Procurement software has a credibility problem. Every vendor in the category claims to use AI. Most of them mean they have a search bar that autocompletes supplier names, or a spend dashboard that color-codes categories you already knew existed. When we looked at Silq eighteen months ago, we were skeptical for exactly this reason. We'd seen too many "procurement intelligence" pitches that were, under the hood, pivot-table wrappers with a chat interface bolted on.
What changed our mind wasn't the demo. It was the data architecture.
What procurement AI actually needs to do
When I ran distribution for a Scandinavian FMCG group across 14 markets, procurement decisions were the connective tissue of the whole operation. Getting the wrong supplier at the wrong lead time didn't just affect cost — it cascaded into production scheduling, inventory buffers, and ultimately whether a product was on shelf. The problem was never a lack of data. We had ERP exports, supplier portals, category managers with years of institutional knowledge in their heads, and a dozen spreadsheets tracking exceptions.
The problem was that none of this learned. Every quarter, a new category manager joined, and the same supplier negotiation patterns had to be rebuilt from scratch. Every procurement event was, in effect, the first one. The organization had memory but no institutional intelligence. There was no system that took the outcomes of past decisions — which supplier contracts had held under pressure, which ones had collapsed at volume spikes, what early signals predicted a supplier going dark — and built that back into the next decision.
That's the gap Silq is filling. Not spend analytics. Not contract repository. A system that builds a continuously improving model of supplier behavior, contract risk, and market dynamics, and surfaces that intelligence at the moment a buyer is actually making a decision.
The model is the moat
Here's what distinguishes Silq architecturally: they built their training loop around outcomes, not documents. Most procurement software ingests contracts, extracts entities, and surfaces clauses. Silq does that too — but the core model is trained on what happened after contracts were signed. Delivery performance against SLA. Price variance over the contract lifecycle. Correlation between supplier financial signals and subsequent performance degradation. They've been ingesting this outcomes data from early design partners for two years, and the model is now at the point where it can flag supplier risk 60–90 days before a procurement team would traditionally notice it.
This matters for defensibility in a way that document ingestion alone never could. If your moat is "we parsed more contracts," a competitor can catch up in eighteen months with a better parser. If your moat is "our model has seen the outcomes of 50,000 supplier contracts across industrial manufacturing, logistics, and FMCG — and it's been right about risk flags 74% of the time in retrospective testing" — that data advantage compounds. New users improve the model. The model justifies more data sharing. You end up in the canonical data flywheel.
We're not saying outcome-based training is easy. It took Silq two years to instrument and clean the feedback loops well enough that the model was actually learning signal rather than noise. That's the work most companies won't do. It's also the work that creates the gap.
Why procurement is a seed bet, not a growth-stage one
Procurement AI is structurally interesting right now for a reason that has nothing to do with AI hype: the software incumbent layer is unusually weak. The Ariba/Coupa duopoly owns the workflow but neither has invested seriously in predictive intelligence. They've added AI features that are essentially LLM wrappers over their existing data models — which is fine for answering "what did I spend on this category last quarter" but useless for "which of my top-10 suppliers by volume is most likely to miss Q3 delivery targets."
The window for a genuinely intelligence-first procurement product is open because the incumbents are protecting workflow revenue, not chasing the intelligence layer. That window won't stay open indefinitely — one of them will either build it or acquire into it within three to five years. The question is whether a company like Silq can get to the point of being the acquisition, not the acqui-hire, before that happens. The path there is the same one it always is: get to real outcomes data at scale, build the model that's genuinely hard to replicate, and be deep enough in a customer's workflow that ripping it out costs more than renewing.
What we diligenced hard
Anders led the technical diligence on Silq. The questions we kept coming back to: How is the training data labeled? Who validates the outcome signals? What's the latency between a procurement event occurring and it feeding back into the model?
The honest answer, going in, was that the feedback loop was slower than we wanted — somewhere between 45 and 90 days depending on the supplier category and how quickly customers pushed outcome data. For high-frequency procurement categories like logistics consumables or packaging materials, that was manageable. For long-cycle categories like capital equipment, a 90-day feedback window is short enough to be useful. That satisfied us.
We also spent time understanding the founding team's domain depth. The two co-founders came out of procurement operations at a large European automotive tier-1 and a Nordic pharmaceutical distributor respectively. They've personally lived the problem of a contract management system that doesn't learn. That background matters because procurement AI sold to enterprise buying teams requires a level of category credibility that you can't fake in a demo. Procurement directors have pattern-matched enough vendor pitches to know immediately whether the person in the room has managed a supplier relationship under pressure.
The investment thesis, plainly stated
We invested at seed because we believe the data advantage Silq is building is real, the founding team has the domain credibility to earn trust in enterprise procurement organizations, and the incumbent weakness in the category creates a durable window. The check size was $3M — at the high end of our seed range, reflecting the depth of the data infrastructure already in place.
The risk we're carrying is typical for enterprise software at this stage: sales cycles are long, enterprise procurement decisions require multi-stakeholder alignment, and there are at least three well-funded competitors in adjacent spaces who could pivot into outcome-based intelligence if they chose to. We think the two-year head start on training data and the outcomes-first architecture make that pivot expensive for competitors — but it's not zero risk.
What gives us conviction is watching Silq's existing design partners use the product. When a procurement director at a growing manufacturing company can say "we flagged a critical supplier risk 10 weeks before their delivery failure — based on Silq's alert — and had a secondary source qualified by the time it happened," that's not a feature. That's institutional memory that didn't exist before. That's what we back.