Algorithmic Management Audit

Algorithmic Management Audit

Attribution

Built and documented by YourVisionYourCreation LLC (YVYC).

Theoretical foundation: Kellogg, K. C., Valentine, M. A., and Christin, A. (2020), "Algorithms at Work: The New Contested Terrain of Control," Academy of Management Annals, which maps the control functions this audit operationalizes.

Attribution: YourVisionYourCreation LLC, yourvisionyourcreation.com

Doctrine Statement

When software manages people, management decisions do not disappear. They get encoded, scaled, and hidden. An algorithm that assigns shifts, scores performance, or nudges behavior is a manager whose reasoning nobody can question in the hallway.

The audit exists because encoding a management decision does not launder it. It removes the one thing that made the decision survivable: a human who could be asked why.

The Universal So-What

Systems that direct human work get built by teams who describe them in technical language, which means the management functions inside them never get reviewed as management. A scoring model gets a model review. It does not get the review a supervisor would get. This audit applies the second review to systems that only ever received the first.

Theory as Mechanism

The six control functions are load-bearing. Their specific value is that they identify what a system does regardless of what it is called. A feature named "smart routing" that concentrates opportunity toward high scorers is performing the reward and restriction functions whether or not anyone involved described it that way. The framework defeats euphemism, which is the primary obstacle to auditing these systems honestly.

The Six Control Functions

Every system is inventoried against all six. Functions are identified by behavior, never by the name the builder gave the feature.

The compounding case is the one to watch: rating plus rewarding in a single mechanism creates a loop where yesterday's score determines tomorrow's opportunity, which determines tomorrow's score. Nobody designs that loop deliberately. It assembles itself from two reasonable features.

The Six Audit Lanes

Lane 1: Function inventory. Which of the six functions does the system actually exercise, regardless of feature naming? Every identified function is named with the specific mechanism performing it.

Lane 2: Asymmetry inventory. What does the system know about the worker that the worker does not know about the system? Information asymmetry is the resource algorithmic control runs on, and it is measured here rather than assumed.

Lane 3: Transparency and contestability testing. For every decision affecting pay, scheduling, or standing, three separate questions:

• Can the worker see that the decision was made?
• Can the worker understand the basis for it?
• Can the worker contest it to someone with authority to reverse it?

All three must pass. Visibility without an appeal path is notification, not contestability. An appeal path to someone who cannot reverse the decision is theater.

Lane 4: Gaming scan. For every metric, state the behavior that maximizes the metric without serving the actual goal. This is not a prediction of bad actors; it is a specification of what the metric rewards. Any metric with a gaming path shorter than the honest path will be gamed, and that is a design finding, not a worker finding.

Lane 5: Override integrity. Where a human override exists, test whether it is real:

An override that fails any of these four is decorative and gets recorded as absent, not as present.

Lane 6: JumpMaster check. Which gray-area judgment calls is the system resolving silently at scale that should escalate to a human? Every gray area an algorithm resolves quietly is a decision nobody made, applied to thousands of people. This lane names them and requires an escalation path for each.

Findings Format

Each finding carries a lane, the mechanism, the affected population, a severity, and a remedy.

Findings are not averaged into an overall score. A Critical finding is not offset by clean lanes elsewhere.

Design-Time Application

The audit runs best before a system ships, where remedies cost design time rather than trust. At design time three additional requirements attach:

• The contestability path is built before the decision mechanism, not after.
• Every metric ships with its gaming path documented.
• Every gray area the system will encounter has a named escalation route or an explicit decision to accept the risk, recorded.

When the Manager Is an Agent

The moment an agentic system assigns, evaluates, or prioritizes human work, it is performing management and this audit applies to it. This includes agents assigning tasks to human reviewers, agents triaging work queues, and agents scoring human output. The JumpMaster check carries extra weight here: an agent resolving gray areas alone at machine speed produces more unreviewed management decisions per hour than any human supervisor could.

Adversarial Evaluator Gate

Before any audit ships:

If a hostile reviewer wanted to prove this audit was captured by the people who built the system, which lane would they say was run softly?

The answer is usually Lane 5, because override integrity is the finding builders most resist. That lane gets re-run against the four tests explicitly before the audit closes.

The Core Question

Every lane descends from one question: if a human manager did what this system does, in the open, would it be acceptable? Encoding a management decision does not launder it.

What This Skill Will Refuse

• Accepting feature names in place of function analysis
• Scoring an override as present when it fails the integrity tests
• Averaging a Critical finding away against clean lanes
• Treating a gaming path as a worker integrity problem
• Auditing only the technical system while excluding its management functions

Pairs Well With

• motivation-architecture for what the control system does to motivation quality
• human-in-loop-escalation for building the escalation paths this audit demands
• accountability-chain for who owns the algorithm's decisions
• de-skilling-guard for what automated direction does to worker capability over time

YourVisionYourCreation LLC, yourvisionyourcreation.com Licensed under CC BY 4.0