Phantom adoption

A procurement analyst opens her company's AI assistant to rephrase a contract clause. The tool presents a splash screen she has dismissed fourteen times, six capability categories she will never explore, and a text box labelled "Ask me anything." She types her request. Five seconds pass. The response rephrases the clause in American English, adds three sentences nobody asked for, and strips the defined terms she needs to keep. She hits undo, rephrases the clause herself, and moves on.

Two floors up, her COO is showing the board a dashboard: AI tools deployed across 2,300 seats, adoption at 91%, training completion at 100%.

A logistics coordinator at a different company searches for a customs tariff code using the AI assistant his employer rolled out in January. The answer comes back confident, well-formatted, and wrong. He learned to double-check on week two. He now uses the tool as a search bar that returns plausible wrong answers more slowly than Google. His manager's weekly report lists him as an "active AI user" because he opens the tool daily.

A financial analyst asks her AI copilot to build a summary table from a quarterly dataset. The table arrives with invented column headers and rounded numbers that don't match her source data. She rebuilds it in Excel. She raised this in the last feedback survey. Nothing changed. Last month she started entering deliberately garbled prompts so the tool's output would be visibly bad enough to raise questions.

Three workers. Three organisations reporting successful AI deployment. None of the three would use the word successful.

The buyer's showroom

The pattern has a name: phantom adoption. An organisation that has deployed AI by every metric visible to the boardroom while delivering no value on the floor. Licences issued. Tools provisioned. Training logged. The sale was made. The adoption was not.

WalkMe's 2026 State of Digital Adoption release says 88% of executives believe employees have adequate tools, while only 21% of workers agree. That 67-point perception gap is the fingerprint of phantom adoption, appearing wherever enterprise AI is sold to one person and used by another.

Both sides know. WRITER's 2026 enterprise AI survey found that 75% of executives said their company's AI strategy was more for show than real guidance. The measurement layer between executive confidence and frontline reality is broken because deployment metrics satisfy the reporting system that reaches the boardroom. The actual product experience does not.

Demo-driven design

How does a product end up working beautifully in a conference room and failing at a desk?

Enterprise AI is sold through demos, dashboards, and RFP compliance. The demo displays novel capabilities. The dashboard shows executive-level analytics. The compliance sheet ticks regulatory boxes. These three artefacts win the deal. None of them represent the daily experience of using the tool.

Demo-driven design is what happens when feature priorities follow what performs well in a twenty-minute sales call rather than what survives the hundredth use. Impressive autocomplete. Sparkle icons. "Powered by AI" badges. Strong first impressions. Meanwhile, the details that determine whether anyone still uses the tool in three months get neglected because they don't demo: latency under load, error recovery, undo behaviour, integration into existing workflow.

A few of those details, specifically:

• Latency on the hundredth use, not the first. The demo optimises cold start. The daily user cares about p95 response under real load. Five seconds is a demo. Five seconds four hundred times a day is a lost afternoon.
• Error recovery. What happens when the AI is wrong? Can the worker undo, edit, and override without losing context? Most enterprise AI treats errors as edge cases. For daily users, errors are the workflow.
• Output predictability. Workers build mental models of their tools. When identical prompts return wildly different results on different days, trust erodes. Consistency matters more than capability.
• Silent defaults. What does the tool do when the worker ignores it? Does it log non-usage? Report to a manager dashboard? Workers detect surveillance faster than product teams expect.

A productivity tool that makes productive work harder has confused its buyer for its user.

Sabotage as signal

The workers who can't fix these problems through official channels have found another feedback mechanism.

WRITER and Workplace Intelligence's 2026 survey of 2,400 employees and C-suite leaders found that 29% of employees admitted to actively undermining their company's AI strategy. Entering proprietary data into public tools. Deliberately generating poor outputs to discredit the technology. Refusing mandated platforms. Among Gen Z, the figure was 44%.

The instinct is to read this as a people problem. It isn't. Sabotage is the feedback mechanism when official channels fail. If a third of your workforce is working against the tool, the tool has a product problem. The workers who sabotage and the executives who admit their strategy is performative are describing the same reality through different channels. The difference is that only one group's description reaches the quarterly business review.

The EHR precedent

Electronic health records walked this path first.

Administrators chose EHR systems based on compliance requirements, vendor specifications, and IT architecture. Clinicians found them so burdensome that they spent one to two hours on documentation for every hour of patient care, while physician users graded EHR usability an F on standardised scales, according to Yale's summary of an AMA-led study. The buyers were satisfied because the system met regulatory mandates. The users were burning out because the system ignored the rhythm of clinical work.

It took nearly a decade for the market to respond with clinician-centred alternatives.

Enterprise AI is on the same trajectory, compressed. EHR vendors had deep lock-in through data migration costs and regulatory entanglement. AI vendors have far less. Capabilities are commoditising fast, and switching costs are low. The correction will come sooner, but the window for getting it right is correspondingly smaller.

The pricing feedback loop

Per-seat licensing completes the misalignment. When the vendor charges per seat, the incentive is deployment breadth, not usage depth. A thousand seats sold. Two hundred active. The vendor still gets paid. No market signal to fix usability.

Consumption-based pricing would create that pressure. If revenue tracked actual usage, vendors would have a financial reason to care about latency, error recovery, and workflow fit. Enterprise sales teams resist this because it makes revenue unpredictable. But per-seat pricing is a subsidy for phantom adoption: it lets vendors profit from tools that look deployed without being used.

Governing phantom value

Executives genuinely need visibility and governance. Some daily friction is a fair price for organisational control and compliance. True. But when MIT NANDA's 2025 GenAI Divide report puts stalled generative AI projects at roughly 95% and PwC's 2026 CEO survey shows many CEOs still struggling to turn AI investment into measurable returns, the current balance between governance and usability is plainly wrong. You are governing phantom value.

What boring gets you

The tools that win long-term adoption will not have the best demo or the most impressive executive dashboard. They will be fast, predictable, woven into existing workflow, invisible when working, recoverable when they don't. Boring.

AI capability is commoditising. Models get cheaper and more interchangeable every quarter. The only durable advantage left is workflow fit: building the tool so precisely around the worker's actual day that switching to a competitor would mean losing something they rely on, not something they were told to use.

Every AI vendor will face the same reckoning the EHR market faced a decade too late: are you building for the person who writes the cheque, or the person who opens the app?