SignedInLanguage

1. One-liner

Records who approved each translated compliance course, so an inspector’s language question has a dated answer.

2. Trend signal — why now?

Three things moved at once.

First, AI dubbing and AI translation got cheap enough that compliance teams stopped budgeting per-language. Professional dubbing still runs $20–$50+ per finished minute with 2–6 week turnarounds per language; AI workflows collapsed that to near-zero marginal cost. The result is not “better translations” — it is more translated content shipping, faster, with fewer humans looking at it. RWS’s own published workflow still insists on three human checkpoints (script review, cultural review, final sign-off) precisely because automated pipelines “can produce plausible-sounding mistranslations that a native reviewer would catch immediately.”

Second, the liability got named. Compliance-training commentary in 2026 is explicit that “in regulated industries (financial services, pharmaceuticals, manufacturing), untranslated or poorly translated mandatory training creates documented evidence of a compliance failure,” and that mistranslated terminology “can undermine legal enforceability.” Quiz questions are called out as the sharpest risk — ambiguous translated wording can make a wrong answer look correct, which corrupts the competency data the company later relies on.

Third — and this is the wedge — auditors started asking a question nobody has a file for. Translation-audit guidance states inspectors ask for “documented evidence of how translations were produced: who translated, who reviewed, what reference materials were used, how versions were controlled and how the process was applied consistently across the programme.” Regulators “often ask which policy applied on a specific date, usually the date of an incident.”

Put those together: companies now ship 12 language versions of a mandatory course generated largely by machine, and when an incident happens in the Polish subsidiary, nobody can produce the record of who signed off that the Polish version said the right thing.

The incumbents solve the adjacent problem, not this one. Every TMS — memoQ, Lionbridge, POEditor — sells in-context review to make translators faster. LMS vendors sell version-controlled completion records tying an employee to a content version. Neither produces the artifact in the middle: a dated, attributable attestation that this human, with this competence, affirmed this translated obligation is legally faithful.

Provenance:

• Signal 1 (demand): Compliance-training localization identified as creating liability — mistranslated terms “undermine legal enforceability,” translated quizzes can make wrong answers appear correct — https://agneselisa.net/blog/the-hidden-risks-inside-poorly-translated-elearning-courses-a-guide-for-compliance-teams/ and https://www.truelanguage.com/?p=121023 — observed 2026-07-21
• Signal 2 (feasibility): AI dubbing/translation now production-viable and cheap, but published best-practice workflows (RWS) still mandate three named human checkpoints — creating many sign-off events that currently go unrecorded — https://www.rws.com/blog/ai-dubbing-in-2026/ and https://www.3playmedia.com/blog/ai-dubbing-vs-traditional-dubbing/ — observed 2026-07-21
• Signal 3 (economic): Enterprise compliance training runs $90K–$250K/yr all-in for a 2,500-person company; regulators demand documented translator/reviewer/version evidence — https://coggno.com/blog/lms/compliance-training-cost-per-employee-2026-industry-benchmarks-construction-healthcare-manufacturing-retail/ and https://sumalatam.com/preparing-translations-regulatory-audits-what-to-document/ — observed 2026-07-21 Category: Tech-unlock

3. The opportunity

Cheap AI translation created a volume problem disguised as a savings win. A compliance team that used to localize one flagship course into four languages now ships six courses into fourteen, because the marginal cost went to nearly nothing. The human review budget did not grow to match. So review became sampling, and sampling became a checkbox.

That is fine right up until an incident. Then the question is not “was the translation good?” — it is “prove who decided it was good, and when.” The company has an LMS record showing Marta in Kraków completed Anti-Bribery v4 (PL) on 12 March. It has no record of who ever confirmed that the Polish v4 accurately conveyed the bribery threshold, or whether anyone did.

The incumbents are structurally uninterested. A TMS optimizes translator throughput and treats review as a workflow state to clear — its economic incentive is to remove review friction, not to memorialize it. An LMS versions the content object but is agnostic about whether the content is correct; it records delivery, not validity. The gap sits between two vendors, which is exactly why neither owns it.

What a focused team does 10× better: make the attestation itself the product. Not faster review — recorded, attributable, defensible review, with AI doing the part humans are worst at (finding which of 400 translated segments actually carry legal weight and deserve a human’s twenty minutes).

4. Target market

• Primary customer: Compliance / L&D operations lead at a 500–5,000-employee regulated multinational — financial services, pharma, medical devices, manufacturing — running mandatory training in 5+ languages. Title is usually Compliance Training Manager, Global L&D Manager, or Ethics & Compliance Operations. Buyer sits under a Chief Compliance Officer who has personal exposure when an audit goes badly.
• Why they buy: Their stated fear is not a bad translation. It is being asked a question in a room with auditors and having to say “I’d have to check.” They already spend $90K–$250K/yr on the training stack; a few thousand dollars to make one recurring audit question answerable is rounding error against that, and against the cost of a finding.
• Rough TAM reasoning: This is deliberately a narrow band. Companies large enough to be multilingual and regulated, small enough to lack a dedicated localization QA function (the Fortune 100 build this in-house or buy Lionbridge’s enterprise tier). Ballpark tens of thousands of firms globally across EU, UK, US, and APAC regulated sectors. At the pricing below, low thousands of customers is a $5M business — I do not need a big share.
• Why now for them: Their translated-content volume roughly doubled in the last 18 months because AI made it cheap, while their reviewer headcount stayed flat. The exposure grew silently. The first big publicized enforcement action that turns on a translated training module will make this a board question — and I would rather sell before that than after.

5. Product sketch (MVP)

• Risk-ranked segment triage. Ingest a course’s source and target versions; surface the segments that carry legal or behavioural weight (obligation thresholds, reporting duties, quiz stems and answer options) and rank them for human attention. The reviewer sees 30 segments that matter, not 400 that don’t.
• Back-translation and divergence flagging. For each high-risk segment, show the source, the target, and a machine back-translation, with meaning drift called out in plain language (“target says may report; source says must report”).
• Named attestation with scope. A reviewer signs off on specific segments, with their name, role, language competence, timestamp, and what they did and did not review. Partial attestation is a first-class state — “I reviewed the quiz, not the narration” is recorded honestly rather than rounded up to approved.
• Version-diff re-attestation. When the source course changes, show exactly which previously-attested segments are now stale, so re-review is scoped to the delta rather than restarting.
• Audit export. One PDF/CSV per course-language-version answering the inspector’s actual questions: who translated, who reviewed, what glossary and reference material was used, what changed between versions, and when each was signed.
• Glossary of regulated terms. Per-client controlled vocabulary for the terms that must never drift, with violations flagged automatically on every future version.
• Standing coverage view. Across the whole catalogue: which course-language pairs are currently attested, which are stale, which were never reviewed by a human at all.

6. AI angle — what’s load-bearing

Remove the AI and the product collapses into a sign-off form — which is what a spreadsheet already is, and why the current process is bad.

The load-bearing work is triage and divergence detection. A 45-minute compliance course is roughly 400–700 translatable segments across fourteen languages. No compliance manager will review 9,000 segments, and asking them to is why review became sampling. The model does three things a human cannot do at that volume: classify which segments are legally load-bearing versus decorative, detect semantic drift between source and target (modal verbs, thresholds, negations, scope qualifiers — the small words that flip an obligation), and detect when a translated quiz has become ambiguous about which answer is correct.

That collapses a genuinely impossible review into roughly twenty focused minutes per language. The AI does not sign anything and does not get a vote — a human attests, and the record says so. That division is deliberate: an AI-generated approval is worth nothing in an audit, so the product’s value depends on a human being accountable at the end. AI creates the leverage; the human creates the evidence.

7. Localization angle (if any)

N/A as a geographic wedge — but note the irony that this is a product about localization sold globally in English. The buyer is a multilingual company, but the buying centre works in English (or the HQ language), so the UI does not need localization to sell. Where geography matters is regulatory templating: EU works councils, FCA/PRA in the UK, FDA/EMA in life sciences, and APAC labour regimes each phrase the “which version applied” question differently. Shipping per-regime export templates is a later moat, not a v1 requirement.

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

• Pricing: $600–$1,500/mo per workspace, tiered by number of active course-language pairs under attestation. Land at the low tier on a single high-risk course family (anti-bribery, safety, whistleblowing), expand by catalogue.
• ACV: ~$12K realistic average, with a long tail at $7K and enterprise-ish accounts at $25–30K.
• Rough math to $1M ARR: 85 customers × $1,000/mo × 12 = $1.02M. That is a reachable number for a two-person team over roughly 24 months at the conversion rates in section 9.
• Rough math to $5M ARR: ~420 customers at the same ACV, or ~250 customers with ACV expanded to $20K via catalogue growth and per-regime export modules. The second path is more likely than the first — this category expands within an account far more easily than it acquires new ones.
• Expansion path: Priced on course-language pairs, so expansion is mechanical — every new market entry and every new mandatory course grows the account without a new negotiation. Secondary expansion into adjacent attested artifacts: translated policies, SOPs, safety signage, supplier codes of conduct. Same buyer, same audit question, same ledger.

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

• The audit-question cold email. Build a list of ~2,000 compliance and L&D leads at 500–5,000-employee regulated multinationals via LinkedIn Sales Navigator, filtered on regulated SIC codes plus 3+ country presence. The email asks one question: “If an inspector asked who approved the Polish version of your anti-bribery course, how long would it take you to answer?” No demo, no deck. That question either lands like a punch or gets ignored — which makes it a cheap, fast qualifier. Expect 2–4% reply, ~25% of replies to a call.
• Free catalogue exposure audit as the wedge. Prospect uploads two language versions of one existing course; get back a report showing which segments carry legal weight and which have drifted. This is the entire sales motion — it converts an abstract worry into a specific list of their own divergences, with names of their own courses on it. Deliver it in 48 hours, live. Target 30–40% audit-to-paid on qualified accounts.
• The LSP and compliance-consultancy channel. Mid-tier language service providers and compliance consultancies already sit in this workflow and are already being asked for review evidence they cannot cleanly produce. They are not competitors at this layer — attestation makes their deliverable more defensible. Sign 5–8 as referral partners at 15–20% first-year revenue share; each brings a warm portfolio.
• Speak where the question gets asked. Compliance and ethics practitioner communities (SCCE/ECI-adjacent events and forums, regional compliance officer associations) — a 20-minute talk titled “What auditors ask about your translated training, and why you can’t answer it.” This is a small, high-trust world where one credible talk produces a quarter of warm inbound.
• The incident-driven inbound bet. Publish plain-language teardowns whenever a regulatory action touches multilingual training or policy communication. Low volume, but it captures the exact buyer at the exact moment their CCO forwards them the news article.

10. Build complexity — justification

Low. Everything material is off-the-shelf: XLIFF/SCORM/xAPI parsing for course ingestion, standard LLM calls for segment classification, back-translation, and drift detection, an append-only sign-off record, and PDF export. No custom models, no novel infrastructure, no research risk. The genuine work is unglamorous: format ingestion across the messy real world of course exports, and getting the risk-classification prompt tuned well enough that reviewers trust the triage. A pair ships a credible v1 in 8–10 weeks; a solo builder in 12–14.

The hard part is not engineering, it is knowing what an auditor actually accepts as evidence. That is a domain problem, solved by a design partner or an advisor, not by more code.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Records attestations; makes no legal determination and gives no advice.
Ethical — no harm / dark patterns	✅	Net positive: makes honest partial review visible instead of letting blanket approval hide it.
Market exists (evidence above)	✅	Documented liability commentary + published audit-evidence expectations + existing six-figure training budgets.
1–5 person team can build this	✅	Off-the-shelf stack, 8–14 weeks to v1.
Launchable with <$50K / ₹40L	✅	Two people, inference costs, list tooling. Well under.

All five pass.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	15/20	Real and consequential, but episodic rather than daily. It bites at audit or incident time, and plenty of teams will keep gambling that neither arrives this year. That gap between real risk and felt urgency is the single biggest thing holding this below a GO-plus.
Demand evidence	15	11/15	Strong indirect evidence: named liability, published auditor expectations, real budgets, incumbents solving the adjacent problem. What I do not have is a verbatim compliance manager saying “I need this and I’ll pay.” That absence is why this isn’t 13+.
Build feasibility	15	13/15	Off-the-shelf throughout. Ingestion-format sprawl is the only real drag.
Distribution clarity	15	11/15	Named list, named channel, a qualifying question that self-selects hard, and a free audit that converts worry into specifics. Marked down because compliance buyers are slow and often need Legal and Procurement in the room.
Revenue mechanics	15	11/15	Pricing sits comfortably inside an existing six-figure stack, and the course-language-pair unit expands mechanically. Marked down because ACV is an estimate — no benchmark product prices this exact thing.
Time to first revenue	7	7/10	8–12 weeks realistically, via design partners paying during build. Not faster, because this buyer rarely signs without a security review.
Defensibility	10	3/10	Honestly weak. A TMS or LMS vendor could bolt on attestation in a quarter if they decided it mattered. What accumulates is the client’s own attestation history and controlled glossary — real switching cost, but only after a year of use. This is an execution-and-focus play, not a moat play.
Total	100	71/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · domain-expertise-required

Needs someone who can ship the ingestion and classification work, paired with genuine access to compliance auditors. Without the second half, the export produces something that looks official and satisfies nobody — which is worse than useless in an audit.

Key assumptions to validate (3–5)

1. Assumption: Compliance leads at 500–5,000-person regulated multinationals cannot currently answer “who approved the Polish version?” within a day. How to test: Ask exactly that question to 40 of them via cold email and LinkedIn; count how many say “quickly” versus “I’d have to check.” A majority answering “quickly” kills the premise.
2. Assumption: They will pay $600–$1,500/mo for a record, absent an active audit. How to test: Take the free catalogue audit to 15 qualified accounts and ask for a paid pilot at list price. Interest without payment is the failure mode to watch for here — this is a category where everyone agrees it’s important and nobody has a budget line.
3. Assumption: AI segment triage is accurate enough that reviewers trust it. How to test: Run 5 real courses past 3 experienced compliance translators; measure whether the model’s high-risk set captures the drifts they independently flag. Below ~85% recall on genuinely material divergences, reviewers will re-read everything and the leverage evaporates.
4. Assumption: This is a distinct budget line, not something the LMS is expected to include. How to test: In discovery, ask directly whether they would expect their LMS vendor to provide it — and whether they would wait for that vendor to build it.

Risk flags

1. Incumbent absorption: The clearest risk. An LMS or TMS vendor adds attestation as a feature and reframes this as a checkbox rather than a product. Mitigation is depth on the audit-export side and per-regime templating — the part a generalist vendor will do shallowly.
2. Urgency gap: The pain is episodic. Most of the market will nod, agree it matters, and not buy until something forces the issue. This lengthens the sales cycle and is the main reason confidence here is Medium rather than High.
3. Evidentiary credibility: If the export is not what auditors actually accept, the product’s core promise is hollow, and the failure only surfaces at the worst possible moment for the customer. Must be shaped with real audit input before scale, not after.
4. Buyer-committee drag: Compliance purchases pull in Legal, Security, and Procurement. Sub-$1,500/mo pricing is deliberate — it aims to stay under the threshold that triggers full procurement, but that will not always hold.

14. Structured verdict

Score:                  71/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder shipping solo, paired with a compliance-audit advisor
                        or a design-partner CCO. Domain access is the binding constraint, not code.
Time to revenue:        8–12 weeks (design partners paying during build)
Capital to launch:      $8–15K
Top 3 assumptions to validate first:
  1. Compliance leads genuinely cannot answer the "who approved this language version"
     question fast — test with 40 cold outreaches asking only that
  2. Willingness to pay absent an active audit — take the free audit to 15 accounts,
     ask for a paid pilot at list price
  3. Triage recall is high enough to be trusted — 5 real courses past 3 professional
     compliance translators, measure against their independent flags
Kill criteria:
  - Abandon if >50% of 40 compliance leads say they could answer the language-version
    question within a day (premise is false)
  - Abandon if <3 of 15 free-audit recipients convert to a paid pilot (interest without budget)
  - Abandon if model recall on material divergences is <85% against professional reviewers
    (no leverage, product is just a form)
  - Abandon if a major LMS or TMS vendor ships equivalent attestation before v1

15. Next step — 1-week validation sprint

• Day 1–2: Build the list of 200 compliance and L&D leads at regulated multinationals with 3+ country presence. Send the single-question email — “If an inspector asked who approved the Polish version of your anti-bribery course, how long would it take you to answer?” Nothing else. No pitch.
• Day 3–4: Take 5 real multilingual compliance courses (publicly available ones, or design-partner supplied) and run source-versus-target divergence detection by hand with model assistance. Send each finding set to 3 professional compliance translators and ask a blunt question: is this drift material, and would you have caught it?
• Day 5: Decide on two hard numbers. Go if ≥8 of the ~200 emails come back saying they could not answer quickly, and ≥2 of those ask to see more unprompted; and if translators confirm ≥85% of flagged divergences as genuinely material. No-go if compliance leads mostly answer “our LMS handles that” — that answer means the buyer believes the problem is already solved, and belief is what I’d be selling against.

The falsifiable core: either these people can answer the language-version question today, or they cannot. One week of asking settles it, and it settles it before a line of production code gets written.