Governance
Governance: control, traceability, and cost under control
At Charla, governance is how the platform works. Changes go through human approval, every action is recorded, and every cent is measurable.
Putting an AI agent to work inside a company raises three practical questions: who approves what it knows and says, how you can know exactly what it did, and how much it will cost. Charla was designed to answer all three without forcing you to build parallel processes: the controls come built in and are part of the day-to-day of whoever operates the assistants.
Good agent governance, according to established industry frameworks (such as Google's SAIF), rests on three pillars: human approval for consequential actions, least-privilege permissions (each agent only does what it was meant to), and observability (you can see what happened). Charla turns those three principles into concrete features.
Nothing changes silently: propose and approve
Changes to an assistant's knowledge, its wiki, and its branding don't go live on their own. They arrive as reviewable proposals, and a person must approve them before they take effect. This propose-then-approve flow means that no one (not an automation, not a hurried colleague) silently changes what the agent knows or how it presents itself.
In practice, the review loop protects exactly the points that most affect what the end customer sees:
- Knowledge: the content that grounds the assistant's answers only takes effect after it is approved.
- Wiki: edits to the internal base that guides the agent go through the same human gate.
- Branding: changes to the brand and presentation are reviewed before going live.
Real governance is when the system prevents the silent change, not when it merely logs it afterward.Charla design principle
Predictable cost: credits and per-assistant budgets
Platform usage is measured in credits, transparently and tied to the real cost of the AI providers. Instead of an opaque invoice at the end of the month, you follow consumption as it happens and understand where the spend is going.
Each assistant can have its own monthly budget, and you choose how it behaves as it nears the limit:
- Soft cap: the assistant keeps running but fires alerts as it approaches the defined ceiling.
- Hard cap: once the ceiling is reached, the assistant is paused automatically, so spend never catches you by surprise.
- Per-assistant view: because each agent has its own budget, you can give more room to a critical case and rein in an experiment without affecting the rest.
Why credits matter for governance
Cost is a form of control. When consumption is measurable and there is a ceiling per assistant, you consciously decide how much autonomy costs, and keep the budget predictable even as the agents take on more uses.
Traceability: a record of what actually happened
Every privileged action and every tool call is recorded in an append-only audit log, meaning events are added, not rewritten. Activity is correlated by request, so you can reconstruct exactly what happened in a given interaction, from the initial request to the tools that were triggered.
Beyond the audit log, the platform offers operational observability through structured logs, traces, and metrics. Together, these signals give visibility into how the assistants behave: not just whether an answer came out, but how it was built and how much it cost.
People in command: human-in-the-loop and the scope gate
Useful autonomy isn't autonomy without brakes. For consequential actions, Charla keeps people in control: the agent proposes or executes within bounds, and whoever decides what is irreversible remains a human being.
To reinforce the principle of least privilege, there is also an optional scope gate, a kind of semantic firewall that refuses requests outside that assistant's purpose. An agent built for one end isn't dragged into another just because someone asked.
Per-assistant isolation
Governance gets simpler when the boundaries are clear by construction. At Charla, each assistant's data and permissions are separate: what one agent can access and do doesn't mix with the others. This isolation makes it easier to apply the least privilege possible to each case and to contain the impact of any change.
Taken together, these controls turn governance from a promise into a default behavior: changes go through approval, actions stay traceable, spend has a ceiling, and people stay in command. That's how Charla keeps agent autonomy in service of the company, not the other way around.