Josh Cavalier

They identify two consequences for managers.

• Teams that build with AI draw on token budgets, so approving more compute is effectively a capital allocation decision, even when managers lack the skills to judge the trade.
• And because models cannot be routed or trained on work that has not been described, organizations must break work into tasks with defined inputs, outputs, value, and risk.

In a post about AT&T, Josh Cavalier recounts Andy Markus, AT&T's Chief Data & AI Officer, speaking at the CNBC AI Forum in Dallas: the company processes about 45 billion tokens a day, and every request passes through an AI gateway that routes it to the best model for the least cost, with 80 to 90 percent cost savings from open models in some applications. They read that IT description as a human-machine workforce strategy and list three parallels for finance, operations, and HR leaders.

Read original

A routing system can compare model prices, but it cannot say what a small drop in quality costs when the output is wrong—and most companies cannot see that value either. For L&D and finance leaders, cost per request is the easy half of the decision; the harder half is assigning value and risk to each task before choosing the cheaper option.

Cavalier cites Mark Austin, the AT&T vice president who oversees AI for employees, telling The New York Times that routing coding work to open models cut costs 56 percent for a 2 percent drop in quality, then asks whether that is a good trade-off. They pair the example with figures from a same-day EY-Parthenon CEO Outlook Survey: of 1,200 CEOs, only 16 percent said they had clear, real-time visibility into AI return on investment, and another 24 percent could track some costs but not the overall return.

Read original

Proprietary corporate data is where the value sits, but undocumented work cannot be routed, trained on, or measured. They treat workforce preparation as a common failure point, not a soft extra.

Cavalier reports that only 15 percent of CEOs are reinvesting AI productivity gains in workforce development, even as 72 percent expect skills shortages to hold back growth more than access to capital. They close by naming the layer Talent Rewire builds: task-level documentation, decisions about the human-AI split, value and risk assignment, and upskilling.

Read original

They distinguish what the assistant did from what a person still had to do, name the first version's defects, and identify a specific fix for the next one. That makes the account more useful to a team considering AI-assisted work than a clean success story would be.

For a 48-second promo for the Association for Talent Development's Applying AI in L&D Certificate in Phoenix, Josh Cavalier describes one brief that produced the script, narration in a cloned ElevenLabs voice, a storyboard, and 13 animated scenes coded in HyperFrames, each timed to the narration, plus a music mix.

• They also document what broke: the hosted video tool could not take the voice file, so the animation rendered without it; the voice and music were added through a browser controlled by the assistant with a Chrome plugin.
• Cavalier calls the voiceover a bit flat, notes minor visual mistakes, and says the next version would add expressive tags for more vocal variation.
• Work they estimate would take 8 or more hours in After Effects came down to roughly 20 minutes of their time plus about 40 minutes of generation.
• The process still required human input.

Read original · Original 2

Josh Cavalier aims to make the AI video production process learnable for L&D teams: the posts describe the pipeline step by step while pointing readers to a certificate session on building explainer videos with generative AI. The demonstration serves as proof for the teaching offer, which is why the production detail is published rather than kept internal.

Both promo posts direct readers to the Applying AI in L&D Certificate in Phoenix on October 26 to 28, with a 15 percent discount code and an invitation to learn explainer video production with generative AI. The video promotes the session, making the artifact and the lesson the same piece of work.

Read original · Original 2

It answers a newcomer’s first question about Josh Cavalier: what does AI-assisted video production look like end to end, and how much of it is automated? This is a complete, self-critical walkthrough, not a claim about AI's potential.

Cavalier describes making a finished 48-second promo for the Association for Talent Development's Applying AI in L&D Certificate in Phoenix without opening a video editor.

• The account details each step: the assistant pulled course facts from td.org and wrote the script; ElevenLabs generated four narration takes using their voice clone, and the assistant picked the best one; it built the storyboard; it coded 13 animated scenes in HyperFrames using HTML, CSS, and GSAP, timing each animation to the narration; and the final voice-and-music mix happened back in ElevenLabs, where the music was transcribed to check for stray vocals.
• The post also explains what did not work.
• The hosted video tool could not accept the voice file, so it rendered the animation itself; the voiceover and music were added through a browser session controlled by the assistant with a Chrome plugin.
• Cavalier calls the voiceover a bit flat, notes minor visual mistakes, and says expressive tags in ElevenLabs V4 would be the next fix.
• The time claim is concrete: work that would take 8+ hours manually in After Effects took roughly 20 minutes of their time plus about 40 minutes of generation.
• The promo advertises an Oct.
• 26-28 event, with a 15% discount code.
• A shorter companion post covers the same project and adds that a manual After Effects build would take 8+ hours; together, the two posts form a before-and-after pair, with the plain explainer first and the detailed production log second.

For brands planning a campaign

For a learning-and-development vendor, this is the strongest available proof point: a named, in-demand course promotion made with an AI pipeline, with the friction and self-critique left in. That makes an honest live-build webinar more persuasive than a polished case study, with a natural sequence: script, voice, storyboard, animation timing, mix, then the parts that still needed a human.

It shows the strategic frame behind the production work: task-level decisions about what humans, humans-plus-AI, or AI should do, and who is accountable for the cost of getting it wrong.

Cavalier reports on Andy Markus, AT&T's Chief Data & AI Officer, speaking at the CNBC AI Forum in Dallas, and reframes the IT strategy as a workforce strategy.

• The details are worth reading in full: AT&T processes about 45 billion tokens a day, with every request passing through an AI gateway that weighs accuracy, cost, and latency before routing it to a model; in some applications, moving to open models cut AI cost by 80-90%.
• Cavalier draws three parallels for executives.
• Routing is task allocation, so organizations must ask of every task: human, human-AI collaboration, or AI?
• Managers now hold token budgets, making them capital allocators for compute without necessarily having the skills to decide.
• Proprietary corporate data is the asset, but models cannot be routed or trained on undocumented work, so workflows must be broken into tasks with inputs and outputs.
• The sharpest passage is a question the gateway cannot answer: citing an AT&T VP who told The New York Times that routing coding work to open models cut costs 56% for a 2% drop in quality, Cavalier asks whether that trade-off is acceptable.
• The gateway knows the price of routing, not the value of the work.
• Cavalier backs this with same-day EY-Parthenon CEO Outlook Survey figures: only 16% of 1,200 CEOs report clear real-time visibility into AI ROI, another 24% can track some costs but not overall return, and only 15% are reinvesting productivity gains in workforce development, while 72% expect skills shortages to constrain growth more than capital access.
• The post closes by naming the layer Talent Rewire builds: documenting workflows step by step, deciding the human-AI split task by task, assigning value and risk, and upskilling the people who run the work.

For brands planning a campaign

This post is a ready-made webinar outline for L&D buyers: token budgets as a management skill, the ROI-visibility gap, and task-level human/AI allocation with value and risk attached. Pair it with the promo build in the same session: strategy followed by a live demonstration, putting the argument and proof together.

The assistant pulled course facts from td.org, wrote the script, narrated it with a cloned ElevenLabs voice, storyboarded the animations and built the video.

• The build ran through HyperFrames, with 13 scenes coded in HTML, CSS and GSAP and each animation timed to the narration.
• They then made a final mix in ElevenLabs with a music bed.
• Cavalier reports the finished video took under an hour of total work—roughly 20 minutes of their own time plus 40 minutes of generation—compared with an estimated 8 or more hours in a conventional editor.

Presents the work as a tool-by-tool sequence with named handoffs: course facts, script, voice clone, storyboard, coded animation, final mix and browser-based finishing.

For brands planning a campaign

A learning and development vendor can sponsor a step-by-step production walkthrough instead of a general AI keynote. Named tools and visible output give the session a concrete artifact and a process an L&D team could reproduce.

Examples: Source 1 · Source 2

The hosted video tool could not accept the voice file, so the animation rendered without it.

• The voiceover and music were added afterward through a browser, with the assistant controlling the web tasks.
• Josh Cavalier calls the voiceover a bit flat, notes minor visual mistakes and says they would add expressive tags for more vocal variation in the next version.
• The posts also distinguish generation time from human supervision time rather than presenting the result as fully automatic.

Cavalier's posts depart from typical AI-tool promotion by stating the limits, defects and manual repairs alongside the time saved.

For brands planning a campaign

The candour gives vendors a credible way to talk about AI adoption. A co-marketing piece could retain an honest limitations segment and a “what we would fix next” segment—more useful to an L&D team weighing the process than a spotless success story.

Examples: Source 1 · Source 2

The post cites AT&T processing about 45 billion tokens a day, an AI gateway that weighs accuracy, cost and latency before routing each request, and 80 to 90 percent cost savings from open models in some applications.

• It draws three implications for finance, operations and HR leaders: routing is task allocation among human, human-plus-AI and AI work; managers now hold token budgets and have become capital allocators for compute; and proprietary corporate data is the asset, but it cannot be routed or trained on unless workflows are documented task by task.
• The post closes with a cost-versus-quality question raised by a reported 56 percent coding cost cut for a 2 percent quality drop.

Turns a vendor and executive presentation into decisions for non-technical leaders, then ends with the question technology cannot answer: what the work is worth when the cheaper option fails.

For brands planning a campaign

This is a strong content fit for a corporate learning vendor selling AI enablement to senior buyers. A webinar or co-marketing piece could start with the ROI visibility problem and task-allocation framing, then show how documented workflows and upskilled staff turn AI savings into measurable value.

Examples: Source 1

It also notes that only 15 percent of CEOs are reinvesting AI productivity gains in workforce development, while 72 percent expect skills shortages to hold back growth more than access to capital. Josh Cavalier uses those numbers to highlight the work their business covers: documenting workflows step by step, deciding the human-AI split task by task, assigning value and risk to each task, and upskilling the people who run them.

For brands planning a campaign

Survey data and a specific operating gap give vendors a ready-made content hook: productivity gains are not being reinvested in the workforce. A campaign could pair the numbers with a practical session on building the task-level documentation and skills plan that make AI spending defensible to a CFO.

Examples: Source 1

Josh Cavalier directs readers to the Applying AI in L&D Certificate in Phoenix on October 26 to 28, offers a 15 percent discount code and says attendees will be shown how to make explainer videos with generative AI. The work sample and teaching offer cover the same subject: the video demonstrates the technique taught in the session.

For brands planning a campaign

For a learning and development vendor, a sponsored workshop or certificate session built around a live or recorded build is a natural co-marketing format; the promo video can double as the proof of concept. The vendor can supply the audience, case study and follow-up resources, while Josh Cavalier demonstrates the process.

Examples: Source 1 · Source 2

The posts do not disclose whether they are an affiliate, instructor, sponsor, or otherwise compensated; ask for their relationship to ATD before booking.

Original disclosure

Save 15% with code PUMPKIN at td.org.

View example

Ask about their ownership, employment, or partnership role at Talent Rewire.

View example

The post does not disclose a sponsorship or commercial relationship.

View example

The post does not disclose a sponsorship or commercial relationship.

View example

The post does not disclose a sponsorship or commercial relationship.

View example

One post walks through a UGC ad built from a product upload, a Gemini-written brief with a matching influencer persona, and a Sora 2 render; another describes a Claude Skill that turns a static ad into motion design through Seedance 2.5 while keeping the text and product intact. Both creators show the pipeline step by step and name the models, so readers who follow Josh Cavalier’s build logs will recognise the format.

Guilhem frames the work as marketing throughput: producing creative at volume for eCommerce ads and keeping assets on brand for far less than hiring creators. Josh Cavalier focuses on learning and development and the workforce layer beneath it, with examples that end in a certificate session and a task-allocation argument rather than an ad-volume claim.

Read original · Original 2

One post argues that the narrator carries as much of a lesson as the visuals; Cavalier’s promo accounts treat narration as one step among many.

White makes the case for the professional human voice as its own craft, from the hidden costs of DIY narration to the narrator’s role in holding learner attention on STEM material.

• Josh Cavalier shows narration generated from a voice clone and critiqued for sounding flat.
• Together, they mark the two sides of the same production choice.

Read original · Original 2

One post argues the key metric for AI video is how much it costs to get one usable clip, and lays out a pipeline that splits concept images, product shots, image-to-video, voice, music, captions, and editing across separate models.

• Another reads Claude usage limits as a continuity and spend problem.
• Both treat model choice as part of production economics.

Liu’s audience is eCommerce teams, agencies, and AI builders choosing between models under usage limits; Josh Cavalier’s is learning and development teams and executive buyers deciding who does which task and what the work is worth when a cheaper model fails. Liu focuses on routing and spend, while Cavalier explores task-level documentation and upskilling.

Read original · Original 2

Two posts show a rebuilt user-generated-ad factory that runs from a product upload and written request through brand research, a creative brief with a matching creator persona, and video generation.

• They also describe a Claude Skill that turns a static ad into motion design while preserving the original text and product.
• The first post makes the comparison directly: hiring fifty human creators costs more than $50,000 a month, versus roughly $100 to $1,000 per campaign on the AI stack, with unlimited variations and the claim that assets stay on brand.
• If the vendor sells AI content production to agencies or ecommerce teams, this alternative makes the economics of that decision more forcefully than a single 48-second promo.

These posts point to a different buyer.

• They do not address learning and development, instructional design, training vendors or a named training organisation, and there is no dated teaching programme to anchor a co-marketing webinar.
• The pipeline also depends on specific tools: one post centres on a named AI agent platform, the other on a named video model.
• The full playbook, prompts and ready-to-use template are offered only in exchange for a comment and a connection request.
• A vendor cannot inspect or reuse a finished public asset as it could with a promotional video made for a named certificate.
• Keep the L&D buyer and the promise of a reusable asset, and the fit weakens.
• Change the buyer to ecommerce and the offer to pipeline economics, and it strengthens.

The two posts make a British voiceover practitioner's case for professional narration in corporate training and K-12 material.

• One argues against do-it-yourself voiceover on the grounds of learner engagement, editing and revision time, brand credibility, and the cost of re-recording.
• The other describes partnerships with publishers, curriculum developers, educational animation studios, museums and edtech companies, and points to a dedicated page for STEM education narration.
• The subject sits within corporate training, so a vendor selling to L&D teams shares the audience, even though the offer differs.

The argument runs counter to part of the AI-enabled pitch.

• The core claim is that a professional voice protects engagement and brand perception, the opposite of cloning a voice and automatically generating narration takes—the exact step the alternative creator's posts showcase.
• There is no AI production build, scripted promotional asset or recorded production process to demonstrate in a session on AI video production, so a webinar built around this choice would have to change subject.
• Keep the L&D buyer and teaching promise, and the fit weakens.
• Keep the audio subject, and the offer must shift from AI content creation to production quality.

The two posts frame the operating problem: once a model is good enough to use in daily work, usage limits, cost and continuity stop being subscription questions and become infrastructure concerns.

• The practical response is to match each task to the right model, stop burning expensive tokens, and move between models without losing context.
• The second post applies that argument to AI video, listing the steps from idea and script through image generation, style testing, image-to-video, multiple shots, voiceover, music and repeated regeneration.
• It argues that the key metric is the cost of getting one usable video, not the quality of a single model.
• That framing of production economics is close to the business case an L&D vendor needs to make to the finance or operations sponsor who signs the budget.

Neither post offers a deliverable: there is no finished asset, course to co-promote, template or blank working resource, and neither names an employer, client or partnership with a training organisation.

• The posts analyse how teams should organise AI production, addressing agencies, ecommerce teams and short-form video teams in general.
• They state an interest in ecommerce and AI-building audiences rather than learning leaders.
• A vendor could borrow the argument, but would still need a practitioner to demonstrate the production process—the role the original choice fills.
• Keep the L&D buyer and teaching promise, and this alternative supplies the executive hook without the build.