Your document-extraction vendor just joined a portfolio of more than 150 software companies that is not built to sell any of them. On September 8, 2026, Valsoft's TAG Software Group acquired Square 9 Softworks, the New Haven intelligent-document-processing vendor whose software reads invoices, purchase orders and contracts for more than 1,000 organizations in healthcare, government, manufacturing, agriculture and transportation. Terms were not disclosed. Stephen Young stays as CEO and the company will operate autonomously.
That is the reassuring version, and most of it is true. The part nobody put in the release is the one that decides your next three renewals: Square 9 sells generative-AI extraction, generative-AI extraction runs on somebody else's model, and models get retired on dates their owners choose. A holdco with no exit horizon is engineered to compound margin on an installed base. Whether it is also engineered to fund a model refresh every eighteen months is an open question, and nobody has answered it for this product. So the question at renewal is no longer who buys them next. It is who pays for the next model, and on what date.
What Valsoft Actually Bought
Valsoft bought a 21-year-old content-management business with a generative-AI product bolted to the front of it. Square 9 was founded in 2005 and sells enterprise content management, AI-powered data capture, digital forms and business process automation as one platform; Wiggin and Dana advised the seller and Kaizen Equity Partners ran the process.
The buyer is a Montreal roll-up. Valsoft says it owns more than 150 companies across 20+ industries and describes its approach as a "permanent capital mindset" — acquire and support for the long term. Software Equity Group counted 16 acquisitions in 2025 and ranked Valsoft the most active strategic buyer in SaaS for the second consecutive year. TAG Software Group is one of seven operating groups Valsoft created on October 6, 2025 to hold and support the portfolio.
An operating group is a management layer, not a product organisation. Its mandate — Valsoft's words — is enabling acquired companies to thrive while preserving their entrepreneurial culture. Nothing in that sentence is about shipping a new extraction engine.
A Perpetual Hold Is Not the Same as No Exit
Valsoft's own language is genuinely different from private equity's, and the difference is real. Its December 2024 announcement of a US$150 million growth equity round — led by Portage Capital Solutions with PROPELR Growth and existing investor Viking Global — says the firm prioritises "long-term value creation without the constraints of predefined investment horizons." If you have spent a decade watching sponsors flip your vendor every four years, that is a better sentence than the one you are used to.
Steel-man it properly: a buyer who never has to dress a company for sale also never has to strip it for one. No cost-out sprint before an exit, no roadmap frozen during diligence, no third owner in six years. That is a real benefit and Square 9's customers get it.
But the exit did not disappear. It moved up a level. BetaKit reported that the same round was an all-equity raise and that Valsoft plans a public markets debut in the future. Portage, PROPELR and Viking are institutional investors, and institutional investors get paid on an event. The holdco has a horizon; your vendor doesn't. What an IPO-bound compounder optimises for is durable margin across 150 businesses, which is precisely the discipline that makes it a good owner of stable software and an uncertain owner of software with a variable cost of goods.
You can see the price discipline in the one Valsoft deal with public numbers. When it took Quorum Information Technologies private, it paid C$0.80 per share, about C$60 million, a 14% premium to the prior close and 12% to the ten-day VWAP. A 14% premium is a cash-flow price. Nobody pays 14% because they intend to rebuild the product.
The strongest argument against all of this is Valsoft's own, and it deserves stating. The company runs AI Labs, a central function whose page says "our AI applications are built in-house and deployed across the Valsoft portfolio" and that "core AI systems are built internally and shared across the portfolio, removing reliance on third parties." Take that at face value and Square 9 gains an AI budget it did not have as an independent company. But read the second half of that sentence again. Removing reliance on third parties means moving extraction onto somebody else's stack — Valsoft's — and for the AP team running validation thresholds, an engine swap is an engine swap no matter which parent funds it. A centrally built replacement also arrives on a schedule you have less visibility into than a public deprecation table, not more. Either way, the checklist at the end of this piece is the same one.
Nobody Will Tell You Which Model Reads Your Invoices
Square 9 has never publicly named the model, provider or cloud service behind its generative-AI capture — and that gap is the single most important thing on your renewal checklist. The company announced TransformAI on January 20, 2023, pitching it as extraction that "completely forgoes the use of template-based data extraction." It followed with InquireAI on May 14, 2025, which CEO Steve Young described as letting customers pull data "simply by asking for it using easily written, natural language prompts." In November 2025, CTO Brian Banet called capture automation "the lynchpin connecting your data to the systems that leverage it".
The product page claims "near-100% accuracy" and the ability to lift headers, footers, line items and handwriting, plus a "SOC and HIPAA-compliant architecture." Those are vendor claims, published without a benchmark, a test set or a named model behind them.
The company's own Ethical AI Policy confirms the dependency while still declining to name it. Square 9 "partners with best of breed AI technology providers," and "where third party AI providers are used, Square 9 will ensure the provider meets or exceeds Square 9's standards for security and confidentiality." So a third party is in the loop by the vendor's own account, subject to a vendor-management program that reviews SOC 2 and HIPAA audits and runs an annual re-review. You simply do not get to know which party, or what it costs.
This matters commercially, not just technically. If the engine is a frontier API, the vendor carries a per-page cost that moves when its supplier reprices, and its gross margin is exposed in a way a 2005 ECM licence never was. Compare a published price list: Amazon Textract charges $1.50 per 1,000 pages for raw text detection and $50 per 1,000 pages for forms in US West (Oregon), with queries at $15 and expense analysis at $10. Those numbers are on a page you can read before you sign. Square 9's are not, because you are buying an outcome and the vendor is absorbing the input. That is fine — until the vendor's owner is measured on margin.
Deprecation Is a Date Somebody Else Picks
The clock on your extraction pipeline is set by a company you have no contract with. Google's Document AI deprecation table shows the pattern plainly: the pretrained-ocr-v1.0-2020-09-23 Enterprise Document OCR processor was slated for discontinuation in a September 26, 2024 release note and was discontinued on April 30, 2025 — about seven months from notice to dark. Google's Gemini deprecation page is blunter still, listing a June 1, 2026 shutdown for gemini-2.0-flash and warning that the published dates are "the earliest possible dates on which a model might be retired."
A model swap is not a version bump for a document pipeline. Extraction behaviour is where accuracy lives: field boundaries move, confidence scores recalibrate, a model that used to return an empty string starts returning a plausible guess. If your AP team tuned validation thresholds against one engine, those thresholds are now describing a different one. We have watched this exact failure mode elsewhere — an API shutdown with a hard date and a fleet running a silently downgraded model both start as somebody else's release note.
The commodity alternatives are not a free hedge either — as our Textract versus Azure versus Gemini teardown found, the right move is usually to split deterministic OCR from LLM extraction so only half the stack is exposed to a model change. Mistral OCR 4, Google Gemini Flash and Azure AI Foundry each publish rates and retirement policies. Your IDP vendor publishes neither, and its new owner has published nothing about which engine Square 9 runs on, or for how long.
Your Change-of-Control Clause Probably Didn't Fire
Before you draft the letter, check whether you actually have the right you think you have — by default, you do not. Square 9 continues as an intact company under new ownership, with the same CEO and, the release says, the same team. Under a standard anti-assignment clause that is a non-event: reaching a change of ownership takes specific words, which is why Tech Contracts Academy publishes separate model language that either deems a change of control an assignment requiring consent, or grants an outright termination right. No such words, no trigger. The contract simply continues with a new parent.
This is the same trap we documented when NVIDIA hired 109 Poolside engineers and no clause fired, and the reason Descartes' Tai customers found their only lever was a 60-day renewal notice. It is also why the Deloitte–Wavicle deal turned on assignment consent rather than sentiment. Go read the clause before you assume you have leverage. If you don't, your leverage is the renewal date and nothing else — so put it in the calendar today.
What to Put in Writing Before Renewal
This Week: Pull your Square 9 agreement and find three things — the assignment and change-of-control language, the notice period before auto-renewal, and whether the contracting entity named on the paper is Square 9 Softworks, Inc. If a change-of-control clause exists, note the window; those windows are short and they run from closing, not from the day you noticed.
This Month: Send one written question to your account team and keep the answer: which model and version powers InquireAI/TransformAI extraction today, where is it hosted, and what notice do we get before it changes? An answer names a model. A non-answer names a capability. Both are useful; only one lets you plan. Ask the same question about SOC 2 and HIPAA scope, and ask for the report rather than the claim.
Before Renewal: Run an export drill and time it. Get your document repository and the trained extraction configuration — field maps, validation rules, workflow definitions — out of the platform and confirm what you receive is usable somewhere else. A repository dump without the configuration is an archive, not a migration. Then re-baseline accuracy on 200 real documents and store the results; that becomes your evidence the day the engine changes underneath you. Price the alternative while you're there, so the renewal conversation has a number in it.
The Bottom Line
The vertical-market roll-up is one of the best-proven models in software, and it was built for a specific kind of asset: mission-critical, low-churn, near-zero marginal cost, and able to stand still for a decade without losing a customer. Square 9's ECM core is exactly that asset. Its generative-AI layer is not. Generative extraction has a cost per page that somebody else sets, an engine that somebody else retires, and an accuracy profile that changes when the engine does. That is a product with a metronome inside it, and the metronome belongs to somebody else.
None of that makes Valsoft a bad owner. A permanent holder is a better landlord than a sponsor counting down to an exit, and Square 9's customers may well get more stability than they had. But stability and currency are different products, and you are now buying one from a company optimised to deliver the other.
Perpetual ownership guarantees your vendor will still be there. It guarantees nothing about the model.
Continue Reading
Textract vs Azure vs Gemini: Split OCR From Extraction Descartes Bought Tai. Your Only Lever Is 60 Days. Nvidia Hired 109 Poolside Engineers. No Clause Fired. Stripe Bought OpenRouter. A Toggle Is Not a Contract. Klaviyo Bought Agency. The Product Dies August 31. Your Assistants API Dies Aug 26. Azure's Exit Is Different. What RAG Actually Costs: $1,308 a Month at 10M Tokens/Day Ray Is Open Source. The Control Plane Above It Isn't.