In a manufacturing plant, the scrap rate presented at the monthly operations meeting may run steadily at around two and a half percent, while on the same site's second shift an operator sets aside the first three units of every batch, discards them without asking anyone, and has been doing so as a matter of routine for years. The operator does not experience this as a loss; knowing the thermal behavior of the machine and having learned through repetition that the first units fall outside tolerance, they have engineered their own remedy. The shift supervisor is aware of the adjustment but does not report it, there being no field in the system in which such a loss could be flagged, and the production target being met in any case. The production manager sees the aggregate, not the dispersion. The investment committee sees only the ratio, and that ratio sits comfortably inside the acceptable band.
The pattern belongs to no single plant and no single sector. On a construction site, the divergence between the progress percentage certified in the payment application and the work physically in place; in a contact center, the divergence between average handling time and the shortcuts an agent deploys to close a call; in a warehouse, the divergence between inventory accuracy and the sequence in which a picker actually walks the aisles — these are surfaces of one mechanism. Operational information, moving upward, is condensed at every handoff, and condensation carries the mean while discarding the variance around it. What arrives at the decision table is therefore not reality itself but a compressed representation of reality, faithful in aggregate and silent precisely where the value sits.
The principle holding that a problem must be observed where the work is actually performed — gemba — addresses exactly this compression loss. Its substance is that the decision-maker maintains access to operational reality through a channel independent of the reporting chain, not because the report is false but because a report may be entirely accurate and still incomplete. Every line of a report carries a phenomenon defined tightly enough to be reduced to a number; whatever remains undefined, whatever has not yet been assigned a field, whatever goes unmeasured and therefore counts as nonexistent, cannot enter the report at all. A substantial share of operational improvement opportunity resides in precisely that undefined territory, for whatever has already been defined is, by construction, already being managed.
To overlook the functional side of this compression would be to misread the principle. Hierarchical summarization is not a defect but a necessary consequence of the economics of attention: no decision architecture at any scale can absorb the micro-decisions of several hundred operators, and drowning a manager in raw data produces nothing beyond the loss of signal within noise. Summarization is a rational shortcut that lowers cost for as long as conditions remain stable. The difficulty lies not in the shortcut itself but in the schema remaining fixed when conditions change — when a new product line comes online, when a supplier is replaced, when the shift structure is rebuilt — because the deviation generated by the new condition finds no field in the old schema and consequently is never reported at all.
A second mechanism compounds the first: the person carrying floor knowledge has no incentive to transmit it. The informal adjustment an operator has developed is evidence not that the system failed but that the operator made the system work, and escalating it would mean converting one's own remedy into a declaration of a problem. The calculus facing the shift supervisor runs along similar lines, the return on reporting a deviation being uncertain while its cost is concrete — an inquiry, a request for explanation, a corrective action record with a name attached to it. Information thus remains outside the institutional field of view without ever having been concealed, simply because it was never transmitted. No record is kept, since keeping one benefits no one; and the record that was never kept vanishes entirely a few years later, when that operator retires.
The balance sheet counterpart of this loss rarely appears under its own name. It accumulates first in inventory, since an organization that has never measured the variability on its floor holds a buffer against that variability, and the buffer, normalized over time, becomes a permanent feature of the working capital cycle. It accumulates second in rework and warranty provisions, and third in the ratio of planned capacity to capacity actually attainable, a portion of the equipment investment having been made against throughput that exists on paper and is never reached in practice. The fourth accumulation is the most expensive: capacity investment directed at a misdiagnosed bottleneck ties up capital and simultaneously keeps the true constraint invisible for a further period, the problem now being presumed solved.
Once a company enters a transaction process, this layer translates directly into the language of valuation. When an operational diligence team from an acquirer or a lender visits the site, the question posed is typically this: where does the gap between reported yield and actual yield originate, and is that gap measured anywhere. Where the answer rests on the personal knowledge of the founder or of a production manager two decades into the role — on the presence of someone who knows the temperament of that line — the performance is priced as repeatability not yet demonstrated. The outcome takes the form of a discount to the multiple, an earn-out structure, or a closing condition tied to the retention of named personnel. What determines a company's valuation is frequently not the performance itself but the demonstrability of that performance independent of its founder, and floor knowledge that was never written down makes precisely that demonstration impossible.
The mechanism that neutralizes this tendency is not managers walking the floor more often; site contact conducted without cadence and without a written trail resolves within weeks into either an audit or a morale visit, and neither produces information. A functional structure has four separable components. The first is binding observation to a fixed calendar, since visits at irregular intervals allow the observer to select the moment and therefore introduce sampling bias. The second is conducting the observation by tracing one process end to end rather than by working through a question set; following a single part, a single order, or a single payment application across its entire path surfaces the waiting and handoff points that appear in no department's own report. The third is that every deviation surfaced during observation is written the same day into a record, ideally a field within the existing decision log. The fourth is that closure of that record becomes the opening item of the following observation round, an item left open being sufficient to discredit the mechanism within a few cycles.
BEIREK's intervention in this layer on capital-intensive projects is built along that same axis of record and cadence. On the projects we manage, site observation runs as a channel separate from and parallel to progress reporting: the difference between the progress certified in the payment application and the work physically completed on site enters a deviation record on the day the observation round is made rather than at month end, and that record opens the monthly meeting between sponsor and contractor as its first agenda item. The same discipline is applied through commissioning, where the sources of any gap between contracted performance and performance observed at first draw — a specific item of equipment, a particular assembly sequence, a given operating procedure — are documented while the warranty period is still running.
The second output of this structure matters more at the point of handover from project to operations. When the observations accumulated during commissioning are made part of the documentation transferred to the operating team, the knowledge of how the asset actually behaves ceases to be personal and becomes an institutional holding, one that steepens the yield curve through the first operating year and that also pre-answers the questions a technical advisor will raise in a later refinancing or share transfer. A lender's independent engineer presented with an uninterrupted series of deviation records alongside the operating data will typically hold reserve account calibration and performance-test-linked drawdown conditions within a narrower band than the alternative would support.
A common misreading treats the principle as a shop-floor practice. The same compression loss operates identically where the work is actually performed at a credit underwriting desk, in a sales conversation, or in a customer support queue, and the loss there is generally larger, since the absence of a physical output means the deviation leaves no material trace anywhere. A consolidated report prepared at a holding company headquarters, aggregating a line item that each subsidiary has defined slightly differently, is the same phenomenon expressed at portfolio scale. The place to observe is not the place where production occurs but the place where the consequence of a decision first meets a physical or commercial reality.
The question in the end is not whether a manager walks the floor but through how many distinct channels an institution reaches its own operational reality. A single-channel organization knows what that channel is capable of carrying and treats everything else as absent; and what is treated as absent does not disappear but accumulates in the inventory buffer, in the rework line, or in the discount conceded at the closing table. A practical way to assess a company's operational maturity is to look not at how clean its reporting is but at the mechanism by which the information falling outside that reporting is gathered.