I learned this on the client side, not the vendor side. When I worked in brand within a Fortune 500, I saw several times how big decisions were made from dashboards no one had re-read with judgment. One of the reasons we built Epical was to solve that gap from the outside.
Today, when we start with a new client in LATAM, we do the same thing as always: we take the last six months of digital conversation about the brand and process it with regional judgment. If the client already has a licensed listening platform, my team works on their stack. If not, we deploy ours. On top of either one goes the same thing: our models trained on regional language, internal classification tools, and senior analysts reading curated samples.
The difference between that reading and the dashboard's raw output is usually significant. Sometimes it changes the decision.
The sentiment that wasn't
A consumer-goods client launched a new line in Argentina. The dashboard reported 71% positive mentions in the first two weeks, and the brand team prepared the case to scale the investment across the entire region. Before making the decision, they asked us for a second reading.
The same post, two opposite readings
The team filtered the positive mentions and re-read them one by one. Things came up like "qué buena la nueva versión eh", "me re encanta esta porquería", "está buenísimo, sigan así". All three carry the lexical markers that a model trained on generic Spanish classifies as positive — buena, encanta, buenísimo. All three are irony.
And it's not just Argentina. The same pattern shows up in Mexico with "qué padre la mamada esta", in Chile with "qué bacán la weá", in Colombia with "qué chimba el man". Irony built on positive lexical markers cuts across contemporary Latin American Spanish, and models trained mostly on Spain Spanish or neutral dubbing Spanish still don't resolve it with the same precision they do in English.
Three real mentions. You decide the polarity.
This isn't a dashboard error, it's the correct output given the model running by default. What's missing is an additional layer on top: curated sampling, re-read by analysts who know the regional variant, polarity recalibration for the local corpus.
When the team re-processed the corpus with that layer, the real sentiment was 34% positive, 41% neutral, 25% negative. And within that original "positive" there was a core of irony that actually anticipated an emerging brand problem.
What the dashboard said vs. what the corpus said
The most uncomfortable part of that project was presenting the CMO with a figure that contradicted the one his own team had shown him two weeks earlier. The obvious question that came up in the room was "so the data we saw was useless?". The answer — the data was fine, what was missing was the reading — is more subtle than it sounds, but it's exactly what they ended up buying. The client decided to pause the regional rollout and redesign the value proposition before scaling the investment.
The sentiment attributed to the wrong subject
A multilateral organization called us because its monitoring reported a spike in negative sentiment associated with a financing operation they were announcing in a country in the region. The client's communications team interpreted the data as public rejection of the loan and started preparing a defensive response — explaining the financial terms, the eligibility criteria, the rate methodology.
Who they were really talking to
When the Epical team re-read the corpus, the pattern was different. The negative sentiment wasn't directed at the loan or the organization. It was directed at the local political official signing the operation. People weren't rejecting the financing — they were rejecting whoever was receiving it. The conversation had proper names, personal hashtags and references to a local political controversy that was underway.
Monitoring had aggregated all the mentions under the organization's tag, which is the expected output. What was missing was resolving the object of the sentiment within each mention: distinguishing whether what people were saying went to the client or to something in its orbit. That requires human reading with local political context — not more data, nor better models.
The defensive response the client was about to publish would have deepened the problem. It was answering the wrong question. The hardest part of that project wasn't diagnosing the problem, it was convincing the client not to respond. For a multilateral organization with a multimillion-dollar budget, staying quiet is costly. But it was the right call, and they got it. We recommended institutional silence from the organization, communication directed at the country's technical team, and monitoring focused on the specific political actor. The spike dissipated within a few weeks.
The missing layer
These cases aren't anomalies. They appear systematically whenever a global platform processes conversation in markets with linguistic, cultural and political codes different from the dominant training corpus. They're the reasonable limits of what a model trained at global scale can deliver by default.
The three layers between the conversation and the decision
What's missing on top is a layer that almost never comes packaged with the tool. Lexical reading that distinguishes the specific variants of Latin American Spanish and Portuguese. Resolution of the object of the sentiment when a mention groups the client together with nearby political or cultural actors. And local context in real time, because a digital conversation in LATAM rarely lives isolated from whatever is happening at the moment.
Between the dashboard and the decision there's a reading space that requires regional judgment, cultural context and review discipline. In most global brands operating in LATAM that space is empty, or covered by a team that has neither the resources nor the specialization to do it well.
That layer is, still today, the difference between scaling a campaign that was underperforming and catching a reputational problem while it's still containable.