
The pristine whiteboard mounted on the corner wall of the fifth-floor incubator suite at one-north was completely bare, save for the ghost of an erased product flow from three months back. In the late nineties, covering the early wreckage of the dot-com boom out of San Francisco, you could smell a founder’s panic fifty feet down the corridor: paper taped to drywall, frantic marker scrawls, half-eaten takeaway cartons, and founders arguing with their lead engineers until their voices gave out. Decisions were hammered out through exhausting, imperfect human friction. Here, in the glass-and-steel corridors of Singapore’s technology hub, the suite had the airless calm of an intensive care unit.
Across the pale birch table sat Julian, a twenty-nine-year-old founder whose logistics SaaS platform had closed an oversubscribed seed round fourteen months prior. His laptop was open to a dark terminal where background scripts ticked along in eerie silence. A constellation of autonomous agents ran continuous loops: one scraped regional competitors’ pricing tiers across ten Southeast Asian jurisdictions, another synthesised thousands of customer support logs into thematic clusters, and a third continuously re-ranked their product backlog based on weighted sentiment scores.
Julian ran his thumb along the aluminum edge of his laptop with the quiet pride of someone who believed he had outsmarted entropy. He told me his team had stopped having roadmap arguments because the pipeline ingested the telemetry, scored feature requests against market velocity, and spit out sprint objectives with confidence intervals attached. Every single goal was mathematically defensible. With a contented smile, he leaned back into his ergonomic mesh chair and told me they had successfully eliminated founder bias and engineered out friction. Three weeks later, his lead venture backer refused to write a Series A check, and by the end of the quarter, two of his four flagship enterprise accounts had walked away.
Over a quarter-century of watching technological shifts swallow industries whole, I have seen the same tragedy stage itself across different eras, changing only its vocabulary. In the early 2000s, automated credit-scoring models promised mortgage desks that algorithms could price collateralised debt without the erratic interference of loan officers’ subjective hunches.
In 2008, when the subprime architecture imploded, the executives sitting before congressional committees looked dazed because their models had pronounced systemic risk nearly impossible, yet they had forgotten how to look out the window and evaluate credit with their own eyes. Their critical faculties had atrophied through decades of offloading responsibility to math.
Medicine calls it atrophy when a limb is locked in plaster: the bone mends, but the underlying muscle tissue starves from lack of resistance, shrinking until it can no longer support the body’s weight. Across the venture corridors of Singapore, Jakarta, and Silicon Valley, an identical pathology is taking hold—strategic AI atrophy, the quiet decay of founder conviction and intuitive discernment brought on by outsourcing ambiguous, high-stakes decisions to synthetic consensus. Julian’s seed round had bought world-class engineering, but his automated operational hygiene had quietly amputated his team’s gut.
Also Read: Dubai is selling trust, not just capital: What it means for Southeast Asian founders
What Julian celebrated as eliminating bias was what researchers at Harvard Business School, MIT, and Boston Consulting Group diagnosed with far more clinical precision in Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.
In studying professionals given access to frontier language models, they documented a dangerous psychological surrender: workers handed smooth, confident algorithmic outputs routinely succumbed to falling asleep at the wheel. Because the machine’s reasoning was synthetically polished and immediately coherent, the human operators stopped interrogating the underlying premises, trading the exhausting labor of critical friction for the effortless glide of automated consensus.
When I asked Julian to walk me through the strategic pivot he had executed that spring, he clicked into an archived project folder. His largest logistics customer in Indonesia—a legacy distribution network operating hundreds of inter-island freight lanes—had repeatedly begged for an awkward, bespoke offline export tool so warehouse dockworkers could manually override automated routing manifests via WhatsApp threads whenever severe monsoon weather grounded small cargo vessels.
Julian explained that the model had flagged the Indonesian request as an anti-pattern burdened with high servicing overhead, low gross margin potential, and bespoke maintenance risk. The competitive synthesis engine had cross-referenced forty public SaaS post-mortems and recommended doubling down on an automated, standardised self-serve dashboard instead, rendering the statistical verdict unequivocal.
When I asked if he had boarded a flight to Jakarta, if he had spent forty-eight hours on the ground at the Tanjung Priok terminal standing in the rain to watch how freight managers route trucks when port gates flood, Julian looked at me with the polite, bemused tolerance technical founders reserve for print journalists. He asked why he would waste three days stuck in Jakarta traffic when an agent could parse ticket logs, run sentiment classification on customer emails, and cross-reference feature sets against regional SaaS benchmarks in under a minute.
He was reciting what tech companies love to call data-driven objectivity. Yet as ethnographer Tricia Wang observed in Why Big Data Needs Thick Data, big data and quantitative telemetry routinely blind companies to critical human shifts because they only measure the variables an algorithm has been instructed to track. In Jakarta, logistics did not run on clean, cloud-native APIs. It ran on personal relationships between dock clerks and independent truckers, sudden cash-on-delivery settlements, and informal workarounds. The clunky custom feature the Indonesian client had begged for was not an operational inefficiency; it was their lifeline in an unpredictable physical environment.
Also Read: If nobody knows you, nobody owes you their trust: How founders can build it anyway
When Julian’s pipeline quietly stripped that feature from the product roadmap, the customer did not file a bug ticket or register a complaint in the system. They simply recognised that the founders in Singapore had no genuine understanding of how commerce operated in the archipelago, cancelled the pilot, absorbed their losses, and went back to spreadsheets.
The seductive hazard of running an enterprise on AI models is that it constantly confuses probability with breakthrough insight. A foundational model is, by its mathematical definition, an engine of historical consensus. In a study published in Science Advances titled Generative AI Enhances Individual Creativity but Narrows Collective Diversity, researchers discovered a striking paradox: while generative AI tools help individuals write cleaner, more persuasive prose faster, the collective pool of ideas produced across entire groups becomes narrow, predictable, and monotonous.
When a founder asks a model to evaluate an early-stage market, prioritise a roadmap, or structure a commercial offering, it searches the statistical centre of the training corpus to select the highest-probability path—merely the average of how hundreds of other software businesses have solved adjacent problems in the past. As computer scientists documented in Nature regarding AI Models Collapse When Trained on Recursively Generated Data, recursive reliance on synthesised data systematically erases long-tail anomalies, which is precisely where unexpected breakthroughs live.
Every transformative technology business in history was born out of an insight that a consensus model would have scored as an anti-pattern. Renting air mattresses on a stranger’s floor was an operational and liability nightmare, selling physical books on the internet in 1995 offered paper-thin retail margins alongside astronomical shipping costs, and manufacturing high-end electric roadsters when lithium cells cost nearly a thousand dollars per kilowatt-hour was an economic impossibility.
Breakthrough ventures succeed precisely because of the eccentric, unquantifiable convictions of human founders who refuse to accept historical averages. When an entire ecosystem of founders feeds its strategic dilemmas into the same few foundational models, the inevitable result is collective homogenisation: everyone builds the same feature sets, adopts the same sanitised landing-page copy, and deploys the same three-tiered pricing strategy. They build companies that are mathematically defensible, structurally sound—and completely devoid of a competitive soul.
As economists Ajay Agrawal, Joshua Gans, and Avi Goldfarb observed in The Economics of Artificial Intelligence, when machine intelligence makes predictions and routine cognitive execution cheap, the economic value of the complementary human asset skyrockets. That complementary asset is not prompt syntax; it is judgment, the willingness to bear the weight of an unhedged decision when historical data offers no clear answer.
Two months after our initial conversation, I met Julian again at an open-plan co-working space on Robinson Road. His company had surrendered its leased incubator suite, and the engineering team had been trimmed from 11 to four. His screen was open to an empty text editor, the terminal of autonomous agents quiet for the first time in a year. Stirring a cup of black coffee that had gone cold, he admitted that the hardest part was realising that when the market pushed back, they had no muscles left to push back with.
For 14 months, whenever they hit a crossroads regarding pricing tiers, enterprise retention, or product positioning, they had fed the variables into the agents. They had debated the summaries, but they had never practised the agonising work of wrestling with the ambiguity themselves. They had treated the prompt output as the oracle, and when the enterprise clients left, the agents had no answers, merely generating four more variations of the same generic SaaS playbooks. Staring out across the dense mid-morning traffic flowing toward Shenton Way, Julian whispered that while they thought they were moving faster than everyone else, they had actually been letting the machine think for them until they forgot how to think for themselves.
The lesson for this generation of operators is not a Luddite call to close the terminal or throw away the tools. AI agents are astonishing instruments for automating baseline operational drag, perfectly suited to parse code, clean datasets, transcribe calls, and summarise telemetry. The machine should manage the operational floor, but it must never be allowed to usurp the strategic ceiling.
If you are building an organisation, you must guard against the silent creep of AI atrophy by using models to cluster user tickets while continuing to board the plane, sit across the table from your customer, and watch the unspoken hesitation in their eyes when they talk about their business. You must study consensus outputs specifically to recognise the average, predictable path your competitors will take, using human intuition to chart the contrarian path they cannot see.
Above all, you must protect the muscle of accountability. When a decision carries existential risk, refuse the comforting anaesthesia of an algorithmic score, sit in the quiet, feel the weight of the ambiguity, and make the call yourself. In an economic landscape flooded with infinite, auto-generated certainty, human conviction is the only asset that cannot be commoditised. The algorithm can calculate the odds, but only a founder can accept the consequences. If you surrender your judgement to the machine, you may build a frictionless company, right until the moment you discover you have automated away the only reason you were ever in the room.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
The post We took founder bias out of the roadmap: Then our customers fired us appeared first on e27.
