What AI Control checks

Every morning AI Control runs 15 checks over your company's data. Each one looks for its own problem using a precise rule. Below is exactly what is calculated and when a card appears. An administrator turns off unneeded checks in the briefing settings, in the What to check section.

None of the audit-based checks count mystery shopper audits: they have their own methodology. Duplicate copies of the same audit are not counted either — the same employee, location and checklist within 3 minutes. “Not applicable” answers count as neither compliance nor a violation.

Urgent

• Failed audits. An audit from the last 36 hours with a score below what's usual for your company — below the score that 20% of audits over 60 days fell under, and below 70% (if there were fewer than 20 audits in 60 days, simply below 70%). Or an important requirement was violated in the audit. The card appears only if there is no task for the audit yet. A task for the manager due tomorrow is suggested.
• HACCP risks. The site has a critical deviation not fixed on time, or 2 or more log entries were missed in a day. Works for companies that have finished setting up HACCP.

An important requirement is a question whose weight is at least twice the average (median) weight of the questions in that checklist. If all questions in a checklist have the same weight, it has no important requirements.

Quality at locations

• Met only some of the time. For a requirement at a location, the last 8 answers over 30 days are taken. A card appears if there are at least 5 answers, at least 2 “yes” and 2 “no” among them, the answer changed 3 or more times, and one of the last three is a violation. It looks like ✗✓✗✓✗: each audit on its own looks normal, but together they show that things are tidied up for the audit. If the last three answers are “yes”, the location seems to have improved, and there is no card.
• Violations that keep recurring. A requirement was violated in 3 of the last 4 audits over 45 days, including the latest one. If a task for this requirement at the location was already closed in the last 90 days, the card says the violation has come back. Requirements that fall under “met only some of the time” are not repeated here.
• Score dropping. The location's average score on a checklist over the last 2 weeks is 7 or more points lower than in the 4 weeks before. At least 3 audits are needed in each of the two periods.
• Audits stopped. The location used to be audited regularly — 6 or more audits in 4 weeks — but there have been no audits in the last 7 days, and the pause is three times longer than the usual interval. If the whole company has gone quiet, there will be one card for the company.

Each location gets one card of each kind, with the most telling requirements.

Audit reliability

These cards are visible only to administrators. The AI never acts on them on its own.

• Repeated photos. The same file from a phone is attached to audits on at least three different days within 30 days. Only unique file names that the phone gives each shot are used. Short names like 4456.jpg are ignored: they can match by chance. A repeat doesn't always mean fraud — compare the photos before drawing conclusions.
• Scores too high. Over 90 days an auditor scores on average 10 or more points higher than colleagues. The comparison is made only at the same locations and on the same checklists that both they and others audited — otherwise they could simply be visiting strong locations more often. At least 6 of their audits and 6 audits by colleagues are needed. A control audit by another auditor is suggested.
• Mystery shopper vs internal audits. Over 90 days the mystery shopper rates the location 15 or more points lower than the location's internal audits. At least 1 accepted visit and 3 internal audits are needed.

Discipline

• Plan not met. An auditor has 2 or more missed scheduled audits in 2 weeks. Statuses are the same as in the planner: an audit is missed if the tolerance window closed and there was no audit.
• Stuck tasks. A location has 3 or more open tasks more than 2 days overdue, or a task was returned for rework 2 or more times. Tasks with a deadline more than 60 days ago are not counted.

People and training

Cards in this group are visible to administrators.

• Knowledge gaps. A checklist section has been violated at 3 or more locations over 30 days, and the share of “no” answers for the section is 25% or more (at least 20 answers are needed). If the same thing isn't done right in different places, what's needed is training, not a task for one location. Creating a test or instruction on the topic with the AI assistant is suggested.
• Tests not passed. For a published test, fewer than 60% of completed attempts over 60 days were passed (from 5 attempts) — the card names the weakest topic. Or 3 or more employees are overdue on taking it.
• Instructions not read. An instruction with mandatory acknowledgement was published more than 7 days ago, and fewer than 70% of the employees it's assigned to have confirmed it (and at least 3 haven't).

Good news

• Score improvement. A location raised its average score on a checklist by 10 or more points (2 weeks versus the 4 weeks before, from 3 audits in each period) and is holding at 85% or higher. There are no more than two such cards.

Frequently asked questions