AI procurement

Harvey sold flat seats on top of a metered model bill, and agent usage drove its gross margin to -50% by June. The margin recovered after it launched its own Kimi K3-based model, with no reported price change, so renewals of agentic AI seats need model-change and usage terms.

A 40-call black-box audit from USTC flagged 7 of 15 OpenAI-compatible GPT services for behaviour consistent with output-token inflation. The flags are not proof, but padded answers pass quality evals and invoice checks alike, so measure output tokens per task against the first-party API.

In-region LLM inference costs about 10% on every platform. What differs is what the promise excludes (abuse monitoring, batch, caching, control-plane metadata) and whether you can enforce it by policy.

Six questions decide an AI vendor security review, and a SOC 2 report answers none of them. How OpenAI, Anthropic, Microsoft, Google and Amazon Bedrock answer each, as of September 2026.

An unrepaired probability sample of 400 MCP servers found only 48.8% complete an initialize handshake, and 37.5% never start at all. Meanwhile 68.8% of BFCL v4's extracted tool definitions and 85.6% of UltraTool's are exact repeats, against 0.4% for real MCP tools.

The Era by Eon Benchmark ran nine models over 33 questions three times each and published the 95% intervals alongside the leaderboard. They are about 30 points wide, and 33 of the 36 pairwise comparisons could not be called. Your internal bake-off has the same problem.

Stilla's founders promised continuity the day Meta's purchase was reported. Meta's own stated reason for buying is merchant messaging on WhatsApp, not the workplace agent you run on Slack and GitHub. The continuity promise came from the party about to stop existing.

Silver Lake already owns both Cegid and Silae, so their €10bn merger fires no change-of-control clause. What moves instead is Cegid's DGFiP e-invoicing registration — number 0007, three years, renewable in 2027, the same year the deal closes and French SMEs must start issuing.

Valsoft's TAG Software Group acquired Square 9 Softworks on September 8, 2026. A perpetual-hold holdco with 150+ companies is built to compound margin on an installed base; whether that extends to funding a generative-AI model refresh is unanswered — and Square 9 has never named the model behind its extraction engine.

A Copilot seat lists at $19 to $30 a month, but agent consumption, a credit pool that expires and per-gibibyte indexing decide the bill. One GitHub Copilot developer running a daily agent session costs $174 a month against a $19 seat.

The NSA, CISA and FBI advisory AA26-251A tells AI providers to flag accounts by 24/7 usage, instant maximum throughput and cache-optimised traffic — an exact description of a production agent fleet. Its recommended mitigation is an undisclosed model downgrade, engineered to defeat quality measurement.

Mistral's €3 billion Series D moved its lead investor to Samsung Electronics without changing one customer contract. The same announcement publishes a four-part sovereignty definition that maps to clauses you can actually test — and the European Commission's own reference rubric weights ownership at 15%.

ERPBench ran 100 identical ERP problems through six model families twice — solo against fixed opponents, then all six competing in one market. DeepSeek won the first, Gemini the second, and the two agreed on 21 of 100 tasks.

M&T Bank's Copilot population has been published three ways in twelve months — 16,000 using, 17,000 deployed to, 15,000+ with access — while its own filings show headcount fell 928. None of those figures is weekly active users, and that is the only one a per-seat renewal turns on.

A study of 14,922 agent trajectories across five coding benchmarks finds Claude succeeds by matching the human patch's file scope while Qwen succeeds by exceeding it at every scale. Equal pass rates buy unequal diffs, and your ticket template is the control.

Two studies published the same day measured what grounds AI software recommendations. Perplexity's third-most-cited domain is a demo-software vendor's marketing blog, 59.8% of its citations sit below Tranco rank 100,000, and 34.7% of numeric citations point at a page that won't open or doesn't contain the number.

Zendesk built a billing tier specifically to avoid charging for containment. Fin still bills 24 hours of customer silence as a resolution. Across seven platforms priced at the same 1M-contact workload, the resolution definition moves your bill further than the rate does.

An NHS AI scribe dropped the word 'null' and a negative test became a demyelination diagnosis. Healthwatch England found 27 scribe products in use and patients, not clinicians, catching the errors — because fluent prose is exactly what defeats a human review step.

McKinsey's 2026 survey found 32% of organizations skipped a software purchase because agentic coding tools could build it in-house — while the share reporting any EBIT impact from AI stayed flat at 37%. The industry spread explains why: insurance and the public sector, at 19% and 17%, buy audit evidence and liability, not code.

Three of the four best-known synthetic data vendors were acquired or shut down between November 2024 and June 2026. Prove synthesis works with a free Apache-2.0 SDK inside your own perimeter before you buy a platform — and for most test-data problems, masking is still the cheaper and more explainable answer.

OpenAI stops serving its models to Cursor on 12 November 2026. Pointing Cursor at your own OpenAI key is not a swap — Cursor's documentation says its Zero Data Retention policy does not apply to custom keys, and every request still routes through Cursor's backend.

Sony Music Publishing and Warner Chappell allege Claude's lyric guardrails are defeated by re-prompting. That exact behaviour trips a condition in every major AI copyright indemnity — Microsoft's, Google's, OpenAI's and Anthropic's own — by four different routes.

Nvidia's 3,400 tokens/sec on Groq 3 LPX and Cerebras' 4,400 on CS-4 are both single-stream figures. On a 256-LPU rack at 100K context, the memory math caps concurrency near a batch of 12 — so any capacity plan sized off a headline token rate is sized for one user.

Nvidia carries $279B in supply commitments, an increase it attributes primarily to memory, and still guided gross margin from 75% down to a 71-72% trough on RAM costs. Your server quote has no such contract behind it, and every self-hosted inference TCO model built on 2025 hardware prices is now understated.

Thomson Reuters' new in-house legal model is a Qwen derivative two hops down, and Harvey's Tenet is post-trained from Kimi K3. Neither lineage appears in a DPA or a subprocessor list — because an open-weight base model provider processes nothing.

Descartes paid about $100M for Tai and named broker transaction, carrier and shipment data as the rationale. Tai's published terms already pre-authorize assignment to a successor by merger and permit pooling of aggregated data — so the only lever a broker holds is the 60-day notice before auto-renewal.

OpenAI cut GPT-5.6 Sol to $4/$20 per million tokens but guarantees the rate only "at least through November 21, 2026," with no successor published. The cut was asymmetric, so no two workloads saved the same amount — and the reversion is +50% on output, not +33%.

StartupBench ran nine frontier agents across 97 workflows taken from products AI startups already sell. A 73.67 rubric average produced an acceptable deliverable on 29.55% of runs — and agents satisfied auxiliary requirements more often than core ones.

Four agent orchestration products shut down, were superseded or repriced in twelve months. A six-criterion scorecard that weights exit cost at 30%, a six-week pilot with a portability test in week three, and the vendor landscape mapped to both.

Priced on one 10,000-conversation workload, the same agentic work costs $5,940 or $20,000 a month depending only on the billing unit. Per-resolution is the only meter where the vendor absorbs the 30x token variance instead of the buyer.

Claude Fable 5 and Mythos 5 are Covered Models: they refuse to run in a zero-data-retention workspace on the Claude API, Bedrock, Google Cloud or Microsoft Foundry. Anthropic's own docs also say flagged prompts can be held up to two years even under ZDR — while OpenAI previews the opposite bet.

Anthropic is in advanced talks to buy Decart for roughly $7 billion, mostly in its own pre-IPO stock, to cut what a Claude token costs it to produce. Every Sonnet through 4.6 still lists at its March 2024 price, and the one cut that did land arrived as a new model number — which is why a flat multi-year rate card is the wrong thing to sign this quarter.

Stripe is reportedly paying over $7 billion for the gateway 8 million developers route prompts through. Its privacy toggles bind the downstream model providers, not OpenRouter itself — and the only DPA anybody countersigns is enterprise-tier.

Seven routes off OpenAI, priced and deployment-matched on 12 August 2026. Residency is already solved almost everywhere — only Mistral, Cohere private deployment and self-hosted open weights clear a jurisdiction rule.

DOE's Genesis Open Models portal wants proprietary scientific data and the first window closes August 14. Applying transfers nothing — but there is no published contributor agreement, so the handling terms are yours to draft.

OpenAI's NextSlide deal closed early in 2026 and only surfaced on August 8, when the founder posted a goodbye note — five months after he joined OpenAI in March. In this wave of AI acqui-hires the product dies in seven to thirty days and the data is deleted weeks later — so the export window has to be contractual, not assumed.

AlixPartners announced its completed acquisition of agentic AI consultancy Artium on 4 August 2026. Your build partner now sits inside a restructuring firm that advises creditors, takes CRO seats, and files Rule 2014 connection disclosures naming its clients.

Airtable bundles 15,000 AI credits into a $20 seat and sells the same allowance standalone for $30. Bending Spoons — which capped Evernote's free tier and moved to cut 75% of WeTransfer — just bought that subsidy. The deal has not closed, which is the only window you have to convert it into contract terms.

DeepSeek shipped new weights behind the same API name on July 31 — identical architecture, identical price, DeepSWE up from 7.3 to 54.4. Any agent evaluation pinned to a model name instead of a build is now returning a stale verdict.

Chinese-built AI models now account for 30-46% of enterprise API token traffic on US platforms — up from 4.5% in early 2025. Coinbase cut AI spend in half. Lindy migrated 100% from Claude to DeepSeek. Congress is investigating. Beijing may restrict access. Enterprise leaders need a structured risk-reward framework before regulation forces their hand.

OpenAI proposed giving the US government a $42.6B equity stake. What government-owned AI vendors mean for your procurement and vendor strategy.

The Pentagon signed AI deals with 8 vendors May 1 and excluded Anthropic over usage policy. What enterprise buyers should learn about multi-vendor AI.

Anthropic eyes $900B valuation, surpassing OpenAI's $852B. Why 46% of Google's Q1 profit comes from its stake — and what enterprise buyers should do.