TL;DR. Every professional carries a working model of their industry: who the players are, what customers want, which skills matter, how work gets done. That model was assembled from evidence, and then it stopped being updated. The problem has never been that people refuse to question their assumptions; it is that nobody tells you which assumptions have expired or when to check. This piece builds that schedule from official data. Using EU27 enterprise death rates, we derive an assumption half-life for each sector, the number of years until half the firms currently in it are gone. It runs from about 5.0 years in travel and tour operation to about 11.5 years in real estate, a 2.32-fold spread, with the whole business economy at about 8.2 years. Then we add a second clock, AI adoption by sector, which more than doubled between 2023 and 2025 in eleven of twelve sectors examined, and show that the two clocks are effectively uncorrelated. You cannot infer one from the other. The audit below tells you which one governs your beliefs, and how often to run it.
The problem is not stubbornness, it is the absence of a schedule
Ask any experienced professional whether they should question their assumptions and they will say yes. Ask when they last did it systematically and the answer is usually never, or after something went wrong.
That is not a character flaw. It is a design gap. Beliefs about your industry do not arrive with expiry dates. The belief “our buyers care most about price” was probably true when you formed it. Nothing in your day tells you the year it stopped being true. And because the belief keeps producing plausible-sounding decisions, it never triggers the error signal that would prompt a review.
So the useful question is not “should I question my assumptions” but “how fast does the ground under them actually move, and can I measure that?” It turns out you can, at least for the structural facts, and the answer differs enough between sectors that a single generic answer would be wrong for most readers.
Clock one: how fast the players change
Eurostat’s business demography statistics record, for each sector, the enterprise death rate: enterprise deaths in a year divided by active enterprises that year. That figure can be turned into something more intuitive.
If a sector loses a share d of its enterprises each year, the share still present after n years is approximately (1 – d) raised to the power n. Setting that equal to one half and solving gives a half-life: the number of years until half the enterprises now in your sector are no longer there. It is the same arithmetic used for any constant-rate decay.
We computed it for twelve sectors using 2019, the last pre-pandemic year, alongside AI adoption from the EU’s enterprise ICT survey.
Table 1: Sector turnover and technology adoption, EU27 (verified data with a derived half-life column)
| Sector | Enterprise death rate, 2019 | Assumption half-life (derived) | AI adoption 2023 | AI adoption 2025 | Change |
|---|---|---|---|---|---|
| Real estate activities | 5.83% | 11.5 years | 8.50% | 24.76% | +16.26 pp |
| Manufacturing | 6.71% | 10.0 years | 6.79% | 17.27% | +10.48 pp |
| Construction | 7.47% | 8.9 years | 3.20% | 10.79% | +7.59 pp |
| Professional, scientific and technical | 7.87% | 8.5 years | 18.66% | 40.43% | +21.77 pp |
| Information and communication | 8.08% | 8.2 years | 29.53% | 62.52% | +32.99 pp |
| Wholesale and retail trade | 8.30% | 8.0 years | 6.74% | 18.62% | +11.88 pp |
| Transportation and storage | 8.44% | 7.9 years | 5.26% | 11.15% | +5.89 pp |
| Accommodation and food service | 9.02% | 7.3 years | 3.81% | 11.98% | +8.17 pp |
| Retail trade, except motor vehicles | 9.09% | 7.3 years | 6.08% | 15.46% | +9.38 pp |
| Scientific research and development | 9.73% | 6.8 years | 27.36% | 52.08% | +24.72 pp |
| Administrative and support services | 11.00% | 5.9 years | 8.33% | 19.86% | +11.53 pp |
| Travel agency and tour operator | 12.99% | 5.0 years | 11.35% | 34.96% | +23.61 pp |
| Whole business economy | 8.06% | 8.2 years | 8.06% | 19.95% | +11.89 pp |
Sources: Eurostat, “Business demography by size class and NACE Rev. 2 activity” (bd_9bd_sz_cl_r2), EU27, all size classes, enterprise death rate 2019; and Eurostat, “Artificial intelligence by NACE Rev. 2 activity” (isoc_eb_ain2), EU27, enterprises using at least one AI technology, 2023 and 2025. The half-life column is derived by CEOtudent as the natural logarithm of 0.5 divided by the natural logarithm of one minus the death rate, and assumes the 2019 annual rate continues unchanged. It is a way of reading the death rate, not a forecast. The whole-business-economy death rate is for the business economy excluding activities of holding companies; its AI figure covers all activities except agriculture, forestry, fishing and the financial sector, so the two aggregates are close but not identical in coverage.
The spread is the point. Someone in real estate and someone in travel operations both hear the same advice about staying current, but the underlying facts turn over at rates that differ by a factor of 2.32. A five-year-old mental model of the real estate landscape is middle-aged. A five-year-old mental model of travel operations describes a population that is already half replaced.
Note also that the coincidence in the bottom row is exactly that: the 8.06% enterprise death rate for 2019 and the 8.06% AI adoption figure for 2023 are unrelated measurements of unrelated things that happen to share a number. We have left both in because removing one would be worse, but no inference should be drawn from the match.
Clock two: how fast the work changes, and why it is a separate clock
The right-hand columns of Table 1 measure something different: not whether the firms survive, but whether the way they work is changing underneath them. Between 2023 and 2025, AI adoption more than doubled in eleven of the twelve sectors.
The obvious assumption is that these two clocks track each other, that churning industries are also the fast-modernising ones. They do not.
Table 2: Are the two clocks related? (CEOtudent derived analysis)
| Relationship tested | Correlation across the 12 sectors |
|---|---|
| Assumption half-life against AI adoption level in 2025 | -0.14 |
| Assumption half-life against AI adoption growth, 2023 to 2025 | +0.11 |
| AI adoption level in 2023 against subsequent relative growth | -0.67 |
Derived by CEOtudent from the values in Table 1. These are Pearson correlations across only twelve sector observations and are descriptive of this set, not inferential; with n = 12, correlations of this magnitude in the first two rows are indistinguishable from no relationship.
The first two rows say the same thing: knowing how fast firms die in your sector tells you essentially nothing about how fast the work inside them is being rebuilt. Information and communication has by far the highest AI adoption in the table, at 62.52 percent in 2025, and a completely unremarkable half-life of 8.2 years. Real estate has the longest half-life at 11.5 years and still nearly tripled its AI adoption. Travel and tour operation is unusual in running fast on both.
The third row is a caution about how to read the growth column, and we include it because leaving it out would flatter the analysis. At -0.67, sectors that started low grew fastest in relative terms, which is what happens whenever you compute percentage growth from a small base. That is why Table 1 reports percentage-point changes rather than relative ones. Construction going from 3.20 to 10.79 percent is a 237 percent relative increase and a 7.59 point one; the second number is the honest one.
The practical consequence is a rule: audit on whichever of your two clocks is faster. For most people in professional services or IT, the technology clock is now running well ahead of the turnover clock, so the belief most likely to be out of date is not “who my competitors are” but “how this work gets done.”
A worked example: an assumption almost everyone holds, and the data that contradicts it
An audit protocol is worth little if you have never seen one catch something. So here is a belief that is close to universal, stated confidently in a great deal of business commentary, and not supported by the official European enterprise statistics.
The belief: 2020 was a mass-extinction event for businesses.
Table 3: What the recorded EU27 enterprise data shows for 2020 (verified data with derived counts)
| Measure | Value |
|---|---|
| Sectors with a recorded death rate in both 2019 and 2020 | 150 |
| Sectors where the death rate fell in 2020 | 124 (82.7%) |
| Sectors where the death rate rose in 2020 | 26 (17.3%) |
| Median change in death rate | -0.50 pp |
| Mean change in death rate | -0.57 pp |
| Business economy death rate, 2018 / 2019 / 2020 | 7.09% / 8.06% / 7.22% |
| Business economy birth rate, 2018 / 2019 / 2020 | 9.69% / 10.01% / 8.85% |
Source: Eurostat, bd_9bd_sz_cl_r2, EU27, all size classes. The sector counts, percentages and median and mean changes were computed by CEOtudent across every sector in the dataset reporting both years.
In the official statistics, the recorded enterprise death rate went down in 2020, in more than four of every five sectors. The larger movement in that year was on the other side of the ledger: the birth rate fell by 1.16 points, more than the death rate’s 0.84-point fall. Fewer businesses were started, rather than dramatically more of them dying.
Now audit the audit, because a finding is not a conclusion. At least three explanations compete, and an honest read holds all of them:
- Support schemes. Widespread state support across the EU in 2020 kept firms alive that would otherwise have closed. The number may be real and temporary.
- Administrative lag. An enterprise “death” in these statistics is a registry event, not the day a business stops trading. Closures in 2020 could surface in later reference years.
- Reversion. The 2019 rate of 8.06 percent was itself elevated, up 0.97 points from 7.09 percent in 2018. Part of the 2020 fall is a return toward the earlier level rather than a pandemic effect at all.
That is what a completed audit looks like. It does not end with “everyone was wrong.” It ends with a belief that has been demoted from fact to open question, plus a specific list of what would settle it. The value is not the contradiction, it is that the belief is now being held at the right strength.
It also demonstrates the most common failure mode: the original belief was not fabricated. Businesses genuinely did suffer in 2020. The belief simply substituted a vivid, available mechanism for the measured quantity, and nobody checked which one the statistics recorded. Most expired assumptions look exactly like this. Our piece on why you accept the first conclusion covers the same trap in a different setting.
The Assumption Audit: a CEOtudent protocol
This is designed to be run in a single afternoon, once or twice a year depending on your clock, and repeated with the same written record so you can see what changed.
Step 0. Set your cadence from Table 1.
Find the row closest to your sector. Divide the half-life by four and audit on that interval, rounded to something you will actually do. Roughly: about every 15 months for a half-life near 5 years, about every 2 years for a half-life near 8, about every 3 years for a half-life near 11.5. If the technology clock in your sector is the faster one, halve the interval. The quarter-of-a-half-life rule is a convention, not a finding; it exists so you review well before the population you formed your beliefs about has meaningfully turned over.
Step 1. Write down twenty beliefs before you look anything up.
Speed matters more than polish here, because the beliefs you can state instantly are the ones actually driving your decisions. Cover five areas: who the customers are and what they want, who the competitors are, what skills the work requires, how the money flows, and what is considered normal practice. Write them as flat declaratives. “Clients will not pay for this without a meeting” is auditable; “relationships matter” is not.
Step 2. Tag each belief with the year you formed it.
Not the year you last repeated it, the year you first had evidence for it. Most people find that a third of their list is older than their sector’s half-life. Those go to the front of the queue. This step alone produces most of the value, and it takes about ten minutes.
Step 3. Classify each belief by what would settle it.
Three buckets, and the bucket determines the effort:
- Checkable. A published statistic, a price list, a regulation, a filing settles it in under ten minutes. Do these immediately, before forming any view about the answer.
- Testable. No public source, but a small deliberate experiment would settle it within one cycle of your work. Write the experiment down now, in one sentence, with the result that would change your mind.
- Structural. A belief about how things fundamentally work. These cannot be settled quickly, and they are the most dangerous, because they are load-bearing for everything above them. They get Step 5.
Step 4. Check the checkables against a primary source, not a summary.
This is where most audits quietly fail. If your belief concerns a fee, read the fee schedule. If it concerns a rule, read the rule. If it concerns a market size, find who produced the number and what they counted. A summary of a statistic tends to carry the writer’s framing, and framing is precisely the thing you are trying to test. Table 3 above exists because the recorded definition of an enterprise “death” is not the definition most people assume.
Step 5. For each structural belief, write the strongest case against it.
Not a list of risks, an actual argument that it is wrong, written by you, in full sentences, as though you believed it. The output you want is a specific answer to: what would the world look like if this were false, and does the world already look a bit like that? If you cannot construct the counter-case, you do not understand your own belief well enough to be relying on it. This is where an AI assistant is genuinely useful, precisely because it has no stake in your being right, though you should read our note on trust calibration before treating its answer as evidence rather than as a prompt.
Step 6. Mark each belief confirmed, revised, expired or unresolved, and record the date.
Four states, no others. “Unresolved” is a legitimate outcome and should be common; forcing a verdict is how audits turn into theatre. The date is what makes the next audit cheap, because you only re-examine what has aged past your cadence.
Step 7. Act on exactly one expired belief.
An audit that changes no behaviour was a reading exercise. Pick the single expired belief with the largest footprint across your decisions and change one concrete thing this month because of it. One. The rest go into the record for the next cycle.
What this audit will not do
The protocol handles structural beliefs about your industry, the kind that leave traces in public data or can be tested cheaply. It will not help with beliefs about specific people, questions of value or ethics, or anything where the honest answer is genuinely unknown to everyone. Do not run it on those and conclude the outcome was inconclusive; they were never in scope.
The data behind Table 1 has limits worth being explicit about, and noticing them is itself an application of the method. The business demography series ends in 2020, so the half-life column reflects turnover as it was, not as it is. It covers the EU27 and does not transfer to other economies as percentages, though the ranking of high-churn against low-churn sectors is likely more portable than the levels. Enterprise deaths are registry events, which is exactly the definitional issue Table 3 turns on. And the half-life itself assumes a constant annual rate, which no real sector obeys; it is a way of making a rate legible, not a prediction about any particular firm.
There is a reflexive point here that belongs in an article about expired assumptions. The best public data on how fast your industry changes is itself several years old. That is not a reason to ignore it. It is a reason to hold the number at the strength it deserves, which is the entire discipline this piece is about.
FAQ
How is this different from a SWOT analysis or a competitor review?
Those inventory the external world as you currently understand it. This audits the understanding itself. A SWOT built on a belief that expired in 2021 will be internally consistent and wrong, and nothing in the exercise will reveal that.
Why not just audit everything every year?
Because you will not, and an audit you skip is worth less than a smaller one you complete. The cadence in Step 0 is set from measured turnover so the effort matches the rate at which your beliefs actually decay. If your sector’s half-life is 11.5 years, an annual full audit is largely re-reading last year’s answers.
Is the half-life a prediction about my company?
No, and this matters. It is a property of a sector’s enterprise population, derived from an annual rate on the assumption that the rate holds. It says nothing about any individual firm’s odds. Read it as “how quickly the landscape I formed my beliefs about is replaced,” not as a survival estimate.
Why does the 2020 example use European data rather than global figures?
Because Eurostat publishes the underlying series in a form that can be checked, sector by sector, by anyone. That is the same standard Step 4 asks of you. A global figure assembled from mixed definitions would have been easier to quote and impossible to audit.
Doesn’t the 2020 finding just mean the statistics are wrong?
It might mean the statistics measure something narrower than the belief does, which is not the same as being wrong. That distinction is the most transferable skill in this article. When data contradicts a confident belief, the first question is always whether the two are talking about the same quantity.
Can I run this with an AI assistant?
For Steps 4 and 5, usefully so. For Steps 1 and 2 you should not, because the value comes from what you can state without prompting, and a model will generate a plausible, generic list that is not actually the one running your decisions. Ask it to argue against you, not to tell you what you believe.
How does this relate to forecasting or foresight work?
It sits before it. Forecasting asks what will happen next; an assumption audit checks whether your description of the present is still accurate. A forecast built on an expired model of today inherits the error and adds to it. Our ranking of foresight methods covers the forward-looking half, and our decade-long scorecard of expert AI predictions shows what happens when nobody checks either.
Sources
Eurostat. Business demography by size class and NACE Rev. 2 activity, dataset bd_9bd_sz_cl_r2. European Union, EU27, enterprise birth and death rates, reference years 2018, 2019 and 2020.
Eurostat. Artificial intelligence by NACE Rev. 2 activity, dataset isoc_eb_ain2. European Union, EU27, enterprises using at least one artificial intelligence technology, survey years 2021, 2023, 2024 and 2025.
Eurostat. Artificial intelligence by size class of enterprise, dataset isoc_eb_ai. European Union, EU27, all activities except agriculture, forestry, fishing and the financial sector, survey years 2021 to 2025.
The assumption half-life column in Table 1, the correlations in Table 2, and the sector counts, medians and means in Table 3 were computed by CEOtudent directly from the published Eurostat series named above and were recomputed independently before publication. The half-life is a constant-rate reading of the published death rate and is labelled as derived wherever it appears.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
