Agent governance and the EU AI Act: what changes for work platforms
As AI agents move from suggesting to acting inside your work platform, 'who approved this?' stops being a nice-to-have and becomes a compliance question. A practical look at why propose-and-approve, immutable audit trails, and per-workspace cost control are becoming table stakes.
By Dmitrii SelikhovFounder
Key takeaways
- As work-platform agents shift from suggesting to acting — assigning issues, changing status, archiving in bulk — the governing question becomes 'who approved this action, and can we prove it,' which is an audit-and-accountability problem, not a model-quality one.
- The EU AI Act's risk-tiered, transparency-and-human-oversight framing pushes vendors toward designs where consequential automated actions are reviewable and attributable by default, rather than logged after the fact if at all.
- A propose-and-approve broker — where an agent's write actions become proposals a human accepts or rejects, each recorded in the same immutable audit trail as human actions — turns 'governed autonomy' from a slide into an enforceable runtime property.
- Per-workspace AI cost ledgers matter for governance too: unbounded agent spend is its own risk, and a transparent, metered budget makes autonomy predictable instead of a surprise invoice.
For most of the last two years, 'AI governance' in work software meant a setting that turned the assistant on or off. That framing is now obsolete. The 2026 generation of agents doesn't just draft text — it acts: it assigns issues, changes statuses, archives in bulk, routes incoming requests, and runs on a schedule while you sleep. The moment an agent can change your data, the governing question stops being 'is the output good?' and becomes 'who approved this action, and can we prove who did?'
Why the regulation points the same direction
The EU AI Act codifies an intuition that good engineering already shared: higher-stakes automated decisions need transparency, human oversight, and traceability proportional to their risk. You don't have to be classified as a high-risk system to feel the gravitational pull — the entire regulatory direction of travel rewards designs where a consequential automated action is reviewable and attributable by default, and penalizes black-box automation that acts without a record.
For a work platform, that translates into concrete properties: a human can sit in the loop on actions that matter, every automated action is attributed to a principal, and the record is immutable. Bolting those onto an agent layer after the fact is hard. Designing for them from the first table is the difference.
Propose-and-approve, as a runtime property
The cleanest way to make this enforceable is a propose-and-approve broker. Read-only agent actions execute inline; any write — or any destructive action like bulk archive — becomes a proposal that a human accepts or rejects, and both the proposal and the decision are written to the same audit trail as human actions, under the same row-level security. 'Governed autonomy' stops being a marketing phrase and becomes something you can grep: every consequential thing an agent did, who approved it, and when.
This is the model Planoda is built on. Destructive agent tools route through the broker; auto-approved operations still record an audit row; and the MCP path that lets external agents (Claude Code, Cursor, ChatGPT) operate the workspace enforces the same gate. The agent is a teammate with a badge, not an anonymous script with database access.
Don't forget the budget
There's a second, quieter governance axis that the cost-conscious will recognize immediately: spend. An agent that can act autonomously can also burn money autonomously. A per-workspace AI cost ledger — a transparent, metered budget with a preflight check before each run and a recorded spend after — turns autonomy from an unpredictable invoice into a controllable line item. Governance isn't only about safety; it's about predictability. The platforms that win the next few years will be the ones where letting an agent loose is a decision you can make calmly, because the guardrails, the audit trail, and the budget are already there.