EVERYTHING WE SHIPPED IN OCTOBER,
YOU CAN TAKE BACK.
Five things went live between the 11th and the 15th. They look unrelated on a changelog — a form, a mail, a research column, an assistant, a label. They are the same decision five times.

Most CRM vendors spent this autumn adding AI to their product. The pattern is consistent: the vendor's model, on the vendor's key, with the vendor's terms, billed through the vendor. It is the fastest way to ship an assistant, and it quietly moves one more dependency from your side of the table to theirs.
We shipped five things in October. None of them hands us anything new. Each one puts a key in somebody else's hand — yours, or your customer's — and leaves them able to take it away without asking us.
Your model, your key
The AI that works in your sheets runs on your own access. Anthropic, or any OpenAI-compatible endpoint — OpenAI and Mistral both work, Mistral tested against its EU endpoint. You enter the key; it is yours, and it bills to you.
There is no platform key as a fallback, and that absence is the feature. A fallback would mean that on the day your key expires, your records quietly travel to a provider you did not choose, under terms you never read. Here, an expired key means the AI stops. That is the correct failure.
The profile page shows which provider, which model and which endpoint are in use, next to a checklist for your own account: training off, zero data retention, EU endpoint, Art. 28 and Art. 30 in place. We do not tick those for you. You hold the contract with the model provider, so you are the one who can.
Your assistant, your token
Since 13 October your own assistant can work in Podma — locally through Claude Desktop, or remotely over https://podma.cmp42.com/mcp for a cloud assistant that accepts a personal bearer token. Both get the same tools: list, read, query, create and write sheets; CRM objects listed separately with their record counts; a read-only HubSpot import.
The access is a personal token, per person and per tenant, with an expiry date. Not an integration you install once and forget, not a shared service account that outlives the person who asked for it. When somebody leaves, their token dies with their account, and nothing else changes.
One thing a token cannot do: delete imported data. That stays with a signed-in person in the application. An assistant that can wipe a table because a prompt went sideways is not a productivity gain.
Your customer's consent, and theirs to withdraw
Since 11 October a form on your site writes straight into Podma: one prospect per enquiry, and the consent stored with the exact wording and its version, as a record with a checksum. Not a boolean that says yes. The sentence the person actually agreed to, kept so it can be produced two years later when somebody asks.
Since 12 October every one of those consents gets a confirmation mail, German or English, with a signed link valid for 24 hours — and a withdrawal link in every mail we ever send on that basis. Sent through our own mail gateway, SPF, DKIM and DMARC passing.
Read that as an architecture statement, not a compliance box. The person who gave the consent can end it from any mail, without reaching us, without a form, without a support ticket. We built the exit before we built the list.
And where a claim came from
The AI in the sheet can research on the web and writes the source URL into a Source column — not into a footnote, not into a chat log that scrolls away, into the row next to the claim. Switched on per tenant, against a list of domains you allow, at most five searches per request.
This one has a limit worth stating plainly: research runs only with an Anthropic key, because the search is Anthropic's tool. With Mistral or OpenAI the sheet works and the research does not.
What the next releases bring
The list above runs today. These five are decided, designed and on the way — marked as announced until the day we can say "as announced" and drop the marker:
- Erasure reaches into the audit trail: the deletion stays visible as an event, the deleted content does not come back with it.1
- Retention and legal hold become an enforced refusal — the system declines, instead of a policy document asking someone to.1
- Tenant separation moves down into the database, below the application that already enforces it.1
- Access control per Model, with field masking, so a role sees the rows it should and the columns it should.1
- Destroying the key of an encrypted value becomes a form of erasure in its own right.1
¹ Planned for a future release.
Labelled test data arrived with the release on 15 October: AI-generated rows are written only on an explicit request and carry the label generated test data — not real. CRM objects never accept invented data at all.
The test is revocability
There is a question worth asking any vendor that put an assistant in your CRM this year. Not which model do you use — that changes every quarter. Ask instead: what can I take away from you without calling anyone? The model key. The access token. The customer's permission to be contacted. If the answer is "open a ticket", the dependency is not where you think it is.
That is what the five October releases have in common, and it is why they shipped together rather than as a feature list. Ten of the twelve design partner seats are still open. If you want to put this against your own data, that is the door.