Portfolio drift: when the unit economics shift under an approved AI strategy.

Most enterprise AI portfolios in 2026 are running on a strategy approved against a workload composition that no longer exists. The board signed off on a program in 2023 that assumed inference dominated the cost surface, retrieval was a bounded lookup against a curated index, and orchestration was a thin layer sitting above a single model provider. None of those three assumptions describes the workload composition of the same portfolio 18 months later, and the operating model the strategy specified has, in almost every case we have studied, failed to absorb the change.

We call the resulting condition portfolio drift. It is not a failure of strategy in the classical sense. The strategy was correct against the workload it was written for. The drift describes the gap between the workload the strategy was approved against and the workload the portfolio is actually running, and the operating consequences of governing the second under the artifacts written for the first.

01The shape of the drift

The shift between 2023 and 2026 has three components, and the operating consequences of each are different. The first is workload composition: the share of total spend devoted to single-shot inference has declined in every portfolio we examined, displaced by orchestration over multistep agentic workflows that issue between 12 and 40 inference calls per business transaction. The second is cost curve: the marginal cost per business transaction is no longer the marginal cost of one inference call against a posted price list, and the procurement model written against the posted price list has, in every case, mispriced the portfolio by a factor of between three and 11. The third is governance surface: the artifact specifying the kill criteria, the escalation path, and the audit posture was written for a system that does not exist, and the system that does exist is, by construction, not governed.

None of these three are recoverable through better forecasting against the original strategy. Each requires a structural rewrite of the operating model, and none of the 17 portfolios we examined had a mechanism in place to trigger that rewrite at the moment the drift began.

02What the 17 portfolios show

The dataset comprises 17 enterprise AI portfolios written between January 2022 and December 2024, drawn from financial services (six), healthcare (four), industrials (four), and technology (three). Aggregate annual run-rate spend across the 17 at the time of original approval was four hundred and 60 million dollars. Aggregate run-rate spend at the close of the most recent fiscal year was one billion two hundred and 10 million dollars, an increase of one hundred and 63 percent against an originally approved increase of 41 percent over the same period.

The dispersion across the 17 is significant, and the median portfolio understates the operating tail. Three of the 17 exhibited drift bounded inside 15 percent of the original cost envelope, in each case because the original strategy contained an explicit kill criterion tied to per-transaction cost rather than to total program spend. The remaining 14 exceeded the original cost envelope by a median factor of two point seven, with a tail at 11.

03What the operating model failed to absorb

The diagnostic question we ran across the 14 drifted portfolios was straightforward. If the workload composition had changed exactly as it did, but the operating model had been instrumented to detect and respond to the change, would the portfolio have drifted? The answer in 12 of the 14 cases was no. The drift was not a failure of the strategy. It was a failure of the operating model to absorb a change in the workload that the architecture, separately, was perfectly capable of executing.

The specific failures cluster into three categories. First, the procurement model priced the portfolio against a single vendor and a single workload class, and was not rewritten when the workload composition shifted into orchestrated agentic systems with calls against three or more model providers. Second, the kill criteria specified in the original strategy were tied to program-level spend rather than to per-transaction unit economics, and the per-transaction economics were never instrumented at the level of granularity required to trigger a kill decision. Third, the escalation path required board-level approval for a strategy rewrite, and the board calendar did not contemplate a rewrite inside the original three-year horizon.

The portfolio was not failing on any business metric the board had approved. It was failing on a unit economic that had not been written into the strategy because the workload it described did not exist when the strategy was written. Engagement note, financial services portfolio, third quarter 2025

04Kill criteria that would have caught it

A kill criterion, in our usage, is an instrumented condition under which the operating model is required to either rewrite the strategy or terminate the program. The strategy artifacts we reviewed contained, on average, two point one named kill criteria, of which the median was tied to total program spend exceeding the approved budget by a fixed percentage. None of the criteria written at the program-spend level would have caught the drift, because total program spend did not exceed the approved envelope until 18 to 22 months after the unit economics had broken.

The three portfolios that did not drift, or that drifted within an acceptable envelope, shared a structural feature. Each had at least one kill criterion specified at the unit economic level, instrumented continuously, and tied to an automatic escalation that did not require a calendar event to trigger. The criteria varied across the three, but each followed the same form: if cost per business transaction exceeds X, the operating model is required to escalate within 72 hours, regardless of total program spend.

05A remediation playbook

For the 14 portfolios in active drift, the remediation sequence we run is the following. The first step is a two-week diagnostic run against the existing portfolio, with the explicit objective of writing the kill criteria the original strategy lacked. The second step is a 72-hour architectural prototype of the workflow most exposed to the drifted unit economics, instrumented to the criteria written in step one. The third step is a 30-day production stand-up of the prototype on your own infrastructure, with the operating model rewritten to absorb the new workload composition. The fourth step is a quarterly review against the kill criteria, with the rewrite trigger explicitly delegated below the board.

The sequence is designed to be reversible at every step. The diagnostic does not commit you to remediation. The prototype does not commit you to production. The production stand-up does not commit you to a multi-year contract. The quarterly review is the moment at which the client decides whether the operating model is absorbing the workload, and the decision authority sits below the board because the board calendar cannot move at the cadence the unit economics now require.

06Implications for the next cycle

The next workload composition shift is, on the evidence of the last three years, between 12 and 24 months away. The portfolios approved in 2025 and 2026 against the agentic orchestration workload that displaced single-shot inference will, on the same trajectory, drift against whatever displaces agentic orchestration. The operating model that absorbs the next drift cleanly will not be the operating model that successfully absorbed the last one. It will be the operating model that contains the kill criteria, the escalation path, and the rewrite trigger, written into the strategy at the moment of approval.

The argument we are making is not that strategy decks are wrong. The argument is that strategy decks describe a workload at a moment, the workload moves, and the operating model is the artifact that decides whether the strategy moves with it. Boards approve strategy. Operating models absorb drift. The portfolios that have done well over the last three years are the portfolios where the second was written with the same care as the first.