UAE employees save 3.6 hours a week with AI, then spend 3.4 hours fixing its output
Workday's UAE research shows weekly AI use at 90%, but 3.4 hours of weekly rework nearly cancels the 3.6 hours saved. The gap is an engineering problem.
What happened
Summary of reporting by Khaleej TimesWorkday opened a Dubai office with dedicated regional operations on 1 October 2026 and, to mark the launch, released local findings from its report "Beyond Productivity: Measuring the Real Value of AI in the UAE". The survey covered 500 full-time AI users at companies with 1,000 or more employees, fielded by Censuswide in June and July 2026. Among UAE employees, 90% now use AI at least weekly, 58% use it multiple times a week and 18% use it on every working day. More than half of directors and C-suite executives use AI weekly. The reporting is covered by Khaleej Times (https://www.khaleejtimes.com/business/uae-employees-save-36-hours-a-week-using-ai-but-spend-34-hours-fixing-its-work) and Middle East AI News (https://www.middleeastainews.com/p/workday-opens-dubai-office-cites).
The productivity picture is thinner than the adoption picture. Respondents average 3.6 hours a week saved through AI against 3.4 hours a week spent clarifying, correcting or rewriting poor-quality output. Workday frames the result as evidence that trust and human oversight are the bottleneck. Two caveats: the two figures are separate survey averages, not a per-worker net calculation, and they describe average behaviour, not a budget.
“nearly 40% of AI time savings are lost to rework, and only 14% of employees consistently see clear positive net outcomes”
Workday, Beyond Productivity global research (3,200 employees)
Where the savings go is the more encouraging part. Only 24% of organisations primarily put the gains to cost reduction. Thirty-three percent use the efficiencies to absorb greater workloads without adding headcount, 32% direct the gains into training and reskilling, and a further 32% invest in workload management, flexibility and wellbeing. Just 0.2% of respondents said AI had produced no meaningful time or cost savings.
The UAE data extends Workday's global study of the same name, a survey of 3,200 employees at US$100m+ revenue organisations fielded by Hanover Research in November 2025, which found nearly 40% of AI time savings lost to rework and only 14% of employees consistently getting clear positive net outcomes.
The Azrty take
UAE organisations have solved AI adoption and now face an output-quality engineering problem that training alone will not fix.
For UAE organisations the persuasion phase of AI is over: nine in ten employees already use it weekly, and more than half of directors and C-suite executives do too. What the Workday data exposes is the second phase nobody budgeted for: assurance. At 3.6 hours saved against 3.4 hours of rework, the net benefit is close to zero for the average worker, and any business case built on raw time savings will not survive contact with these numbers. This mirrors Workday's global study of 3,200 employees (fielded by Hanover Research, November 2025), which found nearly 40% of AI time savings lost to rework and only 14% of employees consistently seeing clear positive net outcomes. The UAE sample sits at the harsh end of that global picture, which should worry leadership teams in a market that expects to be ahead.
The instinctive board response will be to fund training. That is half right and mostly wrong. Rework exists because staff use generic, context-free chat tools and then manually re-inject everything the organisation already knows: brand rules, product facts, policy, code standards. Output quality is low because the input has no grounding, no versioned prompts and no gate between generation and delivery. Workday's own launch release (https://investor.workday.com/news-and-events/press-releases/news-details/2026/Workday-Launches-in-the-UAE-to-Help-Organisations-Transform-HR-Finance-and-IT-in-the-AI-Era/default.aspx) frames the answer as context, guardrails and trusted processes. Its global research release (https://investor.workday.com/news-and-events/press-releases/news-details/2026/New-Workday-Research-Companies-Are-Leaving-AI-Gains-on-the-Table/) makes the point more bluntly: nearly 40% of savings lost to rework, and 77% of daily users reviewing AI output as carefully as human work.
The redistribution figures carry a warning. The 33% of organisations absorbing more work without adding headcount are the most exposed: same reviewers, more volume, same review burden. That is how the rework tax scales silently. Meanwhile only 24% are primarily cutting costs, which means the ROI story most CFOs signed off on (headcount arithmetic) is not the story the data tells. Regulated sectors here, government, banking and telecoms, face a harder version: review is not optional, it is an audit requirement, so every ungoverned prompt that reaches a customer is a compliance surface.
How we would approach it: stop treating the chat window as the product. Route all AI traffic through one gateway so routing, budgets and access control live in one place and every call is attributed. Ground outputs in a curated knowledge base rather than model memory. The canonical result behind this is the RAG paper (Lewis et al., 2020, https://arxiv.org/abs/2005.11401), which found that generation grounded in a retrieved, external document index is more factual and more specific than generation from the model's own parameters. Keep prompts as versioned, shared assets instead of everyone improvising their own. And put blocking gates between AI output and anything a human signs off. FastLLM Proxy (https://azrty.com/software/fastllm-proxy) is the gateway. Kryton (https://azrty.com/software/kryton) is the self-hosted knowledge base your team and your AI assistants read and write together. PromptForge (https://azrty.com/software/promptforge) is the versioned, shared prompt library. Procoder (https://azrty.com/software/procoder) is the quality gate for AI coding work: untested, unformatted or unfinished work does not get to call itself done. A gate set for code looks like this:
# Procoder: work that fails a gate never reaches a human reviewer
gates:
- id: tests-green
blocking: true
- id: linter-clean
blocking: true
- id: scope-complete # no unfinished files or TODO stubs
blocking: true
- id: diff-coverage # every changed line is touched by a test
blocking: true
The same principle applies outside code: for documents, generate against source-grounded context and run a citation check before anything reaches a reviewer. RAGAS (Es et al., 2023, https://arxiv.org/abs/2309.15217, repo https://github.com/explodinggradients/ragas) automates exactly this as a reference-free faithfulness check: it decomposes the answer into individual statements and verifies each against the source context, so a document only passes when its claims are actually grounded.
Where most teams will get it wrong: they will buy more AI seats or switch models, blame staff skill, or both. Neither reduces the 3.4 hours. The single biggest miss is measurement. In our experience, almost every organisation tracks hours saved by AI; almost none tracks hours spent correcting it. Until rework is a tracked metric next to savings, the productivity paradox stays invisible in the dashboards and visible only in the calendar.
What to do now
- Add rework to the dashboard: from this month, log hours spent clarifying, correcting or rewriting AI output alongside hours saved, per team. The 3.6 versus 3.4 split in the Workday data is invisible unless you measure both.
- Route employee AI traffic through a single gateway (FastLLM Proxy is built for this) so routing, budgets and access control live in one place, and you can see which tasks and teams generate the most rework instead of guessing.
- Move the prompts that survive review into a shared, versioned library (PromptForge) so staff stop reinventing context-free prompts that produce low-quality output.
- For AI-assisted code, enforce blocking gates (tests, linting, scope completeness via Procoder) so untested agent output never reaches a human reviewer and never becomes rework.