AI captioning vs. human captioning for L&D: when each workflow is right, what it costs per minute, and how to build the hybrid model that meets WCAG without overspending

The question buyers actually have — and why it is usually framed wrong

L&D teams searching “AI captioning vs. human captioning” are at a specific moment in the procurement process: they have a compliance requirement (ADA Title II, WCAG 2.1 AA, internal policy, or an accommodation request they cannot fulfil with YouTube auto-captions), they have a budget constraint, and they have not yet committed to a workflow or a vendor. They want a single answer that resolves the tension between cost and quality.

Vendors want to give them that single answer, because a single answer makes the sale easier. AI vendors say: AI is accurate enough and dramatically cheaper. Human captioning vendors say: human captioners catch what AI misses and the cost difference is worth it. Neither answer is wrong for the vendor’s best-fit customers. Both answers are wrong when applied as a blanket recommendation to a mixed-content L&D library.

The productive reframe is to treat “AI or human?” as a content-type routing question, not a programme-level binary choice. The architecture question is: given the content types in your library, the accuracy requirements for each type, and the volume and budget available, what is the workflow assignment that minimizes total programme cost while meeting compliance requirements for every video?

That question has a specific, calculable answer for most L&D libraries. This post provides the framework to calculate it.

Why this matters more after the ADA Title II deadline

Before the April 2026 ADA Title II enforcement deadline, the cost-vs-quality tradeoff was a procurement preference. After the deadline, it is a compliance risk calculation. Content that is captioned but not compliant — AI captions below 99% WER on mandatory training video — does not satisfy the requirement. And as the post on why 99% accuracy matters explains, the gap between “captioned” and “compliant” is not a rounding error for most AI workflows on technical content: it is a systematic 10–20 percentage point shortfall on the word categories that carry the most meaning in training content.

The practical implication: an L&D team that routes all content through AI-only captioning and reports 100% caption coverage may have a compliance gap of 30–40% if the AI output on technical content falls below 99% WER. A team that routes content correctly and achieves 99%+ on 100% of mandatory content has a much stronger compliance position even if it has a higher per-minute spend on a subset of its library.

The content-type decision matrix: what determines the workflow threshold

The single most important input to the AI-vs-human workflow decision is content type. AI captioning accuracy varies by approximately 15–20 percentage points across content types, which is enough to determine whether AI-only meets the 99% WER threshold without additional human review. Here is the framework for classifying content and assigning the appropriate workflow.

Tier A: Talking-head training video — AI-only viable

Talking-head training video — a speaker on camera or voice-over narration against slides, recorded in a controlled environment — is the best case for AI captioning. Characteristics: single speaker, controlled acoustic environment, general vocabulary, no heavy product-name density, clear diction, and predictable sentence structure.

Baseline AI accuracy for this content type on modern ASR systems (Whisper large-v3, Google STT, AWS Transcribe): 95–98% WER in English. With vocabulary tuning and a domain glossary, achievable accuracy is 97–99.5% on trained terms. This is within range of WCAG 2.1 AA compliance for a significant portion of the library.

Workflow assignment: AI-only with automated QA sampling. A 10% random sample reviewed by a human QA reviewer each quarter is sufficient to catch systematic failures without reviewing every video individually. The caption vendor accuracy evaluation methodology covers the sampling protocol in detail.

Cost range: $0.006–$0.012/min API cost + $0.05–$0.10/min amortized QA sampling cost = $0.056–$0.112/min total programme cost.

When to escalate: If the speaker has a strong regional accent, if the content includes a significant volume of acronyms or technical terms not covered by the domain glossary, or if quarterly QA sampling shows accuracy falling below 97%, escalate affected videos to Tier B (AI+human-review).

Tier B: Technical training with moderate vocabulary complexity — AI+human-review required

This is the largest category for most enterprise and mid-market L&D teams. It includes: product training for SaaS companies (product names, feature names, UI navigation), engineering onboarding (SDK names, API endpoints, library names, error codes), compliance training with regulatory terminology, and medical or clinical training at sub-specialist level (drug names, procedure names, anatomy terms that are known but not common).

Baseline AI accuracy without vocabulary tuning: 83–90% WER. The failure pattern is highly systematic: general vocabulary accuracy is 96–98%, but product names, SDK symbols, and regulatory acronyms are the high-error categories. A training video that mentions your product’s API endpoints twenty times will have those endpoint names wrong in the AI output for every occurrence, and those are the words learners most need to read correctly.

With domain glossary tuning, accuracy on trained terms rises to 94–99%. The remaining error rate is distributed across untrained terms and edge cases. At the 94–97% total accuracy range, human review is still required for WCAG compliance, but the volume of corrections per video drops from 20–40 corrections per hour to 5–12, which materially changes the cost of human review.

Workflow assignment: Glossary-biased AI captioning followed by human light review. The human reviewer corrects residual errors rather than building captions from scratch. The caption feedback loop post covers how this review process should feed back into the glossary to improve AI accuracy over time.

Cost range: $0.008–$0.012/min API cost + $0.80–$1.50/min light human review (with glossary-aided AI, correction volume is lower) = $0.81–$1.51/min total programme cost. Without glossary tuning: $0.008–$0.012/min API + $1.50–$2.50/min heavy human review = $1.51–$2.51/min, which approaches human-only cost without the quality benefit of a human captioner who understands the domain.

When to escalate: If the human review consistently takes more than thirty minutes per video hour (indicating AI output quality is too low for light review to be efficient), escalate to Tier C. If QA sampling shows that human-reviewed output consistently reaches 99%+ at the light-review cost level, keep at Tier B.

Tier C: High-stakes content with complex or specialized vocabulary — human-primary

This category includes: specialist medical or clinical training at subspecialty level (content where a wrong term is a patient safety issue, not just an accuracy metric), legal compliance training where term precision matters in a regulatory context, live event captions (real-time ASR only; post-production review is not applicable to live), and accommodation-required content for specific learners who have documented that auto-captions are insufficient for their needs.

AI captioning accuracy on subspecialty medical content without heavy domain training: 73–82% WER. This is the WER range where AI output is not a starting point for light review — it is, in places, so wrong that a reviewer who is not a domain expert cannot catch the errors. “Propofol” becomes “pro-pho-fall.” Drug interaction terminology becomes garbled sequences of phonetically similar words. In this content type, AI captioning without domain-expert review is not a cost optimization — it is a patient safety or compliance risk that no organization should be taking.

Workflow assignment: Human-primary captioning, with an AI draft as a starting-point option only if the human captioner has the domain vocabulary to evaluate every correction. For live events, real-time AI CART (Communication Access Realtime Translation) or human CART depending on the vocabulary complexity and stakeholder requirement. The in-house vs. vendor caption team decision post covers how to evaluate whether to build internal specialist captioning capacity or outsource it.

Cost range: $1.50–$3.50/min for professional captioning (domestic US rates); $0.80–$1.20/min offshore with quality controls. Live CART: $1.50–$2.50/min for automated CART, $4–$8/min for human CART at specialist events.

When to consider AI+review: If a domain-specialist glossary can be built (i.e., you have sufficient vocabulary data from prior captioned content in this specialty), glossary-biased AI can raise accuracy enough for Tier B treatment even on some complex medical content. This takes approximately six months of production volume to develop a glossary that covers enough of the vocabulary to make the lift meaningful. New programmes without that history should default to Tier C for specialist medical content.

Special case: screen recordings and software demos

Screen recordings deserve a separate note because they are the most common content type in SaaS company L&D libraries and the worst-performing category for standard AI captioning. The ASR accuracy challenge is structural: screen recordings typically combine narration (high ASR accuracy) with on-screen text that the narrator reads aloud (UI labels, menu paths, product names, error messages). The on-screen text categories are exactly the high-substitution-error categories where AI systems perform worst.

Baseline accuracy on a narrated screen recording for a SaaS product: 84–90% WER. The narrator says “click the Integrations tab then select OAuth provider” and the AI output produces “click the integrations tab then select all with provider.” For a software demo, this is not a minor accuracy issue — it is a product name and UI path error that actively teaches learners the wrong thing.

With a product glossary (all feature names, UI labels, navigation paths, product brand terms loaded as vocabulary bias), accuracy on trained UI terms rises to 93–98%. The residual error rate is on edge cases and untrained terms. This brings screen recordings into Tier B range, but the glossary investment must come first. Without a product glossary, screen recordings default to Tier C even though the narration itself would qualify for Tier A.

The cost model: what each workflow actually costs per minute

The numbers below represent total programme cost, not just API or vendor invoice cost. Total programme cost includes: API or vendor invoice, internal review labour, QA sampling, and the amortized cost of glossary development and maintenance. The hidden internal correction cost is the number most L&D teams are not tracking, and it is the number that most changes the apparent cost advantage of AI-only captioning.

Component 1: AI API or vendor invoice cost

AI captioning API cost (Whisper, Google STT, AWS Transcribe, AssemblyAI) varies by provider and volume tier, but the working range for most L&D programmes is:

• $0.006–$0.009/min for self-hosted or batch-processed API calls at volume
• $0.009–$0.012/min for standard API pricing at low to mid volume
• $0.012–$0.025/min for premium providers or real-time processing

A GlossCap-style glossary-biased captioning service that sits on top of the API layer (handling vocabulary tuning, format conversion, LMS delivery, and QA workflow) typically prices in the $0.03–$0.08/min range at the Team-plan volume level, which includes the glossary infrastructure, the delivery pipeline, and the edit UI. That is still substantially below the human captioning floor.

Component 2: Internal correction labour — the number most programmes hide

This is the cost that changes the apparent AI cost advantage, and it is the cost the hidden half-FTE post identified as the most systematically underreported line in L&D budgets.

Without vocabulary tuning, L&D teams that use AI captioning and then correct the output internally report an average of 45–90 minutes of correction time per video hour. At an all-in L&D coordinator rate of $55–$75/hour (salary + benefits + overhead), that correction time costs:

• 45 minutes correction / 60 minutes video = 0.75 hours correction per hour of video
• At $65/hour: $0.75 × $65 = $48.75 per video hour = $0.81/min
• At 90 minutes correction: 1.5 × $65 = $97.50/hour = $1.63/min

Adding this to the API cost of $0.008/min: total programme cost is $0.82–$1.64/min. That is already within the range of professional human captioning at $1.50–$2.00/min, and the comparison does not include the time cost of the internal QA workflow, the cost of errors that were not caught, or the compliance risk of content that falls below 99% WER despite the correction effort.

With vocabulary tuning (glossary-biased AI), correction time drops to 15–30 minutes per video hour on Tier B content. At $65/hour: $16.25–$32.50 per video hour = $0.27–$0.54/min. Adding API and glossary infrastructure cost: $0.33–$0.62/min total. This is materially below human captioning cost and the correction volume is manageable for a programme at scale.

The caption quality error rate calculator provides a framework for estimating your specific correction cost based on your content mix, error rate profile, and internal labour rate.

Component 3: Human review cost — light vs. full review

When AI+human-review is the appropriate workflow (Tier B content), the human review cost depends on whether the AI output is used as a correctable draft (light review) or as a guide for independent re-captioning (full review):

• Light review (AI output is the primary draft; human corrects residual errors): $0.80–$1.50/min for outsourced professional review; $0.40–$0.80/min for trained internal staff with domain knowledge.
• Full review (AI output quality is too low to trust as a draft; human re-captions with AI as a reference): $1.50–$2.50/min outsourced; $0.80–$1.20/min internal. At this cost level, the AI step is adding minimal value and the programme should evaluate whether to remove it.
• QA sampling only (AI output meets 99% threshold without per-video review; human samples 10% quarterly): $0.05–$0.15/min amortized across the library. This is the appropriate model for Tier A content that has cleared the accuracy threshold.

Component 4: Human-primary captioning (Tier C)

Professional human captioning rates vary by provider type, turnaround time, and content complexity:

• Domestic US professional captioners (NCRA-certified, capable of specialist medical vocabulary): $2.00–$3.50/min at standard turnaround (48–72 hours); $3.50–$6.00/min at rush (same-day or 24-hour).
• Mid-tier professional services (quality-verified, not specialist-certified): $1.50–$2.50/min standard; $2.50–$4.00/min rush.
• Offshore captioning with QA: $0.80–$1.20/min. Lower cost, but quality controls need to be built into the procurement structure — the vendor SLA and contract review checklist covers what to verify before signing an offshore captioning agreement.
• In-house captioners: All-in cost (salary + benefits + overhead + software + QA overhead) typically $1.20–$2.00/min at the volume most mid-market L&D teams produce. Viable for high-volume programmes with stable content types; rarely cost-effective for programmes under 100 hours/month.

The cost comparison, apples-to-apples

Here is the cost summary across workflow types, using conservative and aggressive estimates:

The counterintuitive finding: AI-only captioning on Tier B content without glossary tuning or external review is not cheap — it is expensive, because the internal correction cost makes it comparable to mid-tier human captioning, and the quality is worse because internal correctors who are not trained captioners miss systematic errors that professional reviewers catch. Glossary-biased AI + light professional review is the workflow that achieves both the cost advantage and the accuracy requirement for Tier B content.

For the budget planning framework that integrates these cost components into a multi-year programme budget, see the caption programme budget planning guide.

Building the hybrid model: routing content by tier

The hybrid captioning model is not a compromise between AI and human captioning — it is an explicit routing architecture that assigns each content type to the appropriate workflow based on a set of decision criteria. Most L&D programmes arrive at something like a hybrid model by accident (some content goes to the AI tool, some goes to a human vendor, some gets corrected internally). The difference between an accidental hybrid and a designed hybrid is documentation, consistency, and cost predictability.

Step 1: Classify your library

Before building the routing architecture, classify your existing and future-production content into Tier A, B, and C using the decision matrix above. The classification criteria are:

• Content complexity: Is the vocabulary domain-specific with a high density of product names, technical terms, or regulatory terminology? (Tier B or C) Or is it general vocabulary in a controlled recording environment? (Tier A)
• Stakes: Is this mandatory compliance content, content for regulated industries, or accommodation-required content for a specific learner? (Tier B or C minimum) Or is it general reference content with low compliance exposure? (Tier A acceptable)
• Audience: Does the audience include learners with documented accommodation needs for high-accuracy captions? (Human-primary for at least that content) Or is the audience internal, general-learner, with no specific documented accessibility requirement? (AI+review acceptable)
• Volume: How many hours per month does this content type represent? High-volume Tier A content justifies glossary investment because the savings compound. Low-volume Tier C content may not justify building an in-house captioning capability even if the per-minute cost is high.

The output of the classification is a routing table: content type → workflow assignment → vendor or tool → cost estimate. This becomes the operating document for the caption programme.

Step 2: Build the glossary infrastructure first

The hybrid model only works if the AI tier has vocabulary tuning. Untuned AI on Tier B content produces enough errors that the correction cost eliminates the cost advantage. The glossary is the investment that makes the hybrid model financially viable.

The glossary for an L&D programme at a 200-person SaaS company typically contains 150–400 terms: product name variants, feature names, UI labels, integration names, acronyms specific to the company, names of key customers or partners referenced in training content, and regulatory terms specific to the industry. Building this glossary from scratch takes approximately four to eight hours for a first pass and should be updated quarterly as the product and vocabulary evolve.

As explained in the post on glossary-biased captioning for engineering content, the compounding effect is significant: each captioned hour of training video returns correction data that can be used to extend the glossary, raising accuracy on subsequent hours of similar content. After six months of consistent production through a glossary-biased AI workflow, the correction rate on Tier B content typically drops by 40–60% compared to the first-month baseline.

Step 3: Set the accuracy threshold per tier and monitor it

The routing decision is only useful if the accuracy threshold for each tier is defined and monitored. Recommended thresholds:

• Tier A (AI-only): Must achieve 99%+ WER on DCMP sampling (10% quarterly sample) to remain in Tier A. If quarterly QA shows a content subcategory falling below 97%, escalate to Tier B.
• Tier B (AI+human-review): Human review must bring output to 99%+ WER before delivery to LMS. If light review consistently fails to reach 99% (indicating the AI output quality is too low), escalate to Tier C for that content subcategory.
• Tier C (human-primary): 99%+ WER on delivery. Quarterly QA sampling to verify. If the human vendor consistently falls below 99%, that is a vendor management issue, not a workflow issue — the vendor accuracy evaluation methodology covers how to establish and enforce the threshold in the vendor contract.

The LMS caption analytics post covers how to build the reporting layer that surfaces per-tier accuracy data automatically from your production workflow, so that the tier routing decisions are continuously validated against real data.

Step 4: Document the routing decision for compliance purposes

The hybrid model must be documented in the caption programme policy. An OCR investigation or internal audit that finds different content receiving different captioning treatment — some AI, some human, some reviewed, some not — will ask for the documented rationale. “We route based on content type and vocabulary complexity” is an acceptable answer when supported by a written routing policy and workflow documentation. “It depends on who submits the request and what tool they happen to use” is not.

The programme policy should document: the tier classification criteria, the workflow assignment for each tier, the accuracy threshold for each tier, the escalation criteria, and the QA sampling frequency. This is the documentation that transforms a cost optimization into a defensible compliance programme.

Glossary-biased AI: how it changes the cost model for technical vocabulary

The gap between AI-only and human captioning narrows substantially when AI is vocabulary-tuned, and widens again when the AI system is generic. Understanding how glossary-biased decoding works — and where it reaches its limits — is essential for designing a hybrid model that routes content correctly.

What glossary-biased decoding does

Standard ASR systems use a general language model to select the most probable word sequence given the audio signal. When the audio contains a product name like “Aperture” (a fictional SaaS product with high phonetic similarity to the word “aperture” in photography), the ASR model picks the more common word unless the product name is in the vocabulary. The result is systematic substitution errors on every occurrence of the term.

Glossary-biased decoding works by inserting domain vocabulary into the decoding process: the ASR model is steered to prefer known domain terms when the audio signal is ambiguous between a common word and a domain term with similar phonetics. The effect is not unlimited — the bias can be overridden by audio that is clearly a common word — but for the systematic substitution errors on product names, feature names, and technical terms that are the dominant error type in Tier B content, it is the most effective intervention available.

The numerical effect: without glossary bias, a technical training video with 60 product-specific terms averages 40–55 substitution errors per video hour on those 60 terms (each term appears multiple times, and each occurrence may be wrong). With glossary bias covering those 60 terms, the error count on those terms drops to 3–8 per video hour. The correction effort for those terms falls from approximately 20–35 minutes to 2–5 minutes per hour.

For content types and vocabulary sizes where the glossary covers the dominant error categories, this is the intervention that makes AI+light-review viable as a WCAG-compliant workflow for Tier B content. For content types where the dominant error categories are not vocabulary-driven (heavy accents, poor audio quality, rapid multi-speaker dialogue), glossary bias does not resolve the accuracy gap and the content should be in Tier C.

Glossary coverage and the accuracy ceiling

A glossary that covers 80% of the domain-specific terms by occurrence frequency produces approximately 70–80% of the accuracy gain available from a perfect glossary. The law of diminishing returns applies: the first 100 terms in a well-curated glossary produce the largest accuracy improvement; terms 100–300 produce incremental improvements; terms 300+ produce marginal improvements that may not justify the curation cost.

For a 200-person SaaS company with a moderately complex product vocabulary, a 150–250 term glossary achieves approximately 90–95% of the available accuracy improvement. Building beyond 250 terms is worth doing but should be driven by per-term error rate data from the correction workflow, not by attempting to exhaustively catalog every possible term. The correction log from each video is the primary source of new glossary terms — terms that appear in corrections are by definition terms the ASR is getting wrong, and those are the terms worth adding.

Where glossary bias reaches its limits

Three categories where glossary-biased AI does not close the accuracy gap to the 99% WCAG threshold:

1. Audio quality problems: Background noise, reverb, microphone issues, and recording artifacts reduce ASR accuracy independently of vocabulary. A content creator who records in a room with significant echo will get 80–85% accuracy regardless of glossary quality. The correct fix is recording quality standards, not more glossary terms.
2. Heavy accent + domain vocabulary combination: A non-native English speaker using domain terminology that is also phonetically unfamiliar will produce systematically lower accuracy because the acoustic model and the language model are both working with high uncertainty simultaneously. This is Tier C content even with a strong glossary.
3. Multi-speaker or overlapping dialogue: Speaker diarisation accuracy drops significantly when speakers overlap or when there is a rapid speaker exchange. Training content in this format (panel discussions, Q&A sessions, simulated customer calls) is Tier B at best and often Tier C depending on the acoustic complexity.

The 99% accuracy threshold as a workflow selector

The 99% WER requirement for WCAG 2.1 AA compliance is commonly cited as a target. It functions better as a workflow selector: for content where AI can reliably achieve 99% with vocabulary tuning, AI-only or AI+light-review is the appropriate workflow. For content where AI cannot reliably achieve 99% even with vocabulary tuning, human review is not optional — it is a compliance requirement, not a quality preference.

The specific number matters because the error rate distribution is not uniform. At 98% WER (two errors per 100 words), the errors are random and distributed. At 90% WER (ten errors per 100 words), the errors are concentrated on domain terms and create systematic misconceptions in technical training. At 80% WER, the captions are not just inaccurate — they are actively misleading on the content that matters most to learners.

As the post on 99% accuracy explains in detail, the WCAG threshold is not arbitrary. Below 99%, the caption error rate is high enough that learners with hearing impairments experience a materially different training content than hearing learners — which is the compliance failure the standard is designed to prevent. This is why using YouTube auto-captions (typically 80–90% WER) for mandatory compliance training is a compliance risk, not a cost optimization.

The practical implication for the hybrid model: run a baseline accuracy audit on a sample of each content type before assigning it to a tier. A sample of five to ten videos per content category, reviewed against the DCMP sampling protocol, gives you an accuracy baseline that determines the correct workflow assignment. Without the baseline, tier assignments are guesses.

The feedback loop post covers how to use the correction data from each reviewed video to establish and refine the accuracy baseline over time, so that workflow assignments can be upgraded (from Tier B to Tier A) as the glossary matures and the AI accuracy improves.

Building the business case for the hybrid model

The hybrid model typically produces total programme cost savings of 30–55% compared to a human-only programme and 15–35% compared to an AI-only programme with internal correction labour. The savings argument is not “AI is cheaper than human captioning” — it is “the right workflow for each content type produces the lowest total programme cost while meeting the accuracy requirement.”

The comparison that wins the budget argument

Most L&D teams frame the budget argument as AI cost vs. human captioning cost at the per-minute invoice level. That framing underestimates the true cost of AI-only (missing internal correction labour) and overestimates the true cost of the hybrid model (assuming human review applies to all content, not just Tier B and C).

The correct comparison is:

• Current state: AI-only (or YouTube auto-captions) on all content, with internal correction. Example: 50 hours/month of content, 60% Tier A (talking-head) / 30% Tier B (technical) / 10% Tier C (specialist). Total cost: $0.008/min API on all content + $0.40/min internal correction average = $0.408/min × 3,000 min = $1,224/month in API + labour. Compliance risk: significant — Tier B and C content at 83–90% AI accuracy is not WCAG compliant.
• Hybrid model: 60% Tier A at $0.25/min (AI + QA sample) = 1,800 min × $0.25 = $450. 30% Tier B at $1.20/min (glossary AI + light review) = 900 min × $1.20 = $1,080. 10% Tier C at $2.50/min (human-primary) = 300 min × $2.50 = $750. Total: $2,280/month. Compliance risk: minimal — all content meets 99% WER.
• Human-only: $2.00/min average across all content = 3,000 min × $2.00 = $6,000/month. Compliance risk: minimal.

The hybrid model is 62% cheaper than human-only and has materially better compliance outcomes than AI-only. The business case argument for the CFO is that the hybrid model converts an uncontrolled compliance risk into a documented, measured programme at a cost between the two extremes — and that the compliance risk in the current AI-only state is not theoretical: OCR complaints filed after April 2026 will ask for evidence of 99% accuracy, not evidence of a captioning tool subscription.

For the complete ROI framing for a finance executive audience, including the legal risk quantification and the multi-audience business case, see the ROI framing post.

The cost trajectory: why the hybrid model gets cheaper over time

The hybrid model has an improving cost trajectory because glossary-biased AI accuracy compounds over time. As more content is captioned and corrections are fed back into the glossary, the AI accuracy on Tier B content improves, which reduces the per-video correction volume, which reduces the light review cost. A programme at month six typically produces 30–40% fewer corrections per video than at month one, at the same API and infrastructure cost.

The practical effect: content that starts in Tier B (AI+review) may qualify for Tier A (AI+QA sample only) after the glossary reaches the accuracy threshold required for that content type. The programme cost per minute decreases as the library matures. Human-only programmes do not have this trajectory — the per-minute cost is approximately flat regardless of accumulated production volume.

This is the compounding argument that makes the hybrid model not just the cheapest current option but the best long-term investment: the glossary you build in year one is an asset that lowers your year-two and year-three programme cost. A human-only programme has no comparable compounding asset.

Eight workflow failure modes and how to avoid them

1. Assuming AI accuracy is consistent across content types

The most common error in AI captioning programme design is using the vendor’s quoted accuracy figure (typically derived from a talking-head benchmark) and applying it to all content. A vendor that quotes 97% WER on its marketing materials has almost certainly measured that on general-vocabulary content in ideal acoustic conditions. Your screen recordings and product training videos will perform 7–15 percentage points lower. Measure before you assign.

2. Counting internal correction time as free

Internal correction is the budget ghost of most L&D caption programmes. Because the correction work is done by existing L&D coordinators whose time is not tracked against the caption programme budget, it does not appear as a caption cost. It appears as a capacity constraint (“the team is too busy”), a quality complaint (“the captions are still wrong after we fix them”), and an opportunity cost (“we couldn’t launch that course on time because of captioning”). Making the cost visible is the first step to making the right workflow decision.

3. Routing based on preference rather than content classification

Without a documented classification system, routing decisions are made by whoever submits the captioning request, based on how much they trust the AI tool or how much they dislike sending content to a vendor. This produces inconsistent results: the same type of content gets different workflows depending on who submits it, which creates inconsistent accuracy outcomes and makes the compliance documentation indefensible.

4. Not setting an accuracy threshold before deploying

A hybrid model without documented accuracy thresholds is a hybrid model in name only. The threshold defines when a workflow assignment is appropriate, when content should escalate to a higher tier, and when a vendor relationship should be reviewed. Without the threshold, the programme cannot demonstrate that its routing decisions are compliance-driven.

5. Building the glossary from a master term list rather than from correction data

The instinct when building a vocabulary glossary is to export the product glossary or the documentation site’s terminology list and load it into the captioning system. This produces a glossary that is comprehensive but not prioritized: every term gets equal weight regardless of how often it appears in training content or how likely the ASR is to mis-transcribe it. The correction log is a better source of glossary terms because it tells you specifically which terms are failing and how often. Build from the correction log, not from the documentation export.

6. Using light review for Tier C content

Light review is appropriate when the AI draft is an accurate starting point that needs 10–20 corrections per hour. It is not appropriate when the AI draft has 80–150 errors per hour, which is the typical failure mode on specialist medical or legal content. A light reviewer who is not a domain expert will miss errors they cannot recognize, and the output will not reach 99% WER even after review. If the correction volume is too high for a reviewer to work at the light-review pace, the workflow is Tier C, not Tier B.

7. Not updating tier assignments as the glossary matures

The tier routing table should be reviewed quarterly. As the glossary matures and AI accuracy on specific content subcategories improves, some Tier B content may qualify for Tier A. Some Tier C content, as a domain-specialist glossary is built over time, may qualify for Tier B. If the routing table is never updated, the programme pays for more review than the current AI accuracy requires, which is an unnecessary cost.

8. Treating accuracy audit data as historical record rather than routing input

The quarterly DCMP accuracy sample is most valuable when it directly informs routing decisions. If the sample shows that screen recordings on the new product line are averaging 91% WER despite glossary tuning, that is not just a data point — it is an instruction to move that content subcategory from Tier B to Tier C until the glossary can be extended to cover the new product vocabulary. Programmes that collect QA data without feeding it back into routing decisions are running a quality assurance function without the improvement loop.