Capstone: an AI-assisted production plan and disclosure policy

This capstone brings the course together. You will see how a fictional creator, Priya, builds an AI-assisted production plan for her channel, calculates costs and savings, sets quality gates, writes a disclosure policy and plans for legal and ethical risks. Then you will produce your own plan using the same structure. Priya and her channel are invented for teaching purposes; the steps and numbers show the method, not a recommended set of tools.

The scenario

Priya runs a science explainer channel with 420,000 subscribers. She publishes one 15-minute video a week and three short clips. Her team is a freelance editor (20 hours a week at $45 an hour), a part-time researcher (10 hours a week at $35 an hour) and a thumbnail designer paid $120 per thumbnail. Priya writes and presents every video. Her goals for the next year are to add a French audio track for Quebec and other French-speaking viewers, increase short clips from three to eight a week, reduce errors after publishing, and keep her audience's trust.

Step 1: Map the workflow

Priya maps each stage of a typical week, with hours and pain points:

Step 2: Choose AI uses by value and risk

Using the lesson 8 matrix, Priya sorts candidate AI uses:

• Automate with spot checks (high value, low risk): transcription and text-based rough cuts, silence removal, caption generation with a custom vocabulary, comment spam filtering and theme summaries, chapter markers.
• Use with strict review (high value, high risk): research summaries of papers she supplies, script tightening, French dubbing with her cloned voice, AI-assisted highlight selection for clips, AI-generated thumbnail backgrounds.
• Avoid (high risk or low value): fully AI-written scripts, synthetic footage of real scientists, auto-replies to viewers, generated "experiment results".

Step 3: Estimate costs and savings

Step 4: Set quality gates

Priya adds four gates with owners:

• Fact-check gate (researcher): every claim in the script tagged and verified; all research summaries checked against the original papers; no AI citations unopened.
• Rights and likeness gate (editor): no unlicensed music or footage; no synthetic depictions of real scientists; thumbnails checked for logos and recognizable characters.
• Localization gate (Quebec reviewer): meaning, terminology and tone checked; names pronounced correctly; Quebec French vocabulary.
• Disclosure gate (Priya): platform labels applied where realistic synthetic content appears; French audio tracks described as AI-dubbed with her own cloned voice; sponsorship disclosures present.

She sets targets: caption WER under 3 % on spot checks, a fact-check catch rate above 90 %, and fewer than one correction per video after publishing.

Step 5: Consent, data and contracts

Priya clones only her own voice, using a provider whose terms let her delete the voice model and do not use it to train other systems. She updates her guest release form to state clearly whether guest audio may be translated or dubbed, with a separate opt-in for any voice cloning of guests. She checks that her editor's and researcher's contracts cover ownership of work made with AI tools, keeps all subscriptions on her business account, and adds a clause to sponsor agreements listing which AI tools may be used in sponsored segments and requiring sponsor approval for any generated product imagery.

Because guest voices and viewer comments contain personal information, she limits which tools receive them, turns off training on her uploads where offered, and records these settings in her tool register.

Step 6: Write the disclosure policy

Priya publishes a short AI policy on her website and links it in every video description. A version of it reads:

• What I use AI for: transcription, captions, rough editing, summarizing research papers before I read them in full, finding clip moments, translating and dubbing videos into French with my own cloned voice, and creating some thumbnail backgrounds.
• What I never use AI for: writing my opinions or conclusions, generating experiment results or footage presented as real, imitating real scientists or other creators, or replying to you as if I were writing personally.
• How I label it: realistic synthetic content carries platform AI labels; French audio tracks are marked as AI-dubbed; sponsored segments are disclosed as ads whatever tools were used.
• How I check facts: every claim is verified against original sources by a person before publishing, and corrections are posted in the description and pinned comments.
• How to reach me: email address for concerns about AI use or errors.

Step 7: Plan for risks

Your capstone

Now build your own plan, for your own channel or a realistic fictional one. Your capstone should include: a workflow map with hours and pain points; a value-risk sorting of at least eight candidate AI uses; cost and savings estimates with at least one ROI calculation; at least three quality gates with owners and targets; consent, data and contract measures; a published-style AI disclosure policy; a risk register with scores and responses; and a four-quarter adoption timeline. Keep it to what you could actually carry out, and plan to review it after three months.

Evaluating the plan after three months

At the three-month review, Priya compares results with her targets. Caption spot checks average 2.4 % WER, below the 3 % target. The fact-check gate caught 41 of 44 errors, a 93 % catch rate, and post-publishing corrections fell to 0.4 per video. Short clips rose to seven a week, slightly below the target of eight, because highlight selection took longer than expected; she adds an hour of editor time and keeps the target. Comment summaries revealed strong interest in a series on climate science, which becomes the next quarter's priority.

She also reviews costs. One subscription went largely unused and is cancelled, saving $35 a month. A sponsor asked about synthetic voices in reads; her contract clause and AI policy answered the question in minutes. Viewer feedback on the AI policy was mostly positive, with a few questions about the French voice clone, which she answers in a community post explaining how it works and how French-speaking reviewers check each dub.

The review ends with three decisions: keep the current gates, extend the French pilot to 20 more videos because it met its targets, and revisit the policy in six months or whenever she adds a new kind of AI use.

Presenting your capstone

When you present your capstone, whether to an instructor, a business partner or your own team, focus on decisions and evidence rather than tools. Explain the goals, the workflow map, why each AI use was placed where it was in the value-risk matrix, the numbers behind your cost and ROI estimates, the gates and targets, and how you will know whether the plan is working. Be honest about uncertainties, such as view estimates for a new language or how an audience might react to a twin. A good plan makes its assumptions visible so they can be tested and revised.

Include your draft disclosure policy as an appendix, written in plain language your audience would understand. Ask someone outside your team to read it and tell you what they think you use AI for; if their answer differs from your intent, revise the wording until it matches.