Will AI Accounts Receivable Software Use My Customer and Financial Data to Train Models?

It depends on the vendor, and the answer often sits in details buyers don’t think to ask about, such as how long a model provider keeps the prompts it receives. Trace where your data goes, then get the answer written into the contract.

AI in FinanceBy George Kurdin8 min read

It depends on the vendor, and the answer often sits in details buyers don’t think to ask about, such as how long a model provider keeps the prompts it receives. Trace where your data goes, then get the answer written into the contract.

Monk’s security page states that customer data never trains Monk’s systems or any model provider’s models. The checks below apply to any vendor making a similar promise, Monk included.

What counts as training on your data?

Vendors use the word “training” loosely, so pin down which of these they mean.

Term What happens to your data What to ask
Training or fine-tuning Your invoices or customer emails change the weights of a model Which model, and who else uses it?
Retention by a model provider The provider stores prompts and outputs after the request ends How long, and can the provider use them to train its models?
Inference only Data passes through a model to complete one task and isn’t kept Is that true for every provider the vendor uses?
Logs and telemetry Copies of prompts or outputs are stored for debugging Who can see them, and are they ever used for training?

Promises tend to slip on retention and logs. A vendor can truthfully say it doesn’t train models while a provider it calls keeps your prompts for 30 days.

How do AR vendors handle customer data differently?

Most platforms fall into one of four setups:

  • Pooled training. Payment histories and customer details from every client feed a shared model. Your patterns can shape what other customers see.
  • A model per customer. The vendor trains or tunes a model on your data alone, and that model serves only your account.
  • Inference with no training. Data goes to a model for each task, under provider terms that forbid keeping it or training on it.
  • Hosting in the customer’s cloud. The platform runs inside your infrastructure, so data stays in an environment you control.

Pooling is the setup to rule out first. Our post on whether your AR platform should share your data with other customers covers the risks of pooled models in more detail.

What evidence should you ask a vendor for?

A verbal “we don’t train on your data” isn’t enough. During the security review, ask for:

  • a data-flow diagram showing every point where AR data reaches a model provider
  • the list of model providers, with the retention terms that apply to the vendor’s account
  • a written statement covering whether any customer data, including logs, is used for training
  • the current subprocessor list
  • the latest SOC 2 Type II report, plus an ISO/IEC 27001 certificate if the vendor holds one

A short test during the pilot checks those documents against what the platform does. With a handful of sanitized invoice threads loaded into a pilot account, ask the vendor for the log of outbound model calls covering that window. It should list only the providers named in the diagram. Each of those providers should also confirm in writing, through the vendor, that the vendor’s account is set not to retain inputs or train on them.

The same questions apply to AI features inside tools your AR team already uses, such as email add-ons or ERP assistants, since those can send customer data to a model too.

Watch for a mediating layer between your data and the model. A platform that decides which fields each model receives for a task exposes far less than one that sends whole records.

What should the contract say?

Bring sample language to your counsel rather than relying on the vendor’s standard terms. The clauses that matter most are:

A training prohibition. The vendor and its subprocessors may not use your data, or anything derived from it, to train or fine-tune any model that serves anyone other than you. Naming each prohibited activity matters, since “improve” alone is open to interpretation.

Model provider terms. Every provider that processes your data must be configured not to retain inputs or use them for training, and the vendor must confirm this in writing before production data flows.

Deletion and proof. On request or at termination, the vendor deletes or returns your data within an agreed period and gives you a deletion certificate. Ask how backups are handled, since they’re often deleted on a separate schedule.

Subprocessor notice and audit rights. The vendor gives you notice before adding a subprocessor and lets you audit data handling for cause.

An indemnity that names training. A general indemnity for breach of contract may not clearly cover a dispute about model training, so ask for one that refers to it directly.

Read the license clause in the vendor’s general terms alongside the DPA. Wording that lets the vendor use your content to “improve the Services” should be checked against any training prohibition, since the broader clause can undercut the narrower one.

Which vendor answers should worry you?

Escalate to legal and security if you hear any of these:

  • “We use anonymized data to improve our models,” with no explanation of how the data is de-identified
  • “We keep prompts and outputs for quality,” with no retention limit
  • the vendor can’t name its model providers
  • the vendor refuses to put a training prohibition in the DPA
  • data sharing is switched on by default, and you have to opt out

Any of these is a reason to pause the evaluation until the vendor answers in writing.

What should you check after signing?

Contract terms only help if someone checks them. Confirm the model provider settings before live data flows, and ask for subprocessor change notices as they happen. Agent actions should be logged in a form you can export. Once a year, review the subprocessor list alongside the latest SOC 2 report, and ask for a fresh retention confirmation whenever the vendor switches model providers.

Customer-facing messages need separate controls. Our post on whether AI can email your customers safely covers approval gates and audit trails for outbound email.

How does Monk handle your data?

Monk sits between your systems and the frontier models it uses, controlling what each model receives for a given task, so no model sees your full records. Monk doesn’t train on that data, and neither does any model provider it uses.

Data is encrypted and scoped to your organization in Monk’s AWS environment, with SSO, MFA and role-based access. Every agent action is logged, and customer data is never used to train models. The current subprocessor list is available to customers on request.

Monk is SOC 2 Type II and ISO/IEC 27001:2022 compliant. The SOC 2 report and penetration test results can be requested at trust.monk.com.

Where should you start?

Send every vendor on your shortlist two questions in writing. Is any of our data, including logs, used to train a model that serves anyone other than us? Which model providers process our data, and what retention settings apply to our account with each?

A vendor that answers both plainly, and agrees to put the answers in the DPA, has passed the first test. For the rest of a vendor evaluation, see our list of questions to ask an AR automation vendor. To see how Monk handles your data, book a demo.

Frequently Asked Questions

Does sending data to a model count as training?

No. Inference runs a model on your data without changing the model. The risk comes from what happens afterward, if the provider keeps your inputs or the vendor stores them in logs that later feed a model.

Can a vendor train a model on my data that only my company uses?

Some do. That’s a different arrangement from pooled training, and it should still be written into the contract so the model can’t be reused for other customers.

Does Monk use outside model providers?

Yes. Provider and routing details can be requested during a security review. None of those providers train on your data.

Is a SOC 2 report enough?

No. A SOC 2 report shows that security controls exist and were tested, but it doesn’t say whether your data is used for training. Treat it as a baseline alongside the contract terms.

What deletion terms are reasonable?

Agree on a fixed period for live data and a separate one for backups, and ask for a deletion certificate. Keep any records your accountants must retain out of the deletion request.

When should we require hosting in our cloud?

When data residency rules apply, or when your security policy doesn’t allow customer data to leave your environment. Expect more setup work in exchange for the extra control.

How do we check compliance later if we suspect a problem?

Use the audit rights in your DPA to request the model call logs for the period in question, along with the provider retention confirmations. Bring in a third-party auditor if the two don’t line up.