In-house, vendor, or SaaS? A decision framework for L&D teams choosing how to produce WCAG-compliant captions at volume — when each model is right and what it costs to get it wrong

Why the production model choice is structural

The choice of captioning production model shapes everything downstream: the cost structure, the accuracy infrastructure, the compliance documentation posture, and the daily workflow L&D staff operate within every time a new video is produced. It is the highest-leverage decision in the caption programme architecture — above which LMS to use, above which caption format to export, above which glossary system to implement.

A team that chooses the in-house model builds its programme around one person’s skill and availability. A team that chooses the vendor model builds around a contract, a turnaround window, and a per-minute rate that scales with volume. A team that chooses the SaaS model builds around an edit workflow that only delivers its economics when the glossary is built, the review process is established, and the team treats the AI output as a draft rather than a final product. None of these is a plug-in decision — each requires infrastructure to operate correctly and produces different failure modes when that infrastructure is missing.

The caption programme maturity model frames the production model choice as a Level 2 to Level 3 transition decision. Level 2 programmes (Developing) have a vendor in use or have adopted auto-captions with occasional manual correction, but the production model has not been chosen deliberately against volume and compliance criteria. Level 3 programmes (Established) have a documented production model that matches the organisation’s actual volume, vocabulary complexity, and compliance risk profile — and a written rationale for why that model was chosen and when it should be re-evaluated.

The caption compliance self-assessment covers production model alignment in Domains 1 (coverage) and 2 (accuracy): a production model mismatch appears as strong coverage scores for low-stakes content paired with weak accuracy scores for high-stakes content, or as low Domain 5 (vendor accountability) scores that reflect the absence of a vendor SLA even when a vendor relationship exists in practice.

The three production models

Three production models are available to an L&D caption programme. They differ in cost structure, accuracy control mechanism, compliance documentation posture, and operational flexibility.

Model 1: In-house caption staff — a dedicated employee (caption editor, accessibility specialist, or L&D coordinator with captioning as a substantial component of their role) who produces captions as a primary or major secondary job function. The caption output is the organisation’s own labour, not purchased from a third party.

Model 2: Caption vendor — a third-party captioning service (Rev, 3Play Media, Verbit, AI-Media, or similar) that accepts audio or video files and returns verified caption files within a defined turnaround window. The vendor relationship is typically governed by a contract with a per-minute rate, an accuracy SLA, and defined documentation obligations.

Model 3: SaaS AI captioning — a subscription software tool that applies AI transcription (typically Whisper or equivalent) with an edit-and-review UI, optionally with glossary integration that applies organisation-specific vocabulary at the transcription stage rather than as post-processing. The L&D team manages the review workflow internally.

These three models are not mutually exclusive. Most mature L&D caption programmes use a combination: SaaS for standard new content, vendor for high-stakes or regulated content, and in-house staff for review-and-edit workflow management when volume justifies the hire. The framework below helps identify which model or combination of models fits a specific organisation’s profile.

Model 1: In-house caption staff

What the role looks like in practice

An in-house caption editor is a dedicated employee whose job includes caption production, review, or both as a primary or substantial function. In practice, the role is most commonly titled accessibility specialist or digital accessibility coordinator (common in healthcare and higher education), L&D coordinator or instructional designer with captioning as a major workload component (most common in 50–500-employee organisations), or dedicated captioner or caption editor (rare below 200-employee organisations, usually found in media-intensive or broadcast-adjacent L&D teams).

Few mid-market L&D teams have a role exclusively dedicated to caption production. More commonly, the caption function is embedded in a broader accessibility or L&D operations role. The economic analysis below applies whether captioning is 100% or 40% of the role — the calculation simply allocates the appropriate fraction of fully loaded salary to caption output.

Full cost stack

A dedicated caption editor in the US earns $38,000–$55,000 per year. An accessibility specialist with broader responsibilities earns $55,000–$80,000. An L&D coordinator with captioning as a significant component is typically allocated $40,000–$65,000 per year for the full role. Add 20–30% for employment taxes, healthcare benefits, 401(k) match, equipment, and overhead. A $50,000 salary becomes $60,000–$65,000 fully loaded.

Productivity and per-unit cost

A skilled caption editor working exclusively on caption production can review and correct AI-assisted first-pass output at approximately 4–6 hours of finished video per eight-hour day (50–75 hours per month at full-time). Starting from scratch without AI assistance, a professional captioner produces approximately 1–1.5 hours of finished video per day (20–30 hours per month). The difference is why the SaaS + in-house-editor hybrid is the most common high-volume model: SaaS provides the AI first pass; the in-house editor manages review, improving throughput by a factor of 3–4.

At $60,000 fully loaded annual cost and 60 hours/month output (using AI-assisted review), the cost per finished caption-hour is approximately $83 ($60,000 / 720 hours per year). A caption vendor charges $60–$210 per finished hour at $1.00–$3.50 per minute. The economic break-even between in-house staff and a vendor at $1.00/minute (the standard human-caption floor rate) occurs at approximately 40–50 hours of video per month. Below that volume, the fixed cost of the in-house hire exceeds the vendor rate per unit. Above it, in-house is competitive.

Vocabulary advantage

The primary operational advantage of in-house staff is vocabulary immersion. An L&D team’s in-house caption editor knows the organisation’s product names, SDK symbols, clinical terminology, and internal acronyms accumulated through regular exposure to the organisation’s content. This eliminates the cold-glossary problem that SaaS tools and vendors face at the start of a new engagement. The in-house editor notices when “Kubernetes” is transcribed as “cube net ease” without needing to be told — and knows whether the correct transcription is the brand name or the internal nickname the engineering team uses.

Failure modes specific to in-house staff

Single point of failure. If the in-house captioner takes leave, is out sick, or departs, caption production stops until a replacement is found or a temporary vendor relationship is established. Organisations that build their caption workflow entirely around one person’s skill and knowledge face acute programme risk at every staff transition.

Vacation and leave coverage gap. A professional captioner taking two weeks of annual leave creates a backlog of 30–60 hours of unprocessed video (at 60 hours/month output). Organisations need a coverage plan — typically a vendor relationship maintained at low volume specifically for this use case. The captioning RFP playbook covers maintaining a backup vendor relationship at minimal overhead.

Scaling cliff. In-house staffing scales in increments of one headcount. A programme growing from 60 hours/month to 100 hours/month cannot hire 0.67 of a person. The typical response (overworking existing staff, deferring captions, or allowing a quality compromise) is expensive in each case. A vendor or SaaS supplement removes the scaling cliff by providing variable-cost capacity above the in-house baseline.

Compliance documentation gap. An in-house employee’s caption production is not documented in a vendor contract with accuracy SLAs, audit rights, and records retention provisions. For regulated industries that need to demonstrate to an examiner that caption accuracy was contractually guaranteed and independently verifiable, the in-house model requires supplemental documentation infrastructure: internal accuracy standards documentation, QA protocol records, performance management documentation, and an internal audit trail equivalent to what a vendor contract provides. The caption QA methodology post covers the QA infrastructure that in-house programmes need to produce an equivalent compliance record.

When in-house is the right model

• Volume: More than 40 hours of video per month on a sustained, predictable basis
• Vocabulary complexity: Very high (clinical, pharmaceutical, legal) where immersive domain knowledge is more valuable than a SaaS glossary
• Turnaround requirement: Same-day or next-day (no lead time for vendor turnaround windows or AI processing queues)
• Institutional context: Large enough (typically 500+ employees) that the single-point-of-failure risk can be managed with a backup vendor relationship and coverage protocols
• Budget structure: Headcount budget is easier to secure than variable vendor spend in the organisation’s budget culture

Model 2: Caption vendor

What a vendor relationship involves

A caption vendor is a third-party organisation that accepts audio or video submissions and returns verified caption files within a defined turnaround window. The major vendors serving the L&D and enterprise training market differ significantly in pricing, vocabulary handling, documentation, and LMS integration:

Rev offers human captions at $1.25/minute with two-business-day standard turnaround, plus automated AI captions at $0.25/minute with no turnaround SLA. Rev’s terms of service are consumer-grade; the accuracy claims are not independently audited against a defined methodology. Rev is appropriate for general content at moderate volume; it is not appropriate for regulated-industry content where the vendor contract is the primary compliance instrument. The vendor SLA review checklist covers why Rev’s standard terms are insufficient for regulated contexts.

3Play Media operates at $1.00–$1.50/minute for human captions with five-business-day standard turnaround, with LMS integration plugins for Kaltura, Brightcove, Panopto, and YouTube. 3Play’s searchable transcript product adds SEO and discoverability value for media and higher education content. Better suited for media-production and university L&D teams than for highly regulated enterprise compliance training.

Verbit positions as enterprise-grade, at $1.00–$3.50/minute depending on configuration (human-only, AI-assisted hybrid, real-time CART). Offers BAA execution for healthcare customers, more structured SLA documentation than Rev or 3Play, and dedicated account management. Better suited for financial services and healthcare L&D teams where the vendor SLA is a compliance instrument. The vendor pilot programme design post covers how to evaluate Verbit-tier vendors on compliance documentation rather than accuracy claims alone.

AI-Media has global reach, specialising in live CART and pre-recorded at $0.80–$2.00/minute depending on configuration. Strong in education and government sectors where procurement requires demonstrated government-sector experience. The hybrid ASR + human review model allows faster turnaround at the mid-price tier.

Per-minute pricing reality at L&D volume

The published rates above are entry points. Enterprise volume pricing, LMS integrations, BAA add-ons, and rush turnaround acceleration all affect the effective rate. For a mid-market L&D team producing 30 hours (1,800 minutes) of video per month:

The vendor human-caption rate at 30 hours/month ($2,250+) is 22× the SaaS cost for the same volume. The per-unit economics only favour a vendor for organisations that cannot maintain an effective SaaS review workflow — or for regulated-industry content types where the vendor’s contractual SLA is the primary compliance instrument, not just a performance preference.

The vendor documentation value

The economic case for a vendor relationship at human-caption rates is not accuracy alone — it is the documentation the vendor contract creates. A vendor contract that includes a 99%+ accuracy SLA with a defined measurement methodology, audit rights for processing logs, reference transcript retention for 36+ months, a remediation clause, and a BAA (for healthcare content) creates an externally verifiable compliance record. This is what FINRA examiners ask for. This is what OCR investigators request. This is what GxP audit teams review during inspections.

For FINRA-regulated financial services L&D (Rule 3110 training record requirements), GxP pharmaceutical training (FDA 21 CFR Part 11 electronic record obligations), and healthcare organisations under CMS Conditions of Participation, the vendor contract is part of the compliance evidence pack, not an administrative document. A SaaS tool’s audit log demonstrates that review occurred; a vendor SLA with defined accuracy methodology demonstrates what accuracy standard was contractually committed. These are different instruments. The vendor audit rights and examination evidence post covers the specific documentation provisions that turn a vendor relationship into a compliance instrument rather than just a production service.

When a vendor relationship is the right model

• Volume: 10–200 hours per month (below 10, ad-hoc or SaaS is more flexible; above 200, in-house becomes economically competitive)
• Compliance risk: High or Very High (vendor SLA is required as documentation instrument; internal QA framework alone insufficient)
• Vocabulary complexity: Any tier (vendors handle specialised vocabulary through explicit submission notes and vocabulary sheets, though this is less reliable than a SaaS glossary applied at transcription time)
• Turnaround tolerance: Standard production allows 1–5 business days; rush is available at a premium
• Budget structure: Per-minute variable spend that can be calibrated to actual production volume without a headcount commitment

Model 3: SaaS AI captioning with edit workflow

What distinguishes SaaS from LMS auto-captions

The most important distinction between a SaaS captioning tool and an LMS auto-caption feature is not accuracy at first pass — it is the structured review workflow embedded in the SaaS tool. An LMS auto-caption produces a file and presents no native review mechanism. An L&D team member who wants to correct the file must use a third-party SRT editor or work in the LMS interface (which is often inadequate for efficient caption editing). There is no workflow, no SLA, no accuracy measurement, and no documentation that human review occurred before publication.

A SaaS captioning tool provides: a side-by-side edit UI showing audio waveform, video playback, and caption text; low-confidence transcription segment highlighting; glossary application at the transcription stage rather than post-processing find-and-replace; export to multiple formats (SRT, VTT, TTML, STL) matching LMS requirements; and an audit log recording who reviewed, what was changed, and when the file was approved. The audit log is the compliance instrument that the LMS auto-caption path cannot produce. It is what demonstrates that the AI-generated draft was reviewed by a human before publication. The auto-captions WCAG compliance status post explains why AI-generated captions without a documented review step do not satisfy WCAG 2.1 AA SC 1.2.2 for training content with specialised vocabulary.

Glossary integration: the SaaS differentiation that matters most for L&D

A SaaS tool with glossary integration applies organisation-specific vocabulary at the transcription stage. “TalentLMS” does not appear in any general ASR training corpus in the form L&D teams use it — it produces “talent LMS,” “talent l m s,” or “talent lms” depending on audio conditions. “Workramp” produces “work ramp” or “work ramp.” “Pagerduty” produces “pager duty.” These failures are systematic — they occur on every instance of the term, in every video, across the entire content library. A glossary entry that maps “PagerDuty” to the canonical form eliminates this class of error at the source rather than requiring the reviewer to catch and correct it on every occurrence.

Without glossary integration, a SaaS tool’s accuracy on L&D training content is functionally similar to LMS auto-caption accuracy on domain vocabulary: 80–90% on general speech, significantly lower on product names, SDK symbols, clinical terms, and regulated-industry acronyms. The customer glossary architecture post covers how to build the glossary that moves SaaS first-pass accuracy from 82% to 96%+ on domain vocabulary — and why the glossary investment pays back on every subsequent video processed.

The glossary also compounds over time. Each video processed by the SaaS tool reveals new vocabulary gaps; each gap closed in the glossary reduces correction labour on all future videos of the same content type. An L&D team that builds its glossary systematically over six months of SaaS use ends up with a vocabulary asset that represents years of domain knowledge capture — a switching cost that makes the SaaS relationship more valuable the longer it runs. A vendor relationship does not compound this way: the vendor’s vocabulary handling is reset each time a vocabulary sheet is submitted, and it does not carry forward as an organisational asset.

SaaS subscription economics at volume

Even including the internal review labour (at 20 minutes per video-hour, which assumes a good glossary reducing correction burden), SaaS is substantially cheaper than a vendor relationship at every volume level above 5 hours per month. The savings at 30 hours/month ($1,401) cover more than a year of SaaS subscription cost. At 100 hours/month, the annualised SaaS savings ($56,412) is sufficient to fund an in-house editor salary. The caption programme budget planning guide covers how to model these economics for a multi-year budget request.

When SaaS is the right model

• Volume: More than 10 hours per month (below 10, ad-hoc vendor ordering often provides more flexibility)
• Vocabulary complexity: Medium to High, with a glossary built and maintained; SaaS with a mature glossary handles high-complexity content well
• Compliance risk: Low to Medium (SaaS audit log satisfies ADA/WCAG documentation requirements; very high compliance risk contexts may require vendor SLA as primary instrument)
• Internal review capacity: The team has 15–30 minutes per video-hour available for structured edit review
• Turnaround requirement: AI processing produces draft within hours; review is the bottleneck, not transcription

The decision framework: three dimensions

The decision between production models depends on three dimensions operating in combination. No single dimension determines the answer; the right model is the intersection of all three.

Dimension 1: Volume

Volume is the primary economic driver. It determines whether fixed costs (in-house salary) or variable costs (vendor per-minute, SaaS subscription) produce a lower per-unit cost. Volume thresholds are discussed in detail in the next section.

Dimension 2: Vocabulary complexity

Vocabulary complexity determines how much review labour each model requires and whether a SaaS tool’s first-pass accuracy is sufficient for the content type. Four levels:

Low complexity (general management, company-wide communications, HR policy content): Standard ASR accuracy is typically 92–96% on general speech. SaaS with a basic glossary or even LMS auto-captions with a structured review step can reach 99%+ with 10–15 minutes of review per video-hour.

Medium complexity (technology company product vocabulary, SaaS acronym-heavy content, engineering onboarding): Standard ASR drops to 82–88% on domain vocabulary. Product names, SDK symbols, and internal acronyms fail systematically. SaaS with a 50–150 term glossary recovers accuracy to 95–98% for new content; review takes 15–20 minutes per video-hour with good glossary coverage.

High complexity (clinical training vocabulary, pharmaceutical product names, legal Latin, financial regulatory acronyms): Standard ASR accuracy on domain vocabulary is 60–80%. Novel drug names, clinical procedure terminology, and financial instrument designations may produce near-zero word-level accuracy on first pass. SaaS with an expert-built glossary recovers to 90–95% on known vocabulary; review takes 25–40 minutes per video-hour. Without a mature glossary, review time is 60+ minutes per video-hour — at which point vendor economics (with their domain-specialist captioners) become competitive again. The caption quality error rate calculator lets you measure the actual first-pass accuracy on your specific content before making this determination.

Very high complexity (novel drug names in clinical development, cutting-edge surgical procedures, highly specific regulatory vocabulary with precise term variants): SaaS glossary coverage helps but cannot eliminate all review burden. Human vendor captioners with domain specialisation — medical captioners, legal captioners, financial captioners — have vocabulary coverage built in and produce first-pass accuracy requiring less review. At this complexity level, the vendor premium may be justified on review-time economics alone, independent of the compliance documentation requirement.

Dimension 3: Compliance risk profile

Compliance risk profile determines whether the vendor SLA is a compliance instrument (required) or a quality assurance mechanism (preferred). Four levels:

Low compliance risk: Organisations below the ADA Title I threshold (fewer than 15 employees), or content types that do not trigger a WCAG 2.1 AA obligation. Note: very few L&D content types are genuinely out of WCAG scope — if the content is required training, it is almost certainly in scope for any covered employer.

Medium compliance risk: ADA Title I covered employers (15+ employees) with workforce training obligations. WCAG 2.1 AA SC 1.2.2 applies to all pre-recorded training video with audio. SaaS with a documented review workflow (audit log) satisfies this obligation. The building a caption compliance programme post covers the documentation infrastructure that supports this posture.

High compliance risk: FINRA-regulated financial services L&D (Rule 3110 training records requirements), pharmaceutical GxP training (21 CFR Part 11 electronic record obligations for caption files), healthcare organisations under CMS Conditions of Participation (caption files for HIPAA training and mandatory clinical education subject to documentation requirements), and public entities post-April 2026 ADA Title II enforcement deadline. These contexts require either a vendor SLA (accuracy commitment, audit rights, records retention) or a SaaS tool operated under an internal QA framework that produces equivalent documentation — internal accuracy standards, process documentation, and an audit trail that substitutes for the vendor SLA as the compliance instrument.

Very high compliance risk: Organisations with an active OCR complaint investigation, EEOC disability discrimination charge, civil ADA lawsuit, or similar regulatory action. Caption files at this risk level are potentially discoverable evidence. The vendor’s processing logs, reference transcripts, and accuracy measurement records are exactly what opposing counsel requests and what regulators review. An in-house-only model or SaaS-only model without equivalent documentation creates gaps in chain-of-custody for caption file authenticity.

Volume thresholds in detail

Under 10 hours of video per month

At this volume, no production model generates enough workload to justify a committed relationship with significant fixed overhead. Ad-hoc vendor ordering (Rev, 3Play, or Verbit without an annual contract) and a SaaS solo plan ($29/month) are both viable. The primary constraint is turnaround: if the organisation needs captions within 24 hours of a video being completed, ad-hoc vendor ordering often cannot guarantee rush availability without a premium; SaaS provides AI output within hours and allows internal review on any timeline.

The compliance risk question at this volume is often whether the organisation has an ADA obligation that requires documented accuracy. A 12-employee organisation may have no ADA Title I workforce training obligation. A public university producing 8 hours of training video per month still has an ADA Title II obligation after April 2026 — the volume is low, but the obligation is absolute. For the public-university case at low volume, the SaaS solo plan with an audit log is the right answer: low cost, fast turnaround, documented review. A vendor relationship at this volume adds cost without adding compliance documentation that SaaS cannot provide.

10–30 hours per month

This range is where the SaaS versus vendor decision is most context-dependent. For organisations with medium vocabulary complexity and medium compliance risk, SaaS at $99/month (Team plan, 30 hours covered) is almost always the better choice: $849/month all-in including review labour versus $1,125–$1,688/month for a vendor at $1.00–$1.25/minute. That’s $3,312–$10,068/year saved on a 30-hour-per-month programme.

For organisations with high compliance risk — a hospital L&D team, a financial services compliance training function, a pharmaceutical company producing GxP training video — the vendor SLA may be required regardless of cost at this volume. The economics are irrelevant when the compliance instrument is the vendor contract itself. A hospital producing 20 hours/month of HIPAA workforce training cannot substitute SaaS alone for the vendor BAA and processing log retention the healthcare compliance framework requires.

30–100 hours per month

At this volume, SaaS economics become definitively superior for organisations without high-compliance-risk content types. A $299/month Org plan (unlimited hours) combined with internal review labour ($1,500–$2,500/month at 20 minutes per video-hour across 60–100 hours) produces an all-in cost of $1,799–$2,799/month. A vendor relationship at $1.25/minute for 60–100 hours costs $4,500–$7,500/month — a savings of $2,701–$4,701/month ($32,412–$56,412 annualised) in favour of SaaS.

The exception remains regulated-industry content. Many L&D programmes at this volume run a hybrid: SaaS for the 70–80% of content that is standard training, vendor for the 20–30% of content where the vendor SLA is a documentation requirement. A pharmaceutical company uses SaaS for HR, leadership, and product training; vendor for GxP training where the caption file is an electronic record. The hybrid captures SaaS economics on the majority of the content library while maintaining the vendor relationship for the subset where it is required.

100–500 hours per month

At this volume, the full cost-of-production comparison becomes more nuanced. An in-house caption editor reviewing SaaS first-pass output can process 80–100 hours of finished video per month at a fully loaded cost of $5,000–$5,400/month (one full-time hire). Combined with a $299/month SaaS Org plan, the in-house + SaaS model costs approximately $5,300–$5,700/month for 80–100 hours — competitive with a vendor at $6,000–$7,500/month for the same volume. The in-house hire adds vocabulary knowledge, same-day turnaround, and institutional continuity that a vendor cannot replicate.

The common model at this volume is SaaS as the production infrastructure (AI first pass + edit UI) with an in-house caption editor as the primary reviewer. The SaaS tool eliminates the blank-page production burden; the in-house editor handles review, quality management, and glossary maintenance. The ROI framework for finance executive audiences covers how to present the combined SaaS + in-house case as a cost-reduction initiative rather than a cost-addition.

500+ hours per month

At sustained 500+ hours per month, the in-house model becomes clearly economically competitive. Two full-time caption editors (at $120,000–$130,000 fully loaded per year, or $10,000–$10,833/month) can each process 80–100 hours per month using SaaS-assisted review — 160–200 hours combined. A SaaS Org plan handles overflow and additional content types. A vendor relationship manages regulated-industry content requiring SLA documentation. This three-tier hybrid (in-house primary + SaaS for volume surge + vendor for compliance-critical content) is the model most large enterprise L&D departments converge on at this volume.

The correction-cost trap: what unreviewed auto-captions actually cost

The most common production model mistake is not choosing the wrong vendor or the wrong SaaS tool. It is allowing unreviewed auto-captions to become the de facto production model by inertia — treating “the LMS generates captions automatically” as equivalent to “we have a captioning solution.”

The economics are deceptive. Auto-captions appear free because no line item exists in the budget. But the correction labour is not zero, and the compliance risk is not zero. The hidden half-FTE cost analysis quantifies this for a typical mid-market L&D team:

An L&D coordinator spending 45 minutes per video-hour correcting auto-captions is committing 45 hours of correction labour per month for an organisation producing 60 hours/month of training video. At an L&D coordinator fully loaded rate of $75/hour, that is $3,375/month in correction labour — for a “free” LMS auto-caption feature. A SaaS tool that produces higher-accuracy first-pass output with glossary applied requires only 15–20 minutes of review per video-hour for the same content, cutting that labour to $1,125–$1,500/month. The SaaS subscription plus reduced review labour costs $1,224–$1,599/month all-in — a saving of $1,776–$2,151/month versus “free” auto-captions.

The auto-caption path also produces no compliance documentation. The LMS auto-caption feature does not log that a human reviewed the output. It does not produce a reference transcript. It does not generate an accuracy measurement. It does not create an audit trail. When an OCR investigation or FINRA examination asks for caption compliance documentation — “please provide your accuracy measurement methodology and any records of accuracy review for the past three years of training content” — the auto-caption path produces no records other than the video itself with whatever accuracy it happened to achieve.

The caption error rate calculator provides a framework for measuring the actual accuracy of auto-captions on a representative sample of specific content before deciding whether a production model change is warranted. If the calculation shows 18% word error rate on the organisation’s product-name vocabulary on a 20-minute sample video, that concrete number — not a qualitative observation that “the captions aren’t great” — is what justifies a budget request for a SaaS subscription or vendor relationship.

The correction-cost trap has a second dimension: inconsistency. Individual reviewers correcting auto-captions without a defined QA process produce inconsistent results. One reviewer will correct 95% of errors; another will correct 70% and miss the systematic product-name failures because they are not looking for patterns. The caption QA methodology post covers how to build a review process that produces consistent accuracy across reviewers — and why consistency is what makes a compliance claim defensible, not any individual video’s accuracy score.

Decision matrix

The matrix below summarises the fit assessment across the key decision dimensions. Read across the row for your organisation’s profile; the column with the most checks is the starting recommendation. Most organisations with mixed profiles will end up in a hybrid configuration described in the next section.

Hybrid models in practice

Most mature L&D caption programmes use a combination of models. The single-model approach works well at low volume and low compliance risk; as programmes grow and content types diversify in vocabulary complexity and regulatory exposure, a hybrid becomes the most cost-efficient and risk-appropriate structure.

SaaS primary + vendor for compliance-critical content

The most common hybrid for mid-market L&D teams. SaaS handles 80–90% of content: standard professional development video, product training, HR and leadership content, onboarding modules. Vendor handles the 10–20% of content where the vendor SLA is a documentation requirement: GxP training, FINRA-required compliance training, HIPAA workforce training for covered entities, content produced specifically to address an OCR complaint or regulatory finding. The vendor relationship can be maintained at 5–20 hours/month without triggering the volume economics that would justify a vendor-primary arrangement. The captioning RFP playbook covers how to select and onboard a vendor for this supplemental role without running a full procurement process.

In-house editor + SaaS for first-pass

The model at sustained high volume (80–200 hours/month). SaaS provides the AI first pass; the in-house editor manages the review workflow, glossary maintenance, and quality assurance. This hybrid produces the accuracy of human review at the cost structure of SaaS plus one headcount, rather than the cost of a vendor at human-caption rates for the same volume. The in-house editor can process 80–100 hours of reviewed video per month at this workflow, compared with 20–30 hours per month producing from scratch without AI assistance.

Vendor for new content + SaaS for video updates

When an existing captioned video is updated — a product rebrand, a regulation change, a safety content revision — re-captioning the entire video through a vendor at $1.25/minute is often unnecessary and slow. A SaaS tool with the existing glossary can reprocess the updated video with high first-pass accuracy for unchanged segments, while the internal reviewer focuses on the changed portions. The vendor relationship handles new content requiring SLA documentation; SaaS handles the update lifecycle efficiently. This hybrid is particularly effective for organisations with a stable video library and a regular cadence of minor content updates.

In-house + vendor for back-catalogue remediation

When an organisation undertakes a large-scale back-catalogue remediation, the in-house model alone cannot handle remediation volume on top of new-content production. A vendor handles the remediation backlog at bulk-pricing rates (typically $0.80–$1.00/minute for high-volume remediation agreements); in-house handles ongoing new-content production at a pace the team can sustain. The vendor backlog remediation is time-bounded (typically 6–18 months for a large library); the in-house workflow is the steady-state. Once remediation is complete, the vendor relationship can be reduced to the supplemental compliance-critical-content role described above.

Three trigger points for model re-evaluation

The caption production model is not a one-time decision. Three trigger points typically precipitate re-evaluation, and the framework in this post applies at each one.

Programme launch: initial model selection

The initial model selection happens when an L&D team first commits to a systematic caption programme rather than ad-hoc correction of auto-captions. This is the decision this post is primarily designed to support. The key inputs at launch are: current monthly video production volume (with a 12-month projection), vocabulary complexity of the content types produced, the organisation’s compliance risk profile, and whether any regulated-industry content types require a vendor SLA as the primary documentation instrument. Most mid-market L&D teams launching a systematic caption programme for the first time find that SaaS with a structured review workflow is the right starting model, with a vendor relationship added for any regulated-industry content tier.

Volume threshold crossing

As programmes grow, the economics shift. A team producing 8 hours/month that grows to 20 hours/month has crossed the SaaS subscription sweet spot where a Team plan delivers better economics than ad-hoc vendor ordering. A team producing 20 hours/month that grows to 50 hours/month has crossed the point where the Org plan delivers unlimited hours at a flat rate. A team producing 50 hours/month that grows to 150 hours/month has crossed the in-house break-even threshold where an editor hire is economically competitive. The caption programme budget planning guide covers how to project these crossings in a multi-year budget model so the model re-evaluation is proactive rather than reactive.

Regulatory event: compliance risk profile escalation

An OCR complaint, FINRA examination finding, GxP inspection observation, or civil disability discrimination lawsuit changes the compliance risk profile immediately and materially. A team operating under a medium-risk SaaS-only posture may need to add a vendor relationship and documentation framework within 30–60 days of a regulatory action. Having already evaluated the vendor market and prepared a shortlist before a regulatory event occurs is the difference between a 30-day pivot and a 6-month scramble. The caption vendor pilot programme design post covers how to run a compressed vendor evaluation efficiently when time pressure is acute. Keeping the vendor SLA review checklist current ensures that when a vendor relationship is needed urgently, the contract negotiation does not take another month.

Eight failure modes

Failure mode 1: The auto-caption default

Treating the LMS auto-caption feature as the production model because it requires no procurement decision. The auto-caption default is not a production model — it is the absence of one. It produces no compliance documentation, no accuracy measurement, no review audit log, and no mechanism for systematic improvement. The apparent cost of zero is offset by correction labour that is never tracked, accuracy that is never measured, and compliance risk that surfaces only when a regulator asks for records that do not exist.

Failure mode 2: The in-house hire at below-threshold volume

Hiring a dedicated caption editor or accessibility specialist at 15–20 hours/month production volume when the economics clearly support a vendor or SaaS approach. The in-house hire at low volume produces the highest per-unit cost of any model: fixed overhead divided by low output generates $300–$400+ per finished caption-hour at 20 hours/month, compared with $60–$100 for a vendor or $5–$20 for SaaS + review. The hire also creates a single point of failure before the programme has the institutional maturity to manage it.

Failure mode 3: The single-vendor dependency

Committing entirely to a single vendor relationship with no backup model and no SaaS supplement for volume surges or vendor performance failures. When the vendor’s turnaround time increases during their peak periods, or when the vendor raises rates without notice, or when the vendor relationship ends unexpectedly, the programme has no alternative. Maintaining a SaaS capability alongside a vendor relationship eliminates this dependency for the majority of content types.

Failure mode 4: The SaaS-without-glossary deployment

Deploying a SaaS captioning tool without building the organisation’s glossary first. A SaaS tool without a glossary produces accuracy on domain vocabulary that is functionally similar to LMS auto-captions — 80–88% on product names and domain terms. The improved interface is valuable, but the first-pass accuracy still requires 40–50 minutes of correction per video-hour rather than the 15–20 minutes achievable with a mature glossary. The glossary is not an optional enhancement; it is the mechanism that delivers the economics and accuracy advantages that justify the SaaS subscription.

Failure mode 5: The regulated-industry SaaS overconfidence

Using SaaS as the sole production instrument for content types where the vendor SLA is a documentation requirement. A SaaS audit log demonstrates that review occurred; it does not constitute a 99%+ accuracy SLA with a defined measurement methodology and contractual audit rights for processing logs. For FINRA-examined financial services firms, FDA GxP-regulated pharmaceutical companies, and healthcare organisations under CMS Conditions of Participation, the vendor SLA is part of the compliance evidence, not a preference. SaaS is a production tool within the compliance framework; it is not the compliance instrument itself for these contexts.

Failure mode 6: The volume-only decision

Choosing the production model based on hours per month alone, without factoring in vocabulary complexity or compliance risk profile. An organisation producing 12 hours of pharmaceutical GxP training video per month may legitimately need a vendor relationship at that volume, even though a volume-only analysis would suggest SaaS Solo plan at $29/month is the obvious answer. Vocabulary complexity and compliance risk profile can each independently override the volume-driven economics when they create requirements the cheaper model cannot satisfy.

Failure mode 7: The sunk-cost trap

Continuing with the initial production model past a volume threshold or compliance risk escalation because switching feels disruptive. The model that was correct at 8 hours/month is often wrong at 60 hours/month. The model adequate for medium compliance risk becomes inadequate when a regulatory event changes the risk profile. The caption programme annual review process provides the mechanism for systematic model re-evaluation at each programme review cycle. The decision criteria in this post are what the annual review should apply.

Failure mode 8: The production model mismatch by content tier

Applying the most expensive model to low-stakes content and the cheapest model to high-stakes content — the inverse of what a risk-based approach recommends. A common version: vendor captioning for all-hands communication videos (low vocabulary complexity, low compliance stakes) and unreviewed auto-captions for regulatory compliance training modules (high vocabulary complexity, high compliance stakes, legally required content). The caption compliance self-assessment identifies content-tier mismatches as Domain 1 (coverage) and Domain 2 (accuracy) gaps. The fix is not necessarily spending more — it is redistributing the existing production budget to match compliance risk rather than content prominence.

FAQ

At what volume does it make sense to hire a dedicated in-house captioner?

The economic break-even between an in-house caption editor (at full employment cost including benefits and overhead) and a vendor relationship at standard human-caption rates ($1.00–$1.50/minute) occurs at approximately 40–50 hours of video production per month on a sustained, predictable basis. Below that volume, the fixed cost of the in-house hire is higher per finished caption-hour than the vendor rate. Above it, the in-house model becomes economically competitive and gains the vocabulary knowledge, turnaround, and institutional continuity advantages discussed above. The break-even shifts lower if the organisation already employs an L&D coordinator who can absorb captioning as a portion of their role rather than requiring a dedicated hire — in that case, the marginal cost is the time allocation, not a new headcount.

Can a SaaS captioning tool satisfy the accuracy documentation requirements for regulated industries?

For most regulated-industry contexts, a SaaS tool’s audit log is a necessary part of the documentation framework but not sufficient on its own. FINRA examinations look for vendor-level accuracy SLA documentation tied to a defined measurement methodology. GxP contexts require validated software with version-controlled audit trails. CMS healthcare contexts expect documentation of accuracy methodology with audit rights. A SaaS tool can be the production instrument within a broader compliance framework — but the framework must include internal accuracy standards, QA process documentation, and review workflow documentation that collectively substitute for the vendor SLA’s role as the externally verifiable compliance instrument. Most regulated-industry L&D teams use SaaS for general content and a vendor for the specific content types where the vendor SLA is non-negotiable.

We produce content with different compliance risk levels across our library. Should we use different models for different content types?

Yes, and this is how mature L&D caption programmes typically operate. Establish content tier categories: compliance-mandatory content (legally required training, regulated-industry content requiring vendor SLA) at the highest tier; general professional development and product training at a mid tier (SaaS + review); informal internal communication at a lower tier where SaaS with lighter review is acceptable. Document the tier definitions and the production model assigned to each tier in the caption policy so the mapping can be applied consistently by any team member producing content. The tier structure also creates a defensible response to a regulatory inquiry: “our compliance-mandatory training content is captioned through [vendor] with a documented 99%+ accuracy SLA” is a more credible answer than “we use captions across all content.”

We are a 60-person SaaS company, medium vocabulary complexity, producing about 20 hours of training video per month. What model is right for us?

SaaS with an edit review workflow. At 20 hours/month, the vendor relationship ($1,500–$1,800/month at $1.25–$1.50/minute) is expensive relative to the compliance documentation value at medium risk; the in-house hire is economically wrong (fixed cost divided by 20 hours/month produces $250–$330 per caption-hour at $60,000/year fully loaded). SaaS Team plan at $99/month with 15–20 minutes of review per video-hour produces an all-in of $349–$599/month — a saving of over $1,000/month versus vendor. Build the glossary before the first video: compile your product names, SDK symbols, internal acronyms, and any technical terms the engineering and product teams use. The glossary build takes 2–4 hours and pays back on every subsequent video processed.

How long does it take to build a vendor relationship to the point where it is operationally useful?

A vendor pilot can be established within 2–3 weeks: initial inquiry, compressed pilot evaluation using the vendor pilot programme design framework, agreement on terms and SLA language reviewed against the SLA review checklist, and first production submission. A formal RFP process with multiple vendor evaluations takes 4–8 weeks following the full RFP playbook. If a regulatory event has already occurred and a vendor relationship is needed immediately, a 2-week accelerated pilot is achievable for most vendors. The vendor relationship reaches operational maturity — where the vendor has learned the organisation’s content vocabulary and style — after approximately 3–6 months of regular production volume.

How should we compare the true cost of each model when making the decision?

Build a total cost of ownership comparison that includes four components: (1) direct cost per finished caption-hour (subscription fee or per-minute rate, divided by hours processed per month); (2) internal review labour cost (average minutes of internal staff time per video-hour, multiplied by the number of video-hours per month, multiplied by the loaded hourly rate for the review role); (3) glossary build and maintenance cost (amortised over 12 months of usage); and (4) compliance documentation risk premium (the estimated cost of a regulatory finding if documentation is inadequate, probability-weighted by the likelihood of a regulatory event at the organisation’s compliance risk level). The fourth component is the one most often omitted from budget comparisons, and it is the one that makes the vendor relationship economically rational at high-compliance-risk levels even when the direct cost is significantly higher. The ROI framework for finance executive audiences covers how to structure this comparison for a CFO or VP Finance who sees the SaaS subscription as a cost rather than as a cost-reduction and risk-mitigation instrument.

We started with LMS auto-captions two years ago and have a 300-video library of uncaptioned or poorly captioned content. Should we focus on the backlog or the new production model first?

Both in parallel, but with different priority levels. Establish the new production model for ongoing content first — every new video produced without a reviewed model adds to the remediation burden. Once the new production model is operational (2–4 weeks for SaaS, 4–8 weeks for a vendor relationship), begin a triage-based backlog approach: identify compliance-mandatory content in the existing library (content that is legally required for the workforce, or content directly relevant to an ADA obligation or regulatory requirement) and prioritise those items for immediate remediation. Work through the remainder using the large-scale backlog remediation playbook with a phased approach that keeps the compliance evidence current. Do not delay the ongoing production model change while waiting to resolve the backlog — the backlog is the past problem; the production model determines whether the problem compounds.