LabelLag

1. One-liner

Watches each product’s live recipe against its printed label and flags the allergen, QUID, or ingredient mismatch before the next print run ships — for small EU and UK packaged-food producers.

2. Trend signal — why now?

The recall data has gone from “a known risk” to “the dominant failure mode,” and it traces to one boring gap nobody owns.

• Half of all recalls are now a labelling problem, not an ingredient problem. Per Esko’s 2025 analysis of 251 FDA food-and-beverage recalls, 45.8% were linked to allergen mislabelling — “nearly half of recall events weren’t caused by the ingredient itself, but by failures in the information surrounding it: labels, specs, artwork, claims, and packaging workflows.” Allergens have been the single leading recall cause across both the EU (RASFF) and UK for several years running.
• The UK is recalling for this once every four days. The FSA issued 85 allergen alerts in 2025 (~one per 4 days); its 2024/25 surveillance sampling found 4% of products had undeclared allergens, milk most common. EU RASFF Q1 2025: milk alone cited in 19 recalls, “signalling gaps in label verification.”
• The named root cause is a sync failure, not negligence. Industry teardown (Esko, Food Dive, Meatingplace 2025): “A product formulation changes with a new ingredient, supplier swap, or recipe tweak, the update gets recorded in the specification document, but the label artwork — created months earlier and managed by a different team — doesn’t get updated. The connection between the spec and the label exists in theory, but in practice it depends on someone remembering to check.” A food-safety practitioner on the IFSQN forum puts it plainly: “Food businesses have gotten into recalls simply because they forgot to update one document when a recipe changed.”

Three independent angles — recall statistics, regulator enforcement cadence, and the documented mechanical root cause — all point at the same unguarded seam: the recipe moves, the label doesn’t, and no software is watching the gap.

Provenance:

• Signal 1 (demand): “45.8% of 2025 FDA recalls = allergen mislabelling… failures in the information surrounding the ingredient.” — https://www.esko.com/en/blog/fda-food-and-beverage-recalls-2025 — 2025
• Signal 2 (economic/regulatory spend): UK FSA issued 85 allergen alerts in 2025 (~1 every 4 days); 4% of sampled products had undeclared allergens; EU RASFF Q1 2025 milk in 19 recalls. — https://www.foodhygienecertificate.co.uk/blog/food-recall-statistics-uk / https://qms4.com/q1-2025-eu-uk-food-recalls-trends/ — 2025
• Signal 3 (root cause / workaround): “Food businesses have gotten into recalls simply because they forgot to update one document when a recipe changed.” — https://www.ifsqn.com/forum/index.php/topic/32446-handling-labeling-updates/ — 2025 Category: Workflow automation

3. The opportunity

The market is split into two camps and the most expensive failure falls straight down the crack between them.

Camp A — label generators (ReciPal ~$49/mo, Genesis R&D ~$295–595 one-time, Nutritics, Food Label Maker, MenuCalc). You type a recipe, they spit out a compliant ingredient list, QUID percentages, nutrition panel, and allergen emphasis. Brilliant for making a label once. They have no concept of “the label that’s actually on the boxes in your warehouse right now.”

Camp B — enterprise artwork QA (GlobalVision, Twona AMS). Pixel-and-text diffing of artwork files, built for big-CPG print departments. Priced and scoped for enterprises with a dedicated packaging team. They diff artwork file vs artwork file — they don’t know your recipe changed, because the recipe lives in a different system (or a spreadsheet) entirely.

Nobody connects the two. The small producer’s reality: the recipe lives in a spreadsheet or a recipe-costing tool, the approved label artwork lives in a PDF from a designer 8 months ago, and the only link between them is a human remembering to re-check. That human forgets. Half the recall market is that forgetting.

LabelLag owns the seam: it holds the live recipe/spec and the current approved label content for each SKU, re-derives what the label should say on every recipe or supplier change, and screams when the printed label no longer matches — before you order the next print run or ship the next batch. It is not a better label generator and not a cheaper GlobalVision. It is the watchdog that should sit between them and never existed for businesses under £10M turnover.

4. Target market

• Primary customer: Owner / quality-manager / production-lead at small packaged-food producers in the UK and EU — bakeries, delis, sauce/condiment/snack/ready-meal makers, specialty drinks — roughly 2–50 staff, holding or pursuing SALSA or BRCGS certification and selling into retail or food-service (where a mislabel becomes a recall, not just a refund).
• Why they buy (their words): “Get the math wrong and it’s a recall.” “It depends on someone remembering to check.” They’ve already been burned or watched a peer get burned; their auditor flags labelling control as a corrective action; their retail buyer demands traceable spec-to-label control. The pain is acute, recurring (every reformulation, every supplier swap), and carries existential downside (a recall can end a small brand).
• Rough TAM reasoning: The UK alone has tens of thousands of small food manufacturers; SALSA has certified thousands of small producers and BRCGS covers many thousands more across the EU. Conservatively, 20,000–40,000 UK+EU small producers sit in the sweet spot (cert-holding, retail-facing, too small for enterprise artwork QA, too exposed to run on spreadsheets). Capture a low-single-digit % at £100–300/mo and you clear the $1M–$5M ARR band comfortably.
• Why now for them: Recalls hit a multi-year high in 2025; retailer and auditor scrutiny of labelling control has tightened post-Natasha’s Law (UK PPDS rules, live since Oct 2021, pulled thousands of bakeries/delis into mandatory ingredient+allergen labelling for the first time); and the EU FIC regime keeps layering QUID/allergen-emphasis obligations. The downside got bigger and the regulators got louder in the same 18 months.

5. Product sketch (MVP)

• Single source of truth per SKU: import the recipe/spec (CSV, recipe-costing export, or manual) and attach the current approved label — as typed content or by extracting the ingredient list, QUID, and allergen statement from the label PDF.
• Auto-derive the “correct” label content from the live recipe: ingredient list in descending weight order, QUID percentages, and the 14 EU/UK allergens emphasised per FIC / Natasha’s Law.
• Drift alarm: the instant a recipe or a linked supplier ingredient changes, LabelLag re-derives and shows a red diff — “label on file says X, recipe now requires Y” — with the specific allergen / QUID / order discrepancy called out.
• Supplier-change capture: log supplier allergen statements and recipe versions; when a supplier swap introduces or removes an allergen, the affected SKUs light up automatically.
• Pre-print gate: a one-click “is this label safe to print?” check that blocks/flags any SKU whose printed artwork no longer matches the live spec.
• Audit trail: every change, who made it, and the matching label version — the exact evidence a SALSA/BRCGS auditor asks for on labelling control.
• Export the corrected label content ready to hand to the designer/printer (LabelLag tells you what must change; it need not be the artwork tool itself in v1).

6. AI angle — what’s load-bearing

Two jobs only AI does cheaply enough to make this a <£300/mo product:

1. Reading the messy label back into structured data. The “approved label” exists as a designer’s PDF or a photo of a printed pouch — fonts, columns, bold allergens, marketing copy. Vision-LLM extraction parses that into a structured ingredient list + allergen set + QUID values so it can be diffed against the recipe. Without this, onboarding means hand-keying every existing label and nobody finishes setup.
2. Semantic equivalence, not string matching. “Whole milk powder” in the recipe vs “milk” emphasised on the label vs a supplier spec listing “lactose” — these must be reconciled to the same allergen. A regex can’t; an LLM with a food-allergen ontology in the prompt can decide whether two phrasings are compliant-equivalent or a genuine drift. This judgment call is the entire product.

Remove the AI and you’re back to a spreadsheet that “depends on someone remembering to check” — which is exactly the failure we’re selling against. The AI is the thing that remembers, and the thing that understands.

7. Localization angle

EU/UK-first is the wedge, not a nice-to-have. FIC Regulation (EU) 1169/2011 — allergen emphasis, QUID, descending-order ingredient lists — plus UK Natasha’s Law / PPDS rules define exactly what “correct” means; the ontology of the 14 EU/UK allergens and the QUID maths are region-specific and harder to get right than the US Big-9. A global “label maker” treats these as one of many regimes; LabelLag treats the EU/UK rulebook as the core engine and ships pre-loaded with the SALSA/BRCGS evidence formats auditors expect. Local pricing in £/€ at SMB tiers (£100–300/mo) reaches producers for whom an enterprise artwork-QA seat was never an option. Later expansion to US FDA / Canada / ANZ is additive, but the EU/UK depth is the moat on day one.

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

• Pricing: tiered by SKU count. Starter £99/mo (up to ~25 SKUs), Growth £199/mo (up to ~100 SKUs), Pro £349/mo (unlimited + multi-site + auditor export). Annual prepay discount.
• ACV: blended ~£2,400/yr (≈£200/mo) assuming most paying customers land on Growth.
• Math to $1M ARR: ~420 customers × £200/mo × 12 ≈ £1.0M (~$1.3M) ARR. Roughly 1–2% of the conservative 20–40k SMB pool.
• Math to $5M ARR: ~2,000 customers, OR ~1,000 customers at a higher blended ACV once US/Canada modules and a per-SKU artwork-rendering upsell push ACV toward £350–400/mo. Reaching it means adding co-packers and multi-site groups (10–50 SKUs each), where the audit-trail value justifies Pro.
• Expansion path: more SKUs → higher tier; multi-site / co-packer seats; add-on modules for additional regimes (US, CA, ANZ); a “designer hand-off” / artwork-render upsell; and a verification-stamp/certificate output retailers can demand from suppliers (turns the buyer into a distribution lever — see §9).

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

• SALSA / BRCGS consultant channel (primary). A few hundred independent UK food-safety consultants implement these certs for small producers and personally write the “labelling control” procedures auditors check. Recruit 15–20 as referral partners (rev-share or flat bounty): LabelLag is their labelling-control answer. Each consultant carries 20–80 small clients. Five active consultants → first 100 customers.
• Auditor-corrective-action ambush. Labelling control is a recurring SALSA/BRCGS non-conformity. Target producers who just took a labelling corrective action (visible in consultant networks, LinkedIn posts, IFSQN forum threads) with a “close this finding in a week” pitch. Warm, time-boxed, high intent.
• IFSQN + trade-association direct. IFSQN (large active food-safety forum) and bodies like the Food & Drink Federation / Guild of Fine Food run newsletters and member directories of exactly this buyer. Sponsored how-to content (“the recipe-changed-label-didn’t recall, and how to never have it”) + a directory scrape of small certified producers → personalised Loom showing their own label drift caught in 60 seconds. Expect a 5–8% reply rate on a demo that shows their actual risk.
• Recall-news retargeting. Every published allergen recall is a same-week ad to that producer’s peers (“this was a label that didn’t match the recipe — here’s the watchdog that catches it”). Cheap, evergreen, perfectly timed to the fear.

10. Build complexity — justification

Medium. Off-the-shelf: web stack, vision-LLM for label extraction, LLM-with-ontology for equivalence reasoning, standard auth/billing. Custom work is the EU/UK rules engine (QUID maths, descending-order logic, 14-allergen emphasis) and a reliable diff/alerting layer — non-trivial but bounded, no novel ML, no dataset that doesn’t exist. The honest risk is extraction accuracy on ugly label PDFs; mitigated by a human-confirm step at onboarding. A two-person team ships a credible v1 in ~3–4 months; a single SKU’s “import recipe → extract label → show the drift” demo is buildable in weeks.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Compliance-assist tool; producer retains legal responsibility. Position as decision-support, not a legal guarantee.
Ethical — no harm / dark patterns	✅	Net safety-positive; reduces undeclared-allergen risk to consumers.
Market exists (evidence above)	✅	Recall stats, regulator alerts, paid incumbents in adjacent camps, forum demand.
1–5 person team can build this	✅	2 people, ~3–4 months to v1.
Launchable with <$50K / ₹40L	✅	Off-the-shelf APIs; main cost is the founders’ time.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	16/20	Recall = existential for a small brand; ~half of recalls are this exact failure. Hair-on-fire for the burned; “we have a process” complacency for the not-yet-burned.
Demand evidence	15	12/15	Hard, independent signals: 45.8% recall stat, 85 UK alerts/yr, paid incumbents either side of the gap, explicit forum complaints. Gap = nobody articulating demand for this specific drift-watch product yet.
Build feasibility	15	11/15	Bounded custom rules engine + AI extraction. Real risk is PDF-extraction accuracy; solvable with human-confirm onboarding.
Distribution clarity	15	11/15	Strong named channel (cert consultants) with realistic referral math; conversion through them still to be proven.
Revenue mechanics	15	11/15	Pricing benchmarked vs ReciPal/Genesis/Nutritics; ~420 customers to $1M is achievable. Churn risk if treated as “set up once.”
Time to first revenue	10	8/10	Acute pain + consultant warm intros + a demo that shows real risk in 60s → paid pilots within weeks of a working v1.
Defensibility	10	6/10	Moat = accumulated EU/UK rules accuracy + supplier-spec/label data lock-in + consultant relationships + audit-trail switching cost. Copyable in ~12 months by a label-generator that wakes up; head start + channel is the defence.
Total	100	75/100

13. Qualitative modifiers

Founder-fit tags

domain-expertise-required · technical-heavy

A founder (or close advisor) who has lived FIC/Natasha’s Law/SALSA labelling control is worth more than any feature. The rules nuance is the product; getting QUID or allergen equivalence subtly wrong destroys trust instantly.

Key assumptions to validate (3–5)

1. Assumption: Small producers will pay ~£200/mo for drift-watching they believe their existing process already covers. How to test: 25 discovery calls with SALSA/BRCGS holders; show their own label-vs-recipe drift live; measure how many ask “how do I buy this” unprompted.
2. Assumption: AI extraction is accurate enough on real designer PDFs / pouch photos that onboarding isn’t a slog. How to test: run extraction on 50 real labels collected from pilot prospects; measure field-level accuracy and human-correction time per label.
3. Assumption: Cert consultants will refer/resell rather than see it as competing with their billable hours. How to test: pitch 10 consultants; secure ≥3 signed referral agreements before building beyond MVP.
4. Assumption: This is a recurring need (every reformulation), not a one-time setup that then churns. How to test: instrument pilot accounts for change-events/month over 60 days; <1 meaningful drift-check per month per account is a churn red flag.

Risk flags

1. Liability framing: if positioned as guaranteeing compliance, one miss = lawsuit. Must be decision-support with the producer retaining responsibility; legal review of ToS before launch.
2. Incumbent wake-up: ReciPal/Nutritics/Food Label Maker already hold the recipe; any could bolt on “watch for drift.” Mitigant: ship the EU/UK depth + consultant channel + audit trail faster than a US-centric generator can localise.
3. “We already have a process” complacency: the non-burned buyer underrates the risk. GTM must lead with their own live drift (fear made concrete), not a generic feature list — and lean on the recently-audited / recently-recalled segments first.

14. Structured verdict

Score:                  75/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder + food-safety/labelling domain advisor (ex-SALSA/BRCGS consultant ideal)
Time to revenue:        8–12 weeks from a working v1 (paid pilots via consultant intros)
Capital to launch:      £8–15k / $10–20k (mostly time + AI API spend)
Top 3 assumptions to validate first:
  1. WTP ~£200/mo despite "we already have a process" — 25 discovery calls showing their own drift
  2. AI extraction accuracy on real label PDFs — 50-label accuracy + correction-time test
  3. Consultant channel will refer/resell — 3 signed referral agreements pre-build
Kill criteria:
  - Abandon if <5 of 25 cert-holders, shown their own live drift, ask to buy unprompted
  - Abandon if label-extraction needs >10 min of human correction per label at onboarding (kills self-serve economics)
  - Abandon if a recipe-holding incumbent (ReciPal/Nutritics) ships EU/UK drift-watching before your v1 with the consultant channel locked

15. Next step — 1-week validation sprint

• Day 1–2: Pull 50 real small-producer labels (PDFs/pouch photos) from public listings + 5 friendly producers. Run vision-LLM extraction into structured ingredient/allergen/QUID data. Measure field-level accuracy and minutes-of-correction per label. (Tests assumption 2 — the technical fault line.)
• Day 3–4: Book and run 10–15 calls with SALSA/BRCGS-holding producers and 5 cert consultants. In each, take the producer’s actual recipe + actual label and show the drift diff live. Watch the reaction; ask directly “would you pay £200/mo for this, and would you refer it?”
• Day 5: Decide go / no-go on a falsifiable bar: go only if (a) extraction averaged ≤5 min correction/label AND (b) ≥6 of ~15 producers said yes to £200/mo unprompted AND (c) ≥2 consultants verbally committed to refer. Anything less → iterate the wedge or shelve.