Adept AI

• The site frames data exclusively as a proprietary enterprise asset: 'Proprietary agent training data — Trillions of tokens specific to web UIs and real software usage.'
• Data sovereignty is reduced to customer-data protection ('protecting our customer data, models, and products'), a commercial-confidentiality frame, not a community-rights frame.

• No acknowledgment of Indigenous data sovereignty or the CARE Principles (Collective Benefit, Authority to Control, Responsibility, Ethics).
• No mention of consultation with Indigenous communities or of any non-corporate provenance for the 'trillions of tokens'.
• No recognition of embodied, relational, or oral knowledge; the entire knowledge model is textual/UI-screen interaction data scraped from web software.

The lens is entirely absent. Worse than mere silence: the explicit framing of trillions of tokens as 'proprietary' with no provenance is precisely the extractive data posture the lens is designed to surface. Score floored at 1.

• The only historical gesture is a founding-mythology pull-quote celebrating talent flight from incumbents: 'A wave of AI experts left Google, Deepmind and Meta — the race is on.'
• History is invoked solely as competitive-race narrative ('the race is on'), erasing the material and labor history behind the technology.

• No acknowledgment of colonial or extractive data legacies underpinning web-scale UI corpora.
• No transparency about GPU supply chains, compute geopolitics, or data-labor economics behind 'trillions of tokens'.
• No discussion of regulatory constraints, or of Adept's own contingent history (a startup that lost co-founders and was substantially absorbed via Amazon's 2024 licensing/talent deal).

History appears only as press-clipping triumphalism. The actual deep history (compute political economy, data labor, the company's own near-collapse) is wholly absent and contradicted by the 'race is on' framing. Score 1.

• Page metadata declares 'language: en'; the product is presented monolingually in English.
• 'Localization' appears once, but as a technical agent capability (locating UI elements on a screen), not linguistic or cultural localization: 'localization, web understanding, and planning.'

• No multilingual support, even token-level, is claimed.
• No engagement with culturally specific reasoning patterns; 'web understanding' presumes a universal, implicitly Anglophone enterprise-software world.
• No consultation with cultural or area scholars; 'enterprise workflows' are treated as culturally neutral.

The one occurrence of 'localization' is a false friend — it denotes screen-element location, not cultural/linguistic adaptation. The worldview is monolingual enterprise-Western by default. Score 2, a notch above floor only because 'web understanding' implies some breadth of source UIs.

• Adept publishes named internal benchmarks with numeric scores: 'Adept Locate 93', 'Adept Web VQA 88.2', 'Adept Planning 88' vs 'GPT-4 59'.
• A single comparative datapoint is offered against an external model (GPT-4) on planning.

• All evals are self-defined and self-named ('Adept Locate', 'Adept Web VQA', 'Adept Planning') with no methodology, dataset, or independent replication protocol.
• No third-party audits of training data or bias; weights are closed ('proprietary model'), precluding external verification.
• No disclosure of known limitations, error modes, or failure rates — only success framing ('more accurate and reliable than competitors').

Numbers are present but are marketing benchmarks, not science: self-authored metrics, no methods, no reproducibility, closed weights, zero limitation disclosure. The presence of any quantification lifts this slightly above the indigenous/history floors. Score 3.

• The visual identity gestures at the abstract/affective — 'A top down view of an abstract machine, a ball is seen in a circular groove' — the one moment of non-instrumental imagery.
• Otherwise the register is purely operational: intents, actions, workflows, eval scores.

• No acknowledgment of affective or intuitive dimensions of work being automated.
• No space for ambiguity or uncertainty; everything is rendered as 'reliable', 'future-proof', 'no maintenance'.
• No recognition of the emotional or qualitative texture of the human labor these agents replace.

Aesthetics are decorative, not epistemic. 'User intents directly into actions' collapses all interior ambiguity into frictionless execution. The lone abstract-machine image is mood, not meaning. Score 2.

• The future is modelled explicitly as automation of human work across whole organizations: 'Unlock value across your organization—empowering every department, from HR to Operations.'
• Three concrete deployment futures are sketched (supply-chain, financial services, healthcare), including one keeping 'a human in the loop'.

• Labor displacement is reframed only as upside ('empowering every department'); no engagement with workforce displacement risk despite tagline 'AI that powers the workforce'.
• No environmental or compute-cost disclosure whatsoever.
• No democratic or participatory governance of agentic systems that 'automate any software process'; governance is reduced to a customer 'Trust Center'.

Whose futures? Enterprise buyers'. The displaced worker — central to an 'automate any software process' agent vendor — is invisible except as 'empowered'. One human-in-the-loop mention is the sole safety gesture. No environmental or democratic dimension. Score 2.

• The only named stakeholders are enterprises, customers, departments (HR, Operations), and the founders' prestigious former employers.
• Human presence is limited to 'a human in the loop' as a control checkpoint, not as a represented voice.

• No participatory design with Global South developers or affected workers.
• No disability-community accessibility commitments (ironic for a UI-automation product whose actuation layer could serve accessibility).
• No labor-representative engagement or compensated feedback channels; 'feedback & data collection tools' are framed as model-improvement instruments, not stakeholder voice.

No marginalised constituency appears at all. 'Feedback' means data extraction for model improvement, not a seat at the table. The accessibility opportunity latent in UI actuation is entirely unmentioned. Score 1.

• The site treats its own narrative as exempt from any tension; no contradiction is named or held.
• The richest unspoken irony — 'AI that powers the workforce' is sold by automating the workforce's tasks across 'every department' — is left entirely unacknowledged.

• No willingness to name the displacement contradiction beneath 'empowering every department'.
• No irony, self-test, or inversion; 'future-proof… resilient to changes… require no maintenance' is asserted with zero self-doubt about agents acting on live enterprise systems.
• The company's own contingency (lost co-founders, Amazon absorption) is a glaring inversion of the 'race is on' triumph the page leans on — and is unmentioned.

Zero trickster capacity. Marketing solemnity is total; not one contradiction is permitted to surface, least of all the workforce-empowerment-via-workforce-automation paradox at the page's core. Score 1.