SignOff

1. One-liner

Proves a human really edited the AI-drafted article, so publishers keep the Article 50 labelling exemption.

2. Trend signal — why now?

On 2 August 2026 — two weeks from today — Article 50 of the EU AI Act becomes enforceable. It says that if you publish AI-generated text “with the purpose of informing the public on matters of public interest,” you must slap a visible “this was AI-generated” label on it.

There is one escape hatch, and it is the whole business: the obligation does not apply where the content “has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility.” The guidance is explicit that these checks must be substantive and not limited to superficial matters or cursory approval, and that signatories publishing unlabelled AI-assisted text about public affairs must be able to prove substantive human control.

That word — prove — is the product. Every publisher in Europe is going to claim the exemption. Almost none of them can currently evidence it.

Meanwhile the volume that needs evidencing is exploding. Newsquest, a major UK regional publisher, now runs 30+ “AI-assisted reporters” across its titles using an internal tool to draft up to 30 stories a day each, which are then checked by humans. That is the exact workflow Article 50 is aimed at, at industrial scale.

And the downside is not a slap on the wrist. Article 99(4) puts Article 50 breaches at up to €15,000,000 or 3% of worldwide annual turnover, whichever is higher — the second-highest penalty band in the entire Act.

The tell that this is unserved: the compliance industry has already published the exact schema of the record a regulator will want — AI system inventory ID, Article 50 scenario, notice text, content sample, human review note, approval owner, publication date, change log — and framed the standard as “a reviewer should be able to reconstruct the decision without interviewing five teams six months later.” They’ve written the spec. Nobody has shipped the software.

Provenance:

• Signal 1 (demand/regulatory): Article 50 enforceable 2 Aug 2026; editorial-control exemption requires substantive, provable human review — https://artificialintelligenceact.eu/transparency-rules-article-50/ — observed 2026-07-19
• Signal 2 (feasibility/volume): Newsquest operating 30+ AI-assisted reporters drafting up to 30 stories/day each, human-checked — https://reutersinstitute.politics.ox.ac.uk/journalism-media-and-technology-trends-and-predictions-2026 — observed 2026-07-19
• Signal 3 (economic): Article 99(4) penalties up to €15M or 3% of global turnover for Article 50 breaches — https://artificialintelligenceact.eu/article/99/ — observed 2026-07-19
• Signal 4 (gap): Compliance guidance specifies the required evidence fields and “reconstruct without interviewing teams” standard, but offers checklists rather than tooling — https://www.llmcms.org/guides/the-coming-ai-audit-what-editors-will-need-to-prove — observed 2026-07-19 Category: Regulatory arbitrage

3. The opportunity

The compliance market has split badly, and the gap in the middle is mine.

On one side: C2PA / Content Credentials (Adobe, TrueScreen, ContentCredentials.org). These are excellent at cryptographically marking media — images, video, audio — as machine-generated. They answer “was this pixel synthetic?”

On the other side: GRC consultancies and law firms selling €5–20K “Article 50 readiness assessments” — a PDF, a policy template, an invoice, and no ongoing system.

Neither addresses the actual daily operational question for a text publisher: for this specific article, published on this date, can I demonstrate that a named human did substantive editorial work on an AI draft? C2PA doesn’t do text-editing-effort. The consultants hand you a policy and leave.

Worse, the naive in-house answer is actively dangerous. Most CMSs have revision history, so an editor will say “we’re fine, we have version control.” They are not fine. A revision log shows that a document changed; it does not distinguish an editor who restructured the piece, corrected two factual claims and rewrote the lede from an editor who fixed a typo and hit publish. Under a standard that explicitly excludes “cursory approval,” undifferentiated revision history is evidence against you as easily as for you.

What a focused tool does 10× better: it sits on the draft-to-publish transition, measures the substance of the human edit, captures the named accountable person at the moment of publication, and freezes it all into an immutable per-article record that can be produced on demand. Not a policy. A receipt.

4. Target market

• Primary customer: Head of Editorial Operations / Managing Editor / Editor-in-Chief at EU and UK digital publishers running 20–500 published items per day with an AI drafting step. Regional news groups, B2B trade publishers, and large content/SEO agencies producing bylined editorial for clients. Team size 10–200. Revenue €2M–€80M.
• Secondary customer: Compliance/DPO function at those same organisations — often the actual budget holder, since they already own the GDPR tooling line item and Article 50 lands in their lap.
• Why they buy: Article 50 forces a binary choice on every AI-assisted article — label it, or defend the exemption. Labelling is commercially unattractive (it flags the content as machine-made to readers and, publishers fear, to distribution algorithms). Defending the exemption without evidence is a €15M gamble. They buy the third option.
• Rough TAM reasoning: The EU has thousands of registered news publishers plus a much larger population of trade publishers and content agencies. I don’t need a precise count — I need a few hundred paying accounts. If ~3,000 EU/UK organisations publish public-interest text at meaningful volume with an AI step in the workflow, capturing 400 of them at €500/mo is a €2.4M ARR business. That’s the whole thesis; it does not require the market to be enormous.
• Why now for them: The deadline is 2 August 2026. Generative systems already on the market get until 2 December 2026 for machine-readable marking — but the human-review evidentiary question bites in August. This is a dated, non-negotiable forcing function, not a nice-to-have.

5. Product sketch (MVP)

• Edit-substance scoring — on the draft→publish transition, compares the AI draft against the published text and produces a substantive-edit score with a breakdown: structural changes, factual claim modifications, quotes added, lede rewrite, sourcing changes. Distinguishes real editorial work from typo-fixing.
• Named accountability capture — records which human held editorial responsibility at publication, with role and timestamp. This is the specific natural/legal person the exemption requires.
• Per-article compliance record — an immutable, timestamped record per published item containing the fields the guidance calls for: AI system used, model version, Article 50 scenario, draft and final samples, edit-substance evidence, approval owner, publication date, change log.
• Thin-review alerts — flags articles that went out with an edit-substance score below your policy threshold, before they become an audit problem. This is the feature editors actually love, because it catches process decay in real time.
• Regulator-ready export — one click produces a defensible dossier for a specific article, date range, or title, formatted so a market surveillance authority can reconstruct the decision unaided.
• Policy threshold configuration — set your organisation’s bar for what counts as substantive review, so the standard is documented and consistently applied rather than improvised per editor.
• CMS integrations — WordPress and Arc/Ghost/custom via API at launch; the record is captured from existing workflow, not by making journalists do extra data entry.

6. AI angle — what’s load-bearing

Remove the AI and this product does not exist.

The core function is a semantic comparison of AI draft versus published text that determines whether the human edit was substantive. That is not a diff. A character-level diff scores “changed 4% of characters” identically whether the editor corrected a false claim about a named politician or fixed a comma. The regulation’s own language — “substantive and not limited to superficial matters” — is a semantic judgement, so the tool must make a semantic judgement.

The model has to classify edit types (factual correction, structural reorganisation, sourcing addition, tone change, typo), weight them by editorial significance, and produce a defensible, explained score. It also has to generate the human-readable rationale that goes in the record — “the editor rewrote the lede, corrected the attributed figure from €4.2M to €3.8M, and added two named sources” — because a number alone won’t persuade a regulator.

This is genuinely load-bearing AI doing work no rules engine can do.

7. Localization angle

EU-first by construction — the regulation is the market. But there’s real intra-EU localization work that doubles as a moat:

• Language coverage is table stakes, not a feature. Edit-substance scoring has to work in German, French, Italian, Spanish, Polish, Dutch and Nordic languages, not just English. A US-built competitor will ship English-only and lose the regional publishers who need this most.
• Enforcement is national, not central. Market surveillance authorities are designated per member state, so the export format and the questions asked will vary by country. Knowing that the German and Irish authorities differ in posture is exactly the kind of unglamorous knowledge that compounds.
• UK is an adjacent, not identical, market. UK publishers with EU readership are in scope, and New York has passed a bill requiring both AI disclosure and human review of AI-assisted news — the first attempt to legally define what an editor’s presence requires. If that spreads, the same product sells into US states with a relabelled export.

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

• Pricing: €300/mo (Starter, ≤1,000 articles/mo, 1 title) · €800/mo (Publisher, ≤10,000 articles/mo, multi-title, multi-language) · €2,000/mo (Group, unlimited titles, SSO, custom export, dedicated support)
• ACV: ~€7,200 blended (weighted toward the €800 tier, which is where regional groups and trade publishers land)
• Rough math to $1M ARR: 140 customers × €800/mo × 12 ≈ €1.34M. Achievable with a single well-run outbound motion against a named list — this is a few hundred accounts, not thousands.
• Rough math to $5M ARR: ~500 accounts with mix shifting toward Group tier, plus the two natural expansions below. Requires winning 2–3 large publishing groups as anchor logos and expanding seat/title count within them.
• Expansion path: (1) per-title expansion within groups — a publisher with 40 regional titles onboards 3 and rolls out; (2) adjacent Article 50 scenarios — chatbot disclosure and deepfake labelling records, same evidence architecture; (3) the record archive itself becomes the upsell, because retention of proof matters for as long as the enforcement lookback window, and nobody wants to migrate an evidence archive.

The margin structure is friendly: one model call per published article on a €800/mo account with 10,000 articles is still a comfortable gross margin, and cost per article falls as inference keeps getting cheaper.

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

The deadline does the selling. My job is to be in the room during the panic window of July–October 2026.

1. The named-list outbound play. EU/UK publisher membership directories are public — national press associations, WAN-IFRA’s membership, trade publishing associations. Build a list of ~1,200 organisations publishing public-interest text at volume. For each, check whether they’ve published an AI-use policy page (many have, as a trust signal — which conveniently self-identifies them as AI-assisted publishers). Email the Managing Editor a 90-second Loom showing their own recent article scored for edit substance. Deadline-driven cold outreach with a personalised artifact should beat generic SaaS reply rates; budget for 3–5% and 100 conversations.
2. Ride the consultants, don’t fight them. Law firms and GRC consultancies are actively selling Article 50 readiness assessments right now. Their deliverable ends at “you need a documented editorial workflow with identified responsible persons” — and then they have nothing to install. Partner with 5–10 of these firms as the implementation layer, on referral fee or reseller margin. They have the meetings already booked; I have the thing they can’t build.
3. The trade press and association channel. WAN-IFRA, national publisher associations, and journalism-tech newsletters (Nieman Lab, Press Gazette) are covering Article 50 continuously into August. Publish the single most useful free artifact in the category — an “Article 50 editorial evidence self-audit” that scores 10 of a publisher’s own articles free — and let the coverage carry it. Free scoring is the lead magnet and the demo.
4. CMS ecosystem placement. A WordPress plugin listed against searches for AI compliance puts the product in front of the long tail of smaller trade publishers with zero sales effort. Lower ACV, but it feeds the Starter tier and the logo count.
5. Anchor logo, then cascade. Land one visible regional group (the Newsquest-shaped buyer — many titles, industrialised AI drafting, real brand risk) at a discount in exchange for a public case study. In a herd-like, association-dense industry, one named reference in the trade press does more than six months of ads.

10. Build complexity — justification

Low. No custom models, no novel infrastructure. The v1 is a CMS webhook that captures draft and published versions, an LLM call that classifies and scores the edit delta with a structured rationale, an append-only record store, and a PDF/JSON export. Everything is off-the-shelf.

The genuinely hard parts are not engineering: (a) calibrating the substantive-edit threshold so it matches what a regulator would accept — which is a judgement problem solved by talking to media lawyers, not by writing code; and (b) CMS integration breadth, which is grind rather than difficulty. A technical founder ships a credible v1 in 6–8 weeks, and the WordPress path alone covers a large share of the long tail.

The tight window is the real constraint. Shipping in September beats shipping something more polished in January.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Compliance tooling; no regulated licence needed to sell it
Ethical — no harm / dark patterns	✅	Net-positive: incentivises genuine editorial review over rubber-stamping. Risk of misuse as a “compliance theatre” tool is mitigated by scoring substance honestly rather than optimising to pass
Market exists (evidence above)	✅	Dated statutory deadline, €15M/3% penalties, documented industrial-scale AI drafting
1–5 person team can build this	✅	Off-the-shelf stack, 6–8 week v1
Launchable with <$50K / ₹40L	✅	Well under; main cost is founder time plus legal review of the evidence format

All five pass.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	15/20	Hard deadline, severe penalty ceiling, binary choice per article. Docked because enforcement in month one will be light and many publishers will rationally gamble that nobody audits them in year one — the pain is anticipated more than currently felt
Demand evidence	15	11/15	Regulation, penalties and AI-drafting volume are all hard-sourced. Docked because I could not source a publisher explicitly saying “I will pay for proof-of-review tooling” — the willingness-to-pay is inferred from the penalty, not observed
Build feasibility	15	13/15	Off-the-shelf; 6–8 weeks. Docked slightly for CMS integration breadth
Distribution clarity	15	12/15	Named, enumerable list; consultant channel already holding the meetings; free-scoring lead magnet. Docked because publisher sales cycles run slower than the deadline urgency suggests
Revenue mechanics	15	11/15	Pricing sits comfortably below the consultancy alternative and far below the fine. Docked because compliance tooling faces real churn risk once the initial panic subsides
Time to first revenue	10	8/10	Deadline-driven urgency; pre-selling before August is plausible. Not 9–10 because publisher procurement rarely closes in two weeks
Defensibility	10	5/10	Honest score. The core is an LLM call on a diff — a competent team copies the mechanic in a quarter. Real moat accrues later: the evidence archive is sticky (nobody migrates proof), multi-language calibration is grind, and per-country enforcement knowledge compounds. But at month three this is execution-only
Total	100	75/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · content-heavy

Needs someone who can ship the integration and scoring layer fast, plus enough publishing-industry fluency to speak credibly to editors and write the content that drives the association channel. A media-lawyer advisor on the evidence format is close to mandatory.

Key assumptions to validate (3–5)

1. Assumption: Publishers will pay for proof of review rather than just adding a visible AI label and moving on. How to test: 25 interviews with EU managing editors before writing code — ask directly which they’d choose and why. If most shrug and say “we’ll just label it,” the entire thesis collapses and the answer arrives in week one.
2. Assumption: Editors accept an automated substantive-edit score as fair rather than resenting it as surveillance of their craft. How to test: run the scorer over 50 real published articles with 5 friendly newsrooms and ask editors whether the scores match their own judgement of which edits were substantive.
3. Assumption: CMS revision history is genuinely insufficient, and buyers can be shown this in one minute. How to test: build the “your revision log doesn’t prove what you think” demo and measure whether it converts skeptics in live calls.
4. Assumption: The consultant/law-firm channel will refer rather than build a competing spreadsheet-and-template offering. How to test: sign 3 referral partners with terms before launch.

Risk flags

1. Regulatory drift: The Transparency Code of Practice is still settling, and the AI Omnibus has already moved dates once. If final guidance says revision history is sufficient evidence, the wedge narrows sharply. Watch guidance releases weekly; this is the single biggest binary risk.
2. Enforcement apathy: If national market surveillance authorities do nothing visible in the first 12 months, urgency evaporates and this becomes a vitamin. The business likely needs one publicised enforcement action to reach the €5M case.
3. Platform dependency / incumbent absorption: CMS vendors are already embedding AI into editorial workflows and publicly insisting human control is non-negotiable. Any of them could ship a compliance-record feature natively and reduce this to a plugin. Speed and multi-CMS neutrality are the only defences.
4. Thin defensibility: Scored honestly at 5/10. The idea works as a fast, focused, profitable business — not as a durable franchise unless the evidence archive and enforcement knowledge compound faster than clones appear.
5. Buyer ambivalence: Publishers may quietly prefer not to have a system that documents when their review was cursory. A tool that creates discoverable evidence of thin oversight is a tool some buyers will avoid. The product must be positioned as protection, and the thin-review alert as a chance to fix process before publication — not as a confession log.

14. Structured verdict

Score:                  75/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder with publishing-industry fluency and a media-law advisor
Time to revenue:        6–10 weeks (deadline-driven; pre-sell before 2 August)
Capital to launch:      €8–15K ($9–17K) — mostly legal review of the evidence format
Top 3 assumptions to validate first:
  1. Publishers choose defending the exemption over just labelling — 25 managing-editor interviews, this week
  2. Edit-substance scores match editors' own judgement — score 50 real articles with 5 newsrooms
  3. Consultants refer rather than compete — sign 3 referral partners pre-launch
Kill criteria:
  - Abandon if >60% of 25 interviewed editors say they'll simply apply the AI label and skip the exemption
  - Abandon if final Transparency Code guidance accepts plain CMS revision history as sufficient proof
  - Abandon if a major CMS vendor ships a native Article 50 evidence record before v1 launches
  - Abandon if <5 paying customers by 90 days post-launch, given the deadline should make this the easiest sales window the product will ever have

15. Next step — 1-week validation sprint

• Day 1–2: Build the list — 300 EU/UK publishers with a public AI-use policy page (self-identified AI-assisted publishers). Pull 3 recent articles from each of 20 of them.
• Day 3: Hack a throwaway edit-substance scorer. No product, no UI — just a script producing a scored rationale for a draft/published pair, using AI-drafted-then-edited samples from friendly contacts.
• Day 4–5: Book and run 25 conversations with managing editors and compliance leads. One question above all: “On 2 August, for an AI-drafted article — do you label it, or claim the exemption? And if you claim it, what would you show a regulator?” Say nothing about the product until they’ve answered.
• Day 5: Decide.

Falsifiable outcome: go only if ≥12 of 25 say they intend to claim the exemption AND cannot describe adequate evidence they currently hold, AND ≥5 agree to a paid pilot at €300+/mo before the product exists. Anything less and the market has told me it will just apply the label — which is a perfectly rational choice for them and a fatal one for this business.