What should on-site interaction actually do?

Two kinds of interaction

Activity interaction fills the room with something to do. A poll during a keynote. A wall of messages on the big screen. A lucky draw before lunch. It keeps energy up, and it is easy to count.

Outcome interaction changes what happens to a person because they were there. A meeting booked with someone who offers what you are looking for. A question that reaches the speaker because three hundred people upvoted it. A visit to a stand that becomes a qualified lead with a note attached.

Activity is not worthless. A room that is doing something is a room that is paying attention. But activity that does not lead to an outcome is decoration, and attendees can tell.

The test for any interactive feature is simple: after it happens, is someone better off than they were before, and can you tell who? The five sections below are what that test asks of a platform.

It should lead to a person, not a screen

The most valuable thing at an in-person event is the other people in it. Interaction that ends on a screen — a vote, a message, a like — has missed the point of being in the same building. That means networking cannot be a directory people scroll once and abandon. It needs three things:

• Suggestions that say why. “You should meet Ana” is noise. “Ana is a supplier in payments, and you said you are looking for one” is a reason someone will act on.
• A request, never an auto-connection. Contact details should change hands only when the other person accepts.
• A time and a place. A meeting request without “10:40, 20 minutes, Hall 2 lounge” is just another message to ignore.

One feed, not ten tabs

An attendee on the floor has a phone, thirty seconds and a lot of noise. They do not need an app with twelve sections. They need to know what wants an answer from them right now: a connection request, a meeting about to start, a poll closing, a prize to collect. Once it is handled, it should disappear.

Interaction competes with the event itself for attention. Every extra tap is a reason not to take part.

Consent is a design constraint, not a checkbox

The more interaction connects people, the more personal data moves, and that is where trust gets lost fastest. The defaults matter more than the policy page:

• Nobody listed in the directory until they choose to be, and hiding works both ways.
• Directory cards never show email or phone.
• Exhibitors export only the leads whose owners agreed to share.
• The big screen in the hall shows the room, not individuals.

An attendee who feels watched stops taking part. An attendee who knows exactly what is shared, and with whom, takes part more.

Exhibitors need their own tools, and the result has to come back to them

On a trade floor, interaction is how a stand earns its cost. A stand that runs its own question, its own draw and its own screen, open to the people who visited it, turns a passing crowd into a list of names with context.

But only if the loop closes: every entry is a lead for that stand, every scan carries a rating and a note, and the exhibitor leaves with a file they can actually use on Monday.

It has to work when the network does not

Venue Wi-Fi fails at the worst moment, usually when the hall is full. Badges, scans and stamps that need a live connection stop exactly when the most people are trying to use them. Anything that happens at the door or at a stand has to keep working offline and upload when the signal comes back.

What actually goes wrong

These are the common failures, and none of them is specific to one platform:

• A poll launched mid-talk without the speaker knowing. Half the room looks down at their phone while the speaker makes the key point. Polls work when the speaker asks for them and talks about the result.
• A Q&A that one question takes over. Without upvotes, the loudest or earliest question wins. With upvotes, the room decides what the speaker hears.
• A draw that rewards the wrong behaviour. If a stand's code is printed on a poster, someone photographs it and enters from the coffee queue. A code that rotates on a screen means the visit has to be real.
• A message wall that becomes a billboard. Unmoderated walls fill with sponsor slogans and in-jokes. Moderated questions survive. Free-text walls rarely do.
• A networking directory nobody opts into. If people do not trust what gets shared, they stay hidden, and the directory is empty on the morning it matters.
• Counting the wrong thing afterwards. The vote count goes into the sponsor report. The number of meetings held between buyers and suppliers is the number the sponsor actually paid for.

Where AI helps

AI is good at three things on an event floor: answering the same question for the hundredth time, turning a flood of input into something a person can read, and explaining things in plain language. It is bad at deciding things on someone else's behalf.

Some of this already exists in event software today. In ours:

• Attendees ask the event page. “When is the keynote?” “Which hall is the workshop in?” “Is lunch included?” An assistant answers from the published schedule, speakers, tickets and venue only. It never sees orders or personal data, and it sends anything about a refund back to the confirmation email.
• Organisers ask their own data. “How many people have checked in so far?” “How many tickets did we sell this week?” The answer comes from the live numbers, under the organiser's own permissions, and any change it proposes waits for a person to confirm it.
• The on-site plan is drafted for you. Lanes, staffing and volunteer scripts are drafted from the expected arrivals, with every estimate flagged for a person to check.

Beyond that, this is where we think AI is useful next. These are directions, not shipped features:

• Clustering the Q&A for the moderator. A busy session collects far more questions than there are distinct ones. A moderator who sees them grouped, with the duplicates merged and the upvotes added together, can put the real question to the speaker in seconds.
• Drafting the poll from the session. Given a session abstract, a model can propose three poll questions the speaker can accept, edit or ignore. The speaker still decides whether to ask it.
• Explaining a match in words. A rule can find that two people fit. A model can say why in a sentence the person will actually read, without changing who was matched.
• A recap per person. Sessions starred, people met, stands visited, questions asked. Each attendee leaves with their own summary, not a generic thank-you email.
• The debrief, written from the data. Arrivals by hour, meetings held, polls answered, leads by stand. A first-draft report in the time it takes to close the venue, with the numbers coming from the platform, not from the model.

Where AI should stay out

Just as important is what AI should not do on a live floor:

• Decide who meets whom without a reason you can check. Our own matchmaking is deliberately not AI. It ranks people on rules: what one person offers against what another seeks, role pairs the organiser defines, shared topics. Every suggestion states its reason. A score nobody can explain is a score nobody trusts, and the attendee should know why a stranger is being put in front of them.
• Connect people automatically. An introduction someone did not ask for is spam, however clever the model behind it.
• See private data on a public surface. Anything an attendee can type into is a surface someone will probe. It should only ever see what is already published.
• Invent engagement. Generated messages on the wall, fake activity to make a room look busy, AI-written “attendee” questions. The room can tell, and it costs the organiser the trust interaction depends on.
• Make up numbers. Throughput, attendance, lead counts: the platform has the real figures. The model can explain them. It should never be their source.

Five questions to ask of any platform

The future of on-site interaction is not more screens or more features. It is fewer moments that each lead somewhere: to a person, to an answer, to a lead. For any platform you evaluate, ours included:

• Outcomes. Does each interactive feature end with an outcome I can name and count?
• Attention. Does it put what needs an answer first, or make attendees hunt for it?
• Privacy. Are the defaults ones I would accept as an attendee?
• Exhibitors. Do they leave with something they can use, from the people who actually visited them?
• AI. Where it is used, can I see where its answers come from, and does a person confirm anything that changes?

Interaction done this way is quieter than a wall of messages. It is also the part attendees remember.