Companies tolerated this for fifteen years

PMI published the $135 million figure in 2013. McKinsey published the 530,000 days figure in 2019. Both landed on the desks of people with budget. Neither produced a category of software.

That is worth sitting with before assuming the market was simply slow. Thirteen years is long enough for anyone who wanted to fix this to have tried. The more interesting question is why not fixing it was the correct call for most of that time.

A cost with no owner is not a cost anyone pays

Every instance of this failure looks like something else at the moment it surfaces. The customer escalation is a customer success problem. The blown launch date is a product problem. The rep who quoted an old number is a training problem. The deal that stalled at legal is a contracting problem.

Nobody sees the shape because nobody sees more than one instance at a time, and each one has a plausible local explanation. The aggregate exists only in a consultancy report that describes companies in general rather than this company in particular.

So the cost lands, gets absorbed into four different departments’ ordinary friction, and never appears as a line anyone is accountable for. A budget holder cannot approve spending against a number that does not appear in their own reporting. That is not short-sightedness. That is how budgets work.

And the fixes on offer all had a worse ratio

The available remedies were process. More syncs, wider distribution lists, a RACI, a programme manager whose job is to carry decisions between teams. Each of them works, in the narrow sense that they do reduce the failure rate.

Each of them also costs senior time, every week, forever, and scales with the number of teams rather than with the number of decisions. Add a seventh team and you have not added one more thing to coordinate, you have added six. Companies that pushed hard on process found that the cure consumed more calendar than the disease, then quietly rolled it back.

PMI even measured the upside. Organisations with highly effective communications met their project goals 80% of the time against 52% for the least effective (PMI, 2013). Twenty-eight points is a large prize. It stayed unclaimed because the only route to it was a permanent tax on the people whose time is most expensive.

What changed was the volume, not the rate

Assume the propagation failure rate held steady. Some fixed proportion of decisions fail to reach the teams standing on them. That proportion is a property of how humans hand work over, and there is no particular reason it moved.

What moved is the denominator. Every team now produces more decisions, more artefacts, and more commitments per week than it did three years ago, because the tools got dramatically better at the part that happens inside a team. A steady failure rate applied to a much larger volume is a much larger absolute number of failures.

The rate was survivable at the old throughput. The same rate at ten times the throughput is not.

The 2025 research reads differently once you hold that in mind. The workslop finding, that 40% of employees received AI-generated work in the past month that looked finished and was not, and that roughly half of it changes hands between colleagues, is a measurement of more material crossing the same unimproved seams (HBR, September 2025).

MIT Project NANDA found about 95% of organisations got no measurable return from their GenAI pilots, and attributed it to systems that do not retain what they learn (The GenAI Divide). Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027 (Gartner, June 2025). Those are the sound of companies discovering they optimised the wrong half of the pipeline.

The ratio is what has to change

If the argument above is right, the reason to act now is not that the problem got worse in kind. It is that the arithmetic flipped. The failures got more numerous while the cost of the fix stopped being a permanent tax on senior calendars.

That is the only claim we are making about what we build. Not that it eliminates the failure rate, which we have no basis to assert, but that carrying a decision to the teams standing on it stops being a person’s recurring job. The decision is recorded once where it is already being made. The downstream set is derived rather than remembered. The people affected get told in the tool they are already in, and one human approves anything that reaches a customer.

We should say plainly what we cannot support. We do not know your failure rate, because no public dataset measures it and we have not yet measured it ourselves. The thirteen-year gap between the research and any serious attempt to close it is a fact. The claim that the arithmetic has now flipped is our reading of that fact, and you should treat it as a reading.

The version of this question you can answer yourself is cheaper anyway. Count the decisions your company changed last quarter that affected more than one team. Then count how many had a named person responsible for telling the others. If the second number is much smaller than the first, you are running on tolerance, and tolerance is a function of volume.