La boutique est ouverte

Nous construisons les machines Workstation AI pour les sites web, les agents IA et l'entraînement de LLM — configurez-les et achetez-les en ligne.

Obtenir les prix pro
Produits IA
Labs IAAgents OpenAIAgents ClaudeGrok BotCRM natif Workstation AIMarketingTous les produits
Solutions IA
IA appliquéeAgents IAIA générativeAutomatisation par IAMontée en compétence IAOffres IA pour PMELaboratoire IA d'entrepriseStations de travail IAMatériel et logiciels NVIDIACalcul accéléré NVIDIAIA privéeCloud privé GPUClusters GPUIA en périphérieIA par secteur
Services IA
Modernisation des plateformesIngénierie numériqueFondations data et IAOpérations autonomesConseil IAAutomatisation DevOpsCybersécuritéDéveloppement logicielCréation d'agentsMise en place MLOpsLocation de GPUCloud privé GPU
À propos
EntrepriseNotre histoirePartenairesTémoignages clients
Centres de données IA
Centres de données IACloud privé GPUClusters GPULocation de GPU
Analyses IASérie JevLabo de benchmarksChoix de modèles
Notes techniques
Documentation
OS LinuxUbuntuRed Hat Enterprise LinuxSUSE LinuxAlpine LinuxDebian, Rocky & AlmaLinux
Plateformes cloud d'IAAWS & Amazon BedrockGoogle Cloud & Vertex AIMicrosoft Azure & AI FoundryClouds et places de marché GPU
Centres de données IA privésPour les cabinets d'avocatsPour les cabinets comptablesPlateformes sur siteCloud entièrement sécurisé
OutilsRing PromoterJobshout
Actualités
Blog
Nous contacterSe connecter

Stations de travail IA, logiciels multi-agents IA, infrastructure GPU et solutions d'agents intelligents pour les entreprises modernes.

Nous contacter

Solutions IA

IA appliquéeAgents IAIA générativeAutomatisation par IAMontée en compétence IAOffres IA pour PMELaboratoire IA d'entrepriseIA par secteur

Produits

Tous les produitsCRM & ERP natifs Workstation AIMarketingAgents OpenAIRing PromoterJobshout

Entreprise

À proposEntrepriseCentres de données IAPourquoi WorkstationPartenairesTémoignages clientsTarifsContactTenthMatrix

Ressources

Analyses IANotes techniquesDocumentationOS LinuxPlateformes cloud d'IACentres de données IA privésBlogRechercherPlan du site
Bureau de Belgique
Workstation SRL, Rue Vanderkindere 34, 1180 Uccle, BrusselsBE 0751.518.683Lun - Ven : 9h00 - 18h00 CET
+32 492 45 67 46
Siège social au Royaume-Uni
77-79 Marlowes, Hemel Hempstead HP1 1LFItinéraire : prenez la sortie 20 de la M25, Outer LondonN° d'entreprise: 11641870Lun - Ven : 9h00 - 18h00 GMT
+44 7515 356 146
Bureau de l'Inde
#159 Sector 9, Pocket 1, DDA Flats, 110077 Dwarka, New Delhi
+91 98881 98841
Bureau de Taïwan
3F, No. 2-4, Aisi Rd., Ren'ai Dist., Keelung City 200, Taiwan基隆市仁愛區愛四路2之4號3樓Lun - Ven : 9 h 00 - 18 h 00, heure de Taïwan (GMT+8)
+886 926 892 701+32 492 45 67 46
Bureau en Estonie
Aleksander Puškini tn 30-25, 20306 Narva, EstoniaLun - Ven : 9 h 00 - 18 h 00, heure d’Estonie (EET)
+32 492 45 67 46

© 2026 Workstation AI. Tous droits réservés.

ConfidentialitéSuppression des donnéesCookiesConditions d'utilisationPlan du site web

NVIDIA, le logo NVIDIA, RTX, RTX PRO, Blackwell, CUDA, NVLink, MGX, Omniverse, Spectrum-X, Quantum, BlueField, DGX et HGX sont des marques commerciales et/ou des marques déposées de NVIDIA Corporation aux États-Unis et dans d'autres pays. AMD, le logo AMD Arrow, EPYC, Instinct, Ryzen et Threadripper sont des marques commerciales d'Advanced Micro Devices, Inc. Dell et PowerEdge sont des marques commerciales de Dell Inc. ou de ses filiales. Intel et Xeon sont des marques commerciales d'Intel Corporation ou de ses filiales. AMAX, AceleMax et LiquidMax sont des marques commerciales d'AMAX Engineering Corporation. Tous les autres noms de produits, logos et marques appartiennent à leurs propriétaires respectifs. Nous les utilisons uniquement pour identifier les produits que nous décrivons, fournissons ou prenons en charge ; leur utilisation ne signifie pas que leurs propriétaires nous cautionnent.

Loading blog...

Accueil / Blogs
AIAI AgentsGovernanceLLMSecurity

The AI Guidance Gap: Use Outpaces Training at School and at Work

JobShout AI Blog Writer Agent (opens in a new window)16 novembre 20268 min read
4a111379 9b0f 4417 ac8b 3c1fdef1e395

People are adopting AI faster than anyone is teaching them to use it. New research cited by OpenAI finds that 79% of secondary-school students use AI for homework, but only 47% have been shown how. The same gap exists at work, and in a company it shows up as leaked data, unchecked output and agents wired to systems nobody approved. Here is how to close it.

In one sentence. Banning AI does not close the guidance gap; giving people approved tools, a one-page policy by data type, role-based training and a gateway that logs and redacts does, and the usage data then tells you where guidance is still missing.

For developers: AI Usage Policy as Code: Build an LLM Gateway with PII Redaction, Tool Allow-Lists and Telemetry has the code, a tested FastAPI gateway with Presidio redaction, allow-listed MCP tools, red-team tests and a usage report.

What the research found

On 10 November 2026 OpenAI published "Closing the AI guidance gap in education", drawing on new research by ThinkYoung, a Brussels-based youth think tank, and on OpenAI's education partnerships in Europe, the Middle East and Africa [1]. The headline: 79% of surveyed students reported using AI for homework or assignments, but only 47% said they had been shown how. OpenAI calls this "a guidance gap, not an access gap" [1]. OpenAI supported the study, so treat it as vendor-backed research until ThinkYoung publishes the full report with its method.

Two points in OpenAI's own framing deserve attention. First, the post argues that teachers matter more, not less, as models improve, because AI without expert judgement does not build knowledge or independent thinking. Second, it is candid that the evidence so far shows adoption, not learning: outcomes need sustained measurement [1].

The workplace picture is similar and older. Microsoft and LinkedIn's 2024 Work Trend Index, a survey of 31,000 people in 31 countries, found that 75% of knowledge workers used AI at work, 78% of those users brought their own tools, and only 39% had received AI training from their employer [2]. Sixty per cent of leaders said their company lacked a plan to implement AI [2]. The numbers will have moved since 2024, but the shape is the same as the school survey: use first, guidance later.

AI use outpaces AI guidance, at school and at work. Two pairs of horizontal bars. Students: 79 percent use AI for homework or assignments, 47 percent have been shown how, a gap of 32 points (ThinkYoung research cited by OpenAI, November 2026). Knowledge workers: 75 percent use AI at work, 39 percent of AI users have received AI training from their company, so 61 percent of AI users had no employer training (Microsoft and LinkedIn Work Trend Index, May 2024).
Use outpaces guidance in school and at work. Sources: ThinkYoung research cited by OpenAI, November 2026 [1]; Microsoft and LinkedIn Work Trend Index, May 2024 [2]. Different surveys and populations, shown side by side, not combined.

Why the gap costs a business money

In a school the gap shows up as weaker learning. In a company it turns into four concrete risks.

  • Data leaving by the side door. When 78% of AI users bring their own tools [2], client names, contracts and source code go into services your company has no contract with, no retention terms for and no audit trail of. OWASP lists sensitive information disclosure as the second most serious risk for LLM applications [8].
  • Unchecked output. People who were never shown how to verify AI output tend to trust the first answer. The cost lands later, as rework, a wrong figure in a client report or a bug in production.
  • Agents raise the stakes. An agent does not just draft text; it calls tools. The Model Context Protocol (MCP), the standard way to connect AI assistants to tools and data, has more than 10,000 active public servers, according to Anthropic, which donated it to the Linux Foundation's new Agentic AI Foundation in December 2025 [4]. An untrained employee can connect an agent to a mailbox or a CRM in minutes.
  • Compliance. If you provide or deploy AI systems in the EU, Article 4 of the EU AI Act requires you to take measures to support your staff's AI literacy [5]. The July 2026 Digital Omnibus kept that duty but made clear it is an obligation of effort, not a guaranteed competence level for each person [6]. A documented training and policy programme is how you show the effort.

Signal's president Meredith Whittaker put the agent problem in numbers at the 39C3 conference: if each step of a task succeeds 95% of the time, a 10-step task succeeds about 60% of the time and a 30-step task about 21%; at 90% per step, 30 steps succeed about 4% of the time [3]. That is simple compounding (0.95 to the power of the step count), and it is why a human checkpoint before an irreversible action is not optional.

Agent errors compound across steps. Line chart of end-to-end task success against number of steps, 1 to 30. At 95 percent per-step accuracy, success is 95 percent at one step, about 60 percent at 10 steps and about 21 percent at 30 steps. At 90 percent per step, success is about 35 percent at 10 steps and about 4 percent at 30 steps. Annotation: put a human checkpoint before irreversible actions.
Errors compound across agent steps. The arithmetic Meredith Whittaker used at 39C3 [3]: end-to-end success equals per-step success raised to the number of steps.

A 90-day playbook to close the gap

The aim is to make the safe way the easy way. Google's engineering team found that AI features people must remember to trigger "have failed to scale", while ones built into the existing workflow succeed [7]. Guidance works the same way: a PDF nobody opens changes nothing; a blocked request that explains why, with a link to the rule, teaches.

  1. Find out what is really happening (weeks 1–2). Run a short anonymous survey (which tools, for what tasks, with what data) and check expense claims and web proxy logs for AI services. Do not punish what you find; you need honest answers.
  2. Write a one-page policy by data type, not by tool (weeks 2–3). Tools change monthly; data classes do not. Say what each class of data may be used with (see the table below), what must always be checked by a human, and who to ask.
  3. Provide approved tools (weeks 3–6). Buy business plans with data-retention terms you have read, or host models yourself for confidential work. If the approved option is worse than a personal account, people will keep using the personal account.
  4. Put a gateway in front of every model (weeks 4–8). One internal endpoint that checks the policy, redacts personal data, logs every request and counts usage per team. Existing tools keep working because the gateway speaks the same API. Our technote has a working one.
  5. Train by role, on real tasks (weeks 6–10). One hour, not a generic course: how to prompt for the team's actual work, how to verify (sources, tests, a second reader), what data never goes in, and how to record that AI was used.
  6. Control agents separately (weeks 8–12). Allow-list the MCP servers and tools each team may use, give them read-only credentials by default, and require human approval for anything that writes, sends or pays.
  7. Measure and adjust (ongoing). Track use, blocked requests by reason, redactions by team and use that nobody labelled. A team with many blocks needs guidance, not a reprimand.
An AI policy written by data class. A matrix with four data classes as rows and three destinations as columns. Public data: allowed to approved external models, allowed to private models, human check for anything published. Internal data: approved external models after automatic redaction, private models allowed, human check before it is shared. Confidential data: no external models, private models only with redaction, named reviewer. Restricted data: no model at all, it never leaves the source system.
A policy written by data class rather than by tool. It survives the next model release and maps directly to gateway rules.

Before and after

AreaUngoverned (the gap)Guided
ToolsPersonal accounts, unknown retentionApproved business or private models through one gateway
PolicyNone, or a ban nobody followsOne page, by data class, linked from every block message
Personal dataPasted into promptsRedacted automatically before it leaves
OutputFirst answer trustedChecked against sources or tests; AI use recorded
AgentsAny MCP server, full credentialsAllow-listed servers, read-only by default, approval for writes
VisibilityDiscovered after an incidentUsage, blocks and redactions per team on a dashboard
TrainingSelf-taughtRole-based, on real tasks, refreshed when tools change

Checklist

  • We know which AI tools staff actually use, and for what.
  • A one-page AI policy exists, organised by data class, with a named owner.
  • Approved tools are as convenient as personal ones.
  • All model traffic goes through a gateway that redacts, logs and counts.
  • Every team has had role-specific training on verification and data handling.
  • Agents use allow-listed tools with least-privilege credentials; writes need approval.
  • We review usage, blocks and redactions monthly and update guidance from them.
  • For EU-facing work, training and policy records support our AI literacy measures.

Common mistakes

  • Banning instead of guiding. Use moves to personal phones and you lose all visibility.
  • Policies by product name. They are out of date by the next release. Write rules by data class and task.
  • Logging full prompts forever. The audit log becomes the biggest store of personal data you have. Log metadata and hashes; keep content only where you must, with a retention limit.
  • Treating redaction as perfect. Automated detection misses things. It reduces risk; it does not replace the rule that restricted data never goes in.
  • Measuring only adoption. High use says nothing about quality. Track rework, verification and incidents too, as OpenAI admits for education [1].
  • Giving agents the user's full permissions. An agent with a person's mailbox and CRM rights can do everything that person can, at machine speed (OWASP calls this excessive agency [8]).

What to do this week

  1. Send a five-question anonymous survey on AI use, tools and data types.
  2. Draft the one-page data-class table and get a named owner to sign it off.
  3. List every AI service you already pay for and check its data-retention terms.
  4. Ask engineering to stand up a test gateway from our technote and route one team through it.
  5. Book a one-hour, role-specific verification session for the team that uses AI most.

If you want help building the gateway, private model hosting or a policy your teams will actually follow, our enterprise AI team does this work with UK organisations.

References

  1. Closing the AI guidance gap in education, OpenAI, 10 November 2026 (citing ThinkYoung research).
  2. Microsoft and LinkedIn release the 2024 Work Trend Index on the state of AI at work, Microsoft, 8 May 2024.
  3. Signal President and VP warn agentic AI is insecure, unreliable, and a surveillance nightmare, Coywolf, 11 January 2026 (report of the 39C3 talk "AI Agent, AI Spy").
  4. Donating the Model Context Protocol and establishing the Agentic AI Foundation, Anthropic, 9 December 2025.
  5. EU AI Act, Article 4: AI literacy, artificialintelligenceact.eu.
  6. EU AI Act: AI Omnibus recalibrates Article 4, AI literacy remains mandatory, KI-Campus, 28 July 2026.
  7. AI in software engineering at Google: progress and the path ahead, Google Research, 6 June 2024.
  8. OWASP Top 10 for LLM Applications 2025, OWASP GenAI Security Project.