The company that reads the feedback nobody finishes reading

Somewhere in nearly every company sits a spreadsheet of things customers actually said. Survey boxes filled in at 11pm. One-star reviews with a paragraph attached. Support tickets that trail off into a real complaint. Businesses are very good at collecting this. They are famously bad at finishing it. The comments pile up under a polite technical name - unstructured data - which mostly means nobody has read all of it, and probably never will.

Keatext, a Montreal company founded in 2010, is built on the wager that this pile is worth reading, and that a machine can do it. Its platform ingests open-ended feedback from surveys, online reviews, and support tickets, sorts it into themes and sentiment across more than 50 languages, and then does the part most tools skip: it tells you which issues actually move a number you care about, and what to do first.

That last step is the whole company. Plenty of software can chart how customers feel. Keatext's pitch is narrower and more useful - it hands a team a ranked list of what to fix, phrased in plain sentences a manager can act on Monday morning.

01 / THE FOUNDERA linguist who wanted to read faster

Keatext started with a specific frustration and an unusually specific person to have it. Narjès Boufaden earned a PhD in computational linguistics from the Université de Montréal, studying under Guy Lapalme and deep-learning pioneer Yoshua Bengio. Her research was on how machines comprehend human speech - the messy, context-heavy way people actually talk and write.

She noticed that companies wanted to analyze large threads of customer feedback but the available tools demanded heavy customization and long implementations. There was room, she figured, for something simpler. So Keatext began not as a slick product but as a services shop, taking on custom natural-language projects - including work for the Quebec government - and learning, contract by contract, how businesses actually talk about their customers.

In 2015, the company made the hard turn that trips up most technical founders: it stopped selling its time and started selling a product. Keatext launched a cloud-based text-analytics platform, and Charles-Olivier Simard - an MBA with roughly two decades in enterprise software - joined as partner and CTO. The five services years weren't a detour. They were, in effect, the training data.

02 / HOW IT WORKSFrom a comment box to a to-do list

The mechanics are easy to picture. Feedback arrives from wherever customers left it - a website form, a call center, a store, an app, a social post, a review site. Keatext normalizes all of it into one structured layer, detects sentiment and recurring themes, and surfaces the patterns worth attention. Then it writes them down.

The architecture underneath is quietly smart. Keatext runs its own purpose-built NLP models to do the structured, multilingual understanding, then layers GPT on top to generate one-click, natural-language reports with recommendations. The proprietary engine does the reading; the large language model does the writing. It is a decade-old moat with a fresh coat of polish - and a reason the company can say, with a straight face, that it is not just another chatbot pointed at your data.

Analysts build their own views from widgets - heatmaps, time series, pie charts, pivot tables, quadrant charts that plot where a customer journey is strong and where it sags. Filters apply globally, views are shareable, and the recommendation engine flags exactly which issues most affect satisfaction scores such as NPS and CSAT. A CX manager doesn't have to read ten thousand tickets. They read one report.

03 / THE CUSTOMERSBig names, boring problem

The problem Keatext solves is unglamorous, which is precisely why large organizations pay for it. Its customer list runs to names that generate feedback by the truckload: Intuit, Lenovo, and Hilton are among the current references, and the company has counted American Express, NASA, and Bombardier Recreational Products among past users. These are the sort of buyers - CX, product, marketing, and HR teams - who have more comments than they will ever have people to read them.

The value shows up as time saved and blind spots caught. An emerging complaint that would have surfaced three renewal cycles too late instead shows up as this week's top action item. The recommendation is not "sentiment is down 4%." It is closer to "here is the specific friction, and fixing it is the biggest available lift on your score."

04 / THE MARKETA crowded room, a specific chair

Customer-feedback text analytics is a busy category. Keatext sits among rivals like Thematic, Chattermill, Kapiche, Enterpret, and SentiSum, with heavyweights Medallia and Qualtrics looming over the enterprise end. In a market where everyone shipped an AI "insights" feature the moment GPT arrived, Keatext's edge is unfashionable: it has been doing this since 2010, which means it spent a decade on the boring, hard parts of NLP before the demos got easy.

The guidance the market itself gives is telling: choose Keatext when you want quick insights plus AI-recommended next actions with minimal setup. That is a deliberate lane. It is not trying to be the sprawling enterprise voice-of-customer suite. It is trying to be the tool that answers "so what should we do?" without a six-month rollout.

05 / THE BUSINESSSmall, specific, and renewed

Keatext is a B2B SaaS company that never chased the hype cycle. It raised modest seed funding - roughly US$1.37M across rounds - from Anges Québec, Real Ventures, and Desjardins Venture Capital, capital aimed squarely at the shift from services to product. There is no mega-round in the story, no unicorn valuation. There is a focused tool that big brands kept renewing because it answered a real question.

The company is rooted in Montreal, a city that turned into a world AI-research hub partly on the strength of the same academic lineage Boufaden came out of. Its office sits in Old Montreal, on rue McGill. The CEO, now a Forbes Technology Council member, speaks publicly about inclusion and diversity in AI - a view she frames not as decoration but as a question of who builds the models, and therefore what the models end up being good at.

There is a certain discipline to staying this specific. The temptation, especially post-GPT, is to become a general "ask your data anything" box. Keatext has instead kept insisting that the useful output is not an answer to a question you thought to ask - it is the ranked list of things you didn't. The feedback, after all, is already sitting in the inbox. The company's entire job is to finish the sentence: we collected it, and then.