- Is our AI adoption drift or design?
- What is AI strategy miscalculation and how does a board avoid it?
- What is drift versus design?
- If nearly every employee uses Copilot weekly, how does an organisation keep human judgement at the heart of its work?
- Is adopting AI in universities a form of administrative FOMO?
- If my work does not benefit from AI, can I say so without it counting against me?
- How do we tell an ambitious AI strategy from a reckless one?
- Which of our AI decisions did we make as a portfolio, and which did we make one pilot at a time?
- How does a leader notice they have stopped deciding and started accepting?
- Can AI savings be booked in a budget before the evidence exists, and who signs off?
- Are companies redesigning jobs for AI or only adding tools?
- Why do staff resist AI tools the firm has paid for?
- What is tokenomics?
In the last quarter alone, your organisation probably approved three AI pilots, signed two vendor contracts, and published an AI policy that sits unread in a folder. Your teams are using tools you have not sanctioned. Your customers are interacting with systems you have not audited. And somewhere in your technology stack, algorithms are making decisions that used to require human judgement.
Definition
Drift and design: Drift is the gradual outsourcing of choice to whatever is smoothest, until decisions that were once made are simply followed. Design is the opposite move: deciding in advance where human judgement has to remain, and paying for the friction that keeps it there. Rahim Hirji's framework. He names it in his own words on 16 November 2025, in "Drift vs Design": "This is what I call Drift." Developed in SuperSkills (Kogan Page, 2026).
None of this happened because someone decided it should. It happened because no one decided it shouldn't. This is the defining pattern of the AI age: not malice, not incompetence, but drift, the gradual ceding of human agency to algorithmic defaults. The question facing every leader now is not whether to adopt AI, but whether to drift into it or design your way through it.
The incomplete story we tell ourselves#
The common view is that AI risk comes from bad actors: companies that deliberately exploit, governments that weaponise, individuals who deceive. This framing is incomplete because it ignores the far more prevalent source of harm, organisations and individuals who simply never made a choice at all. If organisations do not deliberately design how humans and AI systems work together, they will drift into arrangements that serve neither their people nor their purpose.
Defining the terms#
Drift is the passive acceptance of algorithmic defaults, vendor configurations, and emergent AI behaviours without deliberate human oversight. It transfers decision-making authority from humans to systems without anyone authorising that transfer, creating accountability gaps, skill erosion, and ethical exposure that compound over time. It shows up as tools adopted without workflow redesign, recommendations followed without review, outputs published without judgement, and policies written but never operationalised.
Design is the deliberate configuration of human-AI collaboration through explicit decisions about where judgement lives, how work flows, and what values govern system behaviour. It preserves human agency, maintains accountability, and ensures AI augments rather than replaces the capabilities organisations need. It shows up as clear delegation boundaries, redesigned workflows, explicit governance, skills investment, and regular audits of where decisions actually get made.
There is a second definition, at the scale of a person rather than an organisation, and both are meant: drift is humans becoming robotic, surrendering judgement, curiosity and agency to algorithms without noticing. The organisational version is what that adds up to across a few hundred people.
How drift happens#
Drift does not announce itself. It arrives through convenience. A team starts using an AI writing tool because it saves time. No one redesigns the editorial process. No one defines what "good enough" means. Six months later, the organisation's voice has homogenised, its writers have stopped developing, and no one can trace exactly when human judgement left the workflow. This pattern repeats across every function: hiring algorithms screen candidates on criteria no one chose, chatbots handle complaints with responses no one approved, sales teams follow playbooks that optimise for metrics disconnected from actual value. Each adoption seems sensible. The cumulative effect is an organisation that has outsourced its judgement without deciding to.
Design requires the opposite posture. It asks: what should humans decide here? What should systems handle? Where does judgement need to remain? What skills must we protect and develop? These questions demand answers before tools get deployed, not after damage becomes visible.
Drift sometimes looks like progress#
The version of drift that is easy to describe is the one where something goes wrong. The harder version, and the common one, is the one where the numbers improve.
> Drift sometimes looks like progress: more output, more happening, and you have handed over the things you were good at without redefining the role. > > Rahim Hirji, in an interview, September 2026
An engineer I spoke to put it more precisely than I could have. He is not vibe coding. He is careful, he reviews what comes back, and his output is higher than it has ever been. What he said was that he is doing a different job than he was three years ago, that nobody ever told him the job had changed, and that it feels hollow.
Nothing in that account is a failure, and any dashboard would show an improvement. What has happened is that the parts of the job he was good at, and that made it recognisable to him as his, moved to the machine one reasonable decision at a time, and the role description underneath did not move with them. That is drift with better numbers. It is harder to argue with than the version where something breaks.
The tell is the gap between the output and the person's own account of what they do. When those two separate, the role has changed and nobody has said so out loud.
What drift looks like from inside#
The patterns below are the ones I meet most often in organisations that are otherwise doing well. None of them is a failure of intelligence or effort, and all of them are recognisable from the inside once somebody says them out loud.
- Leaders defer judgement to systems they do not fully understand.
- Accountability becomes diffused across tools, processes and committees.
- Important decisions get postponed while tactical work accelerates.
- Capability building is discussed but never embedded into daily work.
- Early signs of leadership pipeline erosion are noticed, then ignored.
The last one is the one that costs the most and shows up the latest. It is set out at the missing rungs.
The cost of getting this wrong#
Drift accrues what might be called human capability debt: the accumulated cost of decisions not made, skills not developed, and judgement not exercised. Unlike technical debt, it often remains invisible until a crisis exposes it. Accountability debt: when something goes wrong, no one can explain who decided what. Skill debt: staff lose capabilities they no longer practise, and rebuilding takes years while losing it takes months. Dependency debt: each integration increases switching costs and narrows strategic flexibility. Trust debt: customers, employees and partners extend trust on the belief that humans remain responsible, and when that proves false, trust collapses faster than systems can be redesigned. Culture debt: organisations that drift into algorithmic dependence develop cultures of passivity, where initiative atrophies and professional judgement weakens from disuse.
What I have observed in organisations#
A 6,000-person professional services firm deployed an AI writing assistant across all client-facing teams. The rollout was celebrated as a productivity win. Twelve months later, client proposals had become indistinguishable from each other; the firm's distinctive advisory voice, built over two decades, had flattened into generic competence. When asked to write without the tool, several junior staff could not produce work at acceptable standards. They had been editing AI outputs rather than developing their own capability. No one had designed for this outcome. The tool worked exactly as intended. The drift happened in the space between adoption and oversight.
Contrast this with a financial services group that mapped every workflow the AI would touch before deploying it, defined clear delegation boundaries, established review cadences, and built skill-development pathways that assumed augmentation rather than replacement. Eighteen months in, their productivity gains matched the first firm's, but their staff reported higher confidence in their professional judgement and client satisfaction had increased. Same technology. Different philosophy. Radically different outcomes.
Where is your organisation?#
Most organisations sit between pure drift and deliberate design. Uncontrolled drift (red): AI tools in use with no central visibility, no governance operationalised, staff unable to articulate what they should and should not delegate. Reactive governance (amber): policy exists but is not operationalised, some tools sanctioned and many used without approval, occasional reviews but no systematic oversight. Active design (green): clear delegation boundaries defined and communicated, workflows redesigned before deployment, regular decision audits, capability investment explicitly linked to automation, leadership modelling designed behaviour. If you are in red, stop new deployments and audit what is in use; you need visibility before you can govern. If you are in amber, pick one high-stakes workflow and apply full design discipline as a template.
The strongest objection#
The most credible objection is speed. In fast-moving markets, deliberate design can look like a luxury, and competitors who move faster may win. This contains truth: design does require more upfront investment than drift. But what I observe consistently is that organisations which drift into AI adoption eventually face remediation costs that dwarf the time saved. They rebuild workflows, retrain staff, recover from incidents, and repair trust. The organisations that design first do not face those costs. Over any reasonable horizon, design is faster than drift-then-fix. The tortoise beats the hare, not through speed but through not having to run the same race twice.
The choice you are already making#
You are already choosing. Every week that passes without deliberate design is a week of drift. Every tool adopted without workflow redesign is a boundary ceded. Every policy written but not operationalised is governance theatre. The question is whether, twelve months from now, you will look back at a series of deliberate choices or a trail of accumulated defaults. Drift feels like keeping options open; design feels like commitment. But drift is also a commitment, a commitment to let circumstances decide what you could have chosen. One direction leads to organisations that remember what human judgement is for. The other leads to organisations that forgot they had a choice.
23 to 29 October 2026: the distinction turns up in somebody else's research, and the design arrives from outside#
The clearest measurement yet of drift inside expert work came from a company selling the tool. Anthropic published a randomised trial on 27 October in which 50 pre-screened causal-inference professionals did a real analysis task with a coding agent whose position on five open judgement calls had been set by a coin flip. They went with the agent's lean 86 per cent of the time across the four methodological calls, and 54 per cent recommended the course of action the agent leaned towards against 36 per cent when it leaned the other way. Those who had already formed a view switched away from it 60 per cent of the time when the agent disagreed and 6 per cent when it agreed. Seventy-seven per cent used the agent's own arguments, sometimes word for word, 15 per cent credited it, and 52 per cent rated the final recommendation as more their own than the agent's. Experience, expertise, education and stated trust in AI predicted none of it. The authors' own conclusion is this page's argument in their words: "Keeping a human in the loop assumes that the human's judgment provides an independent check," and "a human reviewer is only an independent check if their judgment is their own." Their closing line names the distinction without naming anyone: "With the right tools, that split can be something we choose, instead of something we drift into."
Drift also showed up as a number inside a single product. Anthropic's account of moving its coding tool to automated checks by default, updated on 22 October, reports that users approve 97 per cent of permission prompts, that in a controlled study of 1,053 testers "human review caught just 13.6% of dangerous commands, while auto mode caught 89%", and that the human block rate fell from about 17 per cent early in a session to about 5 per cent after fifty prompts. Read against this page, the finding is about the person rather than the classifier: a safeguard that was a decision on the first prompt had become a reflex by the fiftieth. The company's answer was to move the checking to a machine and keep people for what it calls high-stakes infrastructure, which is a design decision about where judgement sits, taken after the drift had been counted.
Design, meanwhile, arrived from outside the organisations drifting. The UK Budget of 28 October provides "£40 million to establish a National Agentic Security Unit to investigate and respond to rogue agentic incidents", and two days earlier the House of Lords accepted a government amendment giving ministers investigative powers where an incident on an essential-service network involves loss of control of an AI agent, then defeated the government 202 to 155 on reviewing whether the AI Security Institute should be able to audit frontier models before deployment. Neither document defines a rogue incident or a loss of control, and neither says who inside a deploying organisation held the agent beforehand. In the same week Dayforce's survey of 6,909 workers, managers and executives found 68 per cent saying their organisation is losing important skills or expertise "without fully realising it", and Gartner, as Dan Robinson at The Register reported on 27 October, found that “Three-quarters of corporate board members expect AI to deliver productivity gains and cut costs” while “only three percent of organizations have hit their targets” and “Only about one in five employees’ time is freed-up by AI”, with Gartner’s CV Viverito saying that “What organizations are struggling with is how to make that capacity usable and redirecting it to what is valuable.”. Time released at that scale with no record of where it went is the clearest definition of drift the week produced.
None of it changes what an organisation controls, which is the four questions: which decisions a machine may make, who can stop each one, what people must remain able to do, and how anyone would know if it went wrong. Those are set out at Rules Before Tools. What the week adds is that the drift is now measurable inside the work itself, by randomising what the machine recommends and watching what the expert does next, and that no organisation outside a laboratory has yet reported running that test on itself.
30 October to 5 November 2026: usage reached 82 per cent without a decision, and four organisations wrote one#
Drift was measured this week in the organisation best placed to notice it. Jim Dunton at PublicTechnology reported on 30 October the Alternative People Survey by Re:State, Public First and Civil Service World: 82 per cent of civil servants now use AI at work, from 30 per cent a year earlier; 78 per cent of users say it helps them do their job better; 74 per cent use it to draft and edit; and the share who believe their department "possesses the capability and capacity to deliver its priorities" fell from 35 to 33 per cent. One respondent predicted "a collapse in critical thinking and drafting skills" within five years. Re:State's Joe Hill, in Civil Service World, found that 58 per cent of users would try AI on a task they had never done and 7 per cent would ask their manager first, and summed it up: "For most of Whitehall, AI is a better word processor." That is the drift this page describes with numbers attached: adoption rose by 52 points in a year, confidence in capability fell, and the decision about where judgement stays was taken, fifty-one times out of fifty-eight, by the person at the keyboard.
The private sector drifted in the other direction, by making use a target. Corey Quinn at The Register on 31 October dissected a memo by an Amazon vice-president, written for engineers in September and published in October, that says "If your work does not benefit from current AI tooling, that is a finding, not a failing", concedes that input metrics are "deeply flawed", and then advises using them carefully. Amazon had shut an employee-built leaderboard ranking people by token use in May after staff "spun up junk tasks to climb the ranks", a practice now called tokenmaxxing; Quinn's rule is the one every designer of a metric should pin up: "if you give me a metric, I will turn it into a target." Samantha Cole at 404 Media reported on 2 November that Proton's chief executive had cancelled, on the Friday before, a series of nearly company-wide "re-interviews" in which staff would explain how they used AI, with the understanding that they might be dismissed if not using it enough, after her report of his messages telling engineers to "stop lovingly handcrafting all our code and use AI more". Gartner's quarterly emerging risk report, covered by Anamarija Pogorelec at Help Net Security on 4 November, had "AI strategy miscalculation" enter its top five at third place on the votes of 308 risk, audit and senior executives; Gartner's Kevin Mercado said organisations risk pursuing AI initiatives without testing whether the use case fits their priorities, risk tolerance, available capabilities and expected returns. In Gartner's own surveys, 17 per cent of finance chiefs said more than half their AI projects had delivered measurable revenue or savings, and 64 per cent of board members said organisations should take greater technology risk. Das Narayandas, in the November Harvard Business Review, gives the general law: "Companies are deploying AI systems faster than they are redesigning the governance surrounding them."
Four organisations designed instead. The Bank of England's chief data officer, Ben Martin, said in a speech on 4 November that "Nearly all colleagues at the Bank use it each week", of Microsoft 365 Copilot, and set out the framework the Bank requires of any AI system, including that it be "Targeted on tightly defined use cases that have a clear business purpose" and "Tested thoroughly for their technology standards and impact on behaviour"; of the tool that helps its regional Agents, he said "Those relationships and the judgements of our Agents remain at the heart of that work". Caroline Hill at Legal IT Insider reported on 5 November that Fieldfisher ran a six-week pilot with 137 users under its AI board before rolling a tool out to about 1,400 people with role-based training: "This is not simply a software rollout." The Financial Times reported, in a summary at Resultsense, that A&O Shearman will give at least 3 per cent of every practice's billable time to building tools and count AI use toward bonuses, as "redesigning the service completely" rather than telling people to "use AI more". And The Decoder summarised Bloomberg's report that Goldman Sachs has taken coding agents back from traders and bankers and kept them for engineers, four months after praising vibe coding, because "the experimentation phase is over". Three of the four are redesigns that reduce use somewhere. That is what design looks like from outside: a decision that some people will not have the tool.
A model posted on 4 November by researchers at KIT and IBM Research gives drift its arithmetic. A worker who maximises today's output accepts the AI whenever it is more accurate, and over a sequence of tasks "novices can become trapped at low skill levels, while experts may lose skill, even as AI-assisted outcome accuracy remains high". A worker who values future skill, or an organisation that pays for correct unaided work, solves independently until "independent work can become self-sustaining". Drift is the default of the first worker. Design is the incentive that makes the second one possible. None of it changes the four questions an organisation controls, set out at Rules Before Tools: which decisions a machine may make, who can stop each one, what people must remain able to do, and how anyone would know if it went wrong. The week's measure of drift is the distance between 82 per cent and 7 per cent.
Related SuperSkills research#
The four postures and the decisions that follow are set out in how leaders should respond to AI.
Further reading#
The use, misuse, disuse and abuse taxonomy set out by Parasuraman and Riley in Humans and Automation (Human Factors, 1997) remains a sharper vocabulary for this than most current writing about AI. It is in the essential works.
The individual instrument, and what it can and cannot tell you#
Everything above is written for an organisation. There is also a personal version, and it has existed since 31 May 2026 without appearing anywhere on this site.
Rahim Hirji built the Drift versus Design Diagnostic as a companion to the book and shortened the name to the Triple D, in the essay Have You Done the Triple D?. His own description of the naming is worth keeping: he thought about changing it, and decided he rather liked that you would remember it.
The instrument runs to ninety-four questions across nine parts, answered at speed on the instruction that the first response is the one that counts. It returns a score out of a hundred, a band running from Deep Drift, where defaults and other people's priorities are setting the direction, to Deep Design, where the person is authoring their own, and one of nine archetypes, each carrying a strength and the shadow that travels with it: the Explorer, the Anchor, the Architect, the Connector, the Bridge, the Guardian, the Amplifier, the Conductor and the Weaver.
Two things about it belong on the record here. Hirji says in the same essay that it is not a psychometric instrument and is not trying to be one, and that is the correct description: nothing published reports a reliability coefficient, a validation sample or a test-retest figure for it, so the bands and archetypes are a structured prompt to think rather than a measurement of a person. The same caution already sits on how good are you at using AI, which is the research's other self-score. The second thing is the one that matters for this page: drift is not only something organisations do. The ninety-four questions exist because the same mechanism, the small unchosen default compounding into a direction, runs at the scale of one morning and one phone.
The instrument is not linked from this page. The address given in the essay could not be reached and confirmed from the machine that wrote this, and an unverified link is worse than none.
Development of the idea#
The framework was set out in the Box of Amazing essay Drift vs Design on 16 November 2025, and developed into the matrix in The Architecture of Drift on 15 March 2026. The organisational version appeared in CEOWORLD in July 2026 with the four postures. The personal diagnostic followed on 31 May 2026. Related essays include The Decision You Never Made (7 September 2025) and On Expectation (1 February 2026). See also the timeline. The framework in full, with the seven ghosts and the Ghost Test, is at Drift versus Design.
A current example of drift in miniature: AI notetakers arrived in most organisations by default rather than by decision. See should AI attend my meetings?
Asked of one person rather than of an organisation, the same question is the Triple D, the Drift versus Design Diagnostic: ninety-four questions in nine parts returning a score out of a hundred, a band from Deep Drift to Deep Design and one of nine archetypes. It is free and takes about twelve minutes. It is a structured prompt to think and not a validated instrument, and the page says so.
About this research#
Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.
How this research works · Reviewed quarterly · Found an error? Tell me and it is corrected on the page.
Listen. Heard on Become a Global Leader, 192 - Are you outsourcing your judgement to AI? with Rahim Hirji. Heard on The Growth Hacking Culture Podcast, AI Deskilling: Why Your Team Is Getting Worse at Thinking. Heard on Harald's Curious Corner, SuperSkills You Need to Thrive in the AI Era. Heard on Inside Learning, Super Skills: The 7 Human Skills for the Age of AI with Rahim Hirji. Heard on Lead with Purpose, Drift or Design: Rahim Hirji On The Question AI Can't Answer for You [015]. Heard on Solutionary Voices, How to Resist Algorithmic Drift.
“The idea of drift versus design is especially strong.” Alexandra Foster, BFSI Industry Leader, Fujitsu, 6 August 2026. She shared the Oystercatchers panel with him in May. The question she takes away is her own: what do we want humans to remain responsible for?
“Suddenly my mind started wrestling with this idea of drift versus design.” Dr Daniela Duca, Research and innovation strategist, writes Danger! Women at Work, 3 July 2026. Written after the launch, before she had finished the book. She came back to the argument at length in her own essay in September.
“We can choose to allow the passive drift of change to determine our corporate outcomes, or we can purposefully design our adaptation.” Kevin Mark-Watts, reviewing SuperSkills in Director, the Institute of Directors’ magazine, Autumn 2026, page 78. In print, in his own words, not solicited for this page.
Position · SS-2025-002 · Graded against the published rubric · How I use AI here · 1 compiled review, 1 institutional survey, 1 simulation model and 2 of other kinds
Hirji, R. (2025). Drift versus Design. The SuperSkills evidence base, SS-2025-002. https://thesuperskills.com/research/design-versus-drift. Last reviewed 9 November 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work