Skip to content
← Back to Blog

Your AMS Turned On AI. Nobody Sent You a Contract.

Of seven association platforms we checked, one publishes what happens to member data when its AI features run. The rest are silent, and their contracts were written before the features shipped.

Somewhere in the last eighteen months your association management system added an AI feature. So did your community platform, your email tool and your content management system. Some of them announced it in a release note. At least one of them turned it on by default.

The question a board member will eventually ask is simple and most associations cannot answer it: where does our member data go now, and who decided that.

Where this sits. This is the fourth of five posts on AI and association websites. Everything below is from vendors' own published terms, data processing addenda, subprocessor lists and support documentation, read directly. Where a vendor publishes nothing, we say so.

The Training Question Has a Boring Answer

Start by retiring the fear that is easiest to sell. The major AI providers all commit in writing that they do not train on customer data by default.

  • OpenAI. Data from the API platform and the business and enterprise products is not used for training unless you explicitly opt in to share it.
  • Anthropic. By default it will not use inputs or outputs from commercial products to train models, and its commercial terms state that Anthropic may not train models on customer content from the services.
  • Microsoft Azure OpenAI. Models are stateless, and prompts and completions are not used to train, retrain or improve the base models.
  • AWS Bedrock. Neither AWS nor the third-party model providers use inputs or outputs to train models, and inputs and outputs are not shared with model providers.

Take those at face value. Those are real commitments in real documents. If somebody is selling you a service on the premise that OpenAI is quietly training on your membership roster, they are selling you a fear the paperwork does not support.

Retention Is the Question Nobody Asks

Not training and not retaining are two different promises, and vendors answer the first when you ask the second.

What the terms actually say. OpenAI states it may securely retain API inputs and outputs for up to 30 days to provide services and identify abuse, removing them after that unless legally required to keep them, and that zero data retention is available on request for eligible endpoints and qualifying use cases. Anthropic states it automatically deletes inputs and outputs on its back end within 30 days, but retains data up to two years where its usage policy was violated, and trust and safety classification scores for up to seven years. If somebody at your organization clicks the thumbs-down button, Anthropic states it stores the entire related conversation for up to five years.

And there is human review in the loop. Microsoft stores prompts and completions for human review of potential abuse, accessible to authorized employees through a gated process, and allows customers who meet Limited Access eligibility criteria to apply to modify abuse monitoring. Note the verb. You apply, and you qualify, or the storage happens.

Why it matters anyway. None of this is scandalous and all of it is normal infrastructure practice. It is simply not what an executive hears when a vendor says we do not use your data to train our models. Member records sitting in an abuse-monitoring store for thirty days, reviewable by a human under defined conditions, is a fact your privacy officer should know before it comes up rather than after.

The Notice Period That Should Bother You

This is the finding we did not expect, and it is in Microsoft's own supplier documentation.

Six months for everything, except AI. Microsoft commits to giving customers at least six months notice before adding a new subprocessor that processes customer data. Except, in its own words, for subprocessors that support artificial intelligence functionality, where it provides at least 30 days notice, with the ability to disable that subprocessor for at least six months after notice.

And a worked example of it. There is a live example. Microsoft documentation states that Anthropic onboarded as a Microsoft subprocessor, and that Anthropic models are enabled on by default for most customers in the commercial cloud, excluding the EU, EFTA and the UK. Opting out is an administrator action in the Microsoft 365 admin center that somebody at your organization has to know exists and go perform.

Read that as an association executive. A new company began processing data on your behalf. The notice period was a fifth of the standard one. The default was on. The opt-out was yours to find. Nobody did anything wrong and nobody at your association was in that loop.

What Association Vendors Publish, Which Is Mostly Nothing

We went through the published terms of the platforms associations actually run. The results are lopsided.

  • Higher Logic. The only association-market vendor we found publishing a real AI data-handling statement. It names its provider, states that subscriber data will not be used to train or fine tune models without permission, that any resulting model would be for that customer alone, and that it requires customers to opt in to each AI-enabled feature individually. Its subprocessor list names OpenAI, Microsoft Azure, IBM Cloud and others with the function of each.
  • With one caveat. Those commitments live in a support article. The data processing addendum itself dates from September 2021 and contains no AI or model-training provisions at all. The good behavior is documented in the place that is easiest to change.
  • iMIS and ASI. Its data processing agreement contains no mention of artificial intelligence, machine learning or model training anywhere.
  • MemberClicks and Personify. The published data processing statement has no subprocessor list and no AI provisions of any kind, and dates from May 2022.
  • Aptify and Momentive. The privacy policy markets AI features prominently and contains no provisions on machine learning, automated decision-making, profiling or subprocessors.
  • Fonteva. No mention of AI, machine learning, automated decision-making, profiling or model training, and no subprocessor list.

One fair qualification. These are absences in public documents, not proof that nothing exists in a master agreement or behind a customer login. But the dates tell the story on their own. The contracts were written in 2021 and 2022. The AI features shipped after. Nobody went back.

The Assumption That Is No Longer Safe

Many association executives carry a rule of thumb that state privacy laws do not apply to nonprofits. That came from California, where the statute is built around a for-profit definition of business, and it is now the minority position.

Six states now cover nonprofits. Colorado, Oregon, Delaware, New Jersey, Maryland and Minnesota all reach nonprofits. Oregon's Department of Justice states the law took effect for nonprofits on July 1, 2025. Delaware's Attorney General states plainly that its act applies to both for-profit and non-profit businesses. Several of those states wrote the same narrow carve-out, exempting only nonprofits established to detect and prevent insurance fraud, which tells you the drafters considered nonprofits and chose to cover them.

And the provisions bite. Colorado gives consumers an opt-out against profiling in furtherance of decisions producing legal or similarly significant effects, and defines profiling broadly enough to reach automated evaluation of personal preferences, interests, reliability and behavior. That is a fair description of an engagement score. It also requires a documented data protection assessment before processing that presents a heightened risk, and Delaware's thresholds drop on January 1, 2027, pulling in associations that are out of scope today.

Four Questions for Every Vendor

These are the ones we send, and the answers are usually more revealing than the marketing:

  • Which AI subprocessor did you add, and when did you tell us? Not the model provider's name. The legal entity that processes data on your behalf, whether it is on your subprocessor list today, and how much notice we get before it changes.
  • How long is our data retained by you and by your model provider, and who can read it? Training and retention are different questions and we want both answered separately.
  • Is each AI feature opt-in, and who at our organization enabled it? A feature that is off until we turn it on is a decision we made. A feature that is on until we find the setting is a decision you made for us.
  • Is any of this in the contract, or only in documentation? If the AI commitments live in a support article and the contract is silent, the commitment can change without a contract amendment.

We Will Read the Contracts You Already Signed

The starting point is an inventory, and it is one somebody at your organization can build in an afternoon: every platform holding member data, which of them has shipped an AI feature, whether it is on, and who turned it on. Most associations have never assembled that list and are surprised by at least one row of it.

Send us the list and the agreements, and we will come back with what each vendor actually commits to in the contract as opposed to the marketing, where the retention and subprocessor terms are, which features are on by default, and which of the four questions above your vendors cannot answer. Where a vendor's terms are genuinely fine we will say so, and for one of the platforms above they are. This is part of the ongoing vendor work in our partnership plans, because these documents get revised without notice and a review dated last year tells you very little.

83 Creative

We're a web development studio that works exclusively with trade associations, professional societies, and membership organizations.

← Previous Article Google Is Switching Off Its AI Tools for Students by Default. AI in Search Stays On. Your Families Need to Know Which Is Which.