Every company that has ever run AI in a critical workflow has, somewhere, a document called an impact assessment. It was usually written once, by someone with good intentions, and filed away right after. Nobody goes back to it when the system gets a new version, a new data source, or starts deciding something it used to only suggest. Until now, that was a management failure. Soon it will be an administrative violation, with fines calculated on the company's revenue.
The bill that makes Algorithmic Impact Assessments mandatory for high-risk AI systems is in its final review in Congress, after already clearing the Senate. Alongside it, a separate bill requires prior certification and registration for critical systems, with a one-year adaptation period after publication. The ANPD, consolidated as an independent agency this year, has already announced twenty AI inspections for the current two-year cycle, focused on data subject rights, use by children and teenagers, and the conduct of public bodies. The authority that fines for LGPD failures now also fines for AI failures, using the same yardstick: a percentage of revenue.
What the legislative calendar already makes clear
The impact assessment required by the new regulatory framework follows much the same logic as the data protection impact report the LGPD already asks for. The difference is scope. Instead of asking only what the system does with personal data, an algorithmic impact assessment asks whether the system is biased, whether it is explainable, and what effect it has on protected groups. That means proving, with records, not just what the system was supposed to do, but what it actually did over time.
The requirement itself is not the problem. The problem is the habit most companies already have with any compliance document. Impact assessments, risk maps, data inventories. They all start life as a project deliverable, become a filed document, and nobody can say whether they still reflect the system running today. When the inspection arrives, an outdated document does not protect the company. It records, in writing, how long the failure went uncorrected.
Continuous proof, not a project checklist
The AskLisa thesis on Invisible Work applies to AI compliance too. The most expensive work is not what still needs doing. It is what was done once and never revisited, because there is no routine and no clear owner to revisit it. An impact assessment only works as a defense if it is alive, updated every time the system changes its data source, its decision scope, or the area it serves.
That takes two things most companies still treat as secondary. An auditable history of which source fed which answer, at any point in time. And a named owner per area who knows, without having to ask anyone, whether that source is still the current one. Without both, any formal impact assessment goes back to being just another filed document.
The cost of proving later instead of proving all along
Picture a mid-sized legal department that treats its impact assessment as a one-off project, done once a year or at each external audit. If an inspection finds a high-risk AI system running without an up-to-date assessment, the fine set out in the new framework is calculated on revenue, not a fixed symbolic amount. For a large company, that puts the financial risk of an outdated document in the same order of magnitude as a significant labor violation. And the cost of rebuilding that assessment under inspection pressure, with a short deadline and an already overloaded legal team, tends to be higher than the cost of keeping it alive from the start.
The difference between the two companies is not the size of the compliance team. It is whether the infrastructure behind the AI already generates, as a by-product of daily use, the trail a formal impact assessment will demand. At AskLisa, every answer carries the source it came from and the owner of that source in that area. It is not an impact assessment project. It is what is already there, ready, when someone finally asks for one.

