In an investment committee session, the same figure receives two different treatments depending on where it comes from. When the finance director reads a cash flow projection off a spreadsheet of their own construction, the members around the table typically interrogate the assumptions — which growth rate, which collection period, which currency scenario. When the identical figure is projected from the screen of a planning system, the questions migrate away from the number and toward its presentation, with the discussion advancing on how the output will be implemented rather than on how the output was produced. The substance has not changed; the only variable that has changed is whether the surface delivering the number is human or machine.
The same asymmetry is observable at every operating layer. An application declined by a credit scoring engine generates fewer appeals than an application declined by a specialist citing the identical grounds; a maintenance overhaul deferred by a planning system remains deferred on the calendar notwithstanding the field team's instinctive warning; a stock level recommended by a demand forecasting module, when it conflicts with the purchasing manager's read of the market, is resolved not by the stronger argument but by the fact that the system occupies the default position. The system's recommendation arrives at the table not as a hypothesis to be tested but as a starting point to be departed from, and whoever objects begins speaking already carrying the burden of proof.
The mechanism underlying this behaviour is termed automation bias — the tendency to grant the output of an automated system unexamined precedence over human judgment — and it operates along two distinct axes. The first is the axis of acceptance, in which an incorrect output produced by the system converts into a decision on the presumption that it is correct. The second, and institutionally the more insidious, is the axis of omission, in which a risk the system never flagged is treated as a risk that does not exist, precisely because no warning arrived. The first axis eventually becomes visible under audit, there being a faulty output and a faulty consequence to trace; the second leaves no trace at all, since the only thing recorded is silence.
It is worth recognising that this tendency originates not as an error but as a shortcut. Where a system has produced consistent output across thousands of cycles, the cognitive and temporal cost of re-verifying each individual output rapidly ceases to be economic, and trust becomes a rational investment in lowering the cost of oversight. The difficulty lies not in the shortcut itself but in the shortcut persisting after the condition that justified it has changed. When the demand pattern against which the model was calibrated breaks, when the supplier base is refreshed, when the rate regime shifts, or when the source of the data feed is quietly migrated to another system, the reliability of the output declines while the confidence extended to that output remains at its prior level. Trust, in other words, tends to outlive the conditions that produced it.
A second layer concerns who, institutionally, owns the system's output. Human judgment attaches to a name, and that name, when asked, is obliged to supply reasoning; a system output, by contrast, is institutionally unowned. Who set the model's assumption, when the parameter was last updated, and which version fed which decision are matters left unrecorded in most organisations. This absence of ownership renders the output not vulnerable but, on the contrary, untouchable, because where there is no counterparty to object to, the objection itself remains organisationally shapeless. Whoever contests the system's output in a decision meeting is arguing not with a colleague but with an institutional abstraction.
The balance sheet equivalent of all this rarely appears in a single line item. A demand forecast that drifts systematically upward accumulates not in the inventory figure but in a stock turnover ratio that softens gradually across several quarters, each quarter's deviation remaining within an explainable magnitude while the cumulative effect approaches the length of the entire working capital cycle. Overhauls deferred by a maintenance planning system find their expression not in failure statistics but in the pricing of the insurance premium at renewal and in rework costs normalising into something resembling fixed overhead. A lending or supplier pool narrowed by a model never registers as a cancelled transaction; it exists as a customer channel that was never opened or an alternative supplier that was never approached — present only through its absence.
In capital-intensive and financed projects the cost surfaces more sharply. Where the critical path of a construction programme is accepted as the output of scheduling software, the productivity assumptions underpinning that programme — shift efficiency, weather loss allowance, procurement delivery window — fall outside the scope of discussion; when delay then materialises, the question that emerges is contractual rather than technical, since the date from which the LD cap runs and the party to whom the delay is attributable must now be reconstructed retrospectively. In the same manner, where a DSCR projection enters the covenant package as the output of a financial model, the absence of any record of that model's calibration positions the sponsor unfavourably in the first re-forecast negotiation with the lender following the initial drawdown.
At the valuation table the effect is more direct still. In diligence, the reviewing party asks not about the projection the system produced but about who within the company reviews the chain of assumptions supporting that projection, and at what rhythm; where that chain cannot be demonstrated, the undemonstrable portion is typically priced as uncertainty. The form that pricing takes need not be a reduction in the headline number — more often it appears as an extended earn-out period, a broadened scope of representations and warranties, or an elevated escrow ratio. What determines a company's valuation is seldom performance itself but the demonstrability of the mechanism that produced it; performance resting on system output, where the mechanism is unrecorded, is performance whose repeatability cannot be proven.
What neutralises this tendency is not greater vigilance on the part of the decision-maker but a decision architecture constructed from specific components. The first is that a system output never circulates unaccompanied — each output is presented together with the three to five critical assumptions supporting it and the date on which those assumptions were last refreshed. The second is that the threshold for challenge is defined in advance and in numerical terms: where the gap between the output and a domain-informed independent estimate exceeds a stated band, the burden of proof shifts from the objector to the system, and the divergence is entered into the record. The third is that a named owner sits on the output, so that who set the parameter and who reviewed it travels with the output much as a signature would. The fourth is that the areas about which the system produces no warning are scanned explicitly and periodically, since the axis of omission becomes visible only when it is deliberately sought.
The intervention BEIREK operates on complex and financed projects rests on these four components. For programme, cost and financial model outputs, the assumption record is kept at the moment the output is produced rather than at the moment it is approved; at each revision, which parameter changed and on what grounds enters the record even where the change does not move the result, since an assumption change that does not move the result today is frequently the one that determines the result several quarters out. In parallel, the divergence between model output and the independent estimates of field or commercial teams is measured on a fixed cadence, and once that divergence exceeds the defined band, the decision does not advance until the rationale for the divergence has been committed to writing.
A second mechanism is a structured counter-argument role established ahead of the decision: prior to an investment decision, FID or a major procurement approval, one member of the team is tasked with reporting not on what the model produced but on the areas about which the model remained silent. This role is not a formality, because it generates a record, and that record is itself the evidentiary chain demonstrating to a reviewing party in post-closing diligence that the projection rests on a functioning review discipline rather than on a founder or on a single system. The difference between the presence and the absence of that record produces, in most transactions, an economic outcome larger than any single line item under negotiation.
What automation contributes to institutional decision-making is speed and consistency, and that contribution is genuine; what it tends to remove is the trace of why the decision was taken. The maturity of an organisation is measured not by how advanced its systems are but by whether it has defined, in advance, who resolves the conflict between system output and human judgment, at what threshold, and against what record. Where that definition is absent the conflict does not disappear; it simply closes quietly in favour of the output, and every conflict that closes quietly amounts to a negotiating position surrendered today for a negotiation that will open later.