ThauRoute

1. One-liner

Tells a Vietnamese company which foreign vendor invoices it must withhold tax on, before the 10-day clock expires.

2. Trend signal — why now?

Three things moved in the last fifteen months, and together they turned a niche accounting chore into a recurring monthly liability.

One — the rule changed underneath everyone. Decree 252/2026/ND-CP was signed 30 June 2026 and took effect 1 July 2026, replacing Decree 117/2025/ND-CP. It rewrote how foreign-supplier transactions are taxed and added a brand-new axis to the decision: was this transaction conducted on an e-commerce platform or not? Baker McKenzie’s alert on the change is blunt about the consequence — applying the old Circular 103 three-condition test “is no longer appropriate and carries the risk of incorrectly determining the withholding obligation.” A test that Vietnamese accountants have used for a decade stopped being the right test, in the middle of a fiscal year.

Two — the price of getting it wrong doubled. Circular 69/2025/TT-BTC raised the deemed VAT rate on digital services from 5% to 10%, effective 1 July 2025. On a general service the combined FCT burden is 5% VAT + 5% CIT. The exposure per invoice is no longer rounding error, and Vietnam’s statute of limitations for imposing tax is ten years.

Three — the definitive answer lives in a blog post. Whether you must withhold turns on one binary fact: is the foreign supplier registered on the General Department of Taxation’s portal? If yes, the Vietnamese buyer does not declare or withhold. If no, the buyer must. GDT publishes the list — and the practical reference Vietnamese accountants actually use is a page maintained by the accounting firm Gonnapass, which as of 16 May 2026 carried 247 registered foreign providers, with entries added as recently as 29 April 2026. A list that changes monthly, that determines a ten-year liability, is being hand-curated on a consultancy’s website.

The category is unbuilt in the specific place that matters. MISA AMIS — the dominant Vietnamese accounting suite — automates generating the 01/NTNN declaration and now advertises supplier status checks. But generating the form presumes you already decided the form is required. Guidance for companies without in-house expertise is still, verbatim, to hire a tax consulting firm. The filing is productized; the decision is not.

Provenance:
  - Signal 1 (demand): Vietnamese accountants rely on a hand-maintained list of 247 GDT-registered foreign providers to decide withholding; last updated 29 Apr 2026 — https://gonnapass.com/danh-sach-nha-thau-nuoc-ngoai-dang-ky-thue/ — observed 2026-09-02
  - Signal 2 (feasibility): Decree 252/2026/ND-CP effective 1 Jul 2026 adds the platform-vs-non-platform axis and an electronic notification duty; old Circular 103 test explicitly no longer appropriate — https://www.bakermckenzie.com/en/insight/publications/2026/07/vietnam-tax-administration-rules-effective-july-2026 — observed 2026-09-02
  - Signal 3 (economic): Digital-services deemed VAT doubled 5%→10% (Circular 69/2025, eff. 1 Jul 2025) across 859,048 operating Vietnamese enterprises; IT services market USD 2.63B in 2026 — https://indochinalink.com/foreign-contractor-tax-vietnam-withholding/ , https://en.vietnamplus.vn/businesses-continue-to-drive-vietnam-s-economic-growth-census-post348263.vnp — observed 2026-09-02
  Category: Regulatory arbitrage (a decision duty was rewritten mid-year) + Underserved niche (vendors sell the declaration; nobody sells the verdict)

3. The opportunity

Every incumbent in this market sells the wrong half of the workflow.

MISA, Fast, Bravo and the rest sell declaration generation. You tell the software “this invoice needs FCT at 5/5,” and it produces a correct 01/NTNN and files it through the mTax portal. That is genuinely useful and genuinely solved. It is also the second step.

The first step — the one that actually carries the ten-year liability — is a chain of judgements a human makes per invoice:

1. Is this vendor foreign with no permanent establishment in Vietnam?
2. Is the vendor on the GDT registered-supplier list as of this payment date?
3. Was this transaction conducted on an in-scope e-commerce platform under Decree 252, or outside one?
4. Is this a service (5% VAT + 5% CIT), a software royalty (0% VAT + 10% CIT), IaaS rather than SaaS (5% + 5%), or a hybrid the tax authority will reallocate?
5. If we withheld, have we electronically notified the platform operator so it doesn’t withhold again on the same transaction?

Nobody sells step 1 through 5. The classification questions are genuinely hard — Official Dispatch 3518/TCT-CS holds that “system server and data storage services are NOT software services,” which is precisely why treating AWS like Salesforce is a documented, common error. And step 5 is brand new, has no published deadline, and its failure mode is that your company pays the tax twice on the same invoice.

The disruption is not “AI does accounting.” It’s that a classification problem which used to require a Big-4 memo per vendor category can now be answered per-invoice, in seconds, against a live copy of a list that currently lives in a blog post. The incumbent’s product structurally starts one step too late, and the incumbent has no incentive to move upstream — being wrong about the verdict is a liability they’d rather leave with the customer.

4. Target market

Primary customer: The chief accountant (kế toán trưởng) or finance lead at a Vietnamese company with 20–300 employees that pays foreign vendors monthly — software/IT services firms, digital agencies, e-commerce operators, export manufacturers, and the local subsidiaries of foreign groups. These are companies large enough to buy AWS, Meta Ads, Figma, GitHub, Slack and a dozen SaaS subscriptions, but too small to keep a tax manager on staff or a Big-4 firm on retainer.

Why they buy: They are making a legally consequential decision on incomplete information, repeatedly, on a 10-day clock. The felt pain is not the tax — it’s the ambiguity. The chief accountant knows the Circular 103 test stopped being correct in July, knows the platform axis exists, and does not know which of their 40 recurring foreign invoices changed category. The current workaround is to copy a consultancy’s blog list into a spreadsheet, apply last year’s logic, and hope. Grant Thornton’s Vietnam guidance notes that insufficient documentation is already “resulting in disallowance of input VAT credit/refund and CIT deductibility” — so the downside isn’t only the unpaid tax, it’s the expense being thrown out of the CIT computation years later.

Rough TAM reasoning: 859,048 operating enterprises in Vietnam as of end-2025. The relevant slice is companies with recurring foreign-vendor spend. Even a deliberately conservative 3% — 25,000 companies — at ₫2.5M/month is a market well past the $5M ARR ceiling I care about. I don’t need to be right about the 3%; I need 400 customers.

Why now for them: Their existing answer expired on 1 July 2026. Decree 252 didn’t add a form to file, it invalidated the reasoning they were using. Every foreign invoice paid since then was classified with a test their own advisors now say is inappropriate — and the assessment window on that mistake runs ten years.

5. Product sketch (MVP)

• Vendor verdict — point it at a foreign vendor (or upload the invoice) and get a dated, plain-language ruling: withhold or don’t, at what rates, under which decree, with the reasoning shown.
• Live GDT registry mirror — a continuously refreshed copy of the registered-foreign-supplier list, with change alerts. When a vendor registers or drops off, your obligation flips, and you find out that week rather than at audit.
• Point-in-time snapshots — every verdict stores what the registry said on the payment date, because that’s the fact an inspector will test three years from now.
• Classification assistant for the hard cases — SaaS vs IaaS vs royalty vs hybrid, citing the specific dispatch or circular behind the call, flagging bundles the authority is likely to reallocate.
• 10-day clock — a running list of payments made, days remaining to file, and which ones still have no verdict.
• Article 43 notification tracker — when you withhold on a platform transaction, it records the electronic notice to the platform operator and keeps the receipt, so you can prove you weren’t the cause of a double withholding.
• Audit binder export — per-vendor, per-period evidence file: verdict, registry state, rate basis, filing reference. The artefact you hand an inspector.
• Bilingual throughout — Vietnamese for the accountant, English for the foreign parent company that asks why the invoice was short-paid.

6. AI angle — what’s load-bearing

Remove the AI and the product collapses into the spreadsheet customers already have.

The load-bearing work is classification under ambiguity. A vendor invoice says “Professional Services — Platform Subscription.” A human has to decide whether that’s a software service (0% VAT, 5% CIT), infrastructure (5% + 5%), or carries a royalty component (10% CIT) — and Vietnamese authorities assess substance over form, reallocating income in bundled contracts. That’s reading an invoice, a contract, and a body of dispatches and circulars, then producing a defensible position with citations. It is exactly the shape of work that used to cost a consultancy memo and now costs an inference call.

The second AI-dependent piece is normalising entity identity. The GDT list carries legal entity names; your invoice says “Meta Platforms Ireland” or “AWS EMEA SARL” or a reseller’s name. Matching the payee on an invoice to the right row in a 247-entry registry — across transliteration, subsidiary structures, and payment intermediaries — is fuzzy matching against high-stakes consequences, and it is the step that makes the registry mirror actually usable rather than just accurate.

What is not AI: the filing, the arithmetic, the registry scrape. Those are plumbing. I want the model deciding, and I want every decision it makes carrying a citation a human can check — because a confident wrong answer here is worse than no product.

7. Localization angle

This is a Vietnam-only product and that’s the point, not a limitation.

The moat is entirely local: the GDT portal, the registered-supplier list, Circular 80/2021, Circular 103/2014, Circular 69/2025, Decree 252/2026, and a body of official dispatches that exist only in Vietnamese. A global tax-engine vendor cannot serve this without doing the same country-specific work, and the market is too small to attract them — Vietnam is a rounding error to Avalara or Vertex, and a very good living for a two-person team.

Pricing has to be local. ₫2,000,000–5,000,000/month ($75–190) is a normal software line item for a 50-person Vietnamese company; $500/month is not. That price band is precisely what makes the segment unattractive to the consultancies currently absorbing this work at hourly rates — and per the priced-out-not-unserved pattern, an incumbent publicly declining a segment on unit economics is the strongest form of permission.

Language is a genuine wedge, not decoration. The product has to reason over Vietnamese-language circulars and produce output an inspector reads in Vietnamese, while explaining short-paid invoices to foreign vendors in English. Sitting on both sides of that line is most of the perceived value.

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

• Pricing: ₫2,000,000/month ($75) for up to 25 foreign vendors; ₫5,000,000/month ($190) for up to 100 vendors plus the audit binder and notification tracker; ₫12,000,000/month (~$455) for accounting firms managing multiple client books.
• ACV: ~$1,600 blended, assuming most customers land on the mid tier.
• Rough math to $1M ARR: 625 customers × ~$133/mo × 12 ≈ $1.0M. In a base of 859k enterprises, that’s under 0.1% of the market, and well under 3% of the realistic foreign-vendor-paying slice.
• Rough math to $5M ARR: ~3,000 customers, which almost certainly means the accounting-firm channel is carrying it — one firm seat covering 30 client books is a different and better business than 3,000 direct SMB sales. It probably also means a second country with the same shape (Indonesia’s platform withholding lands 1 Nov 2026 after its August suspension; the Philippines is moving the same direction).
• Expansion path: vendor-count tiers grow naturally as companies add SaaS. Then the audit binder as a paid add-on when the first inspection lands. Then the firm tier, which is where the durable revenue is — firms don’t churn the way a 40-person company churns.

Gross margin is standard SaaS minus inference, and inference here is small: a few hundred classification calls a month per customer, cached hard because the same vendors recur. The registry mirror is one scrape amortised across every customer — the cost of being right does not scale with the number of people you’re right for.

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

1. Take the blog list and make it a product. Gonnapass’s registered-provider page is a load-bearing reference for Vietnamese accountants precisely because nothing better exists. Ship a free, always-current, searchable version with email alerts when a vendor’s status flips — “is AWS registered this month?” is a real query with a real audience. That free tool is the top of the funnel and it costs one scraper to run. Convert on the second question every visitor has: fine, so what do I do about this invoice?

2. Ambush the confusion window directly. The specific searches are known and narrow: thuế nhà thầu, mua phần mềm nước ngoài có phải nộp thuế nhà thầu không, thuế nhà thầu Facebook Google, Nghị định 252. These are transactional queries by people holding an invoice, not researchers. A Vietnamese-language answer page per vendor category — AWS, Meta Ads, Google Ads, Figma, GitHub, Slack, Zoom — each ending in “check your own vendors” is a fast, cheap, well-targeted acquisition surface in a language global competitors won’t write in.

3. Sell through accounting firms, not around them. Small and mid-sized Vietnamese accounting practices are currently answering this by hand for every client and eating the risk. Approach 100 of them with the firm tier: they keep the advisory fee, we take the mechanical work and the liability trail. Each firm that signs brings 10–40 client books. This is the channel that gets to 3,000 customers; direct sales is the channel that proves it’s worth doing.

4. Go where the foreign-parent subsidiaries are. EuroCham, AmCham, KoCham, JCCI and the Korean and Japanese business associations in Ho Chi Minh City and Hanoi are full of exactly the customer — a Vietnamese subsidiary paying group-level foreign invoices, with a chief accountant who must explain FCT to a parent company that has never heard of it. Association newsletters and a 30-minute “what changed on 1 July” briefing convert well because the audience already knows it has a problem.

5. Time the outreach to the calendar. The 20th of each month is the monthly FCT filing deadline. Outreach lands hardest between the 10th and the 18th, when the pain is present rather than remembered.

10. Build complexity — justification

Medium. Roughly 10–14 weeks for a strong pair, or a solo builder with a Vietnamese tax advisor on retainer.

Off the shelf: the web app, auth, billing, invoice OCR, and the classification reasoning itself. None of that is novel.

The custom work is threefold, and it’s where the calendar goes. First, the registry mirror — scraping and diffing the GDT list reliably, with entity normalisation good enough that “AWS EMEA SARL” on an invoice resolves to the right registry row or confidently to none. Second, encoding the decision tree across Circular 103, Circular 80/2021, Circular 69/2025 and Decree 252 in a way that survives the next circular, because there will be one; this is configuration, not hardcoded logic, or the product dies at the next amendment. Third, point-in-time storage, which sounds trivial and isn’t — every verdict must reconstruct what was known on the payment date, years later.

The genuine constraint is not engineering, it’s domain access. This does not get built correctly without a Vietnamese tax professional validating the decision tree, and that relationship has to exist before week one. That’s a real dependency and I’d treat it as the first thing to secure.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Decision-support and record-keeping software. It doesn’t file on the taxpayer’s behalf or give licensed tax advice; the customer or their firm signs. No licence required to ship.
Ethical — no harm / dark patterns	✅	Helps companies pay the tax they actually owe, on time, with evidence. The failure mode we’re removing is both under- and over-withholding.
Market exists (evidence above)	✅	247-entry registry hand-maintained by a consultancy because demand exists; MISA has moved adjacent; guidance still says “hire a firm.”
1–5 person team can build this	✅	Two builders plus a tax advisor. 10–14 weeks.
Launchable with <$50K / ₹40L	✅	Well under. Main line item is the advisor retainer, not infrastructure.

All five pass.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	16/20	Monthly, on a 10-day clock, with a ten-year assessment window and CIT-deductibility loss as the downside. Not quite hair-on-fire — most months the accountant guesses and nothing visibly happens. The pain is deferred, which caps this below 17.
Demand evidence	15	11/15	Strong indirect evidence: a hand-curated 247-entry list, active search demand in Vietnamese, MISA building adjacent, consultancies charging for the answer. What I do not have is a single verbatim Vietnamese accountant complaining in public — the forums where this is discussed are largely closed or Zalo groups. That absence is why this isn’t 13+.
Build feasibility	15	11/15	Standard stack, but registry scraping, entity matching, point-in-time storage and an amendable rules layer are real work. 10–14 weeks, and it doesn’t start without domain access.
Distribution clarity	15	11/15	The free-registry wedge is concrete and cheap, the search terms are known and transactional, and the firm channel is nameable. Docked for the firm channel being unproven — firms may prefer to keep this billable rather than automate it.
Revenue mechanics	15	12/15	Local price band is defensible and sits below the consultancy alternative. 625 customers to $1M is credible. $5M leans on the firm tier or a second country, both of which are assumptions rather than facts.
Time to first revenue	10	8/10	The problem is present-tense and the deadline is monthly. Pre-selling off the free registry tool during the build is realistic; first payment inside 8 weeks of launch is a fair expectation.
Defensibility	10	7/10	Real but soft: accumulated point-in-time registry history nobody else has, an encoded rules layer that compounds with each amendment, and workflow lock-in via the audit binder. Copyable in month 3; harder in month 12 once you hold three years of dated verdicts. Not 9 — MISA could move upstream if it decided to.
Total	100	76/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · domain-expertise-required

This is the rare idea where the domain half is non-negotiable. A brilliant engineer without a Vietnamese tax advisor ships a confidently wrong product, and confidently wrong is worse than absent here.

Key assumptions to validate (3–5)

1. Assumption: Chief accountants at 20–300 person Vietnamese companies experience the withhold/don’t-withhold decision as a genuine recurring uncertainty, not a solved routine. How to test: 25 structured interviews via accounting-firm and chamber introductions. Ask them to classify five real invoices — AWS, Meta Ads, a SaaS subscription, a software licence, a bundled platform contract — and measure disagreement rate against an advisor’s answer. If they agree with the advisor 90%+ of the time, there is no product.
2. Assumption: They will pay ₫2–5M/month for a verdict rather than absorb it into existing advisor fees. How to test: Price the tiers on a real landing page behind the free registry tool and count checkout attempts, not survey responses.
3. Assumption: The GDT registry is reliably scrapeable and changes often enough that a live mirror has standing value. How to test: Scrape daily for four weeks, measure change frequency and structural stability. If it changes twice a year, the alerting wedge is much weaker than assumed.
4. Assumption: Accounting firms will adopt rather than defend the billable hours. How to test: Pitch the firm tier to 15 practices before building it. Two signed pilots is the bar.
5. Assumption: MISA does not ship the verdict layer within 12 months. How to test: Track their release notes and tax-module marketing monthly. This is a watch item, not a one-time check.

Risk flags

1. Incumbent encroachment: MISA already generates the declaration and has begun advertising supplier status checks. It has the customers, the ERP position, and the distribution. The bet is that it won’t take on liability for the verdict — but that bet is the whole business, and it’s the single most likely way this dies.
2. Regulatory churn: The rules changed materially in July 2025 and again in July 2026. That’s what creates the opening, and it’s also a permanent maintenance tax. Every amendment is a re-encoding sprint. Budget for it or the product silently goes wrong.
3. Liability exposure: Selling a verdict means being blamed for a wrong one. This needs careful positioning as decision support with cited reasoning that a human signs off, plus terms that reflect it. Get this wrong and one angry customer is an existential problem.
4. Platform dependency: The registry mirror depends on GDT’s portal remaining publicly scrapeable in its current form. If it moves behind authentication, the free wedge and part of the product need rebuilding.
5. Thin public customer voice: I could not source direct verbatim complaints from Vietnamese accountants. The demand case rests on artefacts and incumbent behaviour rather than quotes. That’s a real gap in the evidence and it’s reflected in both the demand score and the Medium confidence.

14. Structured verdict

Score:                  76/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical founder based in or connected to Vietnam, with a
                        Vietnamese tax practitioner as co-founder or retained advisor
                        from day one. Bilingual VN/EN is close to mandatory.
Time to revenue:        8–12 weeks post-launch; pre-sales realistic during build
Capital to launch:      $8–15K (₫200–380M) — mostly the advisor retainer
Top 3 assumptions to validate first:
  1. Chief accountants disagree with advisors on real invoice classification often
     enough to constitute a problem — 25 interviews, five-invoice classification test
  2. ₫2–5M/month clears against absorbing the work into existing advisor fees —
     live pricing page behind the free registry tool, measure checkout attempts
  3. The GDT registry changes often enough that a live mirror has standing value —
     daily scrape for four weeks, measure change frequency
Kill criteria:
  - Abandon if fewer than 8 of 25 interviewed accountants misclassify at least one
    of the five test invoices — the decision isn't hard enough to sell
  - Abandon if MISA ships a withhold/don't-withhold verdict feature before v1 launch
  - Abandon if the GDT registry changes fewer than 4 times in 8 weeks of monitoring
  - Abandon if zero accounting firms sign a pilot after 15 pitches — the $5M path
    depends on that channel and direct SMB sales alone won't carry it

15. Next step — 1-week validation sprint

• Day 1–2: Build and ship the free registry mirror — scrape the GDT registered-supplier list, make it searchable, add email alerts on status change. This is a day of work, it’s useful on its own, and it starts collecting the audience regardless of what the rest of the week shows. Simultaneously, begin the daily scrape log that feeds kill criterion three.
• Day 3–4: Assemble five real foreign invoices (AWS, Meta Ads, a SaaS subscription, a software licence, a bundled platform contract). Get an authoritative classification for each from a Vietnamese tax practitioner. Then run the same five past 25 chief accountants sourced through chamber and firm introductions, and record where they land.
• Day 5: Compare. Go if 8 or more of the 25 misclassify at least one invoice — that is the falsifiable measure that the verdict is hard enough to be worth paying for. Below that threshold, the decision is easier in practice than the regulations suggest, the product is a vitamin, and I walk away having spent one week and shipped a free tool that cost nothing to maintain.

The sprint is designed so the failure case still leaves something running. If the interviews kill it, the registry mirror stays up as a free utility and I’ve learned the segment’s actual competence level — which is worth knowing before the Indonesian version of this same question lands on 1 November 2026.