Agents do the searching, the screening and the writing. Every figure that decides anything — the score, the should-cost, the total cost of ownership, the award — is computed in code. The language model explains the answer. It never produces it.
Finding suppliers was never the hard part. The hard part is proving the one you picked was the right one — to your CFO, to your auditor, to a regulator — and assembling that proof by hand across spreadsheets, email threads and PDFs while the market moves.
So teams reach for AI, and hit the second problem. A model that generates a duty rate, a cost breakdown or a supplier score produces something that reads like evidence and isn't. In a sourcing decision that commits real money, a plausible number is worse than no number.
Plenty of platforms now say a human approves the output. That is a process control, and anyone can claim it. These are properties of how the system is built — they hold whether or not anyone is watching.
Scores, should-cost, TCO, bid rankings and award allocations are computed in plain
Python and a constraint solver. The model layer merges text only. When
the model is unavailable the numbers still stand and the page says
narrative_status: degraded — it never quietly substitutes a guess.
The award optimizer runs a fixed seed over a model built in sorted order, so identical bids and constraints produce a byte-identical allocation every time. Run the award twice and diff it. That is a claim you can test in the room, not one you have to take on trust.
A missing baseline is excluded from the savings report rather than defaulted to zero. An unquoted price refuses to become a $0 purchase order. The platform would rather return nothing than return something that looks like an answer.
Not "approve everything" — that is the workload you are trying to escape. The guardrails are structural, so the human enters at the one decision that genuinely needs judgment: the negotiation. Everything upstream is gated in code.
One continuous loop. Discovery and enrichment, screening, qualification, RFx, negotiation, award, and everything that happens after the PO is issued.
Suppliers are screened against sanctions and watchlists continuously, not once at onboarding — a status change re-triggers screening on its own. Certificates are validated and their expiry tracked.
Materials are screened too, which is rarer. A global, versioned reference of restricted and dangerous goods — nuclear, export-controlled, chemical-weapon and drug precursors, hazmat, conflict and forced-labour, wildlife — blocks the run before it starts. A single sanctions or export-control violation carries civil exposure of up to $377,700 per violation, or twice the transaction value, whichever is greater.
The cost decomposition is not a set of convenient constants. Ratios derive from Eurostat structural business statistics, US Census ASM and BEA industry accounts — and each component carries a flag saying whether it was measured or modelled, because official statistics do not isolate outbound freight or separate profit from depreciation.
Most vendors would round that inconvenience away. Labelling it is the point: a buyer negotiating from this number knows exactly which parts of it are observation and which are inference.
For every material–supplier pair the platform classifies the posture on a
buyer-leverage versus supply-risk matrix, then computes the economics deterministically:
BATNA, ZOPA, and an anchor < target < reservation ladder with
guardrails. A weak BATNA forces the supply-protective posture. Sparse pricing produces
no fabricated target — confidence drops instead.
Outcomes feed back. Realized savings recalibrate future targets for similar categories, under a five-session floor and an outlier-resistant median, so a single lucky deal cannot move the model. Low-risk categories can be handed to a policy-gated agent that negotiates end-to-end — with offers generated as numbers by a concession-ladder walker that cannot exceed the walk-away price.
A constraint solver allocates demand across compliant bids under your rules — share caps, supplier floors, mandatory dual-source thresholds, country exclusions, lead-time limits, declared capacity. Ineligible bids are removed before the solver runs, so they are never allocatable.
When no allocation is possible it says which constraint broke it, in the buyer's own vocabulary, rather than returning an unexplained failure. And every purchase order binds to the report that justified it, snapshotting the evidence as it stood at award time — rendered as a Basis of Award section inside the PO document, so the justification travels with the artifact.
Pick any purchase order. The platform assembles the whole chain in canonical order — enrichment provenance, screening verdict, due diligence, the scorecard, dispatch, every supplier response and its analysis, the comparative ranking, the award scenario, the decision basis, the approvals and overrides, goods receipt, and the three-way invoice match.
Each stage carries a hash over its canonical content, so identical evidence yields an identical hash and any alteration is visible. A stage with no underlying rows is marked absent rather than invented — a manual PO produces an honestly partial dossier. Export as PDF or JSON.
The panels above are simplified representations of live product surfaces, populated with figures from a demonstration dataset. The arithmetic is the platform's own — components sum to the market average, scenario deltas are differences against the cheapest allocation — but the values are illustrative and are not a customer outcome. Benchmark figures elsewhere on this page carry their source.
Direct materials are not indirect spend with different nouns. A bill of materials, a certificate of analysis, an export-control classification and a dual-source requirement have no analogue in catalogue buying — which is why suites built for indirect spend struggle here, and why the new AI-native sourcing tools stop at the award.
| Capability | Indirect-spend suites | AI-native sourcing tools | MatryxAI |
|---|---|---|---|
| BOM-level sourcingMulti-level BOMs, take-off consolidation, per-component runs | Weak | Strong | Strong |
| Bid optimizationConstraint-based split award across compliant bids | Varies | Emerging | Solver-based, reproducible |
| Negotiation economicsBATNA / ZOPA, concession ladder, outcome calibration | No | No | Yes |
| Autonomous negotiationPolicy-gated agent for low-risk categories | No | No | Yes |
| Continuous sanctions screeningRe-screened on status change, not once at onboarding | Add-on | No | Built in |
| Restricted-material gateScreens the material, not just the supplier | No | No | Yes |
| Tamper-evident award recordHashed evidence chain, exportable | Audit log only | No | Yes |
| Post-PO loopReceipt, COA validation, quality holds, three-way match | Strong | Partial | Yes |
| Pricing modelHow the platform is licensed | Per seat | Per seat | Unlimited users |
Unlimited users is a design decision, not a discount. Sourcing is not a procurement-team activity — engineering specifies, finance approves, quality signs off, legal reads the terms, and an auditor turns up two years later. Charging by the seat makes each of those people a line item, so organisations ration access to the system that is supposed to hold the record. Suppliers are never charged either.
The fastest way to waste your evaluation is to discover in week six that the scope was never right. Here is where MatryxAI earns its place, and where it doesn't.
Demos are easy to stage. These are harder, which is what makes them worth asking for — of us, and of anyone else you are evaluating.
We will not answer that with a number until we have your data, and you should be sceptical of anyone who does. Published research puts the return from competitive, well-prepared negotiation at two to seven percent (Pactum); MatryxAI is where that work gets run, not a separate source of savings on top of it.
What we will commit to is the method: the platform's own value report computes savings only from awards that carry a traceable baseline and lists the ones it excluded. You will be able to check our claim the same way your auditor would.
No. MatryxAI is a sourcing and award layer that integrates with SAP S/4HANA, NetSuite, Dynamics 365, SAP Ariba and Salesforce rather than displacing them. Purchase orders, receipts and invoices flow out to the system of record; supplier and invoice data flows back in.
The analysis takes minutes — a full pipeline run completes in under five. The cycle is bounded by how quickly suppliers reply, not by how quickly you can analyse, and we would rather set that expectation now than in month two. What compresses is the preparation, the comparison and the justification, which is where most of the sixty-day median actually goes.
You do. The ranking is computed from your weights — the default is 40% total cost of ownership, 30% compliance, 30% completeness, and the platform shows you that split. The model writes the explanation. Choosing a supplier other than the recommended one is allowed and requires a stated reason, which is recorded in the decision basis and travels into the audit dossier.
No. The platform learns from outcomes as versioned, attributable data rather than as model weights — realized negotiation results recalibrate numeric targets in the deterministic core, under a minimum-sample floor and an outlier-resistant statistic. Nothing enters a model's parameters, which is also why swapping the underlying model changes no number.
The numbers are unaffected, because the model never produced them. A failed model call leaves the computed payload intact and marks the narrative degraded. A judge scores every agent output and escalates to a human approval request rather than passing questionable work downstream. An agent negotiation that drifts outside its guardrails halts and escalates with the full transcript attached.
Yes. Enterprise supports single-tenant and VPC deployment inside your own AWS account — your infrastructure, your data residency, your identity realm. Source escrow with defined release triggers is available for enterprise agreements.
Not the clean one. The award where the cheapest bidder had a compliance flag, or the single-source part nobody wants to talk about. Thirty minutes, and we will show you what the evidence chain looks like when it is assembled by the system rather than by you.
Schedule a working session