How AI investments turn unprofitable in year two, and how to see it coming
In February 2024, Klarna announced one of the most ambitious enterprise AI deployments of the generative AI era. The Swedish fintech's customer service chatbot, built with OpenAI, had handled 2.3 million conversations in its first month. The CEO claimed it was doing the equivalent work of 700 full-time agents. Cost-per-transaction was falling. Customer satisfaction was holding. The company projected $40 million in profit improvement for 2024. The story was covered globally as evidence that AI displacement of knowledge work had arrived.
Roughly fifteen months later, Klarna began hiring customer service representatives again.
The reversal was not announced as a failure. By Klarna's own accounting, the AI agent had delivered $60 million in cost savings by Q3 2025. But the same earnings release disclosed something else: customer service and operations costs had increased year-on-year, from $42 million to $50 million. The savings were real. They were being absorbed by costs that grew faster than the AI savings could offset. Quality degradation on complex interactions had forced human capacity to be rebuilt. The cost of rebuilding had not been in the original business case.
Klarna's experience is the visible version of a pattern playing out less publicly across hundreds of enterprises. MIT's NANDA initiative analysed 300 enterprise AI deployments in 2025 and found that 95% had produced no measurable financial impact, despite $30 to $40 billion of cumulative investment. McKinsey's State of AI survey of nearly 2,000 companies found that only 5.5% attribute meaningful EBIT impact to their AI initiatives. The technology mostly works. The unit economics do not.
This is not a finance failure. Finance teams are applying the same rigour to AI investments that they apply to ERP rollouts and cloud migrations. The cases are well-built. The numbers pencil. They pass scrutiny. They are also, by the operating reality of AI systems, structurally incomplete in ways that do not become visible until year two.
The structural feature this article names is the Inversion Point: the month at which an AI agent's cost-per-outcome exceeds its value-per-outcome. The month it begins making the enterprise marginally less profitable than not running the agent at all. The Inversion Point is not a single number across all AI deployments. It varies by agent autonomy. For Klarna's customer service deployment, a workflow-class agent that acts within defined boundaries, the Inversion Point arrived at approximately month fifteen. Copilot-class agents that assist human users (Microsoft 365 Copilot, GitHub Copilot, most embedded productivity assistants) typically invert later or not at all, because consumption stays bounded by human acceptance of suggestions and value stays validated at the point of use. Autonomous agents that chain reasoning across steps typically invert earlier. In every class, the Inversion Point arrives invisibly unless someone is measuring for it.
Exhibit 1: The Inversion Point
The month at which an AI agent's cost per outcome exceeds its value per outcome. For Klarna's customer service deployment, the crossover arrived at approximately month fifteen.
• • •
Why the inversion happens
AI agents have a structural feature that no other category of enterprise technology displays in the same way. Their unit economics deteriorate over time, not because of failure, but because three forces operate simultaneously in opposing directions from the day they go live.
The productivity curve flattens. Initial deployment captures the largest gain because the most repetitive tasks are easiest to automate. The marginal task being automated in month twelve delivers less value than the first task automated in month one. The productivity claim is highest at the start, and the steepest part of the realisable gain is captured in the first quarter of operation.
The model consumption curve accelerates. Users learn what the agent can do. They send longer prompts. They use the agent for more complex tasks. They chain multiple agent calls together. Mature usage costs more per request than early usage by a factor of two to five. Costs scale with the success of the deployment, not against it.
The operating tax compounds. Frontier models commonly deprecate within twelve to eighteen months, forcing re-tunes. Regulatory frameworks expand. Evaluation infrastructure grows with the agent footprint. Compliance overhead increases as scope expands. The fixed costs of running an agent at production scale grow faster than the variable productivity gains, particularly after month nine.
The Inversion Point is where these three curves cross. The agent continues to generate volume. It stops generating net value. From that month forward, every additional outcome the agent produces makes the enterprise marginally less profitable than if the agent were retired and the work returned to humans, or restructured against a different operating model.
The agent continues to generate volume. It stops generating net value.
This is the structural feature that traditional software does not have. A licensed enterprise application costs the same per user in year five as in year one. Cloud infrastructure costs scale with usage but predictably. AI agents are the first category where productivity value erodes at the same time as cost compounds, because mature usage drives complexity creep that consumes the productivity gain. The deterioration is not a malfunction. It is the operating reality.
The Inversion Point is observable in retrospect for cases like Klarna's. Its more useful function is prospective: a measurement discipline that lets a finance committee project when crossover will happen, in time to act, rather than discover it through annual results.
• • •
How finance misses it
If the Inversion Point is structural, why has it taken three years and the Klarna disclosure for the pattern to become visible? The answer lies in four habits of technology investment analysis that worked for the last 25 years and do not work for AI.
First, cost is treated as capex plus run-rate, not as variable consumption. Traditional software has predictable run-rates. AI agents have variable consumption that scales with what users feed them. A 50,000-word document costs ten times more to process than a 5,000-word document. The business case shows a flat opex line where the operational reality is a compounding curve.
Second, productivity benefit is calculated as time saved multiplied by loaded cost. This works for back-office automation. It does not work for AI agents in knowledge work. One of the most rigorous large-sample field studies, by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond at Stanford and MIT, examined 5,179 customer support agents at a Fortune 500 firm using AI assistance. Average productivity uplift was 14%, concentrated heavily among novice workers (34%) with minimal effect on experienced workers. This is materially lower than the 25-40% uplift commonly claimed in enterprise business cases.
Third, the denominator is the project, not the agent task. A $5 million business case for a $7 million benefit looks healthy at the project level. The same case at the per-task level might show 60 cents of cost for a task that delivers 45 cents of value. The project envelope absorbs the unit economics problem until the Inversion Point, when the envelope can no longer hide it.
Fourth, year-two costs are assumed to be lower than year one. Traditional software follows this pattern. AI does not. Frontier models commonly deprecate within twelve to eighteen months forcing re-tunes. Regulatory frameworks compound. Evaluation infrastructure expands. Year two is often more expensive than year one, not less.
None of these habits is wrong. They have worked for every previous wave of enterprise technology, from ERP to cloud to SaaS. They simply do not capture how AI agents behave in production. The result is business cases that look healthy at the project level while hiding fragile unit economics at the agent level. Approval is reasonable. Approval without the unit economics is approval of a project envelope, not an investment. An envelope approves a budget. An investment approves unit economics.
• • •
What Klarna teaches us
Return to Klarna. The case is not interesting because Klarna failed. The case is interesting because Klarna's results were largely as advertised, and the deployment still inverted.
By every metric Klarna disclosed in its 2024 IPO filing, the AI agent worked. Cost-per-transaction fell from $0.32 to $0.19 between Q1 2023 and Q1 2025, a 40% reduction. The agent handled the equivalent volume of 700 human agents. The $40 million projected profit improvement was real enough to be disclosed in regulatory documents. None of this was vaporware.
And yet, by month fifteen, the company began rebuilding human capacity. Forrester's principal customer experience analyst Kate Leggett described the strategy as having 'overpivoted on cost containment without thinking about the longer-term impact of customer experience.' By Q3 2025, total customer service and operations costs had grown, not shrunk. The AI savings were absorbed by the operational tax of running the program, the cost of handling quality failures the AI created, and the cost of rebuilding human capacity that had been removed too aggressively.
This is the Inversion Point made visible. Not because the AI broke, but because the structural forces that always operate on AI agents in production caught up with the optimistic case. Klarna is the canonical workflow-class case because it played out under public reporting obligations. The broader pattern is empirical: Forrester's 2026 enterprise AI agent panel found that 22% of agent deployments reported negative ROI at twelve months, and the proportion grew through months thirteen to eighteen before stabilising as kill criteria activated or agents were restructured. Klarna sits within that range. Copilot-class agents typically invert later. Autonomous agents typically invert earlier. The structural pattern is the same; the timing differs. Most enterprises absorb the reversal into general operating costs, identified only when a board member asks the unit economics question.
Exhibit 2: When Inversion Arrives
Distribution of Inversion Points across enterprise AI agent deployments. The 13% cohort that has not yet inverted at 24 months is the only group measuring cost-per-outcome monthly with predefined kill criteria.
Klarna disclosed this publicly. Most enterprises absorb it into general operating costs.
The lesson is not that Klarna made a mistake. The lesson is that the original business case did not contain the analytical apparatus to predict, measure, or manage the Inversion Point. No business case in 2024 did. Few do today.
• • •
The discipline that catches it
The Inversion Point is invisible at the project level where finance reviews investment. It becomes visible at the agent level, monthly, if and only if four operating disciplines are in place before deployment, not retrofitted after. The construct complements emerging vendor and analyst metrics (Gartner's agent value multiple, Forrester's evaluation-coverage benchmarks) by anchoring them to a temporal crossover that boards can ask about directly.
Define success in dollars before deployment. Every agent should have a measurable value-per-outcome established before engineering begins. Not 'productivity uplift,' which is unmeasurable. A specific dollar value per task the business actually values. Without this, the value side of the Inversion Point equation cannot be computed.
Tag consumption at the agent level. Cloud and platform costs must be attributable per agent, per business unit, per use case. Forrester's 2026 enterprise AI agent panel found that organisations without per-agent cost attribution materially mis-forecast their AI infrastructure spend. Without tagging, total cost can be measured, but the cost of any specific agent cannot.
Compute cost-per-outcome monthly. Total fully-loaded cost, divided by number of outcomes that delivered intended business value. Reviewed by a joint forum of business owner, finance, and AI delivery. The cadence matters because year-on-year trends mask short-term cost drift. By the time the inversion is obvious in annual numbers, the program is committed.
Set kill criteria tied to the projected Inversion Point. Every agent should have a predefined threshold below which it is either re-engineered or retired. If the projected Inversion Point falls below the original case threshold for two consecutive months, continuation requires explicit re-approval, not assumed continuation. The kill criteria are predefined because in-the-moment decisions about cost overruns are biased toward continuation.
None of these disciplines is technical. All of them are decisions about how the operating model works. The 13% of enterprise AI deployments that have not inverted at 24 months are not running luckier agents. They are running disciplined agents inside operating models that catch the cost compounding before it overwhelms the value generation.
• • •
The question to ask at the next board meeting
If you are a director or a CFO reviewing the next AI investment case, the question that exposes the unit economics is not 'what is the ROI?' It is more specific.
In what month does this agent's cost-per-outcome exceed its value-per-outcome, and what is the kill criterion when it does?
This question forces three answers the business case must contain. The first is a projected Inversion Point with assumptions. The second is a measurement system that can produce cost-per-outcome monthly. The third is a kill criterion agreed before deployment, not negotiated after the inversion arrives.
If the answer is unclear on any of the three, the business case is incomplete. Approval is reasonable. Approval without these answers is approval of a project envelope, not an investment.
The enterprises that build this discipline now will have a defensible AI capability in 18 months. The enterprises that do not will follow the Klarna pattern. The technology will work as advertised. The savings will be real. They will be absorbed by costs that grow faster than the savings can offset. The board will be told the AI investment is paying off. The board will not be told that the cost-per-outcome inverted six months ago.
The Inversion Point is the structural feature of enterprise AI that the field has not yet learned to manage. It is also the only one that matters for whether your AI investment is profitable or expensive. Whether your organisation measures it now, or discovers it later, is the only variable.
ABOUT THE AUTHOR
Vijayan Seenisamy is a senior practitioner of enterprise AI delivery with more than twenty years of experience shipping enterprise systems at scale. He is the creator of the AI Role Operating Framework (AI ROF™), the author of The AI Delivery Manager Blueprint and The Pilot Trap, and the publisher of The AI Delivery Discipline newsletter. He currently leads group transformation at a top-30 ASX-listed company.
SOURCES
MIT NANDA Initiative, "The GenAI Divide: State of AI in Business 2025" (July 2025); McKinsey QuantumBlack, "The state of AI in 2025: Agents, innovation, and transformation" (November 2025); Brynjolfsson, Li, and Raymond, "Generative AI at Work," NBER Working Paper No. 31161 (April 2023); Klarna Bank AB 2024 IPO filing, Q1 2025 and Q3 2025 earnings releases; Forrester Research 2026 Enterprise AI Agent Panel; Moffatt v. Air Canada, 2024 BCCRT 149.
Vijayan Seenisamy is the author of The Pilot Trap and The AI Delivery Manager Blueprint, and the creator of the AI Role Operating Framework (AI ROF™).
