Harvey Signs at Raffles Place. The Bill Starts Now.
When WongPartnership signed with Harvey, it acquired not just a seat licence but a compounding stack — wrapper-maintenance costs running 30–60% above the sticker, prompt libraries with no clear owner, and a hallucination-liability perimeter across three jurisdictions that Vietnam's new AI statute just made more expensive. The Forrester 400% ROI ceiling belongs to firms with clean, documented processes. Most Southeast Asian mid-market partnerships are not those firms.
What does a Singapore partnership actually own after twelve months on a Harvey contract? Look past the brochure and the pilot deck. What sits on the books, in the workflow, in the associate's browser history at 11pm — that is the honest question, and if you are the finance partner staring at renewal, you have already noticed it does not get a straight answer.
The question comes up quietly at renewal time and gets louder as each invoice lands. In Southeast Asia it has arrived faster than most partnerships were ready for: the region's marquee firms have moved on legal AI in the last four months, and Singapore has published the framework everyone will be measured against. Someone in the exco will have to price what the Guide, the vendor and the deployment together add up to. That is a debt problem before it is a compliance problem.
the guide, the office, the partnership
Three things landed within a hundred days of each other, all inside a two-kilometre radius. On 6 March 2026 the Ministry of Law published its Guide for Using Generative AI in the Legal Sector, a non-binding but very carefully phrased document that names professional responsibility, confidentiality, and transparency as the three tests any deployment must survive. Read the actual PDF, not the write-ups; the annexes are where the sharp edges are. In June, Harvey opened a Singapore office to serve the region from a local base. And WongPartnership signed with Harvey as the first ASEAN firm on the platform, framed as a practice-wide deployment.
Set the narrative aside. What each of these actually created is an obligation. The Guide created an obligation to document your controls. The vendor created a subscription with escalators. The partnership created a client-communication and confidentiality perimeter around a system trained on somebody else's corpus. None of that is Sanjay in accounts' problem. Sanjay in accounts still has to book the invoice.
the shape of the recurring cost
Harvey is not sold like Westlaw. The company is running at roughly $190 million ARR at an $11 billion valuation as of the March 2026 raise, with a client base disclosed at over 100,000 lawyers across 1,300 organisations. Divide those numbers and you get the going rate. It is enterprise SaaS pricing on a 25–50 seat minimum with 12-month terms, and Southeast Asian mid-market partnerships — the fifty-to-two-hundred-lawyer band that most of the region sits in — sit at the least-favoured end of the curve, because volume discounts do not kick in until Am Law scale. A fifty-seat deployment lands in mid-six figures a year before you have added a single integration hour or paid for the change-management consultant your partners will demand three weeks in.
Now overlay depreciation logic. A licensed research tool has a defensible useful life because the product surface has been stable for twenty years. A generative product built on frontier models does not have a stable product surface. The model changes; the safety layer changes; your prompts stop working the way they did in July; your evaluation suite, the one your innovation partner sold you on last year, needs a full re-run because you are re-baselining against a different underlying model. Nothing about that is unique to Harvey. It is what buying a wrapper around somebody else's foundation model looks like, and the maintenance line does not depreciate; it renews.
The recurring cost, in other words, is not the sticker. It is the sticker plus the compulsory maintenance of the wrapper you built around the wrapper. Legal-ops teams running that hidden line honestly are booking it at 30–60% of the sticker on top.
the debt that comes with it
Here is the part your innovation committee did not model. Call it AI slop debt: the accumulating liability of half-finished proof-of-concepts, unevaluated agent flows, retrieval indexes with no clear owner, prompt libraries scattered across four Confluence pages, and orphaned fine-tunes nobody can reproduce. On the day you sign the Harvey contract, the debt starts compounding. You inherit a set of workflows the vendor has optimised for a general Am Law audience, and you keep building around them for Bahasa Malay drafting, Vietnamese counterparty due diligence, Bahasa Indonesia bond prospectus review. Every one of those bespoke pieces sits above the vendor's abstraction layer. Every one has to be re-tested every time the vendor pushes an upgrade.
| What the pitch says | What sits on the books after 12 months |
|---|---|
| A per-seat SaaS license | Sticker + 30–60% wrapper and eval maintenance line |
| Improved associate productivity | Prompt libraries and RAG indexes with no clear owner |
| Fewer research hours | New PI-exposure vector via hallucinated citations |
| A market-signalling partnership | A confidentiality perimeter across three jurisdictions |
The risk on the other side of the ledger has become measurable in a way it wasn't a year ago. Damien Charlotin's independent database of court decisions where a party demonstrably relied on hallucinated citations now counts 1,598 documented cases worldwide as of 9 June 2026, up from 719 in January and 1,227 in early April. Roughly eight new cases a day, most of them in the US but with the frontier moving into Commonwealth jurisdictions your partners practise in. On 31 March alone, seventeen US court decisions in a single day flagged suspected AI hallucinations in filings. That is the frequency now.
Underlying rates are worse than the vendor pitch suggests. The Stanford HAI study on leading legal AI research tools — still the closest thing the field has to an independent audit — reported hallucination rates of 17–33% on real-world legal queries, depending on the tool. Harvey's own internal benchmark claims a materially lower rate. Harvey does not publish an independent audit. A partner signing off on filings does not get to average those two numbers.
the counter-argument, on its merits
The strongest defence of the spend is Thomson Reuters' Forrester Total Economic Impact study on CoCounsel Legal, which put ROI at 400% over three years for a modelled 500-attorney composite firm. That is a serious number and it deserves engagement rather than dismissal.
Two problems, on the merits. First, the composite was CoCounsel, not Harvey, and the two products have different pricing curves and different data-security postures relevant to a Southeast Asian buyer who has just read the MinLaw Guide's confidentiality section. Second, the same publisher's technology desk has been warning that agentic AI pilots will likely be quietly shut down by 2028 unless firms fundamentally change how they prepare for them, with the failure mode being clean-data assumptions colliding with messy real practice. Both statements can hold at once. The 400% is the ceiling for firms that already had documented processes, clean matter management, and disciplined conflict checks. The pilot-failure warning describes the modal outcome for the median firm.
Most Southeast Asian mid-market partnerships are the median firm on those readiness dimensions. The reason is not that they are behind. The region runs leaner, more relationship-driven practices where the tacit knowledge lives in senior associates' heads and never made it into the DMS. A strength when the deal is bespoke; the wrong precondition for a well-behaved agentic deployment.
hanoi's tier system
While Singapore was launching a guide, Hanoi was passing a statute. Vietnam's standalone Law on Artificial Intelligence came into force on 1 March 2026 with a three-tier, risk-based classification: high-, medium-, and low-risk systems, with pre-deployment classification duties on the deployer. If your Singapore partnership has a Ho Chi Minh City office serving cross-border matters, and your Harvey deployment touches Vietnamese counterparty data as part of that work, you now have a compliance perimeter your MSA did not price. Nobody at signing thought about it. Nobody at renewal will either, unless someone puts a number on it.
what I would do this month, if you asked me
If a managing partner in Singapore, Jakarta or Kuala Lumpur put a Harvey renewal in front of me tomorrow, this is the sequence I would insist on before signing.
Cut the seat count by a third for year two, and put the savings into a two-person legal-ops line — one former associate, one data engineer — reporting to the innovation partner, not to IT. Their job is to own the wrapper, the prompt library, and the eval suite against the MinLaw Guide's three principles. Book an eval budget explicitly on the P&L; I use 10% of the vendor spend as a minimum, and treat every model upgrade as a re-baselining event with a written go/no-go. Insist on a written data-residency schedule from Harvey that names Vietnam specifically, and refuse to sign without it. Put a hallucination-sanctions clause into your professional-indemnity conversation now, before the Charlotin database's next hundred cases arrive. And, the one that will annoy your marketing partner, stop putting the Harvey partnership in the client pitch. Start putting your controls in the pitch. That is what the general counsel across the table is buying.
Do those five things and the Harvey line item still hurts, but you own what you paid for. Skip them and, twelve months from now, you will be renewing on a wrapper you don't understand around a model you can't audit, with a slop-debt ledger that grew faster than your revenue. That is a bill you don't want to redline in front of the exco.
Tarry Singh is the founder and CEO of Real AI (realai.eu), an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan (earthscan.io) for Energy AI, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.