Two numbers frame this post. In Counselor's 2026 State of the Industry, 85% of distributors express concern about cybersecurity and ransomware attacks, more than worry about any other risk. In PPAI's research on AI adoption, data privacy is a top-four barrier for both distributors (27%) and suppliers (31%).
Both worries are rational, and they collide the moment an AI agent asks for a login to your ERP. Here is what to ask before you hand one over. These are the questions we would want answered if we were buying.
1. Where does my data live, and is it ever shared?
Ask for the physical answer: which cloud, which region, which subprocessors. Then ask the structural one: is my company's environment separate from every other customer's, or is it a row in a shared database with a filter in front of it?
The answer you want is a separate environment per company, enforced in the infrastructure and the database, not only in application code. If the vendor cannot explain how one client's data is prevented from ever appearing in another client's run, keep asking.
2. Is my data used to train a model?
This is the question PPAI's 27% are really asking. The answer should be an unconditional no, backed by the vendor's contract with its model provider. Client information should reach a model only as the working context of a run, under commercial terms that exclude training, and never be used to improve another customer's agents.
3. Where are my credentials, and who can see them?
An agent that signs into your systems has your passwords. Ask where they are held, whether they are ever stored in plaintext, and whether one user's session can be reused by another. The right answer is a secret store, referenced by pointer, decrypted only inside the environment that uses it, with per-person sessions.
4. Can I see everything it did?
Ransomware is one failure mode. A quieter one is an agent doing something wrong and nobody knowing. Ask whether every run is recorded: each step, a screenshot of every screen touched, every action taken, and who approved what. Ask whether that record can be edited afterwards. It should not be.
5. What is your compliance status, honestly?
commonsku notes that its AI features are SOC 2 compliant, and that is the right bar for the industry. Ask any vendor where they stand, who the audit partner is, and whether they will complete your security questionnaire. Be wary of a vendor that will not say "in progress" when it is in progress.
6. Who carries the liability?
If an incident originates in the platform, who responds, who notifies, and who pays? A vendor that expects you to indemnify your customers for a system you do not control has answered the question.
The right AI vendor does not ask you to trust it. It shows you the architecture and puts the obligations in the contract.
How AltOps answers these
We publish our answers on the AltOps security page: an isolated environment per company and a private execution pod per person, AES-256 encryption at rest, credentials held in a secret store and decrypted only in the pod that uses them, US infrastructure, no training on client data, every run recorded end to end, and a SOC 2 program in progress with Thoropass as audit partner. Bring your security team to the demo. We would rather answer the six questions before you sign than in a questionnaire after.
Sources: ASI, "Counselor State of the Industry 2026: Adopting an AI Strategy," July 2026; PPAI Research, "Promo's AI Challenges and Opportunities"; commonsku, "How Promo Distributors Are Using AI to Sell More in 2026."
Written by Madhavam Shahi, teaching agents to run the back office at AltOps.
