On October 5, 2016, a Southwest Airlines plane was waiting at the gate in Louisville, Kentucky, to fly to Baltimore. A passenger’s phone began to overheat in his pocket. It was switched off and not charging, he said. He threw it to the floor, and it started to give off thick smoke. Everyone got off through the main cabin door before take-off.
It was a Samsung Galaxy Note 7, and it was meant to be one of the safe ones. He had bought it on September 21 as a replacement. Its box, he said, carried the black square Samsung had added to tell the new phones from the recalled ones.
The first Note 7s had started catching fire within weeks of going on sale that August. Samsung recalled them and swapped in phones whose batteries came from a different supplier. Then the replacements started smoking too. On October 11 Samsung stopped making the phone for good. In the US alone, the recall covered about 1.9 million phones. The BBC put the cost at about $5.3 billion.
The answer came in January 2017. Each battery had its own defect. In the replacement battery, poorly controlled welding left sharp, tall bumps on one electrode’s tab, tall enough to reach the opposite electrode, and some cells had been assembled without the protective tape over that tab. Exponent, one of the firms Samsung hired to investigate, reported that the first of these cells it examined showed no deficiencies. Cells made after that first look carried the new defect.
The same day, Samsung announced an 8-point battery safety check. It included an X-ray test and a test in which the battery is taken apart, to look at, among other things, the tab welding and the insulation tape. The case has the full investigation.
Four kinds of cost
Quality professionals split the money a company spends on quality into four piles. The American Society for Quality, ASQ, describes them this way.
Prevention costs are spent to stop problems before they happen: clear specifications, training, good design. Appraisal costs pay for checking: inspecting incoming parts, testing products against their specifications, rating suppliers. Then come the costs of poor quality. Internal failure costs arise when a defect is caught before the customer gets the product: scrap, rework, working out what went wrong. External failure costs arise when the customer finds it: repairs, warranty claims, complaints, returns and recalls.
The Note 7 is the last pile at its most expensive. Samsung’s new checks belong to the second. Taking batteries apart at the factory costs money every day. Two recalls and the end of a flagship phone cost a great deal more.
Further down the line, usually dearer
The same defect usually gets more expensive the further it travels. A 2002 report for the US National Institute of Standards and Technology put it plainly: “the longer a defect stays in the program, the more costly it becomes to fix it.” It backed that with a table of two older studies, from software and electronics, both of which found fixes costing more at later stages. The exact multiples rest on assumptions and differ by industry, so the direction is the part worth keeping.
The Note 7 also shows the limit of appraisal. Exponent’s first look at the replacement cells found nothing wrong, because the defect appeared in cells made later. A check only catches what is there when it runs. Prevention, such as better-controlled welding, works on cells that haven’t been made yet.
The wedding invitations
Imagine a print shop with an order for 500 wedding invitations, and a typo in the date. Caught on the proof, it costs a minute to fix. Caught after printing, it costs a reprint: paper, ink and a lost afternoon on the press. Caught after the invitations have gone out, it costs a reprint, postage, an apology, and phone calls to 200 guests, some of whom may already have booked travel. The typo is the same in each case. Only where it was found has changed.
A check the agent can run
The same arithmetic applies to work done by AI agents, where a mistake can travel fast. Anthropic’s guide to Claude Code puts the appraisal step at the center: “Give Claude a check it can run: tests, a build, a screenshot to compare.” Without one, it says, “looks done” is the only signal available, and you become the verification loop: “every mistake waits for you to notice it.”
Automated checks and human review catch different things. A test suite the agent runs after every change can catch a broken function in minutes. It can also miss a change that passes every test but solves the wrong problem, which a reviewer reading the work against the request may spot before it is merged. Either way, what is caught before delivery usually saves rework that a customer would otherwise trigger. For an agent that writes reports or answers customers rather than code, the same logic holds: a set of known cases with known answers, run before anything goes out, is a factory X-ray for the work.
Samsung’s X-ray could see inside a sealed battery. What it couldn’t do was find a defect in phones that had already gone out to customers.