There is a clear trend emerging where agents will become standard operating tools. Events teams will have ‘agent libraries’, human-in-the-loop will become the norm and small, specialised agents will outperform mega-bots. Soon events teams, instead of asking where’s the template, will ask, where’s the agent?
Our competitive advantage won’t be access to AI, everyone has that. Our advantage will come from the quality of the agentic events systems we build. That will be determined by the clarity of our workflows, the quality of our data, the systems we build and how fast we improve them.
emap was an early mover in the use of agents to build events systems. To start, we focused on improving customer experience.
In autumn 2024, we built an agent that provides all awards entrants with judges’ feedback. We realised there was an opportunity to significantly improve the awards entry experience. Given the time taken to submit an entry, it seemed important to seize the AI agent opportunity and give feedback to all, something that previously was not possible, given the scale of the task.
In a year, our agent provides well over 10,000 personalised entry feedback responses. The customer response has been excellent. We believe it is a significant factor in the increase in volume of entries over the last two years. You know you are onto something when the sales team are chasing event operations for the feedback results.
Shortly after, we built an agent that provides instant live session summaries at conferences. We realised there was an opportunity to significantly improve the conference experience. Conferences have the highest value when attendees feel they took away some actionable outcomes.
We use an AI service to create live session transcripts. As the session is coming to an end, the transcription is uploaded to the agent. The agent has been iterated numerous times to ensure an accurate session summary. The agent generates bullet point key takeaways that are instantly added to the session final presentation slide for the session attendees. The key takeaways are also made available on the event website post event.
The session key takeaways summaries are now embedded in our conference production workflow. The value and novelty of them is high with speakers commenting and attendees reaching for phones to take pictures, the measure many of us still use to assess whether we have produced valuable content.
We have also successfully completed a pilot with a third-party provider of a customer service event website chatbot and email agent. The agent has answered thousands of the ‘when does it start’, ‘what is the dress code’, ‘how do I get my ticket’ questions. Our customers have been able to get instant answers without having to hunt for the information on our event websites, often not as easy to find as it should be. We have saved our expert operations teams answering those simple questions repeatedly.
With the same third-party provider, we are now piloting agents for sponsorship lead identification and nurturing, sponsorship maximisation, sponsorship proposal building and awards entry personalisation. The results of the sponsorship lead pilot are encouraging. The other pilots have just started.
Having chosen to prioritise building agents that would improve our customer experience, we have recently turned our attention to building agents that we can assign our event work to, turning defined workflows into agentic systems.
Brainstorming
We started this work with an all-staff brainstorm. Staff joined function specific (operations, marketing, production, sales) teams and were asked to think of something they do every week, that they would like to see automated.
We had three guiding principles:
- All agents must have a human in the loop
- We should be platform agnostic
- Start with agents that do not require integration with other systems
We agreed that whilst learning what agents are capable of, we must always check what they are doing and use our human expertise to constantly train, refine and iterate.
Assuming that staff would be most familiar with ChatGPT, we wanted to change that. We wanted to learn which models are best for what tasks. That meant considering Claude, Gemini and Copilot.
We decided to start with the simplest approach to agents. Agents that work on the knowledge we give them from documents and data we upload.
Time permitting, groups were encouraged to go beyond thinking of the idea and start exploring tools / building agents.
Having expected some caution / resistance from staff to AI agents, quite the opposite happened. The number of processes / tasks that staff would like to assign an agent to execute was overwhelming.
Sales identified opportunities to build agents to identify, profile, score and enrich leads; scan press releases for sponsorship leads; generate proposals and presentations; monitor competitors; monitor clients and create award entry templates, flashcard and sales pitches.
Event operations wanted agents to write awards scripts; proof awards slides; shorten the analysis of the complex shortlisting process; create judges’ cards / social media assets; recommend venues; recommend judges; generate pre-event attendee communications and produce interactive event essential guides.
Event production hoped agents could generate first drafts of programmes from desk and phone research transcripts and notes; analyse research calls to recommend better questions; populate briefing documents for web briefs, marketing and sponsorship and delegate sales; recommend speakers; automate speaker invites; monitor competitors and track production and sales progress against project plan.
Marketing wanted to explore further a campaign improver; marketing planner; campaign performance debriefs and image format checker and converter.
An event ops group did get beyond thinking of the idea and built a working agent for awards script writing. They uploaded previous year scripts, shortlist, winner and judges’ comments data for the knowledge base. When adding this year’s data, the agent, built in less than an hour, has reduced the time spent on writing the script from a day to 20 minutes. The first script it produced was judged 95% perfect. The agent has been quickly refined since and is now in use.
The next step was to identify where to start with such a long list of potential agents. Fittingly, we fed all into Claude and asked which could be done following our guiding principle of no integration with other systems. Keeping the humans-in-the-loop, the management team reviewed the suggestions and then asked all staff to confirm that was where they felt we should begin.
That led to a two-hour hackathon building the following agents:
- Award shortlisting analysis
- Awards slide proofing
- Judges recommendations
- Awards script writer
- Press release lead generation
- Proposal / sales deck generation
- Marketing campaign performance debriefs
- Briefing documents
The status of the agents since the hackathon, is as we expected.
Some were simple to build and have quickly proved their value – awards script writer and marketing campaign performance debrief.
Only one attempt has been abandoned – awards slide proofing – as we were unable to solve file size and format issues. AI does have limitations. It is good to quickly realise and accept that.
The rest are being developed, with staff believing the time saving they should deliver justifies the continuing effort building, testing and iterating.
What have we learned?
Be realistic about what can be achieved. Agents make mistakes and they require clear instructions, good data, testing, ownership and effort, time and patience. The quality of the agent’s output depends on the clarity of thinking behind it.
A blended approach of build and buy seems wise. Buy for customer facing agents. Those are the most complex agents and need the high degree of accuracy and sophistication an expert agent supplier can provide. Build for simple internal processes where you can be confident there will always be an expert human in the loop.
Given the effort required to build, test, train and iterate, agents have highest potential for high volume, low complexity, repetitive / repeatable processes. Agents are best where a process has clear steps, inputs and outcomes, rather than where creative brilliance is required.
There is enormous potential for releasing operations capacity for higher value tasks or launching more events. The awards script writing agent alone will save weeks of work in a year, across all awards.
These tools don’t just accelerate existing work. They inspire entirely new work. We would never have envisaged feedback to all awards entries or live session summaries. That’s a good thing but comes with a word of caution. We aren’t yet seeing agents help reduce workload, more they are encouraging us to do more, raising the bar of what is possible.
The platforms do perform better at different tasks. ChatGPT seems good at high level strategic analysis. Claude seems better and quicker at understanding organisational context. It requires fewer prompts to complete detailed tasks.
Make sure you use the best model available, not just the default. Dig into the settings of the model picker and select the most capable option, it changes every couple of months.
The learning curve is gentler than any previous technology shift in history. You don’t need to be a programmer. Many agent builders guide your build. They tell you what they need to know. You need curiosity, a willingness to experiment, and the willingness to let a machine handle part of your work.
The single biggest advantage with agents you can have right now is simply being early. Early to understand them. Early to use them. Early to adapt. If you haven’t started, it isn’t too late.
This article was first published in InPublishing magazine. If you would like to be added to the free mailing list to receive the magazine, please register here.
