What should I ask a software vendor before letting their AI assistant read or write into my CRM?
The short answer
Ask five things: what data it reads versus writes, whether it can undo its own changes, where logs live, whether your data trains their model, and what happens on cancellation. Insist on a sandbox test with 50 real records before granting production write access to any AI agent touching HubSpot, Salesforce, or Pipedrive.
A CRM is the one system of record a founder-led company actually has. Letting a third-party AI agent write into it — update deal stages, log calls, merge contacts, send emails on your behalf — is a bigger decision than most vendors make it feel. The sales pitch is speed. The real question is what happens when the agent is wrong, and who notices.
Read access versus write access are different conversations
Most vendors bundle these together in a demo, but they carry very different risk. Read-only access (the agent summarizes pipeline, flags stale deals, drafts a forecast) can't break anything; worst case it gives bad advice. Write access (the agent updates fields, changes owners, logs activities, sends outbound) can corrupt your data at scale in minutes, and most CRMs don't version-control field changes the way code repos version-control commits. Ask the vendor to name, field by field, what their agent writes to — not "CRM data" as a category, but the actual object list: Deal.Stage, Contact.Email, Activity.Log, Lead.Owner. If they can't produce that list on request, they haven't audited their own agent's blast radius.
The five questions that matter
Ask whether the agent's writes are reversible — specifically, is there a change log keyed to each write that a human can roll back without restoring from a full backup? Ask where inference logs and prompt history live, and for how long, since that's often where PII sits unprotected longer than the CRM itself retains it. Ask directly whether your account's data is used to train or fine-tune models shared across other customers — this is the single most common point where vendors get vague, and vague is a no. Ask what the offboarding process looks like: does the agent's access token get revoked instantly, and is there a data deletion certificate on contract termination. And ask for a sandbox or staging environment where the agent runs against a copy of your CRM, not production, for at least two weeks before go-live.
| Question | Why it matters | Acceptable answer |
|---|---|---|
| What specific fields can the agent write to? | Scopes the blast radius of a bad decision | A named list, not "your CRM data" |
| Are writes logged and reversible? | Determines recovery time after an error | Field-level change log with one-click rollback |
| Is my data used for model training? | Determines if your pipeline data leaks into a shared model | Explicit opt-out, in writing, in the contract |
| What's the sandbox testing process? | Lets you catch errors before production damage | A staging CRM copy, minimum 2 weeks |
| What happens to my data on cancellation? | Prevents data lock-in or residual exposure | Deletion certificate within a stated window, e.g. 30 days |
Run the sandbox test before you sign anything
Vendors will often skip the sandbox step if you let them, because it slows their sales cycle. Don't let them. Export 50 real (or realistically anonymized) records, load them into a staging instance, and watch what the agent actually does over a week of normal use — not a curated demo. Specifically check three failure modes: does it duplicate contacts instead of merging them, does it silently overwrite a field a human had manually corrected, and does it send anything externally (an email, an SMS) without a human approval step. If the vendor resists giving you a sandbox at all, that alone is the clearest signal — a mature product should treat write access to your system of record as something customers are allowed to be cautious about, not something they're expected to trust on faith.
Where SalesARC fits
SalesARC's own AI products — the ArcClaw™ prospecting agents and the AI voice agent — are built to populate and act inside a CRM, so we get asked these same questions from the other side of the table. We recommend sandboxing any agent, including ours, before giving it production write access, and we support field-level scoping rather than blanket CRM permissions. If you're evaluating CRM setup and governance before adding AI on top, SalesARC Playbook covers the process side first.
Related questions
- How do I vet an AI-powered lead generation vendor to make sure they're delivering qualified prospects and not just a scraped list?
- How do I know if my CRM is actually being used by my team or just sitting there unused?
- What is an AI SDR, and do they actually work?
- Is it better to buy one connected sales platform or piece together separate CRM, prospecting, and quoting tools?