A Certain Answer

From the tower at Los Rodeos, neither aircraft could be seen. Both were using the same runway.

The KLM jet was preparing to leave. The Pan Am jet still needed to get off the runway.

It was 27 March 1977. A bombing at Gran Canaria had sent both flights to Tenerife. Diverted aircraft filled the taxiway, forcing departing planes to use the runway to taxi.

Fog moved across the airfield. The tower had no ground radar to show where the planes were.

The Pan Am crew asked whether they should turn at the third intersection. The controller answered: "The third one, sir, one, two, three, third, third one."

The crew had asked for clarification. The controller had named the exit and counted it out.

But the tower could not watch the instruction become a turn off the runway. The radio had to carry what nobody could confirm by sight.

For the crews, hearing the words was only part of the work. They still had to know what those words allowed them to do.

At the other end of the runway, KLM's captain began advancing the throttles.

His first officer warned that they did not have air traffic control clearance. The captain replied: "No, I know that. Go ahead, ask."

The first officer asked. The tower gave instructions for the flight after takeoff. Those instructions were not permission to begin it.

The first officer repeated the instructions back. He ended with words recorded as "we are now at take-off."

Nobody knows for certain whether he said "at takeoff" or "taking off." One could describe being ready; the other could describe an action already underway.

The controller did not understand the transmission as notice that KLM was taking off. He answered "OK."

By then, according to the Spanish report, KLM had begun its takeoff roll about six and a half seconds earlier. The tower's reply did not start the movement.

The controller continued: "stand by for take-off ... I will call you."

At the same time, Pan Am radioed that it was still taxiing down the runway. The two transmissions overlapped.

In the KLM cockpit, interference obscured the warning and the instruction to wait. The Spanish report describes a whistling sound lasting about three seconds.

The tower had spoken. Pan Am had spoken.

A later exchange did get through.

The tower asked Pan Am to report when the runway was clear. Pan Am said it would.

KLM's flight engineer heard enough to question whether the other aircraft had left. He asked: "Is he not clear, that Pan American?"

The captain, Jacob van Zanten, answered "Oh, yes", emphatically.

Van Zanten was KLM's chief flight instructor. His photograph appeared in the airline's advertising. The Spanish report suggested his standing perhaps helped explain the lack of further objections.

The aircraft continued. The collision killed 583 people.

The Spanish investigation placed the main responsibility on the KLM captain. Dutch authorities disputed that emphasis, pointing to mutual misunderstanding and the limits of radio communication. KLM ultimately accepted responsibility.

A sequence involving fog, blocked messages, unclear language and authority cannot be reduced to someone failing to speak.

Every party held information someone else needed. The tower knew it had not issued takeoff clearance. Pan Am knew it was still on the runway.

The engineer's question brought the doubt into KLM's cockpit. The answer closed it without establishing that the runway was clear.

The same records

I've met a quieter version in marketing.

Early in my career, I thought our job ended once we generated demand. We handed lead batches to sales, sure the campaign numbers looked strong.

In pipeline reviews, marketing saw qualified leads. Sales looked at the same records and saw noise.

Nobody had changed the leads between those readings. The teams meant different things by them.

I started reading the CRM audit logs and writing SQL against the raw tables. I wanted to see what happened after the handoff, beyond the report I had helped produce.

The buyers were real. They were not ready to buy that quarter.

Both reports could be true because they answered different questions. Marketing's evidence of demand did not give sales evidence of a sale ready to happen.

Calling the leads qualified had allowed that difference to travel unnoticed. The records arrived, but the word attached to them carried no shared test.

I had treated the handoff as the end of my responsibility. Reading the raw records meant following the problem beyond the point where my report stopped.

More copies of the same report would not have resolved the difference. The teams already had the records. What was missing was a shared understanding of what those records meant for the work ahead.

What sending proves

A handoff has two sides. Someone sends information, and someone else must use it to act.

The sender can usually prove the first part.

The message went out. The document was shared. The report contained the right records.

That evidence says little about what the receiver understood.

When sending becomes the evidence that communication happened, I call the defense Outflow. It lets the sender stop at "We communicated that."

The statement may be true. The needed transfer may still have failed.

From a sender who checks only transmission, information that never transfers looks exactly like information that arrived.

Both leave the same sent message behind.

The difference becomes visible only when someone tries to use it.

That is why I distinguish sending from Proof of Transfer. The check has to reach the receiver's understanding, not just the sender's record.

A plan still has to travel through a brief and a handoff. It can lose meaning at each step while remaining perfectly intact as a file.

The uncomfortable part is that nobody has to lie. Each person can describe what they sent accurately, while the next person works from a different meaning.

Where the customer meets it

The customer has no reason to care where that meaning changed.

Your customer does not know where marketing ends and operations begins. A lead left in a queue for days does not experience a clean departmental boundary.

The click was paid for. The wait mostly helps a competitor.

I've seen the same limit in work that brought rankings but no return the client could trace. Sometimes the problem sat beyond the channel: bad reviews, or calls nobody answered.

The report could be right about the work and still miss what happened to the customer.

I want to follow the information until it becomes something the customer actually meets.

A message can be built for the room it was written in and make complete sense to everyone there. That does not mean it gives the next person what they need.

Inside the room, people can fill gaps from what they already know. Outside it, those gaps become part of the work.

The customer eventually meets the result, not the shared background that made the message seem clear.

For me, taking responsibility means following the problem past the channel that produced it. A successful send cannot settle whether the next person could deliver what was promised.

Something to check against

So what did aviation change?

The safety response included standard, concise language and keeping "takeoff" out of exchanges that were not takeoff clearances. Words needed to mark a clear boundary between flight instructions and permission to move.

Crew resource management also addressed the authority inside the cockpit. A concern needed more than a chance to be spoken. It needed a fair chance to change the decision.

For a business handoff, I want a written intent and a check against it.

The written intent gives both people something fixed to consult. The receiver then says back what they understand and what they will do.

Neither part works well alone.

A document without a check can create false confidence.

People believe they understand, so they stop asking questions. The document's existence becomes a reason not to examine its meaning.

A check without written intent has a different weakness. Both people can agree while meaning different things, because nothing settles what they are agreeing to.

The useful test is whether the receiver's account matches the intent closely enough to act.

The sender has to listen for differences, too. A say-back that earns an automatic nod is only another exchange of words.

Cheaper to send

I expect AI to make this distinction more important.

A model can write a plausible plan in seconds. The plans got cheap. Checking them did not.

I expect the same imbalance to grow as more work passes between agents.

Anthropic described an early version of this problem in its multi-agent research system. The lead agent gave short instructions such as "research the semiconductor shortage."

One subagent researched the 2021 automotive chip crisis. Two others duplicated work on 2025 supply chains.

The agents had received instructions. They had not received a clear division of work.

Anthropic's approach was to specify the objective and output format, give guidance on tools and sources, and define task boundaries. The extra detail addressed what the short instruction left open.

The MAST researchers found related failures across multi-agent systems. Proceeding on wrong assumptions instead of seeking clarification accounted for 6.80% of failures in their analysis.

I take that as evidence that the handoff problem exists in agent systems. I do not take it as proof that a better brief will fix it.

My expectation is that cheaper sending will produce more unchecked handoffs. I expect fluent replies to make those handoffs harder to question, because the reply can sound settled before anyone tests its meaning.

That is where I expect Proof of Transfer to earn its cost.

I would be wrong if teams relying on sent instructions and fluent summaries had no more downstream misunderstandings than teams checking the receiver's understanding. At comparable workloads, if the check adds delay without reducing rework or misread intent, I would drop it.

I want to judge that by what happens after the handoff. A convincing plan cannot supply the evidence for its own success.

The cost of asking

A say-back takes time. It can also sound like doubt, especially when the receiver knows the work well.

If every small exchange requires a formal check, the checking can become its own burden. People need room to act without defending every reading.

I would expect a good AI summary to spare people some repeated explanation. I can see why adding another conversation after it would feel wasteful.

Those are real costs. I would put the delay where work changes hands and a wrong reading changes the outcome.

Even there, the check has limits. It cannot stop two people talking over each other. It cannot make a senior person listen.

The MAST researchers also warn that better context and communication protocols alone are often insufficient. I would not treat a detailed brief as protection against every kind of agent failure.

Tenerife shows the failure at its extreme. An agency misreading a brief is not a plane crash. The costs are not comparable.

But the business still needs a way to find misunderstanding before the customer does.

Prevention gets little attention because the avoided failure leaves so little to see. Firefighting offers a visible rescue and someone to thank.

That makes the time spent checking easy to cut. The time spent repairing the handoff comes later, when it is harder to refuse.

Before the work moves

I use a Story Anchor to hold the intent in writing. The test is to give it to someone who was not in discovery.

Ask what problem the client is trying to solve. Their answer should come from the document, without needing the people who wrote it to fill the gaps.

Ask how we will know we succeeded. Agreement on the work is not enough if success still means different things.

Ask what they should not work on. A boundary that remains in someone's head cannot guide the person doing the work.

Then compare the answers with the written intent.

If the answers differ, the gate has found something worth stopping for. Correct the document or the understanding before the work carries the difference further.

"I'll get that information later" does not pass the gate. It leaves the next person to work around the same missing part.

The sender also gets useful evidence. A wrong answer can reveal a weak brief, not a careless reader.

The check is complete when the receiver can use the intent without borrowing the sender's memory. That is more work than forwarding a file. It is less work than discovering the difference after delivery.

At Tenerife, the tower knew it had not cleared KLM to take off. Pan Am knew it was still on the runway. Inside KLM, the engineer heard enough to ask again.

His question crossed the cockpit. The captain's answer came back with certainty. The aircraft kept moving.

An answer can end a question without settling it.

The work continues among people who check what arrived.