ATTENTION · MEETINGS · SYSTEMS
Channel Capacity Is the Real Reason Your Org Is Slow
Shannon's limit applied to standups, Slack, and review cycles. The bit-rate of management.
Shannon’s limit applied to standups, Slack, and review cycles. The bit-rate of management.
Every leader has a theory for why their org is slow. They blame the wrong people. They blame the wrong process. They blame the wrong culture.
The theories share a common shape. Slowness is treated as a moral failure of the system. Motivation needs to be raised. Alignment needs to be sharpened. Ownership needs to be installed. Each of these is a vibes-level intervention. None of them is wrong, exactly. All of them miss the constraint.
The constraint is physical.
An organization is a network of biological communication channels. Each human node operates at a fixed bit-rate. Each channel between nodes is bounded by signal-to-noise ratio. Once you draw the diagram correctly, the math forces a conclusion that most management literature refuses to confront. Most companies are not slow because their people are bad. They are slow because they violate Shannon-Hartley every Tuesday at 10am.
The Math
Claude Shannon published A Mathematical Theory of Communication in 1948. Cover and Thomas built the canonical modern textbook on it in 1991. MacKay added the inferential bridge in 2003. Seventy-five years of mathematical work converged on one of the cleanest constraint laws in science.
For any channel that transmits information across a noisy medium, the maximum reliable transfer rate is bounded by:
C is the channel capacity in bits per second. B is the bandwidth, the physical width of the channel. S is the signal power. N is the noise power. The ratio S/N is signal-to-noise.
Two properties matter. Capacity scales linearly with bandwidth. It scales logarithmically with signal-to-noise. Doubling the channel doubles capacity. Doubling the clarity of the signal adds one bit per second per hertz. Noise is mathematically punishing in a way bandwidth is not.
This is not a metaphor.
It is the same constraint that governs your wifi router, the fiber-optic cables under the Atlantic, the deep-space network linking Earth to Voyager, and every conversation in your standup. The math does not care that the substrate is biological. It applies anywhere a signal travels through a medium that distorts it.
The Speech Channel
Coupé, Oh, Dediu, and Pellegrino published a study in Science Advances in 2019 that should have ended a thousand management debates. They recorded native speakers of seventeen languages reading standardized texts. They measured speech rate and information density per syllable. They multiplied them together.
The result was the same number across every language. Japanese speakers compensate for low-density syllables by talking faster. English speakers compensate for high-density syllables by talking slower. Phonetic structures vary. The product is invariant.
Human speech transmits at 39 bits per second.
This is the biological channel capacity of one human mouth to one human ear. The visual reading channel sits in the same range, somewhere between 40 and 50 bits per second after Shannon’s 1951 entropy calculation of English text and the cognitive load of decoding. The brain throttles incoming information to roughly the same rate regardless of which sensory channel delivers it.
Now run the calculation no one bothers to do.
A one-hour meeting with eight people contains 3,600 seconds. Only one person can speak intelligibly at a time. Simultaneous speech is additive noise that destroys the signal. The total information generated at the source is 3,600 multiplied by 39, which equals 140,940 bits.
That is 17.6 kilobytes.
Eight cumulative hours of expensive human labor produce less data than a single high-resolution emoji. Multiply by seven receivers and roughly one megabit gets downloaded into the assembled brains. That is the optimistic ceiling. It assumes zero filler, zero status updates, zero monologues about a side project, and a speaker operating at peak compression. Real meetings achieve a fraction of this.
The standup, the all-hands, the bilateral sync, the retrospective. Every synchronous meeting in your calendar is the lowest-bandwidth data transfer mechanism your company owns.
The Noise Floor
Channel capacity has two enemies. Insufficient bandwidth is one. Noise is the other. The logarithm means small increases in noise produce large reductions in capacity. The math is brutal.
Microsoft publishes telemetry from its hundreds of millions of Teams users. The 2025 Work Trend Index shows the average worker now receives 153 Teams messages and 117 emails per weekday. By 8am local time, Teams becomes the dominant medium. Mass emails with twenty or more recipients are up 7% year over year. One-on-one threads are down 5%. The distribution is moving from targeted signal toward broadcast noise.
A worker cannot decode 270 messages a day at 40 bits per second. The numbers do not work. They skim, they triage, they ignore. The brain treats the entire incoming stream as background radiation and tunes most of it out. Important messages get missed because they arrive on the same channel as the lunch order poll.
This is not a discipline problem.
Gloria Mark at UC Irvine measured the cost of a single interruption to a knowledge worker. The average recovery time to fully restore a cognitive state and resume the original task is 23 minutes and 15 seconds. The interruption itself can last three seconds. The penalty does not scale with the duration of the interruption. It scales with the act of interrupting.
Compose the daily ledger of a typical knowledge worker. Five meetings of forty-five minutes each. Eighty Slack pings. Twenty emails that demand a reply. Each ping is a 23-minute penalty. Each meeting fragments a focus block. The maker time available for deep work, the kind that actually moves a roadmap, collapses to near zero.
Cross, Rebele, and Grant studied this in 2016 across more than three hundred organizations for the Harvard Business Review. They found that 20 to 35% of value-added collaboration is generated by 3 to 5% of employees. The other 95% generate the noise. The 5% drown.
Adding Slack does not increase your organization’s bandwidth.
It raises the noise floor. The signal-to-noise ratio falls. Channel capacity drops with it. Each new tool added without subtracting an old one is a net reduction in throughput.
The Decoding Ceiling
The speaking side is bounded. The listening side is bounded too.
SmartBear and Cisco ran a foundational study on code review in 2009. They measured the relationship between lines of code reviewed per hour and the rate of defect detection. The result is a hard ceiling that no amount of seniority, motivation, or coffee can move.
Reviewers detect 70 to 90% of defects when they read 200 to 400 lines of code in a sixty to ninety minute session. Above 500 lines per hour, defect detection collapses. Above 1,000 lines, you might as well not review the code. The reader’s working memory cannot hold enough variable state to validate the logic. The decoding channel is saturated.
The same constraint shows up in document review, contract review, design review, and every other place where one human evaluates the output of another. Pretending that a senior engineer can review 2,000 lines of code on a Friday afternoon does not violate process. It violates physics.
The fix is not faster reviewers. The fix is smaller payloads.
Stripe, GitLab, and Google’s engineering productivity research, summarized in Forsgren, Humble, and Kim’s Accelerate, all converge on the same answer. Small batch sizes. High-density written context. Asynchronous review windows. The cognitive constraint is fixed. Only the payload is negotiable.
The 2018 Stripe Developer Coefficient report quantified what happens when this discipline collapses. The average developer spends 13.5 hours per week paying down technical debt. Another 3.8 hours dealing with bad code. That is 17.3 hours, 42% of the work week, consumed by decoding the consequences of previous communication failures. Stripe estimated the global opportunity cost at $85 billion per year.
Technical debt is not a code problem.
It is the physical residue of every meeting where the requirements were ambiguous, every PR that was rubber-stamped because the reviewer was overloaded, every Slack thread that should have been a memo. The bits got dropped at the source. The downstream channel has to encode them by hand, every time, forever.
Topology Costs
There is one more law to put alongside Shannon’s, and it predates information theory in spirit by nearly two decades. Fred Brooks named it in The Mythical Man-Month in 1975. Adding more people to a late project makes the project later.
The math is unforgiving. A team of N people has N(N-1)/2 possible communication channels. Five people produce ten channels. Ten people produce forty-five. Fifty people produce 1,225. Coordination overhead grows quadratically. Productive output of the new headcount grows at best linearly.
Past a certain headcount, the marginal hire produces less than the marginal coordination cost they impose on everyone else. Past that point, hiring slows the company down. This is observable in every late-stage scaleup that doubles its engineering org and ships less than it did the year before.
Conway’s Law completes the picture. Organizations design systems that mirror their communication structure. If two teams cannot exchange high-bandwidth signal, the products they build will have a low-bandwidth interface between them. Microservice architectures are a confession about org chart topology. APIs are contracts drawn at the boundary of a comms failure.
Robin Dunbar set the outer limit. The human brain can track approximately 150 stable relationships at any one time. Past that number, the routing table overflows. The org has to introduce formal protocols, dotted lines, RACI matrices, and middle management. Every layer of management is a router that adds latency and drops packets.
Galbraith named these tradeoffs in 1974. Goldratt formalized the bottleneck logic for physical inventory in 1984. The intellectual machinery to think clearly about this has existed for fifty years. Most organizations operate as if it does not exist.
What the Math Forces You to Do
The implication is uncomfortable.
You cannot solve organizational slowness by adding meetings. You cannot solve it by adding tools. You cannot solve it with a new framework, a kickoff, an offsite, or a culture document. Every one of those interventions either consumes existing bandwidth or raises the noise floor. The math goes the wrong way every time.
The interventions that actually work all do the same three things.
They subtract channels. Shopify killed all recurring meetings with more than two people in January 2023 and reclaimed an estimated 320,000 hours of focus time inside one quarter. They did not add a new tool. They removed an existing one.
They raise signal density. Amazon banned PowerPoint and required six-page narrative memos. The author spends days compressing the data. The meeting starts with thirty minutes of silent reading. Every attendee downloads the same payload at the visual decoding ceiling of 40 bits per second. The discussion that follows is high-mutual-information by construction.
They batch decoding work. GitLab and Automattic moved status, decisions, and arguments to written async documents. The receiver chooses when to decode. Context switches drop. The 23-minute recovery tax stops compounding.
Each of these is a Shannon-Hartley intervention. Each reduces some combination of channel count, noise floor, or payload size. None of them raises motivation. None of them aligns the team. They change the physics of the substrate. The output follows.
The Steelman
There is a real critique of the channel-capacity frame, and it is worth stating.
Shannon’s theory is famously meaning-blind. Floridi and Hayles have both written extensively on this gap. A megabit of nonsense compresses just as well as a megabit of signal. The math will tell you the channel is saturated. It will not tell you whether the saturated bits are doing useful work. An organization can be perfectly Shannon-optimal and still ship the wrong product.
This is true. It is also not a defense of the current state of most orgs.
The companies that struggle with semantic relevance are a small minority. Most companies struggle with raw throughput. Their bits do not arrive. Their meetings produce less information than the calendar invites that scheduled them. Their tools amplify noise. Their reviews skip the defect because the payload was too large. The question of whether the right bits are flowing is downstream of the question of whether any bits are flowing at all.
Get the physics right first. The meaning question becomes tractable after.
The One-Line Version
Your org is not slow because your people are unmotivated. It is slow because you are pushing data through a 39-bit-per-second biological channel and pretending the calendar can scale.
Sources
This piece draws on Shannon (1948), Cover and Thomas’s Elements of Information Theory (1991), MacKay’s Information Theory, Inference, and Learning Algorithms (2003), Coupé, Oh, Dediu, and Pellegrino’s 2019 cross-linguistic speech rate study in Science Advances, Shannon (1951) on the entropy of printed English, Mark, Gudith, and Klocke (2008) on interruption recovery, Cross, Rebele, and Grant’s 2016 Harvard Business Review article on collaborative overload, Microsoft’s 2022 and 2025 Work Trend Index telemetry, the 2009 SmartBear and Cisco code review study, the 2018 Stripe Developer Coefficient report, the 2023 Asana Anatomy of Work Index, Mankins (2017) on meeting cost in HBR, Brooks’s The Mythical Man-Month (1975), Conway (1968), Dunbar (1992), Galbraith (1974), Goldratt’s The Goal (1984), and Forsgren, Humble, and Kim’s Accelerate (2018). The Shopify meeting purge of January 2023 and Amazon’s six-page memo discipline are referenced from public sources.