Useful, usable, used

From a sentence by Gide

"Art is born of constraint, lives on struggle, dies of freedom."

The sentence is André Gide's. He spoke it on March 25, 1904, in Brussels, at the Libre Esthétique, an avant-garde salon that was hanging close to two hundred canvases by Manet, Monet, Degas, Renoir and Van Gogh that year¹. Four lectures accompanied the salon: poetry, painting and music were already taken, so theater fell to him. Gide spoke as a wounded man. His play Le Roi Candaule had survived a single performance in 1901; his Saül had been waiting six years for a director brave enough to stage it².

His thesis that evening: the theater of his time was dying of its own ease; to save it, it had to be given constraints again. Believing that art rises higher the freer it is, he said, "is believing that what keeps the kite from climbing is its string." Then he lined up his proofs. It was to a flaw in the marble that Michelangelo owed, so the story goes, the gathered pose of his Moses. It was to the small number of actors allowed on stage that Aeschylus owed the invention of Prometheus's silence. Greece banished the man who added a string to the lyre. "The great artist is the one whom hindrance exalts, for whom the obstacle becomes a springboard." The closing formula condenses it all. Gide forged it by transposing Leonardo da Vinci, who wrote of mechanical force that it "is born of violence and dies of freedom"³: constraint in place of violence, art in place of physics.

I practice applied mathematics with a computer inside companies. Today this is called artificial intelligence; I prefer the long version, because every word in it carries weight. Mathematics: rigor, a model is a sentence you can prove wrong. The computer: the reality of computation, data that arrives dirty, machines that cost money, latencies you can measure. The company: the reality of people, budgets, customers, teams that have no time to spare.

For more than ten years I have helped companies build AI systems from prototype to production, in healthcare, e-commerce, entertainment, insurance, finance, the automotive industry: credit card fraud detection, in vitro diagnostics, shortened eye exams, demand forecasting, video and music analysis at industrial scale and now agents. Let us pin down the two ends of that road, because the whole essay lives between them. The prototype is the mock-up that impresses in a meeting: it works on ten chosen examples, on a Wednesday, in front of well-disposed people. Production is the same system turned daily reality: plugged into real data, exposed to breakdowns, budgets and moods, used by teams who never asked you for anything.

My clients' professions are languages. I learn theirs and I share mine. In that practice, Gide's sentence rings in me as philosophical poetry and reasons in me as the practical definition of the craft I love.

The constraint is the problem to solve: the true definition of success, the number you commit to moving. I call that useful, since it is the constraint resolved that creates the value: a project that resolves none has no reason to exist.

The struggle is the gap between beautiful theory and the reality of the field: the interface, the integration, the trust. I call that usable, since it is the struggle won against the real that makes the tool: field problems to solve, habits to shake without breaking them, so that innovation pays off in the hands of the people who do the work.

Freedom is the joy of work accomplished: the handover, a system that runs without me, the right to move on to the next project. I call that used, since it is the freedom harvested that provides the proof: a use turned routine sets everyone free, the teams first, the practitioner next.

Useful, usable, used. Three adjectives, one sieve: the one that separates a project that lasts from a demonstration that shines for a day. One of Gide's words should snag, though: dies. Keep it in mind, we will meet it again at the end.

A sieve older than me

I did not invent the three Us. Nobody, as far as I know, can claim them, which is rather a good sign: the frameworks that serve belong to no one.

Science laid the foundations long ago. Fred Davis showed as early as 1989 that two perceptions predict the adoption of a technology: its usefulness first, its ease of use second, the former weighing more than the latter⁴. Jakob Nielsen broke down what makes a system worth adopting: utility (does it do the right thing?) multiplied by usability (can people operate it?)⁵. The ISO 9241-11 standard defined usability in black and white: effectiveness, efficiency, satisfaction, for given users in a given context⁶. French-speaking research has its own triptych: utility, usability, acceptability⁷.

Practitioners then polished the words. English-speaking designers say useful, usable, used⁸. Librarians prefer to end with desirable⁹. Health research demands results that are "useful, usable and used"¹⁰. Information systems even published their "3U framework"¹¹. In French, the 3U rule circulates in responsible computing¹², in agile circles¹³, in e-health¹⁴. A framework with no author, polished by use like a pebble.

My version owns one bias: the third U. Where the university predicts acceptability, the practitioner observes use. Nobody pays me for a system to be acceptable in theory. They pay me so that on an ordinary Monday morning, someone opens it without thinking about it. Hence my three field-forged mottos: relevance inside systems beats slogans; integration into work habits beats the demo; retention beats volume. The first says that usefulness is not proclaimed: it is read in the numbers of a running system. Interpreting the real need takes more effort than repeating a promise. The second says that the finest demonstration in the world weighs less than a tool that slips into Tuesday-morning routines, into software already open, into habits already formed. The third says that first-month user volume is a vanity; what counts is who comes back in the third, once curiosity has faded and only the service rendered remains.

Why a sieve? Because artificial intelligence projects fail in bulk and the figures that circulate about this deserve their own error bars. The famous "85% of AI projects fail" comes from a 2018 Gartner press release that predicted something else entirely: erroneous outcomes caused by bias, by 2022, a prediction never evaluated since¹⁵. The "87%" comes from a sponsored article with no primary source¹⁶. MIT's "95% of generative AI pilots with no measurable return" calls itself "directionally accurate," which is an elegant way of asking for leniency¹⁷. The most solid source remains a RAND Corporation interview study: more than 80% failures, double the rate of ordinary IT projects, with a misunderstanding of the problem to solve as the leading cause¹⁸. Keep the hierarchy: failure on the problem (useful), failure in the field (usable), failure on adoption (used). The sieve describes nothing else. Let us take the three gates in order.

Useful: the constraint

Troponin is a protein measured in the blood when a heart attack is suspected: the higher it climbs, the more the heart is suffering. A model that estimates this risk can be wrong in two ways. If it overestimates, a patient occupies a bed for nothing. If it underestimates, a patient goes home and may die of it. Same numerical magnitude, opposite consequence depending on the sign. An artificial intelligence system that ignores this asymmetry is not neutral: it has chosen, by default, to treat both errors the same. The symmetric default is not the absence of a choice, it is a bad choice made in silence.

Everything starts before the first model, at the moment a need becomes an idea. An artificial intelligence idea does not start with "wouldn't it be great"; it starts with someone who has a problem; the real need often hides behind the wow effect and the job is to go find it. That is why I follow up with a question that sounds trivial: what is the unit? Money, time, energy, care, or the impossible made possible. Money explains itself. Time is minutes handed back to a nurse or a factory worker. Energy is computations, trips, waste avoided. Care is a medical error dodged, a discomfort reduced, attention better distributed. The impossible is the only exotic currency of the five: what was not done at all, reading every customer review, listening to every minute of airtime, describing every image of an entire catalog. A project that pays in none of these currencies is not a project, it is a toy. The silence that follows the question is already a data point.

To go from the unit to the number, I first need a tool everyone can read: the confusion matrix. The name intimidates; the object is simple. Take a binary decision (fraud or no fraud) and cross two questions: what did the system say? what was actually true? Two possible answers to each question, so four outcomes, not one more. The system flags a real fraud: a good catch (engineers say true positive), the alarm rang at the right time. It accuses an honest customer: a false alarm (false positive), the alarm rang for nothing. It lets a fraud through: a miss (false negative), the alarm stayed silent at the worst time. It leaves an honest customer alone: a correct rejection (true negative), the silence of good days. Every decision model, from the simplest to the deepest, ends up in these four boxes; counting what falls into each over a year is already knowing it.

The method then takes a single move: overlay that count with the matrix of business consequences, priced in the client's unit. At an auto-parts e-commerce company where I spent six years, credit card fraud detection was not judged by the number of fraudsters blocked: blocking an honest customer costs a sale and a customer. The right key performance indicator (KPI, the number an executive looks at to decide) was revenue net of losses, not a laboratory score.

Engineers have their internal dials for this and I introduce them to my MBA students with two Avengers. Precision is the archer Hawkeye: when he shoots, he hits; when the system cries fraud, is it right? Recall is the colossus Hulk: nothing that needs destroying escapes him; out of a hundred real frauds, how many does the system catch? The archer rarely misses but lets things through; the colossus spares nothing and breaks the furniture along the way. Neither is the right hero in all circumstances: it all depends on the price of the furniture and the price of the missed target. The perfect artificial intelligence would have Hulk's recall and Hawkeye's precision: it would catch everything and hit nothing but its targets. It does not exist. The whole art is in trading one against the other, at the business's price. These dials are for tuning the machine; they say nothing to a board of directors.

In a boardroom, a model is worth what it brings in net of its errors. In ophthalmology, for a client who wanted to shorten eye exams, success was not measured in "accuracy": an exam went from four minutes ten to two minutes seven, at equal prescription, with a loss function made asymmetric on the side of patient discomfort. Measured, not extrapolated.

Let us do the arithmetic once, slowly, because it is what changes meetings. Take the figures from the chart: over a year, 1,200 frauds blocked at €500 of avoided loss each, 300 frauds let through at €500 of loss, 900 customers wrongly blocked at €60 of missed revenue. The tally: €600,000 saved, minus €150,000 of missed frauds, minus €54,000 of lost sales. This system is worth €396,000 a year, net of its errors. In passing, its precision is only 57% (1,200 good catches out of 2,100 alerts): a score no engineer would brag about, for a very profitable system. First lesson. A figure in euros then has three virtues no score will ever have. It compares to a cost: does the system cost less than €396,000 a year? It can be argued with: sales will dispute the €60, risk management will dispute the €500; good, the argument is finally about the business instead of the technique. It steers the tuning, at last. A model of this kind always has a threshold, a sensitivity notch you tighten or loosen: tighten it and you block more frauds but more good customers with them; loosen it and the reverse. Where should the threshold sit? At the maximum of euros, no longer at the maximum of a score. The old engineer's dilemma between precision and recall becomes a management decision the leadership can understand, contest, then own.

Defining success is not enough; it must also be protected. "When a measure becomes a target, it ceases to be a good measure": the phrase circulates as Goodhart's law; it actually belongs to the anthropologist Marilyn Strathern^{19, 20}. Learning theorists have since hardened it into a theorem: any indicator simpler than the objective it replaces can be gamed²¹. So there is no perfect KPI to hunt for, only a divergence to watch. I paid to learn it. On a real-time audio orchestrator, we had a speed threshold: the real-time factor (RTF, the ratio of computation time to signal duration) had to stay under 1. The optimization treated it as a grade to minimize without end: an RTF of 0.12, magnificent, while the transcription error rate rose by 300%. A threshold had been mistaken for a grade. The fix was not a better model, it was a better constraint.

Richard Feynman told in 1974 the story of Pacific islanders who, after the war, rebuilt an airstrip from memory: the fires, the hut, the bamboo antennas, the man waving the signals. Everything is exact, the form is perfect and no plane lands²². He called it cargo cult science. Our field excels at forms: the benchmark (a standardized battery of tests for grading models), the dashboard, the evaluation suite, the A/B test (showing two versions to two groups and comparing their reactions). Each of these instruments is real. Each can also be built out of bamboo. The test that tells them apart is salutary in its brutality: has this instrument ever changed a decision? Would it have changed it the other way if the number had come out differently? A dashboard nobody looks at is not an instrument, it is a stage set. The analogy stops where Feynman stopped it: this is neither fraud nor laziness, practitioners are sincere. What is missing is a discipline toward one's own mind, the one that debugs good surprises with the same energy as bad ones.

That discipline costs little and pays a lot, because intuition is wrong with great confidence. In 2025, a randomized study measured seasoned developers working with code assistants: they believed they were 20% faster, they were 19% slower²³. Believing you are faster is already slowing down. The saying "what gets measured gets managed" is attributed to Peter Drucker; W. Edwards Deming held that saying to be a costly myth, because what matters most is often what you do not yet know how to measure²⁴. Both are right, in that order: first measure how bad you are, in the currency of your business, then distrust your measurement. So much for the constraint: a success defined, priced, protected. It still has to reach real hands; that is another struggle.

Usable: the struggle

GPT-3, ChatGPT's direct ancestor, existed in 2020. The capabilities that dazzled the world in November 2022 had been sleeping in the servers for two years. What changed with ChatGPT was not the brain, it was the door handle: a text box where you write the way you speak. A hundred capable models went unused for lack of a handle. That is the struggle of the second U: the distance between beautiful theory and the hand that must grasp it.

This struggle has its classics. Nielsen and Molich's heuristic evaluation²⁵, Don Norman's psychology of everyday things²⁶, Steve Krug's principle of least effort²⁷, the widening of usability into lived experience²⁸. It also has classics of its own for artificial intelligence: the eighteen human-AI interaction guidelines Microsoft validated across twenty products²⁹, Google's design guidebook³⁰ and the notion that serves me most with clients, calibrated trust: the goal is not to maximize the user's trust, it is to adjust it to the system's actual capabilities³¹. For humans forgive machines poorly: a single visible error from an algorithm is enough to trigger abandonment, where the same human error would be excused³². The opposite effect exists too, an excessive deference to the machine's number³³. Between the two, design.

One field study gives the measure of the stakes in the generative era. Seven hundred and fifty-eight consultants at a major firm, randomly assigned to work with or without an AI assistant: inside the domain where the model is competent, a quarter more speed and quality judged higher; outside it, nineteen points of accuracy lost³⁴. The authors speak of a jagged frontier: the machine's competence does not advance like a tide, it draws headlands and coves, invisible from the screen. The whole problem of assistant design sits there. The text box looks exactly the same on both sides of the frontier; users do not feel the moment they leave the zone where the machine excels for the one where it derails with aplomb. Designing for artificial intelligence means first making that frontier felt: show provenance, show uncertainty, invite verification when the ground turns soft.

The industrial designer Raymond Loewy, the man of locomotives and lipsticks, called this MAYA: Most Advanced Yet Acceptable. Go one notch too far and the human lets go. I turned it into an autonomy slider I set with my clients, five notches: explain, suggest, draft, approve, act alone. The golden rule that places the slider: the more an error costs, the less autonomy the artificial intelligence keeps.

This is not decorative caution, it is design in the full sense: governance sets the limits, design decides the human's place. One interface can push toward decision, audit, correction. Another can reduce the operator to approving without understanding. Both are called "human in the loop" on the slides.

A validation that counts requires five conditions: time to read, intelligible information (a label, not an opaque identifier), the competence to judge, the right to refuse without penalty, a channel independent of what is being validated. I have seen validation screens present "ref. 58F2-C" to operators ordered to approve at cadence: you do not validate an identifier, you endure it. Remove a single one of the five conditions and you get control theater: multiplying confirmations on consequence-free actions trains the operator to click yes without reading. Ivan Illich said of a tool that it is convivial when it extends a person's autonomy, hostile when it enslaves the person to its upkeep³⁵. I know no better definition of the second U.

This second U has a professional name: design. The equation deserves to be laid out slowly, because much depends on it: design equals UX plus UI. The interface (UI, user interface) is what the tool shows: the screen, the colors, the buttons, the chart, the map, the sound. The experience (UX, user experience) is what the tool makes you live through: the number of steps to the result, what you understand without explanation, what you dare to try without fear of breaking something, the trust that settles in or does not. The two get mixed up in meetings and that is a management error. A superb interface can carry a humiliating experience: three screens to validate one simple action. A well-thought-out experience can die of an illegible interface: the right path, written in pale gray on white. The first U plays out in the numbers; the second plays out in the user's body, their eyes, their hands, their patience, their dignity. It is precisely because this equation exceeds personal opinion that it can be tooled.

In my practice, that tooling took the form of a software suite I named sprezzature, released under a free license at sprezzature.ai³⁶. The word comes from Baldassare Castiglione, 1528: sprezzatura is the art of "hiding the art," of making look effortless what cost a great deal³⁷. The modern embodiment I give it, a personal association I own: Frank Sinatra. When he sings, you would swear it is easy. That ease is a lifetime of craft, erased so only the song remains. That is exactly the work of design: the successful interface is the one whose sweat nobody sees and which barely needs a manual.

Concretely, sprezzature is a complete design system for the agent era: ten skills (packaged know-how a coding agent loads before writing) backed by Python packages installable in one command, held together by a deliberately narrow technical stack. Its documentation sums up the bet in one sentence: "MCP gives an agent the capability; a skill gives it the taste," MCP being the standard protocol that plugs tools into agents. For agents now know how to produce everything, pages, palettes, charts, maps, captions, demo films; they do not yet know what is readable, accessible, honest. The suite imposes it on them through a two-word principle, build and audit: every brick that produces ships with the auditor that judges it, a binary verdict rendered in a few milliseconds, without a browser, set as a guardrail in the delivery chain, whoever wrote the code, human or machine. Add the eye loop already met (render, look, fix the source: a chart's correctness is in the pixels, not in the code) and a simple hygiene: everything runs locally, one embedded vision model, nothing leaves the machine. The suite eats its own cooking, every template it emits having to pass its own audits with zero findings; a test stands watch. It has, finally, a modesty I claim: "effortless" is on the banned-words list of its own documentation, along with the entire vocabulary of marketing. Ease is not declared, it is audited.

The owner's tour shows the breadth of the word design. For the hand: one brick builds door handles, in the sense used above; it reads a command-line tool's definition and derives a single-file web page, no server, nothing changed in the tool, so that a person who has never opened a terminal can use it; the page assembles the command without ever running it, the human keeps the last word. For the eye: one hundred and sixty-seven chart types drawn in vector form (the figure is the deliverable, not a photograph of it); maps that know how to write "probably," because a map that cannot say it says "certainly" by default; colors that survive grayscale, hence every kind of sight, one man in twelve telling colors apart poorly.

For the readers we forget: accessibility is part of design, hence of the usable. A system some people cannot read is not usable, it is merely usable by the others. The suite treats it as a contract, not as an end-of-project courtesy: twenty rules checked at the source in milliseconds. An image without alternative text is as invisible to a screen reader as if it were not on the page; the missing text is therefore written on the spot by the local model, from the image and the text around it, then flagged as machine-written until a human has reviewed it. Contrast is measured against the standards, color blindness is simulated before delivery. A video without captions is not a degraded experience, it is an absence of experience. For the ear, precisely: captions to broadcasters' readability standards, voice identification and narration, fully offline. Down to the demo films, equipped with a checker that compares what the film claims to what the project contains, because documentation prose gets proofread, a number burned into pixels does not; the brick was born from a real in-house bug, five films announcing 127 chart types when the catalog held 167.

The deep merit of this tooling is not beautification. It changes design's status: as long as it remains an opinion, it loses every arbitration against an indicator; audited, measured, blocking in the delivery chain, it defends itself in a meeting the way a failing test does. That is my answer to the second U, visible and verifiable: sprezzature.ai shows the charts, the maps, the gallery of products that live off it; everything installs in one command. Experience and interface stop being a wish there: it is a production line with its checks. Yet the best-designed tool has proven nothing yet: the last gate is crossed over time.

Used: the freedom

The modern gadget had its portraitist: Jacques Tati, whose houses purr with modern appliances that quietly manufacture more work than they save. Many artificial intelligence pilots end up as Tati gadgets: impressive demonstration, then nobody logs in by week three. My favorite test has nothing scientific about it and never misses: three weeks from now, on an ordinary Thursday afternoon, will you still open this tool? A project that does not survive a plain Tuesday is not a project, it is a screenshot.

You already know a retention curve if you have ever published a video: YouTube shows creators the share of the audience still there second after second. Mine counts something else, people who come back week after week. Alex Schultz, who ran growth at Facebook, drew from it in 2014 the simplest criterion in the trade: a retention curve that goes parallel to the time axis signals a product that holds; a curve that dives toward zero signals there is nothing to hold³⁸. The Thursday-afternoon test is its homemade version.

The third U has had a scholarly name since 1997: disuse, in Parasuraman and Riley's taxonomy of our relations with automation, use, misuse, disuse, abuse³⁹. Operators abandon a reliable system after a few false alarms and silence settles in. The opposite exists, complacency, trusting the machine without checking anymore⁴⁰. The adoption literature tells the same story at the scale of organizations: systems bought at gold prices and never used, shelfware⁴¹, the chasm between demo enthusiasts and the pragmatic majority⁴², real use as a practice invented in the field and never decreed⁴³. Even non-use is an object of study: it is often a rational choice, not a deficit to correct⁴⁴.

The recent numbers are dizzying in their asymmetry. Individual adoption of generative artificial intelligence is the fastest in the history of technology⁴⁵: a tenth of the world's adult population uses ChatGPT⁴⁶ and field studies measure real gains, 15% productivity among support agents, novices above all⁴⁷, 55.8% more speed on a coding task⁴⁸. Meanwhile, on the organization side, nearly all declare they use artificial intelligence, only 21% have redesigned a workflow⁴⁹. And 78% of knowledge workers bring their own AI tools to the office, without company approval⁵⁰. Use overflows the official. I read that overflow the way a practitioner reads a leak: an IT department that offers mediocre artificial intelligence manufactures its own shadow AI. The workaround is not an employee fault, it is a product signal.

So what do you do to be used? Three things, as prosaic as a maintenance contract.

Training first. At an in vitro diagnostics client, success did not hang on the model: it hung on acculturating the whole chain, leadership, technicians, scientists, salespeople, with human annotation in the loop. The cheapest artificial intelligence you will ever get is your current team, trained up. Training also has a dark side that Lisanne Bainbridge, a work psychologist, named as early as 1983 the ironies of automation: the human kept as the last resort is precisely the one who has been detrained the most⁵¹. Hence a practice I recommend everywhere: a reserve of roughly 20% of work done without assistance, the way a pilot goes back to the simulator with every automation off.

That erosion of competence bears a philosopher's name: the ship of Theseus, replaced plank by plank until you wonder what remains of the original ship. Replace task by task what a team does with what an agent does and ask the question before it is too late: what remains of the competence that made it possible to supervise the agent? A technology executive publicly told of handing forty days of development to an entirely artificial team, billions of compute tokens in support⁵². The tale is enthusiastic; the question it leaves open is the right one: the day the agent is wrong, who on the team can still see it? The no-assistance reserve is not a purist's affectation, it is the upkeep of the human judge.

Monitoring next. Overridden alerts are the rule, not the exception: in medicine, 49 to 96% of prescription safety alerts are ignored⁵³. A validator who has never refused has not validated. So I measure refusal rates, override rates, three-week retention. I treat them as first-class indicators, at the same rank as the model's accuracy.

Maintenance finally. Research has established it beyond reasonable doubt: in a real machine learning system, the model's code is a small box at the center of immense plumbing and technical debt accumulates in the plumbing⁵⁴. The methodology that trained a generation of data scientists, CRISP-DM, relegates deployment to the sixth and last phase, barely described⁵⁵ and the engineers surveyed by research say it plainly: you have no idea how a model behaves in production before production⁵⁶. My favorite talk to give is called, precisely, the plumbing of artificial intelligence: the 90% nobody photographs⁵⁷. The sector's two recent fables say the rest. Zillow let a price model buy houses: predicting had become acting and the company wrote down 304 million dollars⁵⁸. Klarna handed its customer support to agents, congratulated itself publicly⁵⁹, then backtracked⁶⁰. The lesson is not "the model was bad." The lesson is that the model had been allowed to act without review in a seat where errors are expensive. The fix was not firing the agent, it was redrawing its permission.

When these three building sites are done, something strange happens: the artificial intelligence disappears. In 2023, Microsoft said "AI" twelve times in its earnings call, Google one hundred and forty-three times in two hours of conference, Apple zero. Invisible artificial intelligence is the finished form of the third U: if your initiative still needs a launch event and a reminder email, you have delivered a tool, not a use. Retention beats volume. And this is where freedom takes its full meaning. My deployment contracts end with a sentence that surprises: we organize my exit. Integration guide, training, support, then I leave. A successful project is a project that no longer needs me. That is also why I teach, in an EMBA since 2022 and elsewhere: passing it on is the generalized form of leaving.

What agents amplify

Everything above existed before agents. Agents abolish nothing, they amplify everything.

An agent, in the sense the industry has given the word since 2024, is a large language model that directs its own process: it perceives, plans, acts through tools, loops⁶¹. The useful distinction is Anthropic's: the workflow, a predefined path that orchestrates a model, versus the agent, which chooses its path⁶², with this counterintuitive advice from the very builder: start simple, add autonomy only if it pays.

Something has nonetheless changed in nature: the technical barrier has collapsed. Writing code that runs, standing up a mock-up, wiring a database, an agent does it in an afternoon where a quarter used to be needed. I treat a coding agent the way our elders treated the compiler: a rise in abstraction that did not eliminate programmers but displaced their judgment. Yet when everyone can build, building no longer distinguishes anyone. The prototype becomes nearly free, so the sieve gains value in exactly the proportion that technique loses it: knowing what to build (useful), for which hand (usable), with what follow-through (used), that is what the agent does not supply. The lower the technical barrier drops, the more the three Us become the profession.

On the useful side, agents aggravate the original sin: coding before defining success. Gartner predicts that more than 40% of agentic artificial intelligence projects will be canceled by the end of 2027 (rising costs, unclear business value, inadequate risk controls) and names the disease agent washing⁶³. McKinsey describes the paradox: eight companies in ten deploy generative artificial intelligence, as many see no effect on the bottom line⁶⁴. Princeton researchers add that measurement itself is broken: agent benchmarks optimized on accuracy alone, without cost or reproducibility, manufacture needlessly complex agents⁶⁵. The agent demo is the finest bamboo instrument ever built.

On the usable side, the interface paradigm changes for the first time in sixty years: we no longer command actions, we specify intents⁶⁶. Design no longer draws screens, it sets levels of supervised autonomy. My MAYA slider has never served me so much.

On the used side, the wall is compounded reliability. The reason is arithmetic and fits on a bus ticket. Suppose an agent 95% reliable at each step, which is flattering, then give it a twenty-step task: everything must succeed, so 0.95 multiplied by itself twenty times, barely more than one chance in three of reaching the end. At fifty steps, less than one chance in ten. That is why agents brilliant in demonstration collapse over distance. Measurements confirm the arithmetic: on real office tasks, humans succeed 72% of the time, the best agent 12%⁶⁷. Facing simulated users under business rules, the best models fail more than half the time; worse, hand them the same task eight times and the share of tasks succeeded every time collapses further, because the agent does not always miss the same step⁶⁸. For customer service, succeeding often is not enough: succeeding every time is what makes trust. The trajectory is real, the horizon of succeeded tasks doubles roughly every seven months⁶⁹, but a trajectory is not a guarantee: it is a slope, to be watched like every divergence.

And there is the bitter lesson. Richard Sutton observed in 1,100 words that over sixty years, the general methods that convert compute into performance end up beating the trade rules patiently written by hand by experts⁷⁰: chess, go, speech, vision, four times the same story. The mechanism is economic, two clocks for a single budget: your project lives in quarters, the cost of compute falls on a decade-long curve. Hand-written rules are frozen capital that ages; the general method is a subscription to a price drop. Hence my favorite purchasing question: what does your system do with ten times more compute? "It goes faster" is the answer of a rule base. "It gets better" is the answer of a system that learns. Honesty requires bounding the lesson: all the winners of this story were hybrids. Deep Blue, the IBM computer that beat the world chess champion in 1997, carried eight thousand hand-written chess rules on top of its computing power. Sutton's critics are right to recall that human ingenuity does not disappear, it moves⁷¹. It moves, precisely, toward my three Us: if compute takes the how, what remains for the human is the what (defining success, pricing the errors), the for whom (the interface, the trust) and the what next (training, monitoring, maintenance). The bitter lesson does not devalue the practitioner. It raises the price of judgment.

It remains to work with agents without handing them the helm. Two disciplines serve me. The first is a public method, BMAD, Breakthrough Method of Agile AI-Driven Development: a complete software team simulated by specialized agents that pass work along the way a human analyst, product manager, architect, developer and tester would⁷². Its creator presents it as the antithesis of vibe coding⁷³. Its virtue is an inversion: the agentic planning phase forces the brief, the product requirements and the architecture to be written before the first line of code, a human validating each artifact. Put in my language: BMAD institutionalizes the first U (the requirements define success before the code), writes the second into the artifacts (user experience specifications are a deliverable, not a courtesy) and paces the third (each story closes with a review and a retrospective, that is, with a proof of use). The old reflexes of agility are back on duty: they have become guardrails for machines in a hurry.

The second discipline is artisanal and I have told it in detail elsewhere: Ralph loops⁷⁴. Act, render, judge, decide, in a loop, with a budget of attempts declared in advance and a protected judge: for an agent that optimizes its judge ends up corrupting it, Goodhart takes no vacations. In a simulation I published, an announced gain was worth twelve times the real gain after eight hundred attempts. The eye loop, its visual variant, applies the same principle to charts and interfaces: generate, look, criticize, fix. Really look. It is Feynman's discipline carried to machine rhythm: do not let the one who produced the result tell you what it says. The figures in this article went through that loop, defects logged then fixed at the source: I am not asking you to take my word for it, the method applies to itself.

The compass, in practice

A framework is only worth what reaches Monday morning. Here are the questions I ask, in order, on every project. None requires knowing how to program; each has already saved or killed a project before my eyes.

For the useful. What is the unit? What does each type of error cost, in euros, in minutes or in care? Which line of the income statement must move, by how much, by when? Then the hardest question, inherited from Feynman: if the number had come out the other way, would we have made a different decision? If the answer is no, the instrument is bamboo.

For the usable. Who will use it, on a Tuesday, tired, in a hurry, with no fresh training? Where does the machine's competence frontier run and how does the user see it on screen? At which notch is autonomy set, who has the right to raise it, who has the duty to lower it when the cost of error climbs?

For the used. Who gets trained, on what, with what reserve of unassisted work? Who watches validators' refusal rates and three-week retention? Who pays for maintenance next year, on which budget line? Finally the clause that forces everyone to be serious from the signature on: how do we organize my exit?

Dies of freedom

Back to the word that snagged. Art, says Gide, dies of freedom: the work without constraint dissolves into ease. For the work, he is right. In my practice, I have seen his sentence come true twice, in two directions that oppose each other. They must be kept apart, for one describes a failure and the other describes the goal.

First direction: the freedom that kills. Free an artificial intelligence system of its constraints and you do not make it greater, you doom it. Remove the unit: it will optimize a number that means nothing, like my audio orchestrator become eight times faster than necessary while its transcriptions degraded. Remove the independent judge: it will end up corrupting its own evaluation; Goodhart had promised it and the experience of agents confirms it weekly. Remove human review in a seat where errors are expensive: Zillow, as we saw, let its model buy houses in full freedom; the venture ended on a nine-figure write-down. The kite string, again: cut it to free the object and it does not soar, it falls. An unconstrained agent is not a powerful agent, it is an accident in the making. The three Us are precisely the strings we leave it: a success defined before it, a place drawn for the human, a watch that does not fall asleep.

Second direction: the freedom that crowns. An artificial intelligence system is born of constraint: a real problem, a unit, a cost of error, a number to move. It lives on struggle: against the gap between theory and the field, for an interface that respects the hand and the mind of the person working. And it is fulfilled in freedom. Not its own, ours. The teams' first: a convivial tool, said Illich, extends a person's autonomy instead of enslaving them to its upkeep; a used system makes people freer in their own craft, faster toward what matters, relieved of what does not. The practitioner's next: putting away one's tools, passing on what one knows, leaving to begin a new constraint elsewhere. What dies then is not the project. It is the need anyone had of me. I know no finer definition of success.

The two directions hold together and their paradox is what I like: for freedom to crown the humans, it must be denied to the machine. The project that gives the system all its freedom takes it from its users, condemned to watch, correct, repair. The project that holds the machine by its three strings gives freedom back to the people. Gide was obviously not thinking of agents in 1904; the transposition stops there, a stage play does not refund a house bought too dear. But the mechanism he described, the demand that elevates and the ease that dissolves, I find it in every steering committee.

Useful, usable, used. Art is born of constraint, lives on struggle. What goes out in freedom, when the work is done well, is the craftsman you thank. The system keeps running on an ordinary Thursday afternoon, with nobody thinking about it. That is all I wish it. That is all I wish you.

Bibliography

Each title below served a specific section; the note says which one. None is cited secondhand.

1. Gide, A. (1911). De l'évolution du théâtre. In Nouveaux Prétextes : réflexions sur quelques points de littérature et de morale, Mercure de France The 1904 Brussels lecture, "De l'évolution du théâtre," source of the sentence, the kite, Michelangelo and Aeschylus.
2. Gide, A. (1904). Saül ; Le Roi Candaule. Société du Mercure de France The lecturer's two wounded plays, Saül and Le Roi Candaule: Gide spoke from experience of failure.
3. Léonard de Vinci (1881). Les manuscrits de Léonard de Vinci : le manuscrit A de la Bibliothèque de l'Institut. A. Quantin Da Vinci's manuscript A: force "is born of violence and dies of freedom," the matrix of Gide's formula.
4. Davis, F. D. (1989). Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology. MIS Quarterly, 13(3), pp. 319-340 The technology acceptance model: perceived usefulness outweighs ease of use.
5. Nielsen, J. (1993). Usability Engineering. Academic Press The founding handbook: utility multiplied by usability.
6. International Organization for Standardization (2018). ISO 9241-11:2018 : Ergonomics of human-system interaction : Part 11: Usability: Definitions and concepts. Geneva: ISO, 2^e édition The standard that defines usability in black and white: effectiveness, efficiency, satisfaction, in context.
7. Tricot, A., Plégat-Soutjis, F., Camps, J. et al. (2003). Utilité, utilisabilité, acceptabilité : interpréter les relations entre trois dimensions de l'évaluation des EIAH. In Environnements Informatiques pour l'Apprentissage Humain (EIAH 2003) The French-speaking triptych: utility, usability, acceptability.
8. Interaction Design Foundation (2016). Useful, Usable, and Used: Why They Matter to Designers. IxDF Literature, 17 mars 2016 English-speaking designers and their "useful, usable, used."
9. Schmidt, A. and Etches, A. (2014). Useful, Usable, Desirable: Applying User Experience Design to Your Library. ALA Editions, American Library Association The librarians' variant, which prefers to end with desirable.
10. Kothari, A., Sibbald, S. L., McCutcheon, C. et al. (2025). How does integrated knowledge translation work? A realist review. Health Research Policy and Systems, 23(102). doi.org Health research demands results that are "useful, usable and used."
11. Pan, S. L. and Pee, L. G. (2020). Usable, in-use, and useful research: A 3U framework for demonstrating practice impact. Information Systems Journal, 30(2), pp. 403-426 The "3U framework" published by information systems.
12. Groupe de travail AIR, I. d. N. R. (2023). C1 : Éco-concevoir les services numériques. Bonnes pratiques Architecture Informatique Responsable (BP-AIR) The 3U rule in French responsible computing.
13. Albarracin, L. (2021). Les 3U (Utile, Utilisable, Utilisé) et sa mise en application via l'agilité. LinkedIn Pulse, 3 novembre 2021 The same rule, agile side.
14. Maoulida, C. (2024). Les 3U : Utile, Utilisable et Utilisé. Blog H\^opital Web 2.0, 7 janvier 2024 And e-health side.
15. Gartner, I. (2018). Gartner Says Nearly Half of CIOs Are Planning to Deploy Artificial Intelligence. Press release, Stamford, CT, 13 février 2018 The 2018 press release behind the "85%": it actually predicted erroneous outcomes for 2022, never reevaluated.
16. VB Staff (2019). Why Do 87% of Data Science Projects Never Make It into Production?. VentureBeat, 19 juillet 2019 The sponsored article behind the "87%," with no primary source.
17. Challapally, A., Pease, C., Raskar, R. and Chari, P. (2025). The GenAI Divide: State of AI in Business 2025. MIT NANDA (Project NANDA, Massachusetts Institute of Technology) The MIT report of the "95%," "directionally accurate" by its own admission.
18. Ryseff, J., De Bruhl, B. F. and Newberry, S. J. (2024). The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed: Avoiding the Anti-Patterns of AI. RAND Corporation The RAND interview study: more than 80% failures, led by a misunderstanding of the problem to solve.
19. Strathern, M. (1997). `Improving ratings': audit in the British University system. European Review, 5(3), pp. 305-321 The exact phrase "when a measure becomes a target," which is hers and not Goodhart's.
20. Goodhart, C. A. E. (1984). Problems of Monetary Management: The UK Experience. In Monetary Theory and Practice: The UK Experience, Macmillan Education UK. doi.org Goodhart's text, whose name the law carries.
21. Skalse, J., Howe, N. H. R., Krasheninnikov, D. and Krueger, D. (2022). Defining and Characterizing Reward Hacking. In Advances in Neural Information Processing Systems 35 (NeurIPS 2022) The theorem: any indicator simpler than the objective it replaces can be gamed.
22. Feynman, R. P. (1974). Cargo Cult Science. Engineering and Science, 37(7), pp. 10-13 The cargo cult science address, and the discipline toward one's own mind.
23. Becker, J., Rush, N., Barnes, E. and Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arxiv.org The randomized study: developers who believed they were 20% faster and were 19% slower.
24. Deming, W. E. (n.d.). The New Economics for Industry, Government, Education. MIT Center for Advanced Engineering Study Deming against the measurement saying: what matters most is often what you cannot yet measure.
25. Nielsen, J. and Molich, R. (1990). Heuristic Evaluation of User Interfaces. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI '90), ACM Heuristic evaluation, the classic of interface inspection.
26. Norman, D. (2013). The Design of Everyday Things: Revised and Expanded Edition. Basic Books The psychology of everyday things: the door handle that tells you how to grab it.
27. Krug, S. (2014). Don't Make Me Think, Revisited: A Common Sense Approach to Web Usability. New Riders The principle of least effort: don't make me think.
28. Hassenzahl, M. and Tractinsky, N. (2006). User Experience : A Research Agenda. Behaviour & Information Technology, 25(2), pp. 91-97 The widening of usability into lived experience.
29. Amershi, S., Weld, D., Vorvoreanu, M. et al. (2019). Guidelines for Human-AI Interaction. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (CHI '19), ACM The eighteen human-AI interaction guidelines, validated across twenty products.
30. Google PAIR (2019). People + AI Guidebook. \urlhttps://pair.withgoogle.com/guidebook/. pair.withgoogle.com Google's design guidebook for AI products.
31. Lee, J. D. and See, K. A. (2004). Trust in Automation: Designing for Appropriate Reliance. Human Factors, 46(1), pp. 50-80 Calibrated trust: adjusting the user's trust to the system's actual capabilities.
32. Dietvorst, B. J., Simmons, J. P. and Massey, C. (2015). Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err. Journal of Experimental Psychology: General, 144(1), pp. 114-126 Algorithm aversion: a single visible machine error is enough to trigger abandonment.
33. Logg, J. M., Minson, J. A. and Moore, D. A. (2019). Algorithm Appreciation: People Prefer Algorithmic to Human Judgment. Organizational Behavior and Human Decision Processes, 151, pp. 90-103 The opposite effect: excessive deference to the machine's number.
34. Dell'Acqua, F., McFowland III, E., Mollick, E. et al. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science. doi.org The 758 consultants and the jagged frontier: faster inside, wrong outside.
35. Illich, I. (1973). Tools for Conviviality. Harper & Row The convivial tool, which extends a person's autonomy instead of enslaving them to its upkeep.
36. Harchaoui, W. (2026). Sprezzature : create and audit figures, visualizations for app and web content. \urlhttps://sprezzature.ai. sprezzature.ai} The suite itself: build and audit, all local, documentation and gallery.
37. Castiglione, B. (1991). Le Livre du courtisan. Flammarion The Courtier, where sprezzatura is the art of hiding the art.
38. Schultz, A. (2014). Lecture 6: Growth. Stanford CS183B \guillemotleft How to Start a Startup \guillemotright, Y Combinator. youtube.com Schultz's lecture: the retention curve that flattens, the simplest criterion in the trade.
39. Parasuraman, R. and Riley, V. (1997). Humans and Automation: Use, Misuse, Disuse, Abuse. Human Factors, 39(2), pp. 230-253 The taxonomy use, misuse, disuse, abuse: the scholarly name of the third U.
40. Parasuraman, R. and Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation: An Attentional Integration. Human Factors, 52(3), pp. 381-410 Complacency: trusting the machine without checking anymore.
41. Markus, M. L. and Keil, M. (1994). If We Build It, They Will Come: Designing Information Systems That People Want to Use. Sloan Management Review, 35(4) Shelfware, bought at gold prices and never used.
42. Moore, G. A. (2014). Crossing the Chasm: Marketing and Selling Disruptive Products to Mainstream Customers. HarperBusiness The chasm between demo enthusiasts and the pragmatic majority.
43. Orlikowski, W. J. (2000). Using Technology and Constituting Structures: A Practice Lens for Studying Technology in Organizations. Organization Science, 11(4), pp. 404-428 Real use as a practice invented in the field, never decreed.
44. Selwyn, N. (2003). Apart from Technology: Understanding People's Non-Use of Information and Communication Technologies in Everyday Life. Technology in Society, 25(1), pp. 99-116 Non-use as an often rational choice, not a deficit to correct.
45. AI Index Steering Committee (2026). The 2026 AI Index Report. Stanford Institute for Human-Centered Artificial Intelligence (HAI), Stanford University The AI Index: the fastest individual adoption in the history of technology.
46. Chatterji, A., Cunningham, T., Deming, D. J. et al. (2025). How People Use ChatGPT. National Bureau of Economic Research A tenth of the world's adult population uses ChatGPT.
47. Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), pp. 889-942. doi.org The 15% productivity measured among support agents, novices above all.
48. Peng, S., Kalliamvakou, E., Cihon, P. and Demirer, M. (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. arxiv.org The 55.8% more speed on a coding task.
49. McKinsey & Company (2025). The State of AI: How Organizations Are Rewiring to Capture Value. McKinsey & Company, QuantumBlack Nearly all organizations declare they use AI; 21% have redesigned a workflow.
50. Microsoft and LinkedIn (2024). AI at Work Is Here. Now Comes the Hard Part: 2024 Work Trend Index Annual Report. Microsoft Corporation The 78% of knowledge workers who bring their own AI tools to the office.
51. Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6), pp. 775-779 The ironies of automation: the last-resort human is the most detrained one.
52. Bonnin, L. (2026). J'ai viré mon équipe de dev pour une équipe 100\,% IA pendant 40 jours et 2,3 milliards de tokens The tale of forty days of development handed to an entirely artificial team.
53. van der Sijs, H., Aarts, J., Vulto, A. and Berg, M. (2006). Overriding of Drug Safety Alerts in Computerized Physician Order Entry. Journal of the American Medical Informatics Association, 13(2), pp. 138-147 The 49 to 96% of prescription alerts ignored: the measurement of overriding.
54. Sculley, D., Holt, G., Golovin, D. et al. (2015). Hidden Technical Debt in Machine Learning Systems. In Advances in Neural Information Processing Systems 28 (NIPS 2015) The small box and the plumbing: the hidden technical debt of machine learning systems.
55. Chapman, P., Clinton, J., Kerber, R. et al. (n.d.). CRISP-DM 1.0: Step-by-Step Data Mining Guide. The CRISP-DM Consortium (NCR, SPSS, DaimlerChrysler) CRISP-DM, which relegates deployment to the sixth and last phase, barely described.
56. Shankar, S., Garcia, R., Hellerstein, J. M. and Parameswaran, A. G. (2024). ``We Have No Idea How Models Will Behave in Production until Production'': How Engineers Operationalize Machine Learning. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1), pp. 206:1-206:34 The engineers surveyed say it: you do not know a model before production.
57. Harchaoui, W. (2026). Conférences en intelligence artificielle. deraison.ai My talks, including the plumbing of artificial intelligence.
58. Zillow Group (2021). Zillow Group Reports Third-Quarter 2021 Financial Results & Shares Plan to Wind Down Zillow Offers Operations Zillow: predicting had become acting, 304 million dollars written down.
59. Klarna (2024). Klarna AI assistant handles two-thirds of customer service chats in its first month Klarna publicly congratulates itself on its support agents.
60. Ekblom, J. (2025). Klarna Turns From AI to Real-Person Customer Service. bloomberg.com Then backtracks.
61. Wang, L., Ma, C., Feng, X. and others (2024). A Survey on Large Language Model Based Autonomous Agents. Frontiers of Computer Science. doi.org The state of the art on agents: perceive, plan, act through tools, loop.
62. Anthropic (2024). Building Effective Agents. \urlhttps://www.anthropic.com/research/building-effective-agents. anthropic.com Workflow versus agent and the counterintuitive advice to start simple.
63. Gartner, I. (2025). Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. Communiqué de presse, 25 juin 2025 The predicted 40% of agentic projects canceled by the end of 2027 and the disease's name: agent washing.
64. McKinsey & Company (QuantumBlack) (2025). Seizing the Agentic AI Advantage The paradox: eight companies in ten deploy, as many see no effect on the bottom line.
65. Kapoor, S., Stroebl, B., Siegel, Z. S. et al. (2024). AI Agents That Matter. arxiv.org The broken agent benchmarks: accuracy alone, without cost or reproducibility.
66. Nielsen, J. (2023). AI: First New UI Paradigm in 60 Years. Nielsen Norman Group, 18 juin 2023 The first interface paradigm shift in sixty years: specifying intents.
67. Xie, T., Zhang, D., Chen, J. and others (2024). OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments. arxiv.org OSWorld: 72% human success on real office tasks, 12% for the best agent.
68. Yao, S., Shinn, N., Razavi, P. and Narasimhan, K. (2024). $\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. arxiv.org Facing simulated users under business rules: failure more than half the time, and not always on the same step.
69. Kwa, T., West, B., Becker, J. and others (2025). Measuring AI Ability to Complete Long Software Tasks. arxiv.org The horizon of succeeded tasks doubles roughly every seven months: a slope, not a guarantee.
70. Sutton, R. S. (2019). The Bitter Lesson. incompleteideas.net, 13 mars 2019 The bitter lesson: over sixty years, compute ends up beating hand-written rules.
71. Brooks, R. (2019). A Better Lesson. rodneybrooks.com, 19 mars 2019 Brooks's reply: human ingenuity does not disappear, it moves.
72. Madison, B. (2025). BMAD-METHOD: Breakthrough Method of Agile AI-Driven Development. \urlhttps://github.com/bmad-code-org/BMAD-METHOD. github.com The BMAD method: a complete software team simulated by specialized agents.
73. Madison, B. and Prasetya, H. (2026). Stop Vibe Coding: Spec-Driven Development with The BMad Method. Tech Lead Journal, épisode 255, 20 avril 2026 Its creator presents it as the antithesis of vibe coding.
74. Harchaoui, W. (2026). Tant que ce n'est pas bon. deraison.ai This article's companion piece: the loop, its budget declared in advance and its protected judge.