“Edison is the go-to for a new investor to get full understanding of the investment case and business. Edison’s distribution reaches investors that we want to connect with directly.”
Why is the market suddenly asking whether AI pays off?
For two years, enormous AI budgets have been waved through on faith rather than evidence, such that spending and returns have stopped matching up. Big Tech’s capital expenditure on AI infrastructure reached $427bn in 2025 and is projected to increase to $562bn in 2026 according to RBC Wealth Management, while Deloitte’s 2025 survey found that 85% of organisations increased their AI investment over the past year and 91% plan to increase it again. Yet the payback has proved elusive. MIT’s Networked Agents and Decentralized Architecture (NANDA) 2025 initiative found that 95% of enterprise generative AI pilots have delivered no measurable return, and McKinsey reports that for 2025 more than 80% of companies see no material contribution to earnings from generative AI, with just 1% regarding their strategies as mature. The mood has shifted from what AI can do to what it returns. The first wave of this boom rewarded the infrastructure that AI runs on; the next will reward whoever can turn AI into a measurable return.
What is the EU AI Act and why does it matter?
The EU AI Act, in force since August 2024, is the world’s first comprehensive law governing AI. It takes a risk-based approach, where the more consequential the use, the greater the obligations for transparency, documentation, data governance and human oversight. Its roll-out is staggered, and it has just been softened. Under the 2025 Digital Omnibus agreed by EU institutions, the toughest obligations on high-risk systems were pushed back to late 2027 and 2028, with lighter requirements for smaller companies applying from late 2026. The direction of travel, though, is unchanged. The Act signals a wider shift towards AI that must be explainable, auditable and accountable, not simply impressive in demonstration. For enterprises, this reinforces a preference for suppliers that understand regional rules rather than relying solely on US hyperscalers like Amazon, Meta and Microsoft Azure. This is where firms like Asseco stand out as beneficiaries. Asseco is one of the largest software vendors in Central and Eastern Europe, embedding compliant AI into the core banking, healthcare and utility systems its regional clients already run.
Why does a problem with returns create investment opportunity?
The headline failure rate hides a more useful finding. The problem traces back to execution rather than model quality, where poor data, weak integration and no reliable way to measure results lead to failure. S&P found that the share of companies abandoning most of their AI projects before deployment rose from 17% to 42% in a single year. That reframes where the value sits. If the models themselves are becoming a commodity, the advantage passes to the companies that solve the harder problem of making AI work inside a business and proving the return on investment that customers now demand. MIT found the clearest returns in back-office automation and that specialist vendors succeeded roughly twice as often as in-house builds. Broadly, the value is captured by companies that secure the data AI depends on, automate the high-volume tasks where savings are countable, embed AI into the core systems a company already runs and serve the specialised niches where a return is physical, immediate or legally required.
Why do the returns start with data rather than algorithms?
A model is only as good as what it is fed, and most enterprises cannot yet measure what their models produce. Data quality is consistently named the single biggest obstacle to AI returns as many enterprises hold vast, fragmented and largely ungoverned data, much of it unstructured. Deloitte found that one in four organisations blame inadequate data and infrastructure directly, and according to Forbes, 39% of firms report that measuring the return is itself one of their biggest problems. Two issues follow: AI trained on messy data underperforms, and without a clean baseline a company cannot prove whether the investment worked at all. This is where the least visible companies matter most. Celebrus Technologies, formerly D4T4 Solutions, captures first-party customer data in real time and applies it to identity and fraud prevention, supplying the compliant, high-quality data foundation that enterprise AI needs to function. ActiveOps takes the measurement side with its software to track how back-office teams in banking, insurance and healthcare perform, giving managers the numbers to show whether AI is making them more productive. A return is impossible to claim without both a trustworthy data foundation and a reliable way to measure the result.
Where is AI already paying its way?
The strongest returns are coming from the back office, where the work is high in volume, repetitive and easy to cost. Despite many budgets still being aimed at sales and marketing, operational efficiency was found to be the single most common way companies reported a return in 2025, with 64% citing it, according to Forbes. The pattern is consistent: automating customer service, document processing and similar workflows produce visible savings. Companies like Netcall apply this through low-code, contact-centre and robotic process automation tools, injecting AI into narrow, high-volume tasks where cost reduction is measurable rather than theoretical. Expert.ai takes a deliberately different route from the giant general-purpose models, building smaller, domain-specific language systems for document-heavy work such as insurance claims and legal review, where the cost saved per document is explicit. In each case the AI is narrow, embedded and measurable.
Why does AI depend on the systems beneath it?
The highest-value uses of AI sit inside a company’s core processes, and those are the hardest to reach. McKinsey found that only one in 10 AI projects move beyond a pilot and that the ones with real economic potential are those embedded in core operations rather than bolted on. The obstacle is integration. This is why enterprise resource planning (ERP) software, the backbone that runs finance and operations, was the fastest-growing area of technology investment in Deloitte’s survey, with almost half of AI investors also putting money into ERP. Aptitude Software positions its Fynapse platform as AI-native finance software, providing the audit-grade, real-time ledger that lets AI operate within strict financial controls. Kainos, a digital services firm and Workday implementation partner, moves AI out of experiments and into the systems employees use every day. Wavestone, a management and technology consultancy, is brought in earlier still, to fix data strategy and governance so that AI investment stands a real chance of delivering a return.
Which niches make the return easiest to see?
Some returns are simply easier to see than others. In transport and logistics, where the payoff is physical and can be counted in time, fuel or money saved, the case is unusually clear. Tracsis supplies software, data analytics and hardware to the rail and traffic-management industries, where the return on AI shows up directly in delays reduced and journeys optimised. Microlise provides telematics and fleet-management technology to logistics operators running on thin margins; its AI has to translate into lower fuel bills and better-used vehicles, or the customer simply will not renew. These are not speculative deployments. The value is concrete, the measurement is built in, and the buyer knows within a quarter whether the software has paid for itself, which is exactly the discipline the wider market is now demanding.
Where is the return harder to measure but just as real?
Not every return shows up as a saving. In regulated and trust-dependent settings, the payoff is a direct cost avoided or a crisis averted. Idox supplies software to local governments and the public sector, where budgets are tight and risk tolerance is low, so any AI feature has to earn its place through concrete administrative efficiency rather than novelty. Fintel provides software, compliance and data to the UK financial-advice market. Its value lies in risk mitigation, automating regulatory checks and research so that advisers avoid the far larger cost of getting compliance wrong. Trustpilot depends on proprietary machine learning to detect fake and fraudulent reviews, where the return is the integrity of the platform itself, the asset its entire business rests on. In each case the discipline is the same as elsewhere: the AI must be justified against a cost that is real, even when it does not appear on an invoice.
Edison insight
The market has moved into a more sceptical, secondary phase of AI investment. The first phase rewarded the infrastructure that AI runs on; this phase belongs to the software, data and automation companies that can turn AI into a measurable return. With most AI pilots still failing to prove their worth, and regulation such as the EU AI Act pushing towards AI that is accountable as well as capable, demand is consolidating around particular types of companies: those that can fix data, automate the measurable, embed into core systems, or serve niches where the payoff is easy to see. The opportunity is real but selective. The discipline that separates enterprise value from marketing hype is now the whole investment case.
Megatrends: Disruptive technologies, digital economy, automation & industrial innovation, changing nature of work
In this note, we introduce the Edison risk-adjusted margin (ERAM), which we will update on a regular basis. The ERAM compares banks based on the relationship between the net interest income (NII) earned and the credit risk taken. It is designed to screen a large selection of banks and identify cases where the risk-adjusted margin is likely to change, which will in turn affect profitability and, hence, valuation. Looking both at the ERAM and the relative valuations helps to suggest investment opportunities in the banking universe. We start by looking at the key drivers of banks’ valuations and find that profitability, expressed as return on equity (ROE), is the main driver for banks operating in mature markets (ie the higher the profitability, the higher the valuation). The relationship between profitability and valuation is strong. In total, we look at three factors: the ERAM, cost efficiency and capital levels, which in combination determine a bank’s profitability. In this report, we focus on the yield on the loan book compared to the credit risk taken (the risk-adjusted margin). The relationship here is even stronger than that found for valuation. We will publish separate reports on cost and capital efficiency at a later date.
