By Christophe Monget, CEO at MoiraCorp. McKinsey says over eighty percent of corporate AI adopters see no earnings impact. HBR’s latest analysis explains why — and what to do about it.
Generative AI promises to transform industries like manufacturing, financial services, and professional advice — automating expertise and optimising processes at a scale we have not seen before. But the data on actual outcomes paints a different picture from the marketing.
A general-purpose technology, not a quick win
HBR’s January-February 2026 article on systematic AI experimentation frames it well: AI is more like electricity than like the internet. Both are general-purpose technologies. Both took decades to deliver their productivity gains, because both required complementary changes — to workflows, to skills, to organisational structure — before the technology could pay back.
Adoption typically hits the “productivity J-curve”. An initial dip in measurable productivity as the firm absorbs the new tools, retrains its people, and rebuilds its processes. Then a sustained gain — but only for firms that make the complementary changes alongside the technology investment. Firms that just deploy the tools and wait usually find themselves sitting on the bottom of the J indefinitely.
The hard truth from the data
A 2025 McKinsey survey of corporate AI adopters showed something striking: over eighty percent reported no measurable earnings impact from their AI investments. Not negative impact — no impact. The firms had bought the tools, deployed them, and the bottom line had not moved.
Setting aside hallucination concerns and the well-documented quality issues with generative AI outputs, the deeper problem is that most firms ignore the complementary shifts AI requires: redesigning tasks, building new skills, reworking the flow of decisions. They drop a chatbot onto an existing workflow and wonder why nothing changed.
If your operations run smoothly on traditional methods — whether manual or legacy systems — they may still outperform an AI-augmented version of themselves without the hassle. The question is not “should we use AI?” The question is “would adopting AI in this specific workflow create more value than it costs?”
The cost side of the equation
That question only makes sense if you take the cost side seriously. AI adoption has three cost layers, all of which are usually underestimated:
- Upfront investment in the tools, the integrations, and the infrastructure
- Talent and upskilling — both hiring new capability and bringing existing staff to a standard where they can use the tools safely and effectively
- Ongoing runs — model costs, cloud bills, monitoring, maintenance, and the institutional overhead of governing AI use responsibly under Consumer Duty and Article 32
If the total cost across those three layers does not significantly outweigh the value created, the deployment is a net loss. And the value calculation has to be honest — not “this saves an adviser two hours a week if everything works perfectly” but “this saves an adviser two hours a week on average, after we account for the time spent reviewing outputs, correcting errors, and managing the audit trail.”
Where the data does show AI paying back
Where AI does deliver, the gains can be substantial — but the pattern is consistent. It pays back in places where structured experimentation has been done, not where adoption was assumed to work.
GitHub and Google have published trials showing twenty-one to fifty-five percent improvements in coding task completion time when AI tools are well-integrated into a developer’s workflow. The variance is itself instructive: how well integrated, into which workflow, with which kind of work — these decisions determine whether you land at the 21% or the 55%.
A Fortune 500 staggered rollout produced fourteen percent productivity gains for experienced staff and thirty-four percent for novices — the larger gain for less experienced staff being a finding repeated across multiple studies. AI is often a more powerful equaliser than it is an accelerator for already-strong performers.
The common factor in both: structured experimentation. Pilot, measure, learn, scale. The firms that get AI to pay back are the firms that test rigorously before they commit.
How we approach this at MoiraCorp
We do not start with the AI conversation. We start with the operational layer underneath it. Where is the time actually going? Where is friction costing the firm? Once that map exists, the AI question becomes precise: in which specific workflows would adopting AI deliver more value than it costs, given this firm’s specific cost structure, skill base, and regulatory context?
If the honest answer is “this is not a workflow where AI will pay back yet” — we tell you. Operational efficiency first, AI where it earns its place. Sometimes the right answer is “fix the process, do not automate it yet”. Sometimes the right answer is “this is exactly where AI will deliver”. The work is figuring out which is which.
The honest evaluation
Drawing on twenty years across investment banking, financial services, and fintech — plus the AI strategy work we have done at MoiraCorp with regulated firms — what I would offer is this.
AI is not a question of “if”. It is a question of “where, when, and how”, asked workflow by workflow, with the cost side honestly measured against the value side. The firms that get this right are the firms that resist the urge to commit before they have tested. The firms that get it wrong are the firms that buy the tool first and ask the questions afterwards.
If your AI agenda is stalling, or if you are about to commit to a programme and want a second opinion on whether the value case adds up, that is the conversation we exist to have.
Sources referenced: HBR, “A Systematic Approach to Experimenting with Gen AI” (Jan-Feb 2026); McKinsey, 2025 corporate AI adoption survey; published productivity studies from GitHub and Google.
