Salesforce split agent metering, while the production evidence reviewed came from narrow workflows

2026-09-01 · OpenAI · Healthcare · Frontier Lab · Unknown

ChatGPT for Healthcare with Epic records connector

Read-only Epic retrieval plus a Healthcare Public Data plugin over PubMed, DailyMed, CMS Coverage, ClinicalTrials.gov, and RxNorm, in a workspace with role-based access, single sign-on, and audit logs

Healthcare CIOs should scope this as a retrieval-and-synthesis pilot, not an autonomous clinical workflow: the first release cannot write to the chart, and individual accounts cannot use the EHR integration at all. Validate patient-context scoping, citation fidelity, break-glass behavior, and audit export now, because those are the controls a future write-enabled release would inherit. Launch coverage puts Epic's footprint at data for more than 325 million patients; OpenAI disclosed no active-customer, accuracy, or savings figure.

Sources OpenAI, TechCrunch

2026-09-02 · Delhivery · Operations · Incumbent Saas · Unknown

TransportOne autonomous transport-management agents

Six specialized freight agents spanning shipment coordination, planning, carrier procurement, freight audit, and analytics, driven by chat and voice inside an existing transport-management workflow

Freight buyers should convert the projections into a paid pilot with pre-agreed baselines and credits for missed savings. Delhivery projects 5-8% freight-spend reduction, 20-40% lower detention charges, and 100% invoice-audit coverage, but named no launch customer and published no observed result, pricing, or false-positive rate. Six named agents are packaging; the economic test is recovered overcharges net of platform fees and the human exception work the agents create.

Sources Delhivery, CXO Today

2026-09-01 · OneRail · Commerce · Startup · Unknown

OmniStar delivery decision layer

Vendor-reported per-order carrier and delivery-mode selection built on OneRail's logistics data, with routing decisions reported down from 20 minutes to 2.5 minutes

Retail and last-mile buyers should demand savings normalized per order and split between model-driven routing, renegotiated carrier rates, and operational redesign. The reported $40 million three-year figure belongs to an unnamed large tire distributor and is a savings run rate, not realized audited savings; the $6 billion Q4 gross merchandise value (total order value moving through the platform) is a forecast. The proprietary decision network, not the model, is the moat, so secure data portability and carrier-selection explainability in the contract.

Sources CNBC

2026-09-03 · Atira · Sales · Startup · Unknown

Atira industrial sales agents

Agents that read technical requests for quotation, identify requirements and exceptions, and generate configurations, documentation, and pricing proposals through customers' existing CRM, ERP, and CPQ systems

This is the week's clearest evidence that a narrow vertical workflow can cross from pilot to production without replacing the system of record, which is exactly why industrial buyers should baseline quote-cycle time, expert-intervention rate, win rate, and pricing errors before signing. Atira reports about 15 customers in full production, five above 100 users, with deployment starting in one to two days on monthly cancellation. The customer-reported outcomes are unaudited and one customer also invested, so treat them as a disclosure shape to copy rather than a benchmark to inherit.

Sources Fortune via Yahoo Finance, Tech.eu

2026-09-01 · Norbert Health · Healthcare · Startup · Unknown

Norbert autonomous nursing-assistant control system

Vendor-reported contactless sensing, protocol execution, electronic-health-record documentation, and patient interaction for daily skilled-nursing rounds

Skilled-nursing operators should evaluate reimbursable workflow economics and protocol adherence rather than robot capability. Norbert reports 96% patient acceptance and 82% monitoring compliance since August 2025, nearly double its claimed manual baseline, which is a usable denominator; facility count, intervention accuracy, adverse events, hardware cost, and reimbursement realization are all undisclosed. Ask for facility-level denominators, adverse-event handling, uptime, and payback after hardware and staffing before scaling past one wing.

Sources Norbert Health

2026-09-03 · Salesforce

Core, Advanced, and Max editions for Agentforce Sales and Service

Three bundles fold Agentforce, Slack, Tableau Next, security, support, and Flex Credits into one seat price, raising the entry floor while expanding top-tier included consumption

Renewal comparisons have to move from seat price to expected actions per seat, because the AI capacity is no longer separably priced from collaboration, analytics, security, and support. Require a modeled post-credit overage curve and written clarification on whether credits pool at org, product, or user level; Salesforce's own 'value' percentages are vendor calculations, not realized savings.

Sources Salesforce, CIO

2026-09-02 · Salesforce

2026 Customer Success Awards deployment disclosures

Workflow-level outcomes published: Tottenham Hotspur reports 80,000 minutes saved in month one, The Grout Guy reports quotes falling from three to five days to 20 minutes

This is the disclosure shape buyers should demand in pilot acceptance criteria — baseline time, post-deployment time, volume, escalation rate, staffing capacity — because it is materially stronger evidence than counts of agent interactions or users enabled. Award submissions are positively selected and customer-reported, and the write-up omits implementation cost, credit consumption, exception rate, and whether time savings became lower total cost.

Sources Salesforce

2026-09-01 · GitHub

Copilot included-credit pools

The promotional credit expiry announced in W35 took effect and standard included pools are now what organizations are billed against

The prior issue documented the credit math; what is new is the effective date. The reset landed as Claude Fable 5.1 and Gemini 3.8 Flash entered Copilot and enterprises gained per-team default-model settings, so model routing is now a budget instrument rather than a developer preference. Engineering leaders should set the routing matrix and budget hierarchy together.

Sources GitHub Docs

2026-09-02 · Other · $550M at $5B valuation

Wonderful

$550 million Series C at a $5 billion valuation led by Insight Partners with Salesforce participating, after expanding from customer-service agents into a horizontal enterprise operating layer

The financing validates forward-deployed implementation — embedded engineering pods, including on-premises work — as part of the product rather than services overhead, which is a different moat than either vertical corpus or model quality. CIOs comparing platforms should separate software gross margin from embedded engineering cost and require a transition plan that leaves internal teams able to operate the deployment. Valuation rose from $2 billion in roughly six months on disclosed customer and market counts with no revenue, retention, or margin data.

Sources Wonderful, TechCrunch

2026-09-03 · Healthcare · $30M Series C

PatientIQ

$30 million Series C on a patient-reported-outcomes network that grew from 200 healthcare organizations in 2022 to more than 870, covering 17 million-plus patients and 50-plus EHR integrations

Healthcare application moats can still be data-network and workflow moats rather than model moats, which is the counterweight to the week's frontier-lab integration story. Buyers should value registry interoperability, longitudinal completeness, and regulatory utility ahead of model capability. The release does not isolate AI-attributable revenue, current recurring revenue, pricing, or same-customer expansion, so the network counts are the evidence and the economics are not.

Sources PatientIQ

2026-09-02 · Healthcare · $6M

Konko AI

$6 million raise with vendor-reported volume across 90 institutions and 60 specialties in Mexico, Colombia, and Costa Rica

An end-to-end automation rate is only procurement-relevant if 'end to end' has a fixed definition and a stated escalation denominator, so buyers should request per-workflow automation rates, escalation and safety rates, no-show recovery, and incremental collections by institution. The customer-side claims of up to 30% revenue growth and 50% productivity improvement are unaudited company figures with no ranges or sample sizes and should not enter a business case unmodified.

Sources Konko AI

2026-09-03 · Operations · $800M pre-money, down from a $1.2B prior deal

PlusAI

New business combination with Texas Ventures Acquisition III at an $800 million pre-money equity value, filed the same day PlusAI withdrew its earlier $1.2 billion transaction

The reset is public evidence that commercialization risk still materially discounts vertical physical-AI valuations, which investors should apply to the rest of the autonomy cohort. Diligence must separate current simulation and development-platform revenue — PlusAI cites $25 million from HyperFoundry and targets $40-50 million contracted in 2026 — from the driver-as-a-service revenue that depends on a 2027 commercial launch. The $1 billion-plus recurring-revenue opportunity is a forecast, and earlier SEC materials described part of that $25 million as milestone-linked development funding, so reconcile revenue recognition before using the figure.

Sources Business Wire via FT, SEC filings

2026-09-03 · Seat Based → Hybrid

Salesforce

Core at $195 per user per month includes 500,000 Flex Credits, Advanced at $395 includes 1 million, and Max holds at $550 while included credits rise from 1 million to 2.75 million. At list rates of $500 per 100,000 credits and 20 credits per standard production action, a standard action is about $0.10, so the 5.5-fold included-credit spread from Core to Max is the real negotiating surface. Model expected actions per seat before choosing an edition; high-use teams may be cheaper at the highest seat price, and industry-edition prices change later in the fall.

Sources Salesforce, CIO

2026-09-01 · Unknown → Usage Based

Anthropic

Enterprise Frontier Safeguards carries no Anthropic charge; customers pay their own cloud provider for storage, reads, writes, and egress on the monitoring logs they now hold. The safeguard is free and the evidence is metered, which moves an unbudgeted line into the buyer's cloud bill and turns retention window into a cost decision. Anthropic has not disclosed alert volume, detector precision, or rolling-window duration, so the bill cannot be sized before the phased rollout later this fall.

Sources Anthropic

2026-09-01 · Hybrid → Hybrid

GitHub

The promotional Copilot credit pools expired and standard included pools took effect; the prior issue carries the per-tier math. The procurement change is that Gemini 3.8 Flash entered Copilot under introductory provider pricing, so current per-task economics are promotional and should not be carried into a 2027 run-rate assumption.

Sources GitHub

Selective metering: entry floors rise while top-tier consumption is subsidized

The simple read that promotional agent credits snap back is only half right. Salesforce raised the committed seat floor on its two lower editions while increasing Max's included credits 175% at an unchanged seat price, which consolidates heavy users onto one bill instead of pushing them to overage. Procurement should model the utilization curve first and pick the edition second; a 5.5-fold included-credit spread means the wrong tier is either stranded inventory or a metered surprise.

Examples

Salesforce three-tier credit spread across Core, Advanced, and Max (integers in pricing shifts), GitHub Copilot standard included pools taking effect September 1

Sources Salesforce, CIO, GitHub Docs

Frontier labs buy vertical distribution through the system of record, not through a vertical model

Neither lab trained a vertical foundation model this week, and both moved into regulated verticals anyway — one by attaching to the dominant clinical record, the other by making its control plane portable across three hosting platforms. Vertical-AI founders should assume integration-acquired distribution is now a standing competitive threat and should compete on the operating denominator they can prove. Buyers get a second sourcing option, but read-only retrieval with no accuracy disclosure is not yet a substitute for a workflow product.

Examples

OpenAI ChatGPT for Healthcare connecting read-only to Epic, Anthropic Enterprise Frontier Safeguards planned across Bedrock, Google Agent Platform, and Microsoft Foundry

Sources OpenAI, Anthropic

The control boundary becomes the purchasable unit in regulated verticals

In regulated buying, the commercial unit is shifting from a vertical-tuned model to the boundary drawn around it: who holds the logs, who owns the keys, who reviews a flag, and how long anything is retained. Regulated buyers should add custody of monitoring data, reviewer identity, retention period, and cross-cloud control equivalence to model procurement scoring this quarter. Neither vendor has published detector performance or an audited control report, so score the architecture and require the evidence at contract.

Examples

Anthropic customer-owned monitoring storage, keys, and review, OpenAI healthcare workspace with role-based access, single sign-on, audit logs, and business associate agreement configuration

Sources Anthropic, OpenAI

Narrow specialists reach production across the record systems rather than displacing them

The three application-layer entrants with production evidence this week all sat on top of somebody else's system rather than replacing it, and all three own a decision or data layer the incumbent does not. That is a cheaper sale and a faster deployment — Atira reports starting in one to two days — but it also means the incumbent keeps the record and the renewal. Buyers should insist on data portability and explainability for the decision layer, because that is the only asset the specialist actually holds.

Examples

Atira agents running across customer CRM, ERP, and CPQ, OneRail OmniStar as a decision layer over its own carrier network, Norbert Health control software layered onto partner robots

Sources Fortune via Yahoo Finance, CNBC, Norbert Health

Oct 1

UK Sovereign AI R&D Procurement Scheme first application cutoff

Four competitions covering NHS productivity, compute efficiency, defence mission integration, and agent security testing, with contracts of £250,000 to £10 million and suppliers retaining created intellectual property. Watch whether outcome-specified, milestone-paid demonstrators become a template enterprise buyers copy.

Fall 2026

Salesforce industry-edition pricing takes effect

The Core, Advanced, and Max comparison is incomplete until industry-cloud prices and their credit allocations post. Buyers in regulated industry clouds should hold renewal decisions until the full rate card is visible.

Fall 2026

Anthropic Enterprise Frontier Safeguards broad availability

Regulated buyers cannot size reviewer staffing, liability, or cloud storage cost without detector precision, alert volume, and rolling-window duration. Broad availability without those disclosures should hold the control out of a scored RFP.

Dec 2

EU AI Act legacy machine-readable marking deadline

The AI Office sent first-stage information requests to more than 30 companies on September 1, covering safety, security, copyright, and transparency. Vertical agent outputs in legal, finance, and healthcare need marking duties written into vendor contracts before the legacy grace period closes.

Q4 2026

Whether OpenAI extends the Epic connection beyond read-only

A write-enabled release would move ChatGPT for Healthcare from retrieval into clinical workflow and would change the risk profile entirely. Validate patient-context scoping, break-glass behavior, and audit export now, while the blast radius is still bounded.

• Big read leads with selective metering — Salesforce raising the entry floor while lifting Max included credits 175% — which revises the simpler W35 read that promotional agent credits snap back.
• Announcements and measured production outcomes are separated explicitly throughout; every measured figure this week is vendor- or customer-supplied and labeled as unaudited.
• Healthcare treated as the week's three-layer vertical: frontier-lab integration (OpenAI/Epic, read-only), regulated data custody (Anthropic, one clause with Agent Techniques as fact home), and three specialists with operating denominators.
• Contradictory thesis evidence stated in the big read rather than buried: OpenAI/Epic strains corpus-and-workflow, Atira contradicts system-of-record absorption, Wonderful contradicts pure verticalization.
• GitHub Copilot credit reset carried as a one-clause effective-date reference to W35 rather than restating the credit math.