What people inside the system fear
The first fear: a machine error will become an accusation. The employee imagines that an algorithm has flagged his purchase and that he must prove his innocence himself, with no explanation of why the system decided so.
The second fear: oversight will turn into surveillance. If it is unclear who sees the signals and what is done with them, any flag looks like a file on a person, not a reason to check a document.
The third fear: the people who carried out the work will be hit, not those who made the decisions. A procurement officer sees that the system looks at his signature, and does not see whether it looks at the person who approved the terms.
The fourth fear is quieter than the rest: AI will show how many errors there have always been. A manager fears not the machine but the number in the first report, which he will be asked to explain.
The fifth fear concerns data: information on staff and their relatives will fall into the wrong hands. A simple rule removes it: the system gets only the data needed for the chosen scheme and runs on the organization’s own servers.
All five fears are removed not by promises but by rules. The Open Contracting Partnership, which described 73 red flags for all stages of procurement, says it directly: a flag means a risk, not proof of a violation.
China: Zero Trust
The Zero Trust system matched big data on public servants in 30 districts and cities of China. According to the South China Morning Post, it flagged 8,721 officials.
The experiment met resistance from officials. For the topic of this page, the resistance matters more than the number: even a working system stops if the people inside do not understand its rules.
This gives a practical lesson for any country. Matching data is technically easy. The hard part is to agree in advance what counts as a signal, who sees it, how an employee learns the reason and whether he can dispute it.
So rollout starts not with a model but with rules for handling signals. Without such rules, an accurate system shares the fate of an inaccurate one: it is stopped by those it gets in the way of and by those who simply did not understand it.
Transparency for honest staff
For an honest employee, checking everything is better than spot checks. When a machine looks at every purchase by the same rules, the check no longer depends on who chose this person and why.
A signal raised by the rules can be closed with a document. If the system flagged a tender with one bidder, the employee attaches a justification, the reviewer removes the flag, and the log keeps a record that the question was settled.
Even oversight also protects against false accusations. When risk signs are counted the same way for everyone, a complaint about a specific person is checked against the same data, not against rumors.
Transparency protects the budget. Money that did not go out under a doubtful contract does not have to be recovered through the courts, and an honest supplier does not have to compete with one who made a deal in advance.
There is a benefit for the manager too: he learns about a problem from his own system and has time to fix it, instead of reading about it in an outside audit report.
The European Commission describes its Arachne+ system the same way: it looks for signs of fraud, conflicts of interest and irregularities with the EU budget using 78 risk indicators and data on company owners and public officials, but it makes no automatic decisions.
Rules before launch: who sees the signals
The rules are written before the first signal, not after the first scandal. The minimum set fits on two pages and is approved by the head of the organization together with the security service and a lawyer.
1. What counts as a signal: a list of risk indicators with thresholds, for example the share of one supplier at a buyer or a price rise after the contract is signed.
2. Who sees the signals: a named group of reviewers. The head of the unit whose purchase is flagged does not get the list of flags on his staff before the reviewer decides.
3. What the reviewer does: requests documents, checks the grounds and closes the signal or passes the materials on. Every action is logged.
4. What the employee knows: which sign was triggered and which documents settle the question. A flag without an explanation is not allowed.
5. What the system does not do: it does not punish, does not publish names and does not pass data beyond the approved group. Arachne+ states this principle directly: there are no automatic decisions, people decide.
6. How long a flag is kept: a closed signal does not stay on an employee for years. The retention period and the deletion procedure go into the same rules.
How not to turn a flag into an accusation is covered in detail on the page on risks and errors.
How to show value in 90 days
Trust appears when people see the result on their own data. So it is worth spending the first quarter on one scheme of violations, not on the whole system at once.
First month: choose one scheme and a few signs from the 73 Open Contracting flags, run last year’s purchases through them and compare the flags with what reviewers already know.
Second month: turn on daily checks of new purchases and review each signal by the rules. False flags are counted too: their share shows where the rules need to be refined.
Third month: a report to the head in numbers. How many purchases were checked, how many signals reviewed, how many closed with documents, how many led to an inquiry, and how much time each one took.
Such a report convinces more than a presentation: it shows that honest staff closed their flags in minutes, and doubtful cases were found before payment. A step-by-step plan is described on the page on how to introduce AI oversight.
Trust of citizens and business
Citizens and business do not need the system itself but confidence that tenders are fair. The Open Contracting Partnership builds its 73 flags on open procurement data: anyone who can read tables can check them.
For business, transparent oversight means clear rules. A supplier knows which signs raise questions and can attach a justification of the price or terms in advance.
For citizens it matters that a machine does not decide their fate. A person reviews a risk signal, and the grounds of a flag can be checked against documents.
Open rules matter as much as open data. If the list of risk indicators is published, a supplier sees that the check looks at signs in procurement data and does not pick inconvenient bidders.
The more data is open, the less suspicion that oversight is selective. The registers such checks run on are described on the page on open data.