Your Thursday briefing · AI series
Last week we looked at what leaves the company when employees use public AI assistants. In September, the fourth episode gave you ten questions to ask your vendors about their use of AI. This Thursday we turn the table around: your customers are going to ask you those questions.
Our position: on AI, your customers won't settle for a promise, they will ask for evidence. The good news is that if you have followed this series, you have already built most of that evidence. What is left is putting it together.
Why these questions are arriving now
For a long time, AI took up one line in a security questionnaire, when it appeared at all. Three published texts change that for a Canadian vendor.
A standard questionnaire exists. On July 9, 2025, the Cloud Security Alliance published its AI Controls Matrix: 243 control objectives across 18 domains, with a companion questionnaire, the AI-CAIQ, designed for self-assessment and for evaluating vendors. The matrix is mapped to ISO/IEC 42001, the NIST AI 600-1 profile and the EU AI Act. Since November 20, 2025, a scored AI-CAIQ combined with an ISO/IEC 42001 certification earns Level 2 of the STAR for AI program. Status: confirmed by the CSA.
What it changes: your customer's procurement team no longer has to invent its AI questions. It can attach a published questionnaire to its next request for proposal.
Financial institutions will have to assess your models. On September 11, 2025, OSFI published the final version of Guideline E-23 on model risk management, effective May 1, 2027. It covers federally regulated banks, insurers, and trust and loan companies. Its definition of a model explicitly includes AI and machine learning methods, the framework covers models from third-party vendors, and externally developed models must receive their own risk rating. Status: confirmed, published guideline.
What it changes: if you sell to a federally regulated financial institution and your product relies on a model, it will have to rate that model before May 2027. It can only do so with your answers.
In Quebec, automated decisions must be explained. Section 12.1 of Law 25 requires a business that makes a decision based exclusively on the automated processing of personal information to inform the person concerned and, on request, to tell them the information used and the reasons and principal factors and parameters behind the decision. Status: law in force.
What it changes: the obligation falls on your customer, but if your system prepares the decision, they can only meet it with your help. Before signing, they will ask whether you can provide those factors.
The eight questions to prepare
They mirror the ten questions from our fourth episode, from the other side. They apply both to AI built into your product and to AI your people use with your customers' data.
Use and data
- Which features of your product or service rely on an AI model, and which are enabled by default?
- Do your employees use AI assistants with our data, and under what rules?
- Is our data used to train or improve a model, and how long does the model provider keep it?
Supply chain
- Which model providers process our data, in which region, and will you tell us before you switch?
Controls
- How do you test your AI features against prompt injection and data leakage, and do you log the model's inputs and outputs?
- Do AI agents or automations have access to our systems or data, and with what permissions?
Governance and decisions
- If your system makes or prepares a decision about a person, can you give us its principal factors in time for us to meet our obligations?
- Who answers for AI in your company, and how do you know your rules are followed?
Question 2 often surprises vendors whose product has no AI in it. Yet it applies to all of them: your customer wants to know whether their contract or customer list ended up in a chatbot opened with a personal account.
Three levels of answer: promise, description, evidence
Take question 3. Three possible answers, all three fictional.
The promise. "We take the protection of your data very seriously." It doesn't answer the question. An experienced analyst reads it as a "no".
The description. "Your data is not used to train our models." Better, but there is nothing to verify, and it says nothing about retention.
The evidence. "Your data is not used for training and is kept for 30 days by our model provider: section 7 of our enterprise agreement, excerpt attached. Setting checked in the admin console on October 15 by our security lead. Next check: January." A document, a date, a role, a review date.
And when the honest answer is no, the rule from our response kit applies: a "not yet" with a date beats a "yes" without evidence. A "yes" you cannot back up becomes a problem the day of an incident or an audit.
The AI evidence file: what this series has already built
Each episode of this series produced one piece. Put together, they answer the eight questions.
- The AI use register (episode 1), with the four columns added last week: questions 1, 2 and 3.
- The one-page rule with three data classes and the account rule (October 29): question 2.
- The list of AI subprocessors and the contract clauses (episode 4) you required from your own vendors: questions 3 and 4.
- The tests specific to language models (episode 3): question 5.
- The inventory of agents and their permissions (October 15): question 6.
- The AI owner and the acceptable use policy (episode 2), then the indicators and the quarterly page (October 22): question 8.
Question 7 is the only one that often needs new work: for each feature that prepares a decision about a person, document the data used and the factors that weigh most. If no feature of your product does that, write it down, with the date you checked.
On top of these pieces, one page: "How we use AI". What AI does in your product and in your company, the model providers, the commitments on data, the owner, the date of the last review. It goes with every questionnaire, and the appendices follow on request.
The three traps
Answering "no" without checking. "Do you use AI?" Many answer no by reflex, while their email, their CRM or their support tool has switched on an assistant by default. The fourth episode showed it: your own vendors added AI without asking you.
Promising more than your model provider. You cannot guarantee your customer a retention period or a processing region that your model provider's contract does not guarantee you. Read that contract before you answer.
A different answer for every customer. Sales answers one way, support another, and both answers end up in the same file one day. One source, one owner, the same page for everyone.
This week's test
Take thirty minutes and two people: whoever receives your customers' questionnaires (often sales or leadership) and whoever looks after security.
- Pull out the last security questionnaire you received, or failing that, the eight questions above.
- For each AI question, write your answer and name the document that proves it, dated within the last twelve months.
- Ask sales what they answered, in writing, to the last three customers who asked about AI.
Three results should worry you. An answer with no document: it is a promise. A question nobody inside can answer: an owner is missing. Two different answers given to two customers: one of them is wrong, and you don't know which.
What you should have in a month
A dated "How we use AI" page signed by its owner, its appendices kept in one place, every "not yet" paired with a date, and one simple rule: every answer about AI comes from that page. Not a reassuring speech: a file that holds when your customer's analyst asks the next question.
Where FortaRisks fits
To answer on your own, our response kit is free: a self-assessment, a response template for "yes", "partly" and "not yet" with the evidence to attach, and the deeper questions sorted by theme. To keep the evidence current from one customer to the next, Forta Compliance assesses ISO 42001 and the NIST AI RMF among 40+ frameworks mapped to the same control baseline: evidence gathered once serves several requirements, and several questionnaires. And the full roadmap from this series is available as an eight-page guide.
Next Thursday: the AI tools your teams installed themselves. Automation platforms, model servers, test interfaces: the ones that answer the internet, and why nobody put them on the inventory.
Sources: Cloud Security Alliance, AI Controls Matrix (July 9, 2025) · Cloud Security Alliance, STAR program levels, including STAR for AI · OSFI, Guideline E-23, Model Risk Management (2027) · LégisQuébec, Act respecting the protection of personal information in the private sector, section 12.1 · OWASP, LLM01:2025 Prompt Injection