t be “moved” freely: growth and run off are constrained by customer behaviour, market capacity, risk appetite, and operational limits. As a result, decision-makers need a practical way to identify where incremental assets should be grown (and where low-return assets should be allowed to roll off) to improve NII, without creating unrealistic month-to-month shifts.

    This challenge requires teams to design a data-driven balance sheet optimisation engine that reallocates (or steers the growth of) interest earning assets towards the most productive business units while respecting realistic constraints. The solution should combine:

    a) a method to measure relative performance/efficiency across decision-making units (e.g. branches, portfolios, regions)

    b) a forecasting component that estimates feasible next-month asset movement (growth/run off “caps”)

    c) an allocation mechanism that produces scenario paths (base vs optimised) and quantifies the impact on NII.

    Key target outcomes are as follows:

    1) improve projected NII versus a business-as-usual baseline

    2) keep reallocations within realistic movement limits

    3) demonstrate stability and transparency of the approach (clear drivers, guardrails, and audit trail)

    4) provide a monitoring view (single run log / performance dashboard) that tracks both computational performance and business outcomes over time

    5) Should be considered in the context of a 5 year forecast over which we have a clear (but sensibly fluid) plan to migrate assets / funding to the most efficient efficient business & financial instruments

    Further detail and materials to follow to support discussions

I

• IBM

  • Challenge number 20

    As agentic systems gain access to a wider range of enterprise data pertaining to individuals, how can this data be managed with the concept of 'memory' whilst honouring the usage profile for the data e.g. when can it be used and when must it be forgotten?

    In my work in the area of cognitive assistants and now agentic systems one area that I've found a potential challenge is the space of memory management / recall. In particular how this relates to more "structured" data present in underlying business processes and organisational data stores. There are approaches available today which look to address this to some level (mem0 for example) but I feel they miss the subtle aspects of data managent particularly around temporal and viabliity scopes as well as more complex structures that are needed for down stream actions but that the user sees in a different way e.g. bank account detials which the user tends to deal with in terms of the name of the account. I have explore this a bit as covered in the Medium post https://medium.com/musings-of-a-software-architect/adding-memory-when-calling-agents-from-assistants-27d92ee4098a

• ITZA Media Ltd.

  • Challenge number 8

    How can we create human-machine contexts, particularly when using AI, that are most conducive to using AI to enhance learning and cognition rather than create incentives for cognitive offloading/laziness?

    We are a digital learning company focused on independent learning for young people. For us, success looks like young people in secondary education who feel excited about learning, who are more confident because they know that they are good at learning, and who enjoy the independence and sense of progress that they feel when they can learn the things they wish. We are exploring uses of AI to try to help foster critical thinking, curiosity, and a self of agency/self-determination. However, through our testing and product insights, it is clear that one of the most critical success factors to AI becoming a force for learning (rather than a force for cognitive decline) is to create the right contexts for its use. In the current modalities of education, where students are evaluated by teachers/supervisors on discrete tasks, there is too much incentive to simply use AI to shortcut the process, even though it could be used for much more. For example, instead of having AI write an essay or report, a young learner can have AI help them break down the larger task into discrete blocks that can be woven together, with AI providing guidance and critical evaluation. Another related example that comes to mind is preparing for exams or interviews using guides where there is an answer key or guide that can be used to *check* work after it has been done, but cannot actually replace the initial effort.

K

• Kongsberg Precision Cutting Systems

  • Challenge number 22

    Can machines learn to recognise manufacturing excellence in fibre based packaging?

    Humans are remarkably capable at assessing quality in fibre-based packaging. Through a combination of visual inspection and tactile feedback, experienced operators can identify material inconsistencies, printing defects, cutting & creasing imperfections, and issues in the geometric or structural design of the final product.

    At Kongsberg Precision Cutting Systems, we are committed to reducing waste and improving manufacturing efficiency across the packaging industry. As production volumes increase and tolerances become tighter, relying solely on human inspection becomes increasingly challenging. Even highly skilled operators may struggle to detect subtle defects or inconsistencies across every product in a production environment.

    This raises an important and challenging question:

    Can we develop intelligent systems capable of recognising manufacturing excellence by detecting defects, errors, and quality deviations through the interpretation of multi modal data?

    Such a system would combine information from multiple sources—including machine vision, process parameters, material characteristics, geometric measurements, and operational data—to evaluate quality with a level of consistency, scale, and responsiveness beyond that achievable through human inspection alone.

    • Attachment 1
    • Attachment 2

N

• NCC Group

  • Challenge number 57

    How can we detect and govern Unmanaged AI Agent Identities in Enterprise Environments?

    Background

    The rapid adoption of large language model (LLM)-based AI agents within enterprise environments has outpaced the development of standardized protocols for their integration and oversight. While the Model Context Protocol (MCP) has emerged as one open standard for connecting AI models to enterprise tools and data sources, its adoption is neither universal nor mandated. Organisations continue to deploy AI agents through heterogeneous mechanisms, direct API integrations, proprietary orchestration frameworks (e.g., LangChain, AutoGPT, CrewAI), vendor-embedded copilots, and ad hoc scripting. Many of these fall entirely outside of existing Privileged Access Management (PAM), Identity Governance and Administration (IGA), and Cloud Access Security Broker (CASB) frameworks.

    This creates a growing class of unmanaged non-human identities: AI agents that acquire credentials, access sensitive systems, and take autonomous or semi-autonomous action, without being provisioned, monitored, or governed through established identity security controls. Unlike traditional “shadow IT,” these agents can independently generate further access requests, chain permissions across systems, and operate at machine speed, characteristics that existing detection models for human or service-account behaviour are not designed to capture.

    Research Problem

    There is currently no robust, protocol-agnostic methodology for enterprises to reliably detect, classify, and govern AI agent activity within their environments, particularly where such agents do not conform to standardized frameworks like MCP. Existing detection approaches (network egress monitoring, credential auditing, SaaS discovery) offer partial visibility but are fragmented, reactive, and easily evaded by agents using direct API calls or novel orchestration patterns.

    Research Questions

    1. What behavioural, network, and identity-layer signatures reliably distinguish autonomous AI agent activity from human and traditional service-account activity in enterprise systems?

    2. Can a protocol-agnostic detection framework be developed that identifies AI agent presence regardless of the underlying integration method (MCP, direct API, custom orchestration)?

    3. How should identity governance models (PAM/IGA) be extended or redesigned to natively account for non-human, autonomous, and self-escalating identities?

    4. What organisational and technical controls minimize the risk of unmanaged AI agents while preserving the operational benefits driving their adoption?

    Anticipated Contribution

    This research would produce a novel detection framework and identity governance model for AI agent oversight, contributing to both the academic literature on non-human identity security and practical enterprise capability for managing an emerging and largely ungoverned attack surface.

Q

• Qinara

  • Challenge number 16

    How can recurring, reusable sub-circuit patterns ("gadgets") be automatically discovered from large collections of quantum circuits, so that a circuit optimiser can learn its own library of high-value building blocks rather than relying on hand-crafted rules?

    Almost all quantum circuit optimisers today work by peephole rewriting: they scan for small fixed patterns and swap them for cheaper equivalents drawn from a hand curated list of identities. Production compilers do this, and so do the rewrite engines behind the ZX calculus, the phase polynomial methods, and Clifford normalisation.

    The catch is that the rule library only ever contains what its authors thought to put in it. That is a limitation of substance, not just of engineering effort. Hand written rules transfer badly between native gate sets and hardware platforms, they capture only the equivalences someone anticipated in advance, and they tend to miss the motifs that actually dominate circuits from specific ansatz families, such as the excitation preserving blocks in unitary coupled cluster chemistry, the repeated structure of Trotterised lattice Hamiltonians, or the mixer and phase layers in QAOA.

    What we want to discuss is whether the rewrite vocabulary itself can be discovered from data rather than written by hand.

    We treat gadget discovery as unsupervised structure learning over a corpus of circuits. Each circuit becomes a directed acyclic graph whose vertices are gate instances and whose edges are the qubit and classical wires that fix the partial order of operations. Three things make this much harder than ordinary frequent subgraph mining, and these are where we think expertise would be most valuable.

    1. The equivalence we care about is semantic, not syntactic.

    Two fragments can look completely different in their gate sequence, their basis, or their local Clifford dressing and still implement the same unitary, whether exactly, up to a global phase, or up to single qubit rotations before and after (local unitary equivalence). So the mining has to happen over equivalence classes. We need to quotient out global phase, local unitary conjugation, and the commutation relations that let the graph's linear extensions be reordered freely, and only then ask what recurs.

    This runs straight into canonical form questions: Weyl chamber coordinates and local invariants for two qubit blocks, phase polynomial normal forms for the CNOT and phase fragment, and the practical problem of which of these invariants gives a canonical key that is cheap to compute and collision resistant enough to index a large corpus.

    2. There is no single notion of how interesting a candidate is, and what there is depends on the target hardware.

    Support count alone tells you almost nothing. The real value of a motif combines how often it appears with how much you actually save when you re express it near its optimum, and for two qubit blocks that saving is well defined, because the minimum entangling gate count is fixed by the non local content of the unitary and the gain is the gap between what the circuit uses and that lower bound. It also depends on the motif's depth and its routing cost under a given connectivity graph, and on whether it composes with or absorbs other motifs you have already found.

    We are most interested in formulations that score the gadget library as a whole, in the spirit of minimum description length or dictionary learning over the corpus, rather than judging each candidate on its own, and in the combinatorial trade offs that come with that.

    3. Scale, made worse by the partial order.

    The number of connected subgraphs explodes with size, the qubit line dependencies mean the usual graph mining canonicalisation does not apply directly, and testing equivalence class membership is itself expensive.

    One direction we want to push on is whether learned embeddings of circuit fragments can place functionally equivalent pieces near each other in a vector space, so that approximate nearest neighbour search recovers motifs that are the same operation in different clothing, which exact matching would never line up. We also want to be honest about where that breaks down, and about how you certify that a fragment retrieved by proximity really is equivalent rather than merely close.

    Concretely, the things we would hope to come away with:

    1. Which invariants (local unitary invariants, phase polynomial normal forms, spectral or graph theoretic signatures of the circuit graph) give canonical keys that are efficient and collision resistant, and where the tractable boundary sits.

    2. How frequent subgraph mining and graph grammar induction can be adapted to ordered, qubit indexed graphs with commutation symmetry, and in particular whether a grammar induction view can recover a hierarchical, compositional set of gadgets instead of a flat list.

    3. How to write the discovery objective as a multi objective or minimum description length problem that couples frequency, achievable simplification, and connectivity aware cost, and what the resulting optimisation landscape looks like.

    4. How to learn representations of circuit fragments that respect semantic equivalence, and what a rigorous bar would be for promoting an embedding retrieved candidate to a verified gadget.

    We can supply sizeable corpora of benchmark circuits across several application domains and native gate sets to keep the discussion grounded in real data.

    • Attachment

S

• Scottish Government

  • Challenge number 39

    How can generative AI agents be used robustly to enable government to take a systemic view of policy?

    Governments departments often work in insular environments, publishing plans and strategies with a focus on their immediate deliverables. Systems thinking techniques offer a framework for connecting factors across a wider system, revealing dependencies, overlaps and gaps, and supporting alignment across the policy landscape. Scottish Government analysts are interested in developing an agentic AI tool that will create system maps that identify cross-plan connections. How can agentic AI be utilised effectively and robustly in this context to support government policy making decisions? Scottish Government seeks expert advice on the development and implementation of this agentic tool, with particular consideration of risks and mitigation mechanisms.

• Scottish Government

  • Challenge number 7

    How could emerging technology improve the incentives for competitive development of AI on small-medium data centres in Scotland, which can make better use of green energy, reuse their waste heat within communities, and keep raw data closer to its owners?

    We have concerns about market failure for developing infrastructure technologies around distributed AI, to match the investment in centralised approaches. Even infrastructure companies who benefit from any approach to AI, whether centralised by Hyperscalers or otherwise, there there may not be sufficient incentives for the green energy use / heat reuse / data privacy benefits of decentralised AI training on small-medium community data centres. Meanwhile, I know of existing R&D at Edinburgh looking at managing incentives and remuneration around this kind of distributed AI, and companies like TemplarAI have recently announced impressive performance for a ~10Bn parameter model, which academic contacts seem to consider legitimate https://www.tplr.ai/research

    As well as our Green Data Centres Strategy (https://digitalconnectivity.campaign.gov.scot/green-data-centres), we have a more explicit commitment to distributed AI data centres in our recent new AI Strategy (https://www.gov.scot/publications/scotlands-ai-strategy-2026-2031/pages/8/):

    “Scotland is recognised as a leader in sustainable data centre development, based on renewable energy supply and water-resilient design. Heat reuse, distributed compute and frontier energy solutions are integrated into new developments.”

• SSE

  • Challenge number 53

    How can AI, computer vision and sensor fusion be used to develop the next generation of fish-counting systems for hydroelectric environments?

    SSE Renewables operates a network of fish-counting systems across its hydroelectric assets to support fish passage monitoring and environmental management. This network is key to monitoring the health and migrations of fish populations and helps dam operators manage the operations of their dams in a way that minimizes impact on fish populations. Historically, these have relied on resistivity-based counters, a technology originally developed in the 1950s that has been refined through multiple bespoke hardware generations. However, this technology relies upon specialist components and knowledge that present long-term maintainability challenges. Alongside this, SSE has recently developed and deployed an AI-based fish-counting system at a number of sites. Early deployment has demonstrated the potential of modern computer vision and AI techniques for fish monitoring, creating opportunities to improve accuracy, automate data collection and reduce operational effort. We are interested in exploring how developments in AI, computer vision, sensor fusion and digital signal processing from other sectors could be applied to fish monitoring. In particular, could AI-based approaches be refined through the use of legacy sensor data, or could traditional resistivity counters and AI systems be combined to provide a more accurate, robust and resilient counting solution than either technology alone? We are seeking insights into future-proof monitoring approaches, methods for validating and improving count accuracy, and opportunities to leverage advances already being applied in other industries.

• SSEN TRANSMISSION

  • Challenge number 50

    How can artificial intelligence be used to develop predictive white-box models from intellectual-property-protected manufacturer models that can be applied across a range of power system studies without infringing proprietary information?

    Power system studies often depend on manufacturer-supplied simulation models that provide accurate behaviour but offer limited visibility into their internal structure. This restricts their applicability across different studies and can hinder model sharing and knowledge transfer between stakeholders. We are interested in understanding whether artificial intelligence and machine learning techniques can be used to develop predictive equivalent white-box surrogate models that accurately represent the behaviour of proprietary models while preserving intellectual property. We would like to understand how such models can be validated, generalized across operating conditions, and used confidently in planning and operational studies.

T

• Talking Medicines

  • Challenge number 5

    How can organisations securely combine and utilise data across sectors while preserving privacy, provenance and trust?

    Many of today's most valuable AI applications rely on bringing together data from multiple organisations, disciplines and technologies. However, differences in data standards, governance, interoperability and commercial sensitivities often limit collaboration and innovation. We would like to explore technical approaches that enable trusted data sharing and integration across organisational boundaries, while maintaining data provenance, privacy, security and confidence in the resulting AI-driven insights.

V

• Vinley (NextGenValueInvesting Ltd)

  • Challenge number 12

    How can we use a natural language system (AI, ML, NLP, etc.) to review financial reports and earnings calls from listed companies and identify obfuscation, red flags, and potential accounting issues hidden in the text?

    It is difficult to catch accounting red flags in financial statements beyond statistical measures like Beneish M-score, Benford's law (related to made up number vs natural occurence of numbers), Sloan ratio.

    The world has been full of accounting scandals that could have been caught by reading the financial statements. Examples are : WeWork with related party transactions, changes in accounting policies including the 2 typical ways of manipulating earnings which are changing revenue recognition policy that inflates non-cash earnings through accruals for example and also extending the useful life (depreciation period) of long lived assets.

    We want to go beyond the bag of words related to the following accounting red flags:

    1. Non-GAAP Measurements: The frequent use of Non-GAAP metrics is often aimed at portraying financial results more favorably compared to GAAP-compliant figures. Discrepancies between GAAP and Non-GAAP results should always be closely scrutinized to assess their validity.

    2. Changes in Accounting Policies: Adjustments in key accounting policies, such as depreciation methods, inventory valuation, or revenue recognition, can significantly impact reported financial results. These changes may be regulatory or opportunistic, potentially obscuring the company's true financial performance.

    3. Changes in Revenue Recognition: Shifts in revenue recognition methods, such as moving from point-in-time recognition to a percentage-of-completion model, can alter the timing of revenue reporting. Such changes may artificially inflate or defer revenues, impacting financial transparency.

    4. Marked-to-Market Accounting: The use of fair value accounting for volatile assets can lead to significant earnings volatility. This approach may also mask the company's core earnings stability by reflecting market-driven fluctuations rather than operational performance.

    5. Unusual/Exceptional/Non-Recurring Events (UENE): Frequent classification of certain items as "non-recurring" can be a tactic to smooth earnings. Such practices may also signal underlying operational weaknesses or attempts to manipulate financial perceptions.

    6. Capitalization of Expenses: Improperly capitalizing routine expenses as long-term assets can inflate short-term profitability. This approach, however, increases future amortization expenses, potentially masking financial strain in the short term. A known example is WorldCom who capitalized billions of dollars in line costs, which were operating expenses, spreading them over several years, artificially boosting earnings and concealed its true financial performance.

    7. Related-Party Transactions: Transactions involving related parties that are not conducted at arm's length can distort financial results. Such arrangements may manipulate reported performance or unfairly transfer value within entities. A known example in the context of related-party transactions is Adam Neumann, co-founder of WeWork, who sold the trademark for the word "We" to the company for approximately $5.9 million. This transaction sparked controversy due to potential conflicts of interest, as Neumann was both the CEO and a shareholder. He later returned the money after public backlash.

    8. Off-Balance Sheet Items: The use of leases, Special Purpose Entities (SPEs), or other mechanisms to keep liabilities off the balance sheet can significantly understate a company's leverage and associated risks. A known example in the context of off-balance sheet items is Lehman Brothers who used off-balance sheet transactions (e.g., Repo 105) to temporarily reduce its leverage and debt ratios. However, subsequent events, including exposure to risky mortgage-backed securities, revealed the company’s true financial health. Another famous example is Enron who extensively used off-balance sheet Special Purpose Entities (SPEs) to hide its debts and inflate its profits.

    9. Subsequent Events: Material events occurring after the reporting period, such as lawsuits, acquisitions, or market disruptions, can have substantial implications for a company's future financial health. These events should be factored into financial analyses and forecasts.

    10. Marked-to-Market Adjustments: Marked-to-market accounting highlights how volatile market-driven valuations affect earnings and asset values. This approach requires careful evaluation to separate operational performance from market dynamics.

    11. Unusual Frequency or Volume of Red Flags: A cumulative increase in red flags, such as the ones listed above, across multiple periods or statements suggests systemic issues in a company's reporting practices. This trend warrants deeper investigation to uncover potential risks.