AI for accessibility: why human expertise still matters

Posted on 6th May 2026

Written by Alan Blanchard, Business Development Director, TSO

At Digital Leaders AI Week 2026 TSO shared how combining AI with expert oversight can both improve productivity and manage the risks that undermine trust in AI.

An expert‑in‑the‑loop approach ensures organisations gain the productivity benefits of AI without compromising accuracy:
AI handles tasks at speed and scale
Specialists validate outputs and apply professional judgement to avoid hallucinations
Corrections and feedback continuously improve model performance
Decisions remain transparent, auditable and defensible.

Why AI alone isn’t enough for PDF accessibility

Document accessibility is a good example. Accessibility is both a legal requirement and a moral responsibility, yet many PDFs remain inaccessible. Automated tools can process and tag documents faster than humans, but accuracy and usability still cannot be guaranteed.

When remediating PDFs for accessibility, there are critical checks that only humans can reliably make, including:

Ensuring alternative text accurately describes images
Checking the tag tree has a logical reading order and matches the visual flow of the page
Confirming links are active and point to the right location with the right description
Verifying text contrast on gradients or images, not just on solid backgrounds
Ensuring table headers correctly describe and associate with the data beneath them
Manually tagging foreign‑language phrases so screen readers switch language correctly
Correcting mis‑tagged headings, paragraphs or artefacts
Validating form field tooltips and instructions for clarity and relevance
Checking whether specialist terminology and acronyms are understandable to end users, adding screen‑reader call‑outs where needed

Intelligent PDF accessibility at scale with RemediAd

Our AI‑driven accessibility tool, RemediAd, applies structural tagging and identifies accessibility issues across large document volumes in seconds. A 50‑page PDF can be processed in under a minute, accelerating remediation at scale.

AI performs the initial tagging and issue detection. Accessibility specialists then review and refine the output, focusing on areas automated tools cannot reliably assess. This includes validation through real-world user testing, with individuals with vision impairments (ranging from partial to complete blindness) using screen readers to ensure content is genuinely accessible. Our teams are expert in document accessibility remediation, with deep knowledge of global accessibility standards.

RemediAd is designed specifically for enterprise documents, with customised templates available for different document types. This enables efficient batch processing while maintaining quality and consistency across large collections of PDFs.

Our approach shows that combining AI efficiency with expert validation, organisations can achieve:

Accurate and meaningful alternative text
A logical reading order that reflects the visual structure of the page
Clear, descriptive link text
Appropriate colour contrast, even on complex backgrounds
Correct table header associations
Accurate language tagging
Clear and relevant form labels and instructions

The result is accessible content that works in real‑world assistive technology use – not just content that passes automated checks.

Productivity without compromise

AI presents a clear opportunity to remove repetitive, time‑consuming work and allow skilled professionals to focus on higher‑value tasks. At organisational scale, this improves efficiency, consistency and outcomes.

However, risks such as hallucinations, bias and poor data quality must be actively managed. Left unchecked, they erode trust and without trust, responsible AI adoption stalls.

In our approach, this risk is mitigated by using purpose-built AI designed specifically to understand the structure of complex digital documents. Rather than generating or altering content, the AI focuses solely on analysing and organising existing information, identifying elements such as headings, lists and tables, even in inconsistent formats. This ensures accuracy, maintains the integrity of the original content, and supports secure, reliable downstream processes.

TSO’s expert‑in‑the‑loop model ensures AI supports people rather than replaces judgement, delivering document accessibility and PDF accessibility that is accurate, compliant and reliable.

To find out more about how AI and expert human oversight can improve other processes such as transcription and information extraction, you can watch our Digital Leaders webinar AI and Human in the Loop.

Originally posted here