When a sales meeting establishes that funnel conversion has fallen below its level of two years earlier, the discussion tends to travel one of two familiar routes: either the team has lost discipline, or competitive pressure has intensified. The third question, rarely raised, is whether what was counted as a qualified opportunity two years ago and what carries that label today are in fact the same object. In the intervening period the sales organisation has grown, CRM fields have been redefined at least once, a share of marketing-sourced records now drops into the funnel automatically rather than by judgement, and opening a record early has become quietly advantageous to the individual opening it. Conversion has fallen because the composition of the denominator has changed; because the measured quantity continues to travel under its original name, the change never appears anywhere as a discussable item.
The same pattern repeats in the budget cycle, in the credit decision, on the hiring panel and in supplier performance review. A corporate buyer's payment-delay threshold was set when the supply chain carried a materially shorter average tenor than it does now; a hiring panel's definition of a successful first year continues to measure the behaviours that were observable when the work was performed in a room; a developer's permitting-duration estimate still rests on a historical median established before the queue logic governing interconnection was restructured. In each instance the decision rule remains in place, inputs continue to arrive on schedule, and the output looks entirely plausible — the only thing that has changed is the identity of the outcome the rule was built to anticipate.
The mechanism at work here is concept drift: the movement, over time, in the meaning of a predicted target, in its determinants, and in the relationship between that target and the inputs used to estimate it. What distinguishes it from ordinary measurement error is that nothing about the model is broken. The rule remains an accurate representation of the world in which it was constructed; the difficulty is that the world in question is no longer the one being operated in. Drift arrives through three distinct surfaces — the operational definition of the measured outcome changes, so that the same label counts a different object; the distribution of inputs shifts, so that the same variable now represents a different population; or the causal link between input and outcome weakens, so that the same signal no longer announces the same thing. All three produce an identical observational symptom, an unexplained erosion in performance, while requiring entirely different interventions, which is why any correction applied before diagnosis merely accumulates a new layer of calibration debt.
The persistence of this pattern becomes legible once it is read as a function rather than a failure. Holding the definition of a decision rule fixed is a shortcut that lowers institutional cost dramatically: it produces comparable time series, establishes a common vocabulary across teams, allows board reporting to remain internally consistent, and removes the organisational expense of relitigating definitions every period. Where the environment moves slowly, stability is unambiguously the correct choice. The difficulty lies not in the shortcut but in its continuation once the rate of environmental change has overtaken the rate at which the rule is refreshed — and that crossing point, by its nature, never surfaces as a line in any report.
The institutional cost rarely appears first in the metric itself. Drift accumulates initially in exception volume: the number of cases in which a manager manually overrides the answer the rule produces rises, and because each override is individually defensible, the aggregate is never read as a systemic signal. The decision cycle lengthens next — an approval of a type that closed in two meetings six quarters ago now requires three meetings and a supplementary analysis, as participants work to close the widening gap between what the rule outputs and what experienced judgement expects. Only in the third phase does forecast error finally register in reporting, and by that point the historical data available for correction is already a blend of two definition regimes, leaving no clean baseline on which a recalibration could rest.
The corresponding effect on the balance sheet and on cash is similarly delayed and indirect. Where inventory policy remains calibrated to a superseded demand pattern, the consequence shows up not in the absolute size of the inventory line but in a turnover ratio that erodes slowly across quarters; where a pricing rule holds constant through a period in which the composition of input cost has shifted, the consequence is not a single visible break in gross margin but an attrition distributed across individual line items and invisible once aggregated. A few days of slippage in the working capital cycle and a few points of margin drift trigger no threshold on their own; compounded across several consecutive years, they produce a structural gap that becomes difficult to explain when the company is measured against its own historical series.
At the diligence table this gap is not labelled drift; it is recorded as a data-quality finding, and findings of that kind are priced. When an acquirer requests funnel conversion or customer acquisition cost as a three-year series, what is actually being tested is not the level of the series but whether it was produced under a stable definition throughout. Where the definition moved between periods — as, in a growing company, it almost invariably has — the resulting picture supports a judgement about the maturity of management's own forecasting apparatus: this organisation does not maintain a record of what its metrics measure. The customary consequences follow in the documents rather than in the debate: earn-out metrics are fixed word for word in a schedule, the representations and warranties package is extended to cover management reporting, and the discount applied to the projection series is raised independently of any uncertainty inherent in the model itself.
The structural conclusion is that concept drift cannot be managed through individual attentiveness, precisely because its defining characteristic is the absence of a signal for attention to fasten onto. Making it manageable requires three separate institutional components. The first is a definition register, in which the operational definition of each measured outcome, its counting rule, its exclusion criteria and the date of its last amendment are held in a single place, with the metric chart and the register read together as a matter of course. The second is an explicit validity assumption, written at the time the rule is drafted, recording which environmental condition is being treated as constant — demand composition, channel mix, payment tenor, a regulatory threshold — so that a change in that condition automatically signals that the rule is due to be reopened. The third is a recalibration schedule that ties review to a predetermined rhythm rather than to observed deterioration, since any system that waits for failure is structurally late.
The mechanism BEIREK applies in capital-intensive project and portfolio work rests on these three components. Every assumption carrying project economics — construction duration medians, permitting calendars, unit cost indices, observed off-taker behaviour patterns, the escalation curve applied to operating expenditure — is registered individually; the register holds not only the value but the period and environmental conditions from which it was derived, together with the specific observation that would invalidate it. The variance between modelled and realised outcomes is read each period against that register, and where the variance points to the validity of an assumption rather than merely its level, the correction is made in the assumption itself rather than absorbed into a coefficient.
A second layer records the moment before the decision rather than the decision itself. Ahead of an FID, or in a portfolio restructuring, the rationale supporting the decision is written at the point of proposal rather than at the point of approval: the expected outcome, the mechanism through which that outcome is expected to arise, and the observation that would falsify the expectation are all fixed in advance. Deviations emerging in later periods therefore cannot be retold backwards from the result. The same logic governs the stakeholder pre-mortem, in which the environmental assumptions whose movement would render the decision wrong are named before the decision is taken, and those named assumptions are carried forward as standing items on subsequent review agendas.
What these mechanisms share is that none of them asks anyone to be more attentive. The definition register, the validity assumption and the recalibration rhythm convert drift from a matter of intuition into a matter of calendar and file — and an institution's forecasting capability accumulates exactly there, not in the acuity of individual judgement but in the documented record of the conditions under which that judgement holds. To the extent this accumulation can be demonstrated in an acquisition review, it constitutes the most concrete available evidence that performance is repeatable independently of the founder and of the incumbent management team.
The question worth putting, ultimately, is not how accurate an organisation's forecasts have been, but how long the thing being forecast has remained the same thing. The number of institutions that know when a given decision rule was last written comfortably exceeds the number that know which environmental conditions it was written for; and that difference is frequently the same difference that surfaces, years later, in competitive position, in margin, and in the multiple a buyer is prepared to pay.