AI and agents
Agents propose. People approve.
And the database refuses to let it be otherwise. Scrivino was designed in 2026 for a world where AI does real work on product content — so the governance came first, and the agents inherit it.
AI didn’t arrive late here.
Every PIM you can buy today was designed before large language models existed. Their audit logs were built for people making dozens of edits a day, not agents making thousands — so bolting AI on means either trusting it completely or switching it off.
Scrivino starts from the other end. Every change, human or machine, is a versioned, attributed changeset on an append-only record, enforced by the database itself. So an agent is just another author. It can be granted rights, never more than the person who invoked it. It can propose changes, and structurally cannot approve its own. Every edit it makes is reversible and attributed to the person who authorised it.
You don’t have to trust our AI. You have to be able to check it. That is a much easier thing to build a system around — if you start early enough.
Four rules every agent works under
Never more than its person
An agent acts on behalf of someone, with that person’s rights or fewer. Consent is per agent and revocable within 30 seconds.
Proposes, never applies
An agent’s change arrives as a proposal. A person applies it. An agent cannot apply its own proposal — the system will not let it.
Never writes silently
Every agent action is a versioned changeset with an author: the agent, and the person behind it. Nothing it does is invisible in the history.
Always reversible
Because history is append-only, anything an agent changed can be compared and restored — like anything a person changed.
What runs today, and what is in build
Governed agent access
LiveConnect Claude or any MCP client with per-agent consent, revocable within 30 seconds. Every action is attributed to the person behind it.
Proposals
LiveAn agent proposes a change; a person applies it. Agents cannot apply their own.
Vera, the in-app assistant
LiveAnswers questions about your catalogue within your own permissions. Any change it suggests is a card a person approves. Bring your own model.
AI budgets
Next · Q4 2026Budgets per workspace that stop rather than surprise you, with EU-resident model options.
Agents on lifecycle states
Next · Q4 2026An agent bound to a state either advises — findings and a vote — or proposes changes. Every run keeps a transcript and can be reverted.
Built-in agents
Next · Q1 2027Tags and summaries, alt text, spelling and style, brand-guideline checks, duplicate finding and language detection. Proposed, never silently applied.
Review agents for regulated content
Next · Q1 2027Specialist checkers read a leaflet against your rules and report findings for a person to decide on.
Claims library
On the roadmapApproved claims as records, matched to the assets that use them.
AI actions on assets
On the roadmapBackground removal, crops, subtitles and translation, with a cost preview before each run.
Governed autonomous agents
On the roadmapAgents with their own triggers that move from dry-run to propose to act as they earn it, with the whole run revertible.
Bring your own model, keep your data where it is
Vera, the in-app assistant, runs on the model your workspace chooses: Azure OpenAI, Anthropic, or any OpenAI-compatible endpoint, including a model you host yourself.
A zero-retention setting tells the provider not to keep what it is sent. Budgets per workspace — with EU-resident model options — are in build, so spend stops rather than surprises.
Vera reads and reports within the permissions of the person asking. When it suggests a change, the suggestion is a card a person approves, like any other proposal.
See an agent work under the rules
We are looking for design partners: manufacturers and brands with a real catalogue and a real deadline. You get the system, our time, partner terms, and a say in what we build next.