TagProof — AI-content disclosure trail for localization & video agencies

1. One-liner

Records what an agency made with AI, then emits the label and audit file its client’s MSA now demands.

2. Trend signal — why now?

Article 50 of the EU AI Act — the transparency and content-marking rules — becomes binding on 2 August 2026. The final Code of Practice on marking and labelling AI-generated content was published 10 June 2026, seven weeks before the deadline. Every organisation that uses AI professionally to produce audio, image, video or text is a “deployer” (and often, when it wires a model in via API, also a “provider”) with disclosure duties. Marketing, media and localization workflows are named explicitly.

The pain doesn’t come from the regulator — early enforcement is complaint-driven, and SME fines are capped at the lower of the fixed sum or 3% of turnover (a €2M-revenue agency’s worst case is ~€60K, not €15M). The pain comes downstream: enterprise clients carry their own Article 50 duties, so they are writing AI-disclosure and provenance requirements straight into the MSAs and POs they hand their agencies. As one legal write-up put it, “It becomes a line in your MSA, not their problem.” An agency that can’t answer “which of these 40 localized assets used AI, where, and can you prove it?” loses the renewal.

Meanwhile the technical primitives just matured: C2PA 2.1 + soft-binding watermarks (finalized 2026) survive the transcoding that used to strip provenance metadata, so per-asset provenance is finally durable across the distribution chain.

Provenance:

• Signal 1 (demand): Enterprise clients pushing AI-disclosure/provenance duties into agency MSAs — “It becomes a line in your MSA, not their problem.” — https://www.arkived.io/ — 2026-07-06
• Signal 2 (feasibility): Final Code of Practice on marking/labelling AI content published 10 Jun 2026; C2PA 2.1 soft-binding watermarks now survive social-platform transcoding — https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content — 2026-07-06
• Signal 3 (economic): EU language-services market USD 10.36B in 2026, Europe holds 43.9% share, top-100 vendors control only 15% (thousands of small/mid LSPs); Article 50 binding 2 Aug 2026 with €7.5M–€15M fine tiers — https://www.businessresearchinsights.com/market-reports/localization-services-provider-services-market-119644 — 2026-07-06 Category: Regulatory arbitrage

3. The opportunity

Two kinds of tools exist, and neither fits the agency in the middle:

• Generation platforms (3Play Media, CAMB.ai, Perso, ElevenLabs, Adobe Firefly) mark their own output. But a localization/video agency stitches together five or six tools plus human editors per project — Firefly for a thumbnail, ElevenLabs for a voiceover, a human colourist, a subtitle pass, an in-house dubbing model via API. Each vendor’s manifest covers only its slice. Nobody produces the per-deliverable, cross-tool record the client asks for.
• Governance/provenance tools (Arkiv, enterprise AI-Act compliance suites) exist, but Arkiv is design-agency-focused and blockchain-anchored (heavy, wrong shape), and the enterprise suites are €2K+/mo GRC platforms built for the client’s legal team, not the 8-person dubbing shop delivering to them.

The gap: a light, agency-native disclosure layer that watches a project’s assets, infers/records what was AI-touched at each step, applies the correct visible label + machine-readable mark, and hands the agency a one-click “disclosure pack” per deliverable to attach to the invoice. The incumbent doing this badly is the spreadsheet — agencies are literally tracking “AI: yes/no/partial” per asset in Google Sheets right now.

4. Target market

• Primary customer: Owner/ops lead at a small-to-mid EU localization, dubbing, subtitling, or video-production agency — 5–50 staff, serving enterprise/brand clients, producing multilingual AI-assisted deliverables at volume. Secondary: in-house content/localization teams at mid-market brands who are the deployer on the hook.
• Why they buy: “My biggest client’s new MSA says every deliverable must carry an AI-disclosure statement and I have to be able to produce provenance on request. I use six tools and freelancers. I’m tracking this in a spreadsheet and I know it won’t survive an audit. If I get it wrong I lose the account, not just a fine.”
• Rough TAM reasoning: EU language-services market is USD ~10.4B (2026) and highly fragmented — top-100 vendors hold only 15%, meaning thousands of small/mid LSPs, plus video-production shops and brand in-house teams. Capturing even 2,000 agencies at €250/mo is ~€6M ARR.
• Why now for them: The 2 Aug 2026 deadline plus the first wave of MSA clauses hitting their inbox this quarter. This is a “solve it before the renewal” purchase, not a nice-to-have.

5. Product sketch (MVP)

• Project workspace — one per client engagement; drop in the deliverables (video, audio, subtitle files, images) and the tools/freelancers involved.
• AI-touch capture — reads embedded C2PA/IPTC manifests where present; where absent (freelancer handoff, tool without credentials), prompts a 10-second “what did this step use?” tag, and stores the answer. Builds a per-asset chain: source → AI-voiceover (ElevenLabs) → human edit → subtitle-MT → human QA.
• Auto-labelling — applies the EU common “AI” icon in the required fixed position for visible disclosure, and writes the machine-readable mark (metadata + soft-binding watermark) so it survives upload.
• Disclosure pack — one-click, per-deliverable PDF/JSON: what’s AI-generated, where, which tool, human-edited or not, ready to attach to the invoice or hand to the client’s legal team.
• MSA clause reader — paste the client’s AI-disclosure clause; TagProof maps it to which of your assets are in/out of scope and flags gaps.
• Provider-vs-deployer helper — for each tool in the chain, tells the agency whether they or the vendor holds the marking duty (the genuinely confusing part of Article 50).
• Team + freelancer tagging — freelancers get a lightweight link to declare AI use on their handoff, so the chain is complete without the owner chasing.

6. AI angle — what’s load-bearing

Two places. First, classification: reading heterogeneous manifests and, more importantly, inferring likely AI involvement from artefacts when no manifest exists (synthetic-voice detectors on audio, MT-smell detection on subtitle text, generative-fill detection on images) so the agency isn’t hand-tagging 200 assets from scratch. Second, the clause reader: turning a free-text MSA clause into a structured scope map against the project’s asset list. Remove the AI and you’re left with the spreadsheet — which is exactly the status quo we’re replacing. The AI is what makes “tag 200 multilingual assets across six tools” a 20-minute job instead of a half-day.

7. Localization angle (if any)

This is a localization play, and the geography is the moat: EU-first, Article-50-native. The visible-label placement rules, the EU common icon, the provider/deployer split, and the member-state supervisory nuance are all EU-specific. A generic US content-provenance tool doesn’t encode any of this. Multilingual by nature (the customers are localization shops), and priced in euros at SMB tiers where a €2K/mo GRC suite can’t reach. Natural expansion: UK (post-Brexit transparency guidance), then any market that copies the Article-50 template.

8. Business model — path to $1M–$5M ARR

• Pricing: €99/mo Solo (freelance localizer / 1-person shop), €249/mo Studio (5–20 staff, unlimited projects, freelancer links), €499/mo Agency (multi-client, API + client-portal disclosure delivery). Usage add-on for high-volume asset classification.
• ACV: ~€3,000 blended (weighted to Studio tier).
• Rough math to $1M ARR: ~330 agencies × €249/mo × 12 ≈ €1M. Very reachable within the fragmented EU LSP base.
• Rough math to $5M ARR: ~1,400 agencies blended, or a smaller count once brand in-house teams (higher tier) and a UK expansion are added. Doesn’t require category dominance — just a few % of thousands of fragmented LSPs.
• Expansion path: per-seat freelancer links, per-project volume, then a “client portal” tier where the agency’s clients self-serve disclosure packs — which turns the client into the next lead.

9. Go-to-market wedge — first 100 customers

• Directory scrape + deadline hook: GALA and Nimdzi publish LSP directories; scrape EU members, send a personalized “your client’s MSA now says X — here’s your disclosure pack in 2 minutes” Loom before the 2 Aug deadline. Deadline-driven cold outreach converts far above baseline.
• Ride the confusion: every localization consultancy and law firm is publishing “Article 50 for LSPs” explainers right now. Offer TagProof as the do-it tool at the bottom of their guide (affiliate/referral). The audience is already reading about the pain.
• Association channels: GALA, ELIA, and ATC (UK) run member webinars — a “how to survive your first AI-disclosure MSA clause” session lands directly on buyers.
• Freelancer viral loop: every freelancer who gets a disclosure-tagging link sees the product; freelancers work for many agencies, so each becomes an inbound seed.
• LinkedIn owner outreach: dubbing/subtitling agency owners are active and identifiable; DM the ones posting about the AI Act.

10. Build complexity — justification

Medium. The workspace, tagging UI, PDF/JSON disclosure pack, and C2PA/metadata read-write are standard web + off-the-shelf provenance libraries (the C2PA SDK and watermarking tooling are open). The genuinely custom work is the artefact-based AI-detection classifiers (synthetic-voice, MT-smell, generative-fill) and the MSA-clause reader — both doable on off-the-shelf models but needing tuning and a labelled eval set. Domain accuracy on the provider/deployer mapping matters more than any single feature and needs a legal advisor. Small team, ~3–4 months to a credible v1.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	The product helps comply; no legal grey zone.
Ethical — no harm / dark patterns	✅	Increases transparency of AI content — squarely pro-disclosure.
Market exists (evidence above)	✅	Hard deadline + MSA-clause pull + thousands of fragmented LSPs.
1–5 person team can build this	✅	Medium build, mostly off-the-shelf provenance libs.
Launchable with <$50K / ₹40L	✅	SaaS; main cost is a legal advisor + eval data.

All five pass.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	15/20	Real and dated (2 Aug), but pain is “protect the account,” not daily hair-on-fire; some agencies will bluff it with a spreadsheet.
Demand evidence	15	12/15	Hard deadline, sourced MSA-clause pull, an incumbent (Arkiv) validating the category. Docked: no direct verbatim customer complaint found yet.
Build feasibility	15	11/15	Standard stack + provenance libs, but custom classifiers and a legal-accuracy bar add real work.
Distribution clarity	15	11/15	Named directories, association channels, freelancer loop, deadline hook — concrete, but conversion on cold LSP outreach unproven.
Revenue mechanics	15	12/15	Euro SMB pricing, fragmented base, clear €1M math. Retention risk once the deadline scramble passes.
Time to first revenue	10	7/10	3–4mo build lands right around/after the deadline; pre-sell possible now but revenue is 8–12 weeks out.
Defensibility	10	5/10	Execution + EU-regulatory-encoding + freelancer network as soft moats. Copyable; C2PA is open.
Total	100	73/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy (classifiers + provenance plumbing) · domain-expertise-required (Article 50 provider/deployer nuance — needs a legal advisor on call).

Key assumptions to validate (3–5)

1. Assumption: EU agencies are actually receiving AI-disclosure/provenance clauses in client MSAs now, not “eventually.” How to test: interview 25 LSP/video-agency owners; ask to see a redacted recent clause. Need ≥8 to produce one.
2. Assumption: Owners will pay €249/mo to avoid losing an account rather than keep using a spreadsheet. How to test: pre-sell a €249 annual plan to 15 of the interviewed agencies before build; ≥5 conversions = go.
3. Assumption: Artefact-based AI-detection is accurate enough that agencies trust it over hand-tagging. How to test: run detectors on a labelled set of 500 mixed assets; target ≥85% precision on “AI-touched.”
4. Assumption: The pain persists past the Aug-2026 scramble (retention). How to test: check whether MSA clauses are recurring/renewing obligations vs one-time — read 10 real clauses.

Risk flags

1. Market timing / one-time-scramble risk: demand could spike into Aug 2026 then soften if enforcement stays toothless and clients stop pushing clauses. Mitigate by anchoring on the recurring MSA obligation, not the fine.
2. Regulatory drift: the Code of Practice is voluntary and evolving; label formats or the provider/deployer line could shift, forcing rework. Platform-dependency on a moving standard.
3. Incumbent expansion: a generation platform (Adobe, 3Play) or Arkiv could add cross-tool disclosure packs. The freelancer network + EU-SMB pricing is the defence.
4. Trust risk on auto-detection: a wrong “human-made” label that a client later disputes damages the agency and us. Human-approval gate required before any pack is finalized.

14. Structured verdict

Score:                  73/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder + localization-industry advisor + on-call EU AI-Act lawyer
Time to revenue:        8–12 weeks (pre-sell now, ship v1 ~Sept 2026)
Capital to launch:      ₹8–15 lakh ($10–18K) — mostly legal advisor + eval-data labelling
Top 3 assumptions to validate first:
  1. Agencies are receiving AI-disclosure clauses in MSAs NOW — 25 owner interviews, need ≥8 clauses
  2. €249/mo willingness-to-pay over a spreadsheet — pre-sell to 15, need ≥5 conversions
  3. Auto-detection ≥85% precision on "AI-touched" — labelled 500-asset eval
Kill criteria:
  - Abandon if <5 of 25 interviewed agencies can produce a real AI-disclosure MSA clause
  - Abandon if <5 of 15 pre-sell targets convert to a paid annual plan
  - Abandon if a generation-platform incumbent ships cross-tool disclosure packs before your v1

15. Next step — 1-week validation sprint

• Day 1–2: Pull GALA/Nimdzi EU LSP lists; line up 25 owner interviews (dubbing, subtitling, video-loc shops). Draft the interview script around one question: “show me the AI clause your last enterprise client sent you.”
• Day 3–4: Run interviews. Count how many can produce a real clause and how they track disclosure today (spreadsheet? nothing?). Present a €249/mo pre-sell to warm ones.
• Day 5: Decide. Falsifiable go signal: ≥8/25 produce a real MSA disclosure clause AND ≥5 commit to a paid annual pre-sell. Below either threshold = the pain isn’t landing yet; shelve until closer to the deadline or pivot to the brand-in-house deployer buyer.