Most teams assume a California risk assessment is something you write about a product that is live. Section 7150(b)(6) of the California Privacy Protection Agency (CPPA) regulations moves the trigger earlier. If you process consumers' personal information that you intend to use to train an automated decisionmaking technology (ADMT) for a significant decision, you need the assessment before the training run, not after the launch.
The part that surprises engineers is what the trigger depends on. It is not the architecture, the model size or whether the data is "sensitive." It is what you intend the trained model to be used for, and "intend" is defined broadly. This post reads sections 7150, 7153, 7155 and the definition of "train" in section 7001(fff) as an engineering lead has to read them: which training work is in, which is out, and what to record at the dataset level so the answer is not a reconstruction later. It is a builder's reading, not legal advice. Counsel decides what is covered.
What does section 7150(b)(6) actually say?
The regulations, effective 1 January 2026, list the processing that presents significant risk to consumers' privacy in section 7150(b). Paragraph (6) covers "processing the personal information of consumers, which the business intends to use to train an ADMT for a significant decision concerning a consumer," and also training a facial-recognition, emotion-recognition or other technology that verifies identity or does physical or biological identification or profiling of a consumer.
The ADMT half only bites for the five significant decision categories in section 7001(ddd): financial or lending services, housing, education, employment or independent contracting, and healthcare. Section 7001(ddd)(6) says advertising is not a significant decision. So a model trained to rank loan applicants is in scope. A model trained to pick which banner to show is not, under this paragraph, whatever else governs it.
Section 7150(a) then says the business "must conduct a risk assessment before initiating that processing," and section 7155(a)(1) repeats the timing: conduct and document it before initiating any processing activity on the section 7150(b) list. For training, the processing starts when the data goes into the run.
What counts as training?
Section 7001(fff) defines "train" for the purposes of sections 7150(b)(6) and 7153 as the process through which a technology discovers underlying patterns, learns a series of actions, or is taught to generate a desired output. The examples are adjusting the parameters of an algorithm used for ADMT, improving the algorithm that determines how a machine-learning model learns, and iterating the datasets fed into ADMT.
Read that against your pipeline, and my reading is as follows. Pretraining from scratch is plainly in. Fine-tuning that adjusts weights on your customers' records is also in, because it adjusts parameters. Retraining on a refreshed dataset is covered by "iterating the datasets." The grey zone is retrieval: feeding documents to a model at inference time does not adjust parameters, but a team that keeps swapping the dataset behind a screening tool should ask counsel whether that is "iterating the datasets fed into ADMT." I would not stake a position on that one without a lawyer's view.
Why "intends to use" is wider than it sounds
The regulation defines "intends to use" in section 7150(b)(6): the business is using, plans to use, permits others to use, plans to permit others to use, is advertising or marketing the use of, or plans to advertise or market the use of the trained technology for a significant decision.
Two clauses in that sentence carry most of the risk for a software vendor. "Permits others to use" means you can be in scope without ever making a decision yourself, because a customer uses your model to make one. "Advertising or marketing the use" means a landing page that says your model screens candidates is evidence of intent, even if the training data was collected for something else. A team that trained on support tickets and then added "talent screening" to the sales deck has changed the answer without changing a line of code.
That is also why the dataset needs a recorded purpose. The regulation's own wording makes intent the test, so intent is what an examiner will ask you to show, and it should exist in writing from before the run.
What section 7153 adds for vendors. Section 7153 is short and aimed at software vendors. Section 7153(a) says a business that makes ADMT available to another business to make a significant decision must give that recipient business all facts available to it that are necessary for the recipient to conduct its own risk assessment. Section 7153(b) limits the section to ADMT trained using personal information.
So the vendor has two jobs. It has to assess its own training processing, and it has to hand customers enough about the model for the customers' assessments, which under section 7152(a)(3)(G) must describe the logic, assumptions and limitations of the ADMT and its output. A customer that cannot get training-data facts from you cannot complete its own document, and it will ask. The vendor contract terms under the ADMT rules are where that obligation should show up in writing.
A worked example: one dataset, two intended uses
Picture a 24-person Series A company that sells scheduling software to home-care agencies. Its team fine-tunes a model on three years of shift records and caregiver notes to predict no-shows. Nobody involved thinks of it as an "AI hiring tool."
In the first version, the model only flags a likely no-show so a coordinator can call ahead. That is operational forecasting. It does not decide a significant decision about a caregiver, and section 7150(b)(6) does not attach, though other triggers could, such as the automated inference from systematic observation of employees in section 7150(b)(4).
In the second version, a product manager proposes using the same model to decide which caregivers get offered extra shifts. Allocation of work is in the employment category of the significant decision definition (see the ADMT definition post for how wide that list reads). Now the model is an ADMT for a significant decision, the records are personal information, and if the company intends to use or permit use of it that way, the assessment is due before the next training run. Same dataset, same weights, different intended use. The assessment trigger moved because a sentence in a roadmap changed.
If the model is already running, the timing rule in section 7155(b) says processing initiated before 1 January 2026 that continues must be assessed and documented no later than 31 December 2027. Training work started after that date has no grace period.
What to record so the answer is not a reconstruction
None of this needs a platform. It needs a short record per training dataset and run, kept somewhere that survives staff turnover.
- 1.Intended use, in a sentence. What decision, about whom, will the trained model feed? Include uses you plan to permit customers to make, and any marketing copy that names the use.
- 2.Significant-decision flag. Does that use fall in one of the five section 7001(ddd) categories? A yes starts the assessment; a no is recorded with the reasoning, because the no is the thing you will be asked to defend.
- 3.Data provenance. Which fields, which consumers, which sources. Section 7152(a)(2) and (a)(3) ask the assessment to name the categories of personal information and how it was collected, so the training manifest is the first draft of those sections.
- 4.A gate before the run. A required field on the training job or the data-access request that blocks a run on personal information until an assessment ID or a recorded "not in scope" decision exists.
- 5.A trigger for intent changes. A roadmap or sales-copy change that adds a significant-decision use to an existing model reopens item 2. The 45-day update rule in section 7155(a)(3) applies when a change is material, and a changed purpose is one of the regulation's own examples.
Teams with unregistered tools face a sharper version of this problem, because an assessment cannot cover a training run nobody knew about. That is the shadow AI gap, and the dataset record is how you close it for training.
When this is the wrong answer
If you train only models that do not feed a significant decision, such as recommendation, ranking of content, churn or ad targeting, and you do not touch the identification technologies in section 7150(b)(6), this paragraph does not apply and you should not build a gate for it. Say so in a one-page decision record and move on. The same goes if you only call a vendor's model and never train on your own customers' data. Your obligations then sit elsewhere, such as the ADMT notice and opt-out rights that begin on 1 January 2027, and our ADMT checker is the quicker first read on whether those apply.
Outside help starts to pay when you have several training pipelines, customer data from multiple products, or customers asking for the section 7153 facts and no one owns the answer. That is the registry, the pre-run gate and the evidence trail, which is what our ADMT compliance engineering work covers. If that is your situation, tell us what the model decides and we will tell you what we would instrument first. For what the finished document has to contain, see what actually goes in a CCPA risk assessment.
Sources
- California Privacy Protection Agency: CCPA regulations, Title 11 Division 6 Chapter 1 (effective 1 January 2026), sections 7001(ddd) and (fff), 7150, 7152, 7153 and 7155, read in full text on 19 October 2026
- California Privacy Protection Agency: CCPA updates, confirming the effective date
