We Measured How Much Information Dies in Every Handoff

An ambulance backs into the bay at a Level I trauma center. The crew has spent twenty minutes with a critically ill patient. They have seen the scene, felt the pulse, watched the breathing change, made decisions on the way. They roll the patient inside and tell the emergency department what they know.

A 2017 study published in Prehospital Emergency Care recorded what actually crossed that threshold. Across ninety critically ill or injured patient transfers, the chief complaint made it 94 percent of the time. The history of present illness, 84 percent. Pertinent physical exam findings, 47 percent. The patient’s gender, 40 percent. The patient’s age, 19 percent (Carter et al., Prehospital Emergency Care, 21(1), 2017).

That last number is worth holding for a moment. Nineteen percent. Of the most basic identifier a doctor needs to dose a drug, weigh a prognosis, or order an imaging study, four out of five times it does not arrive.

This is not a story about EMS crews failing at their jobs. It is a story about handoffs — and one that turns out to be uncomfortably similar wherever it has been studied. Information dies at every transition between people, teams, or systems. We have decades of measurements across medicine, manufacturing, construction, aviation, software engineering, and laboratory psychology. The numbers cluster around a pattern most professionals will refuse to believe applies to them: roughly 20 to 30 percent of information dies at the first handoff, 5 to 10 percent at each subsequent hop, and after five to seven handoffs you are down to about half. Across every domain. Almost regardless of what is being handed off.

The Cambridge Lab, 1932

The first person to measure this carefully was Frederic Bartlett, a Cambridge psychologist who ran what he called serial reproduction studies. He gave his subjects a 330-word Native American folk tale called “The War of the Ghosts” and had them retell it to a second subject, who retold it to a third, and so on down a chain (Remembering: A Study in Experimental and Social Psychology, Cambridge University Press, 1932). After roughly seven retellings, the story had compressed to about 180 words — a 45 percent reduction in raw volume.

But the loss was not random. Bartlett identified three reliable distortions. Leveling: unfamiliar details — the supernatural elements, the non-Western cultural references — dropped out first. Sharpening: the details that survived got more vivid, often more dramatic than in the original. Assimilation: the story warped to fit the listener’s existing schema. By the seventh telling, the ghost story had become a quietly ordinary war anecdote, and every participant in the chain was confident they had told it accurately.

Sixteen years later, Claude Shannon proved why this had to be true. His 1948 paper “A Mathematical Theory of Communication” (Bell System Technical Journal, 27(3)) established that every channel has a finite capacity, and that the information content along a chain of transmissions can never increase — it can only stay the same or decrease. A noisy verbal handoff is, in Shannon’s terms, a channel running near capacity. The sender encodes the rich, multidimensional state of a patient or a system into a narrow stream of words. The receiver decodes through their own schema, which adds noise. The decay is not a failure of professionalism. It is a property of channels.

What the Hospital Studies Found

The most-studied handoff domain is healthcare, because the consequences are easy to count.

A landmark synthesis in the AHRQ patient safety handbook compared three handoff methods (Patterson et al., “Handoffs: Implications for Nurses,” NCBI Bookshelf, 2008). A preprinted sheet plus verbal report retained 96 to 100 percent of information. Note-taking plus verbal report retained 31 to 58 percent. Verbal-only handoffs retained 0 to 26 percent. The most common method in many clinical settings was also the most catastrophically lossy.

In 2014, Amy Starmer and the I-PASS Study Group at Boston Children’s Hospital published the cleanest controlled experiment we have on whether structure can recover what unstructured handoffs lose (Starmer et al., New England Journal of Medicine, 371(19), November 6, 2014). They taught a mnemonic — Illness severity, Patient summary, Action list, Situational awareness, Synthesis by receiver — and required the receiver to read back. Medical errors fell by 23 percent. Preventable adverse events fell by 30 percent. The intervention added zero new information to the system. It just stopped destroying what was already there.

Why does this matter beyond hospitals? Because when nursing researchers actually classified the content of handoffs (Collins et al., CIN: Computers, Informatics, Nursing, 40(1), 2022), they found that only 13.6 percent of what was communicated qualified as “knowledge” — interpreted, contextual, actionable. The rest was raw information or undigested data. Yet nurses on either side of the handoff believed they had transferred knowledge. The illusion that handoffs work is itself the strongest reason they don’t.

The Same Curve Shows Up Everywhere

If this were a healthcare-only finding, we could blame healthcare. It isn’t.

In a study of pharmaceutical manufacturing handoffs at the CRB Group (Kulkarni, “Measuring Information Changes at Handoff Points”), the first handoff lost about 20 percent of information, and each subsequent handoff lost an additional 6 to 7 percent. After five hops, roughly half the original information was gone. That curve — front-loaded, then a long compounding tail — has the same shape Bartlett measured in 1932.

Construction has its own version. When a building project transfers from the construction team to the operations team — the moment called commissioning — industry analyses estimate up to 30 percent of project data is lost at that single transition (FMI/Autodesk industry report, summarized by MSUITE 2020). The construction team holds rich contextual knowledge: why a pipe was routed that way, what compromise the subcontractor accepted, which tolerance was tighter than spec. The handover documents capture a fraction of it. Studies of rework find that 22 percent of redone work is caused specifically by inaccurate or inaccessible information (Building Knowledge Base, 2024), with rework consuming 5 to 10 percent of total project costs in the classic Love et al. studies (International Journal of Project Management, 2005), and reaching 15 to 20 percent on the worst projects.

Aviation noticed this first. After years of accidents traced to crew miscommunication — including the August 2013 UPS A300 crash and the Asiana 214 crash at San Francisco earlier that summer (NTSB AAR-14/01) — the industry developed Crew Resource Management. The core insight was not that pilots needed to talk better. It was that unstructured verbal communication in high-workload environments will always lose information. The fix is to redesign the channel: read-backs, checklists, structured briefings. CRM later became the template for I-PASS in medicine. Aviation’s protocols flowed into medicine’s protocols, because the underlying problem turned out to be the same problem.

Software engineering’s version is the bus factor — the minimum number of team members whose loss would halt a project. A bus factor of one means there is critical knowledge that has undergone zero successful handoffs. It exists in one head, and the compression cost has never been paid because the transfer has never been attempted. Industry estimates suggest organizations lose around 42 percent of project-specific knowledge when annual turnover exceeds 20 percent (CAST Software, vendor report, 2023). The number deserves the salt that all vendor numbers deserve, but the structure — knowledge loss as a step function tied to personnel transitions — matches every other domain.

What Survives, What Dies

Across every studied domain, the same hierarchy of information survival emerges. Narrative survives: what happened, who, the high-salience events, the dramatic incidents. Quantitative data partially survives: vital signs, measurements, dosages — preserved if written, mostly lost if verbal. Context dies first: why this matters, the timestamps, the secondary identifiers, the hedges and qualifications, the spatial awareness, the tacit gut feeling about whether something is wrong.

This is not random forgetting. It is exactly what Bartlett and Shannon predict. High-entropy information — the unexpected, the schema-violating, the precisely numerical — requires more channel capacity to transmit faithfully. Low-entropy information — the predictable narrative, the schema-consistent, the dramatic — survives because it costs less to encode. The receiver’s schema does the leveling automatically. Nobody chooses what to forget. The channel chooses.

This explains the EMS finding that opened this essay. “Chief complaint” is high-salience and narrative-friendly: abdominal pain, sudden onset, last hour. It survives 94 percent of the time. “Age” is a bare number with no narrative scaffolding to hang it on. It survives 19 percent.

The Illusion of Transfer

The most disturbing finding across the literature is not that information dies. It is that nobody on either side of the handoff can tell.

Sign-out sheets in medical handoffs contained errors in 67 percent of cases — missing allergies, wrong weights, incorrect medication details — and physicians rated the handoffs as adequate (Patterson et al., 2008). The Joint Commission attributes roughly 80 percent of serious medical errors to communication lapses, with handoffs the dominant mode. The receivers did not feel like they were missing 80 percent of what they needed. They felt confident.

Three cognitive mechanisms keep the illusion intact. The fluency heuristic: when a message feels easy to process, we infer that we have understood it. A smooth handoff narrative produces the cognitive ease that is functionally indistinguishable from comprehension. Confirmation bias: receivers attend to information that matches their expectations and discount what doesn’t, which means the schema that helps them organize the handoff also filters it. And the curse of knowledge (Camerer, Loewenstein & Weber, Journal of Political Economy, 97(5), 1989): senders cannot un-know what they know, so they systematically underestimate what needs to be made explicit. They compress because they cannot perceive the compression.

Where the Pattern Breaks

Some of this loss is good.

A perfectly lossless handoff would overwhelm the receiver. If the EMS crew transmitted every observation they made — the smell of the apartment, the way the spouse glanced at the bottle on the counter, the seventeen previous calls to that address that the dispatcher remembered — the emergency department would drown. The receiver’s schema doing the leveling is, in Bartlett’s terms, partly a feature. Sharpening makes the surviving details more memorable and useful. The point is not that all compression is bad. The point is that uncontrolled compression is bad — when what survives is decided by cognitive salience rather than operational importance.

The cross-domain numbers also need humility. Medical handoffs losing 74 percent and manufacturing handoffs losing 60 percent are not strictly comparable, because what counts as “a unit of information” differs. The literature has not agreed on a normalized measure. The pattern that holds across domains is structural — large first-hop loss, compound decay, systematic bias in what survives — not a single universal percentage.

And technology is not the answer some people hope it is. In medicine, 84.6 percent of handoff information could be documented in the electronic record (Patterson et al., 2008). It isn’t. The information lives in the sender’s head and the system to record it exists, yet the human-to-system handoff remains lossy in the same way the human-to-human one does. Better tools change the channel. They do not eliminate the cost of crossing it.

What to Do With This

If you are building any system that involves handoffs — between people, between teams, between phases, between agents — budget for 20 to 40 percent information loss at each unstructured transition. This is not a pessimistic estimate. It is the consistent finding across every domain that has measured carefully.

Three concrete moves follow.

Standardize the channel for high-stakes handoffs. This is what I-PASS did for hospitals, what Building Information Modeling did for construction (case studies show 70 to 85 percent reductions in rework-related time wastage; Shaqour, Results in Engineering, 23, 2024), what Crew Resource Management did for aviation. The common pattern is the same in every domain: a fixed template with named fields, separation of narrative from numerical data, and a forced read-back step where the receiver confirms in the sender’s presence. For an agent-to-agent message in a multi-agent system, this looks like a structured schema with required fields rather than a free-text prompt — the difference between

{ "task": "...", "constraints": [...], "prior_context": [...], "verify_back": true }

and “hey can you handle this.” The structured version costs more tokens. It loses less information.

Make the loss visible. Failed handoffs are usually invisible until the failure. Build feedback loops that surface what was missing — error rates that get attributed back to the originating handoff, audit logs that compare what was sent to what got acted on, post-incident reviews that ask which handoff dropped the load-bearing context. In software, this is what observability gives you when you trace requests across service boundaries. Each boundary is a handoff. The trace is the read-back.

Treat documentation as a handoff to your future self. The same decay patterns apply. The narrative of why a system was built will survive in your head for a few months and then start to level. The specific reasons — the constraint you discovered halfway through, the rejected alternative, the subtle invariant that made the design work — die first, because they are the highest-entropy parts of the picture. Write them down at the time, in a structured place, and read them back to yourself as if you were a receiver. The you of next year is a different person on the other side of a long channel.

The EMS crews from the 2017 study were not bad at their jobs. They were operating an unstructured verbal channel near capacity, and the channel did to them what every unstructured channel does to everyone. Once we know what every handoff costs, the question is no longer whether information will die. It is which information we will choose to lose, and which we will design to keep.

That is a question worth asking before the next handoff.

Sources: Carter et al., “Information Loss in Emergency Medical Services Handover of Trauma Patients,” Prehospital Emergency Care 21(1), 2017; Frederic Bartlett, Remembering (Cambridge, 1932); Claude Shannon, “A Mathematical Theory of Communication,” Bell System Technical Journal 27(3), 1948; Patterson et al., “Handoffs: Implications for Nurses,” in AHRQ Patient Safety and Quality Handbook (NCBI Bookshelf, 2008); Starmer et al., “Changes in Medical Errors after Implementation of a Handoff Program,” NEJM 371(19), 2014; Collins et al., “Information vs. Knowledge in Nursing Handoff,” CIN 40(1), 2022; Kulkarni, “Measuring Information Changes at Handoff Points,” CRB Group; FMI/Autodesk industry report on construction handover loss (summarized by MSUITE, 2020); Love et al., “Rework in Civil Infrastructure Projects,” International Journal of Project Management, 2005; NTSB AAR-14/01 (UPS A300 / Asiana 214 cycle, summer 2013); Camerer, Loewenstein & Weber, “The Curse of Knowledge in Economic Settings,” Journal of Political Economy 97(5), 1989; Shaqour, BIM rework-reduction case studies, Results in Engineering 23, 2024; CAST Software industry report on knowledge loss and turnover, 2023.