In this article we're going to look at how we have been eintegrating AI into our work practices.
We deploy AI to automatically gauge how people feel about brands, campaigns and content by mining social media, forums, and other text data. This helps us track reputation and campaign impact in real time.
We have tested large language models (LLMs) like OpenAI’s GPT4 suit of models as “sentiment analysts” alongside traditional tools. For example, our Insights team’s Mapper tool traditionally used an off-the-shelf platform (Brandwatch) to categorise sentiment (positive/negative/neutral), followed by laborious manual checks for nuance. To improve this, we fed sample social posts into GPT-4o to see if it could classify sentiment more accurately out-of-the-box. The AI was very fast and often agreed with automated tools on straightforward cases, but we found nuance is a challenge: without guidance, even advanced AI can mis-read context or sarcasm. In one trial, a post about “love for musical theatre” was tagged as “negative” by both Brandwatch and GPT-4 due to phrasing, even though in context the sentiment was positive. The model needed additional context (“zero shot” vs. “few-shot” prompting) or instructions to correctly interpret the tone.
To address such limitations, we now use a hybrid approach: AI does the first pass, then humans validate edge cases. We are also exploring fine-tuning AI on our domain-specific data. In practice, an analyst might run a corpus of tweets through an AI sentiment classifier, then spot-check posts where the AI seems uncertain or uses ambiguous language. By doing this, we’ve significantly reduced manual workload while maintaining accuracy. In internal tests, this approach cut down the manual review by ~50% – AI confidently handles generic praise or complaints, and our team steps in for the trickier “it’s complicated” posts (e.g. ironic humour or mixed sentiments). We also ensure human oversight of any automated sentiment analysis to catch errors; as industry experts note, AI tools must be guided by human judgment to be effective in PR.
For a national public awareness campaign, we needed to monitor sentiment daily during launch. Our AI-driven process flagged a spike in negative sentiment one evening, detecting angry red emoji and sarcastic comments. On inspection, our team found a legitimate issue (a misinterpreted message) and quickly adjusted the campaign messaging for the next day. The result was a proactive fix before the negativity snowballed – an outcome made possible by AI’s speed and our team’s insight. We’ve seen that AI can surface sentiment shifts within seconds, giving us more reaction time than traditional manual monitoring.
From a broader perspective, we are aligning our sentiment analysis metrics with emerging best practices. In fact, PR industry trends predict that “AI visibility metrics” – like how often a brand appears in AI-generated answers or the sentiment AI detects – will become standard KPIs alongside share of voice and engagement. We are preparing for this future by updating our measurement frameworks. For instance, we track not just raw sentiment (percent positive/negative) but also an “agreement score” between AI and human sentiment ratings to continually gauge our AI’s reliability. This helps in measuring the effectiveness of our AI work: we look at how closely AI sentiment analysis predicts real-world outcomes (sales uptick, surveys, etc.) and refine our models accordingly. By combining machine speed with human sense-checking, we ensure sentiment analysis is both fast and reliable.