Most software hands a person a tool and waits. It records, displays, and reports. The work of noticing, deciding and acting stays entirely with the human — even when the work is routine, and even when there is no time to do it.
Live events are where that fails hardest
A live event has no pause button. Questions arrive faster than they can be read. Content that needs a decision arrives while three other things also need a decision. Attention in the room moves, and the window to respond to it is minutes wide.
A dashboard is not much help here. It shows an operator that 186 questions arrived, which they already knew. What they need is for the routine part of that volume — the duplicates, the policy lookups, the obvious moderation calls — to already be handled, so their judgement goes to the parts that need judgement.
Bounded work, not autonomy
The interesting question about AI in an operational product is not how capable it is. It is how narrowly its authority can be defined.
An agent that can do anything is an agent nobody can safely turn on during a town hall for five thousand employees. An agent that can hold a contribution and record why — and can do nothing else — is something a communications team can actually deploy.
We would rather ship three agents with narrow jobs than one that claims to run the event.
Why specialist agents
Engage watches the room. Answer answers from the event’s own material. Protect moderates. Each one has a single responsibility, a defined set of actions, and a mode you set independently. Report is deliberately outside that set: it reads the event and selects what the team should see, and it cannot act at all.
That separation is not an implementation detail. It is what makes the system explainable to the person accountable for the event, and what makes it possible to trust one part of it without trusting all of it. A team can keep Answer in Assist and run Protect in Auto, because those are different risks.
Human judgement is the point, not the fallback
A deliberate human decision must not be quietly overridden by automation. Actions have to be visible — what was suggested, what was done, why, and who approved it. If something cannot be evaluated safely, the right behaviour is to wait for a person rather than guess.
These are not concessions to nervous buyers. They are the conditions under which anyone would let software act in front of their entire organisation.
Trust is earned one event at a time
Every agent starts in Assist, recommending and nothing more. Teams move an agent to Auto after several events where its suggestions were consistently right — and usually only for a specific class of work.
That progression is the product. A tool that demands trust on day one has misunderstood what it is asking for.
Live audiences are a good place to start
Live audience participation is unusually well suited to this model. The work is high-volume and time-critical. Much of it is genuinely routine. The consequences of a mistake are visible immediately rather than discovered in a quarterly review. And the human in the loop is present, engaged and already watching.
If bounded agents can be made trustworthy anywhere, it is here.
What we are building
An AI event co-pilot: specialist AI agents that help your team engage the audience, answer questions, protect the room, and understand what happened — while humans stay in control.
Not a dashboard with an AI feature. An operating layer for the live room.