1. There is no baseline, so there is nothing to compare against
In an executive survey with 1,719 respondents, the statement "we have a defined process for quantifying the impact of our AI initiatives" was claimed by only 33% of the companies that are not high performers. Two thirds of them have no such process.
Countermove. In the two weeks before the rollout, record three numbers, all three from the system and not from memory: the count, the turnaround time from timestamps, the error rate. If the system gives no timestamp, the project starts with timestamping, not with AI. This is the one step that cannot be made up afterwards.
2. The tool gets swapped, the process stays
In the same survey, close to three quarters of the high-performing companies had fundamentally redesigned their workflows, against one quarter of the rest. And in a pre-registered field experiment with 776 participants, an individual working with AI produced a solution of the same quality as a two-person team without AI — meaning the effect showed up in the team structure, not in the properties of the tool.
Countermove. The first step is not picking a tool but drawing the process out step by step and deciding which step disappears — not which one gets faster. The evidence for this, and the counter-tests, is what redesigning the workflow gathers up.
3. There is no human check on the output
Three documented cases, with three different lessons.
An airline's chatbot told a passenger that the bereavement fare discount could be claimed back retroactively. No such rule existed. The company's defence was that the chatbot is "a separate legal entity that is responsible for its own actions". The Canadian consumer tribunal rejected that in February 2024, and one sentence of the reasoning has been the most quoted line in the case ever since: it makes no difference whether the information comes from a static page or from a chatbot. The company is responsible for all the information on its own surface.
A large consultancy left non-existent references and a fabricated court quotation in a report prepared for a government department in October 2025; a researcher spotted the errors, and the firm partially refunded the fee.
The AI serving the email support of a developer tool invented a non-existent rule in April 2025, according to which the product could not be used on several machines. The co-founder publicly denied it: they have no such rule. The company's follow-up measure was to clearly label the email replies given by AI.
Countermove. Every AI output that goes outside either gets human approval or is labelled as machine-written — and the policy says in writing which is which. The company is liable even when a human did not write it.
4. Shadow AI: the team uses it from personal accounts and the company does not know
According to an international employee survey of 48,340 people across 47 countries, 38% of respondents have used AI at work without knowing whether they were allowed to — 41% of them have no rule about it, and 19% do not know whether one exists.
Banning it as a solution is an old idea, and it has an instructive afterlife: in the spring of 2023 a large electronics manufacturer banned generative AI tools on corporate devices after engineers pasted confidential source code into a public chatbot. Three years later the same company rolled out enterprise AI tools group-wide. The lesson, then, is not the ban but the official route.
Countermove. A single tool running on a company account, and a written rule about what may be put into it — and let the rule come before the ban. The method behind the survey and the numbers is what mapping shadow AI goes into in more detail.
5. Handing customer service over to the machine entirely
In May 2025 a large fintech company publicly reversed its AI-first customer service strategy and started hiring human agents back.
But you have to be careful with the generalisation here, because a series of randomised experiments shows the opposite as well: on a global online marketplace, the pre-sales chatbot lifted sales by 16.3% — the largest measured effect among the seven applications tested.
Countermove. Draw the line in advance: in the pre-sales stage the chatbot demonstrably works, in complaint handling it does not. Getting through to a human should be one click, not a maze.
6. Automating HR and recruitment without a bias review
In the first AI-related case of the American equal opportunity authority, a company's application-handling software automatically rejected female applicants over 55 and male applicants over 60; the case closed in a settlement in August 2023. In another American lawsuit, still in progress, the question is whether the vendor of the screening software can be liable for discrimination — there is no judgment in that one yet.
Countermove. If the screening is automated, it is the output that has to be measured, broken down demographically, rather than the software's promise believed. And every rejection needs a route to human review.
7. The pilot trap: the trial never becomes operations
According to the large executive survey, 44% of respondents say AI is scaling at enterprise level in their organisation (a year ago it was 38%) — while scaling agents stalled at 22% in smaller organisations, at the same time as it grew from 27% to 40% in large enterprises.
Countermove. The pilot needs an exit condition written down in advance, with a date: what number takes it live, and what number stops it. This has to be written before the pilot starts, not at the end.
8. The feeling gets measured instead of the output
This is the best documented mistake on the list. In a randomised experiment in 2025, experienced developers felt a 20% speed-up while the measurement showed they had become 19% slower — by the research group's own summary, people overestimated the effect of AI on their own time by an average of 40 percentage points.
On the corporate side the same gap looks like this: in the executive survey, 80% feel an improvement in individual productivity, 37% see any effect on results, and 6% a significant one.
Countermove. No number that somebody estimated should appear in the weekly report. Only numbers that can be read out of a system. The whole measurement scheme is what measuring AI ROI sets out in full.
9. There is no training, and the rule is not written down either
According to the EU statistical office, the biggest obstacle to AI for European companies is the lack of expertise (7.8%), not the cost (4.2%) — while in 2024 only 22.3% of EU companies with more than 10 employees provided any kind of digital skills training to their staff.
And since August 2026 this is no longer only a business question: the AI literacy obligation of the EU AI regulation has been supervised and enforceable since then.
Countermove. Four sessions, a written rule, an attendance sheet. What the legislation actually demands — and what it does not — is what the team's AI literacy plan walks through.
10. The cost side is only counted up to the subscription
An industry survey of nearly 5,000 people: AI adoption remains negatively associated with the stability of software delivery, and without strong controls the increased volume of change leads to instability. And the study on Danish payroll data shows that the effect did not appear in pay and hours even at workplaces that invested substantially.
Countermove. The calculation has to include review time, correction time and rollout hours. And the most important rule: an hour saved is only a saving if somebody genuinely spends it on something else, or if it means you do not have to hire somebody.
Bonus: "the AI will decide"
In an experiment with 2,234 participants, those working with AI delegated 17% more and did 62% less direct editing — and their output became measurably more homogeneous. The authors call this a "collapse in diversity".
Countermove. Where being distinguishable is business value — brand voice, a proposal, a negotiation — the amount of delegation has to be capped separately. Not every step is a good candidate for the machine; there is a one-minute method for working out which ones are: can AI do this task?
What to take away
Seven of the ten mistakes are not technological: baseline, process, checking, rules, training, exit condition, cost calculation. The decisive part of an AI rollout is organisational work, and that is the good news — you do not need to call in an expert for any of it.
The three-month schedule that heads off all ten is here: in AI strategy for a small company, broken down week by week. How all of this looks in real, documented cases is covered by what the real rollouts show, five companies at a time.