Narjès Boufaden and the trouble with listening

Narjès Boufaden has a soft spot for an android who struggles with jokes. She is a Star Trek fan, with a declared preference for The Next Generation. Its promise of machines that can converse with people belongs to a future she has spent much of her working life trying to understand. The comedy is appropriate. A machine can master an impressive vocabulary and still miss the point of a perfectly ordinary conversation.

Her subject has never required a spaceship. It can fit inside a telephone call, an open-ended survey response or a customer complaint. People hesitate. They refer to something they mentioned three sentences earlier. They change subjects halfway through a thought. A person listening usually repairs these little gaps without noticing. Giving software the same ability is a considerable engineering problem.

Boufaden built her academic career around that problem, then founded Keatext in Montreal in 2010. Five years later, she moved the business from professional services into a software product. The company analyzes written feedback so organizations can understand the experiences behind it. That progression, from studying conversations to selling a way of understanding them, supplies the through-line of her career.

The attractive part of science fiction is often the machine that answers. Boufaden’s work has repeatedly returned to the machine that listens. There is a good deal of less glamorous labor in that distinction: finding a relevant word, recognizing a topic, connecting an opinion to the thing being discussed. The future, in this version, contains rather more reading than flying.

02 /Ninety-five conversations, no tidy sentences

One early piece of her research gives the story a concrete starting point. In 2003, she published a paper on a semantic tagger for information extraction. Its material was a collection of 95 manually transcribed telephone conversations, approximately 39,000 words, concerning maritime search and rescue. Callers and operators discussed incidents, mission plans and debriefings. Some conversations mixed those purposes together.

The task was to mark relevant pieces of language with concepts from that particular domain. Her method combined an ontology, a structured collection of domain concepts, with the Wordsmyth dictionary-thesaurus. The system had to use context to help decide which words mattered. A transcript preserves what was said; extracting useful information requires another layer of work.

This is a revealing choice of material. Spoken language brings interruptions and imperfect phrasing into a field that often prefers clean inputs. The research addressed small collections of conversational text, where simply acquiring an enormous pile of examples was not the answer. It treated the irregularities of speech as part of the task.

A 2004 paper with Yoshua Bengio and Guy Lapalme continued the examination of informative words in spoken texts. In 2007, Boufaden co-authored research with Truong Le Hoang and Pierre Dumouchel on predicting user satisfaction in human-machine dialogues. The experiments involved 1,027 dialogues in a travel-agency setting. The question had moved toward something a business would recognize immediately: how can a system detect whether an interaction is going well?

03 /A laboratory education meets a customer

At Université de Montréal, Boufaden’s doctoral supervisors were Guy Lapalme and Yoshua Bengio. The connection places her inside the city’s research community before the commercial excitement around deep learning. Her academic work was followed by research roles in universities and at the Centre de recherche informatique de Montréal, known as CRIM.

Bengio later recalled that she had been among the early researchers working with him on representation learning for language. His recommendation describes a direction that attracted little interest at the time and credits her with believing in it. That recollection matters because it records a professional judgment made before the approach became fashionable. She had already committed time to the difficult work.

CRIM also brought her into contact with organizations facing practical information-management problems. She encountered an opening for applying the research outside a university. Keatext’s initial work included clients such as the Quebec government, Wajam and JLR. In 2012, a Quebec government project earned an Octas innovation award. The business began with specific organizational needs and the work of meeting them.

There is a useful humility in that starting point. An algorithm can be interesting for reasons that have little bearing on a buyer’s day. A buyer has an existing process, a budget and a problem that needs solving. Boufaden’s career crossed that border. Scientific questions remained central, while the answers acquired a second test: whether somebody could use them.

04 /Five years to change the business

In 2015, Boufaden concentrated the company’s efforts on a cloud-based text analytics platform. Charles-Olivier Simard joined as partner and chief technology officer that year, bringing experience in enterprise software. The partnership joined her conversational-language expertise with his background in building and delivering software for organizations.

The change in business model was substantial. A professional-services company solves problems through projects. A product company needs something that can be delivered repeatedly. Keatext’s platform offered organizations a way to interpret feedback across channels, including surveys, reviews and other written interactions. Its company history names American Express, NASA and Bombardier Recreational Products among the organizations it attracted.

The founders also entered FounderFuel. The accelerator put their company strategy under scrutiny while they continued running the business. A contemporary team account describes long hours, meetings and late-night brainstorming through the 13-week program. It ended with a Demo Day presentation to community members and investors. The laboratory skill of explaining a result had acquired a new audience.

Boufaden subsequently became a mentor at Techstars Montreal AI, helping early-stage businesses with technology and business models. The two subjects belong together in her own history. A founder can have a sound method and an awkward company, or an appealing pitch and a weak method. Her path had required attention to both. The work of translation was becoming something she could pass on.

05 /The complaint beneath the score

By 2019, Boufaden was speaking at Collision in Toronto, where she gave a Growth Summit keynote. Her subject was the money and effort companies were spending to hear from customers, and the uneven results of those efforts. A company can collect feedback in quantity and still struggle to decide what deserves attention.

The distinction is familiar to anyone who has completed a survey. A number is efficient. The explanation takes longer. A low rating may accompany a problem with delivery, a confusing instruction or an unanswered question. Grouping those responses together without their context can leave a manager knowing that people are unhappy while remaining uncertain about what to change.

Boufaden’s product work follows that gap between measurement and explanation. In an interview about customer-service applications, she discussed both routing support communications to the right place and finding patterns across customer feedback. These are different uses of language analysis, but both depend on understanding enough of a message to make it useful to someone else.

Her ambition also extends beyond a retrospective description. She has discussed recommendations about where a business should improve and what those improvements might affect. That is a demanding next step. The organization needs a connection between what people say and the decisions it can take. A beautifully arranged dashboard still leaves somebody with the rather inconvenient responsibility of doing something.

06 /When the conversation leaves a trace

In a 2022 podcast conversation, Boufaden described the value of customer interactions that now appear in chat logs. A question once asked on a shop floor might disappear when the exchange ended. A written chat leaves material that can be examined later, alongside surveys, reviews and emails. It can contain requests for something a company has not yet offered.

“If you have a bad score, it’s already late,” she said, discussing periodic Net Promoter Score surveys. Her concern was the delay between an experience and an organization’s awareness of it. More frequent analysis of the words people use could help reveal a change earlier. The appeal is practical: shorten the distance between a problem occurring and someone noticing.

That line of thought has accompanied her through changing technology. Her questions about conversations predate the current enthusiasm for generative AI. New models supply additional ways to work on them, yet the basic problem remains recognizable. People speak in context. An organization needs to understand that context well enough to make a decision.

“If you have a bad score, it’s already late.”

NARJÈS BOUFADEN · 2022

07 /A Trekkie asks for the numbers

Boufaden’s January 2025 writing about generative AI begins with her affection for The Next Generation and the excitement of seeing science fiction approach reality. The enthusiasm is accompanied by reservations. She praises generative AI’s ability to summarize noisy conversations, while asking what a customer-experience professional needs beyond a readable summary.

Her argument is about the distance between a general impression and measurable evidence. A model may offer an account of what customers feel. A feedback analytics solution must also help people quantify patterns and compare them over time. She uses illustrative figures to explain that distinction. They are examples of the kind of evidence a manager might need, rather than a report of Keatext’s results.

The combination is appealingly human: delight at a new technical possibility, followed by questions about its usefulness. Her interest in robots and conversational machines has survived years spent working on the details. So has her insistence that the technology needs to earn its place in an organization. A fan of the future is allowed to ask for the spreadsheet.

08 /Who teaches the machine to listen?

Boufaden has also spoken publicly about bias and the people who build AI. In her WomenTech Global Awards 2020 talk, she examined how choices about data, features, training and validation can shape outcomes. She argued for more diverse communities of AI professionals, bringing different perspectives to the problems systems are asked to solve.

That concern sits naturally beside her commercial work. Listening is a selective activity even when a machine performs it. Someone chooses the material, defines the categories and decides what counts as a useful answer. A company studying its customers and a researcher assessing fairness both need to question the assumptions built into the process.

Her connection to the university community has continued, too. In November 2025, she served on an alumni jury for Université de Montréal’s Impact Millénium Québecor entrepreneurship competition. The researcher who became a founder was helping assess another generation’s proposals. It is a modest, concrete continuation of the same bridge between academic knowledge and a workable venture.

Across that career, the recurring object is a conversation: a transcript with relevant details hidden among hesitations, a customer exchange that leaves clues about an experience, a new model that can summarize but still needs evaluation. Boufaden keeps returning to what can be understood and what can be acted upon. Even an android with impeccable diction has to learn that much.