(As corporate AI budgets balloon, a former consulting executive who now builds the technology argues that the industry’s real product is not answers but proof, and that most tools on the market fail the test.)
By James Whitfield
Every enterprise AI demonstration has the same moment. The tool produces a fluent, confident answer. Heads nod. Then someone senior asks the question that decides whether a check gets written: says who?
According to Daniel Cohen-Dumani, a 30-year veteran of professional services who now runs an AI company, the shrug that usually follows that question is the central problem in enterprise artificial intelligence. “An answer without a source is a guess wearing a suit,” he says. “It sounds confident. It’s well dressed. And you have no idea if it’s telling the truth.”
Cohen-Dumani’s position is unusual because he is not a skeptic on the sidelines. He founded Experio after three decades inside consulting, where he began his career in Switzerland at the firm now known as Accenture, later built and sold the technology consultancy Portal Solutions, and led transformation work inside a firm of 2,000. His company depends on the technology he is criticizing. That, he argues, is exactly why the standard matters.
The rule his company operates under is blunt: cited, or it doesn’t count. Every answer Experio’s intelligence layer, IQ1, produces arrives with the source document, the exact passage it drew from, and an audit trail a skeptical reviewer can follow in one click. When the underlying knowledge is not there, the system is designed to say so rather than improvise. The company reports 95 percent retrieval accuracy and treats fabricated answers as a zero-tolerance failure, not a rounding error.
That stance runs against the grain of a market built to impress. General-purpose assistants are engineered to always produce something, Cohen-Dumani observes, because a blank response feels like a broken product. In consumer settings that trade-off is harmless. In a law firm, an accounting practice, or an advisory engagement, the calculus inverts. “You’re the one signing it,” he says. “Your name is on the advice.” An error in that context is not an inconvenience; it is a liability with the firm’s letterhead on it.
He is equally pointed about a failure mode he says almost nobody discusses in vendor pitches. Run a language model across a large knowledge base at scale, he warns, and meaning quietly erodes. Context drifts. Similar-but-different concepts blur together. A system that dazzled on a five-document demo degrades, invisibly, by document one hundred thousand. He calls the phenomenon semantic decay, and much of his company’s engineering, from knowledge graphs to retrieval discipline to keeping a human in the loop where judgment matters, exists to fight it. “Accuracy isn’t a launch-day number,” he says. “It’s a thing you lose.”
The implication for buyers is a due-diligence checklist most procurement processes skip. Cohen-Dumani suggests three questions for any vendor: Where did this answer come from? Can my team check it in one click? What does the system do when it isn’t sure? Tools that cannot answer all three cleanly, he argues, do not belong anywhere a mistake becomes a client conversation.
His broader read of the technology is notably restrained for a founder in the space. He is publicly doubtful that general intelligence is imminent, describing today’s models as extraordinary pattern machines rather than minds after examining them, as he puts it, under the hood. He does not expect the frontier labsto solve every industry’s problems either; some, he says, live so deep inside one vertical that a general model never reaches them, and someone with domain knowledge has to go build there. A better base model makes his product stronger, he adds. It does not make it unnecessary.
Where he sees the market heading is toward a bifurcation. Horizontal tools will keep winning the horizontal work: drafting, summarizing, cleaning up an email. Cohen-Dumani uses them himself and calls them brilliant. But the high-stakes questions, the ones where being generic means being wrong, will migrate to systems built for one industry, one firm, and one standard of proof. He frames the choice for executives as the difference between AI that is impressive and AI a firm can build on.
The winning firms, in his view, will not be the ones that trusted the software most. They will be the ones that made verification effortless. Confidence, he likes to say, is cheap. Receipts are expensive. His bet is that the enterprise market is about to start paying for receipts.
