Communicative Load: Quantifying the Share of Outcomes That Rides on Communication Quality

Abstract

Two tasks. In the first, a system converts a spreadsheet to a different file format. In the second, a manager tells a long-tenured employee that their role is being eliminated.

Both are work. Both can go well or badly. But only one of them can go badly because of how it was said. The conversion either produced a valid file or it did not; no phrasing improves it and none ruins it. The conversation’s entire outcome — whether the person understands what is happening, retains dignity, accepts what follows — rides on communication and almost nothing else.

This article defines the variable that separates them. Communicative load (L) is the share of outcome variance in a task attributable to the quality of communication between the parties, expressed on the unit interval and unsigned by construction. L is a property of tasks, not of people or systems. It has been implicit in four separate literatures — information theory, pragmatics, the psychology of dialogue, and organizational information processing — without ever being named as a measurable quantity or estimated as one.

The article defines L formally, distinguishes it from the constructs it is most likely to be confused with, specifies four estimation strategies with their respective weaknesses, addresses the objection that any such quantity is hopelessly context-dependent, and states what L is for: it is the moderator that determines where accuracy about persons converts into benefit, and without it, claims about personalization have no denominator.

Keywords: communication, task analysis, variance decomposition, media richness, grounding, human–AI interaction

---

1. A Variable Hiding in Four Literatures

The observation that tasks differ in how much their outcomes depend on communication is not new. It is so unsurprising that it has never been formalized — which is precisely why it has never been measured.

Four traditions have circled it.

Information theory established the frame: communication is the reproduction of a message from a source at a destination, and the fidelity of that reproduction is an engineering quantity with a mathematical structure (Shannon, 1948). What the framework brackets by design is whether the destination’s particular characteristics matter — the channel is agnostic about who is listening. For human recipients, that bracket is exactly where the interesting variance lives.

Pragmatics established that human communication vastly exceeds literal content, running on cooperative inference between parties who assume each other’s cooperation (Grice, 1975). Implicature, indirection, and inference are not decoration on the message; they are how meaning is transmitted, and they depend on shared assumptions between specific parties.

The psychology of dialogue established that mutual understanding is not automatic but achieved, through grounding — the collaborative work of establishing that what was said was understood, at a cost that varies with the medium and the participants (Clark & Brennan, 1991). Grounding is work, and work has a cost that some tasks must pay and others need not.

Organizational information processing established that tasks differ in equivocality, and that equivocal tasks demand richer communication channels while routine tasks do not (Daft & Lengel, 1986). This is the closest existing approach to L, and the difference is instructive: media richness theory prescribes channel selection given a task, while L quantifies how much of the outcome rides on communication in the first place. Richness theory answers “which medium?”; L answers “how much does this matter?”

Each literature supplies part of the construct. None supplies a number.

2. Definition

L is the proportion of outcome variance in a task attributable to the quality of communication between the parties.

Four properties are stipulated.

Unsigned. L is a variance share on [0, 1]. It cannot be negative. Communication quality can be bad — that is a property of the communication, not of the load — but the share of outcome that depends on communication is a magnitude. This distinguishes L cleanly from any quantity capturing communication’s valence, which belongs to a different variable.

A property of tasks, not of people or systems. L characterizes the work, not the worker. “Delivering a termination decision” is high-L regardless of who delivers it. Skill at high-L work is a separate variable; L is the demand, not the response to it.

Defined relative to a specified outcome. This is the definition’s most consequential requirement. “Teaching fractions” has no L in the abstract. Teaching fractions such that the student can solve novel problems a week later has an L; teaching fractions such that the lesson finishes on schedule has a lower one. L must always be stated with respect to a named outcome, and disputes about a task’s L usually turn out to be disputes about which outcome is meant.

Defined for a specified party pair. Communication happens between parties. A task’s L may differ for a novice and an expert recipient, since the grounding cost differs (Clark & Brennan, 1991). Where L is reported, the recipient population is part of the specification.

3. What L Is Not

Four adjacent constructs are close enough to cause confusion, and each differs from L in a specific way.

Not task difficulty. Proving a theorem is hard and low-L: the proof is correct or it is not, and no phrasing rescues an invalid one. Telling a family that a treatment has failed is not intellectually difficult and is very high-L. Difficulty and communicative load are orthogonal.

Not social or emotional content. Small talk is saturated with social content and typically low-L, because little outcome rides on it. A technical handoff between engineers on a safety-critical system may be emotionally flat and high-L, because misunderstanding produces failure. What matters is consequence, not affect.

Not communication volume. A task can involve extensive talk that carries little outcome variance — status meetings are the canonical case — or a single sentence carrying almost all of it. L measures dependency, not quantity.

Not channel richness. Media richness theory concerns the capacity of a channel to carry equivocal content (Daft & Lengel, 1986). L concerns how much the outcome depends on that content arriving intact. The two are complementary: high-L tasks are those for which channel choice matters most, which is the bridge between the frameworks.

4. Estimating L

The construct is useless without estimation strategies. Four are available, in ascending order of rigor and cost.

Expert task rating. Domain experts rate tasks on anchored scales against a stated outcome, with agreement reported as an inter-rater reliability. Cheap, scalable to large task inventories, and limited by the fact that it measures beliefs about L rather than L. Suitable for constructing ordinal task taxonomies; insufficient for the estimates that theoretical claims should rest on.

Variance decomposition from field data. Where a task is performed many times with measured outcomes, model the outcome with communication-quality indicators — comprehension checks, clarification-request rates, grounding-repair frequency — and take the variance attributable to those indicators as an L estimate. This is the most naturalistic approach and carries the standard hazard: communication quality is confounded with competence, since capable performers communicate better and also do everything else better.

Experimental manipulation. Hold task and performer constant; degrade communication quality deliberately — remove tailoring, constrain the channel, block clarification — and measure outcome change. This estimates L directly and is the only strategy that supports causal claims. It is also the most expensive, and it is bounded by ethics: the highest-L tasks are frequently those where deliberately degrading communication is indefensible, which means field estimates will be needed exactly where experimental estimates are most wanted.

Counterfactual substitution. Compare outcomes when the same task is completed through communication versus through a non-communicative route — a form submission, an automated pipeline. The difference bounds the communicative contribution. Clean where both routes genuinely exist, and unavailable for tasks that cannot be performed without dialogue, which are the high-L cases.

Two reporting requirements apply to all four. L should be reported with the outcome and the party pair to which it refers, since neither is recoverable from the number alone. And L should be reported with an interval, because every strategy above yields an estimate with error, and a point estimate invites the false precision the construct is designed to avoid.

5. Ordinal Before Cardinal

The construct’s usefulness does not wait on precise cardinal estimation, and pretending otherwise would delay its application indefinitely.

For most purposes, an ordinal scale is sufficient: a small number of levels separating tasks whose outcomes barely depend on communication from those whose outcomes are almost entirely constituted by it. Ordinal placement is defensible from expert rating alone, is stable across raters for tasks at the extremes, and supports the primary theoretical use — establishing that a predicted effect appears in high-L tasks and is absent in low-L ones.

Cardinal estimation matters for a narrower set of questions: how much benefit a given improvement in communication quality should be expected to produce, and how to compare tasks near the middle of the range where ordinal placement is genuinely contested. Those questions are worth the expense of experimental estimation. Most are not.

A companion article develops the ordinal taxonomy in the occupational domain, scoring categories of work by communicative load. The present article’s claim is prior to any particular scale: L is a real, definable, estimable quantity, and the levels a taxonomy assigns are approximations of it rather than definitions of it.

6. Objections

“This is unmeasurable in principle — everything depends on communication somehow.” The claim is not that communication is irrelevant to low-L tasks but that its outcome share is small. Some communication is required to initiate a file conversion; almost none of the conversion’s success depends on how well that request was phrased, beyond a threshold of intelligibility. The threshold structure is part of the construct: below a floor of basic comprehensibility, all tasks fail; above it, tasks differ enormously in how much further communication quality buys them. L measures the region above the floor.

“Variance shares are sample-dependent.” Correct, and this is a property of variance decomposition generally rather than a defect of L. A variance share depends on the range of communication quality actually present in the sample: where all performers communicate well, communication explains little variance, exactly as a study of adults finds that height explains little variance in reading ability. L must therefore be reported with its estimation context, and cross-context comparison requires comparable ranges. The construct is not thereby uninformative; it is conditional, like every variance-based quantity in the behavioral sciences.

“Outcomes are multiple and contested.” Yes, which is why the definition requires a named outcome. The apparent ambiguity of a task’s L is usually the ambiguity of which outcome is being optimized, and forcing that specification is a benefit of the construct rather than a weakness. Where stakeholders disagree about a task’s L, they are frequently disagreeing about what the task is for.

“This just restates that some jobs are people-facing.” The restatement is quantitative and that is the entire point. “People-facing” is a categorical label applied to occupations; L is a continuous property of tasks, defined against a specified outcome, estimable by at least four strategies, and usable as a moderator in a formal model. The difference between a label and a variable is the difference between a description and a prediction.

7. What L Is For

Three uses justify formalizing the construct.

As a moderator. The primary use: L is the quantity that determines where accuracy about a person converts into benefit. A framework developed in a companion article proposes that benefit is multiplicative in accuracy and load, which predicts null personalization effects on low-L tasks regardless of how well a system knows the user — a prediction that is only testable if L can be estimated independently of the outcome it predicts. Without L, “personalization works” and “personalization doesn’t work” are both defensible from the existing literature, because the two camps have studied tasks at opposite ends of a range nobody was measuring.

As a task-selection instrument. Where a system’s person-knowledge should be applied, and where applying it is theater, becomes an empirical question with an answer. This has direct consequences for product design and for the honest scoping of capability claims.

As an economic variable. Task-based frameworks in labor economics decompose work by content in order to predict which tasks technology absorbs (Autor, Levy, & Murnane, 2003), and the arrival of large language models has been measured occupation by occupation in exactly those terms (Eloundou, Manning, Mishkin, & Rock, 2024; Felten, Raj, & Seamans, 2021). Meanwhile, the labor market has increasingly priced social and interpersonal skill (Deming, 2017). Read together, these literatures describe a redistribution of value along a dimension none of them names. L names it, and makes the implied prediction explicit: as low-L cognitive work is absorbed, the residual value of human work — and the leverage of systems assisting it — concentrates in high-L territory.

8. Conclusion

The file conversion and the termination conversation differ in a way everyone recognizes and nobody has quantified. That gap has consequences beyond tidiness. It is why the personalization literature contains both strong effects and silence, with no framework explaining the difference. It is why capability claims about “understanding users” cannot be scoped to the tasks where understanding pays. And it is why the economic argument that interpersonal work is appreciating remains a hunch rather than a measurement.

L is proposed as the missing variable: a share of outcome variance, unsigned, task-specific, outcome-relative, and estimable by ordinary means. It is not a new discovery about communication. It is the formalization of something four literatures have each half-stated, offered in a form that can enter a model, take a value, and be wrong.

---

References

Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. Quarterly Journal of Economics, 118(4), 1279–1333.

Clark, H. H., & Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine, & S. D. Teasley (Eds.), Perspectives on socially shared cognition (pp. 127–149). American Psychological Association.

Daft, R. L., & Lengel, R. H. (1986). Organizational information requirements, media richness and structural design. Management Science, 32(5), 554–571.

Deming, D. J. (2017). The growing importance of social skills in the labor market. Quarterly Journal of Economics, 132(4), 1593–1640.

Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308.

Felten, E., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal, 42(12), 2195–2217.

Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.), Syntax and semantics: Vol. 3. Speech acts (pp. 41–58). Academic Press.

Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27, 379–423, 623–656.

---

Citation compliance note: all 8 references above were verified against live sources on August 16, 2026, in the session that produced this draft series. No reference is cited from memory.