The gap that defines the moment
Capital has moved faster than results. Global corporate investment into artificial intelligence reached roughly $582 billion in 2025, more than double the prior year and well past the previous record set in 2021 — a figure reported in Stanford’s AI Index. Adoption followed: close to nine in ten organisations now use AI somewhere in the business. Yet only a minority report a measurable effect on group profit.
That distance between deployment and value is not a reason to disbelieve the technology. It is the investment thesis. The capability is largely settled; the operating model is not. Over the next three to five years, returns will not accrue to the companies that adopted AI earliest, but to those that redesigned how work is done around it — and to the businesses that sell them the means to do so. Investors who understand this are underwriting execution, not innovation.
From tools to workers
The first wave of enterprise AI sold assistance: a drafting aid, a summariser, a smarter search box. The current wave sells completion. Agentic systems now carry a defined workflow from start to finish — reading a document, checking it against a rule, updating a record, escalating the exception — with a human reviewing outcomes rather than steps.
The commercial consequence is larger than the technical one. When software completes work rather than supporting it, the unit of sale changes from a licence per user to a price per outcome. That expands the addressable market of software companies well beyond IT budgets and into labour budgets, and it compresses the traditional services model that billed for the hours now being absorbed. Investors should watch for a single signal above all others: has the company’s pricing metric changed? Where it has, the economics are usually about to change too.
Where the value actually lands
AI does not distribute its benefits evenly. It rewards industries that are document-heavy, rule-bound, already digitised, and burdened by high-volume repetitive judgement.
Financial services sits at the front. Underwriting, client onboarding, transaction monitoring, reconciliation, collections and claims are all pattern-recognition problems dressed as administrative ones. The sector is expected to account for around a fifth of the growth in global AI spending through 2028, and the gains show up quickly in cost-to-income ratios rather than in press releases.
Healthcare benefits most on the administrative side rather than the clinical one. Documentation, coding, prior authorisation and claims processing consume enormous professional time; early deployments have cut documentation time by roughly 40% in reported pilots. The constraint here is regulatory confidence, not capability.
Manufacturing and supply chain gain through prediction. Maintenance scheduled by condition rather than calendar has reduced unplanned downtime materially, and demand planning improves as models absorb signals that spreadsheets cannot. Asset-heavy businesses convert this directly into utilisation.
Retail and consumer businesses convert AI into revenue rather than cost, through personalisation, assortment decisions and demand-led pricing. This is where AI most clearly builds top line instead of trimming expense.
Professional services — accounting, legal, consulting — face the sharpest re-rating of all, because the input being automated is the very thing that has historically been billed. Value migrates upward, away from preparation and toward judgement, assurance and accountability. Firms that price for time will feel this as compression. Firms that price for responsibility will not.
Energy and real assets form the quiet adjacency. The compute buildout consumes power, land, cooling and grid capacity, creating a long-duration infrastructure opportunity that is less crowded than the semiconductor trade and easier to underwrite on conventional terms.
Four layers of opportunity
For capital allocation, the market separates into four layers with very different risk profiles.
The infrastructure layer — compute, data centres, networking, power — carries the highest capital intensity and the most crowded positioning. Returns depend on sustained utilisation and pricing discipline, not on demand headlines. It is a cyclical business wearing a structural story.
The applied vertical layer — narrow software built deep into a regulated workflow — is where the durable margins are most likely to sit. Defensibility here comes from data access, integration depth and switching cost, not from model quality, since models are converging and commoditising. This is the layer most often mispriced, because it looks unglamorous next to the frontier.
The governance and assurance layer — model risk management, audit trails, controls, data lineage, AI-specific compliance — grows regardless of which model or vendor prevails. It is the toll booth on the whole system, and it is early.
The deployment and adoption layer — the capability to redesign a process, retrain the people, and prove the result — addresses the single largest failure point in the market. A widely cited finding is that the great majority of enterprise AI pilots deliver no meaningful return, not because the technology underperforms but because the organisation is not ready to absorb it. Whoever closes that readiness gap captures a disproportionate share of the value created.
What separates the winners
Four tests are worth applying to any business claiming an AI advantage.
First, proprietary data: does the company hold information its competitors cannot obtain? Access to a general model is not a moat; access to a unique dataset is. Second, process redesign: has the workflow been rebuilt, or has an old process simply been automated faster? Automating a bad process produces the same output sooner. Third, measurable linkage: can the benefit be traced to a specific line in the accounts — headcount, cycle time, loss ratio, gross margin? Where it cannot, it usually does not exist. Fourth, governance from the start: in regulated sectors, deployments without controls, audit trails and clear accountability tend to stall at pilot stage and never reach scale.
What could break the thesis
Three risks deserve honest weight. The first is the return gap itself: a substantial share of the capital deployed in 2025 has yet to show up in measured productivity at economy level, and if that persists, funding conditions will tighten before the applications mature. The second is concentration — a narrow set of suppliers, customers and geographies carries much of the current spend, and correlated exposure is easy to accumulate without noticing. The third is that the surplus may accrue mainly to consumers and to buyers of AI rather than to its producers, through falling prices and rapid capability commoditisation. That would be excellent for the economy and disappointing for a large part of the current capital stack.
The positioning view
Three principles follow. Prefer businesses that use AI to widen a moat they already possess over businesses whose only asset is AI. Prefer cash-generative deployers to pre-revenue model builders at this point in the cycle, because the scarce resource is now execution rather than capability. And underwrite the operating model rather than the technology: the durable question is not what the system can do, but whether the organisation around it can change fast enough to be paid for it.
The opportunity in front of investors is not to buy the future of intelligence. It is to fund the far less glamorous work of turning it into an operating profit.
Frequently Asked Questions
Is AI investment in a bubble?
Parts of it almost certainly are, and parts of it almost certainly are not. The distinction that matters is between capability and pricing. The technology works and is being used at scale; what is stretched in places is the assumption that every business touching it will earn an economic return. Infrastructure and frontier model valuations carry the most cycle risk. Applied software embedded in a regulated workflow, and the services that make deployment work, are priced far more conservatively relative to the cash they generate.
Why do so many AI projects fail to produce a return?
Because the failure is organisational rather than technical. Most stalled programmes automate an existing process instead of redesigning it, which delivers the same output slightly faster and leaves the underlying cost structure intact. The other common causes are poor data quality, no clear owner, and no agreed measure of success — meaning nobody can prove a benefit even where one exists.
Which industries will be affected first and hardest?
Industries that are document-heavy, rule-bound and already digitised move first: financial services, insurance, healthcare administration, logistics and professional services. Industries where physical constraints dominate — construction, hospitality, primary production — move later and more gradually, because the bottleneck is not information processing.
Does AI mainly cut costs or create revenue?
Today, mostly costs. The clearest, fastest and most defensible gains sit in processing, servicing and administration. Revenue effects are real but slower and harder to isolate, and they concentrate in consumer-facing businesses through personalisation, pricing and assortment decisions. Any business case resting entirely on projected revenue uplift deserves closer scrutiny than one resting on cycle time or headcount.
How should a company decide where to start?
Choose the process with the highest volume, the clearest rules and the most measurable output — then check whether the data behind it is clean enough to trust. Starting with the most strategically exciting use case is a common and expensive error; starting with the most measurable one builds the credibility needed to fund everything after it.
What does a credible AI business case look like?
It names a specific process, states the current cost and cycle time, sets a target, and identifies the line in the accounts where the benefit will appear. It includes the cost of change management and ongoing model oversight, not just licences. And it defines in advance what result would justify stopping.
Is proprietary data really a competitive advantage?
Yes, and increasingly it is the main one. Model capability is converging and becoming cheaper, which means access to a model confers no lasting edge. Unique, well-structured data about customers, transactions, assets or operations does — because competitors cannot buy it, and it improves the quality of every application built on top of it.
What are the main governance and compliance risks?
Four recur: decisions that cannot be explained to a regulator or a customer; personal or confidential data flowing into systems without a lawful basis; no audit trail linking an outcome to the logic and data that produced it; and unclear accountability when an automated process gets something wrong. In regulated sectors these are usually the reason a promising pilot never reaches production.
How does AI change the accounting and reporting picture?
Mainly in three places. Capitalisation and amortisation judgements around development spend and infrastructure need to be deliberate rather than assumed. Impairment considerations arise as capability commoditises and useful lives shorten. And where AI materially changes a control environment, auditors and lenders will expect the controls over the automated process to be documented and testable.
Will AI reduce headcount?
It reduces certain kinds of task, which is not the same thing. The consistent pattern so far is redeployment rather than straightforward elimination: volume-processing roles shrink while review, exception-handling and oversight roles grow. Organisations that plan for redeployment tend to realise benefits faster, because staff cooperate with a change that does not threaten them.
What should investors actually look for in a target?
Whether the advantage is proprietary or rented; whether the pricing model has shifted from selling access to selling outcomes; whether customers renew and expand; whether gross margins hold after inference costs; and whether governance was built in or bolted on. A company that can point to a redesigned process and a moved number is materially more valuable than one that can point to an impressive demonstration.
What is the biggest risk to the current thesis?
That the surplus flows to buyers rather than producers. Rapid commoditisation could push the benefit of AI to customers through lower prices, leaving much of the invested capital earning a poor return even as the technology succeeds completely. That outcome would be good for the economy and painful for a meaningful part of the current capital stack — which is precisely why the operating model, not the technology, is the thing to underwrite.
How SRC can help
At SRC, we work with businesses and investors at exactly the point where an AI ambition has to become a number.
We help clients build and stress-test the business case — sizing the expected benefit, testing the assumptions behind it, and modelling the capital and cash flow impact before commitments are made. On transactions, we support due diligence on technology-led targets, with particular attention to whether the claimed advantage is durable, measurable and properly reflected in the accounts.
We also work on the control side: designing governance, documentation and audit-ready processes around automated workflows, so that efficiency gains survive scrutiny from regulators, auditors, lenders and boards. And we advise on the structuring, tax and reporting consequences of technology investment, so that the treatment is deliberate rather than discovered late.
For growing businesses, our most common contribution is narrower and more practical: helping management identify the two or three applications that will actually pay, and letting the rest wait.
If you are weighing an AI investment, evaluating a target, or simply trying to separate the opportunity from the noise, we would be glad to talk.