| The Agentic Enterprise |
AK · Thu, Jul 9, 2026 · 7 min |
|
|
Thursday, July 9, 2026
Your AI vendor wants to move in.
The model stopped being the product. The outcome is, and the biggest vendors want to own it from inside your building.
OpenAI streams its biggest ChatGPT for Work update today, hours after previewing the GPT-5.6 family. A week ago Microsoft committed $2.5 billion to embed 6,000 engineers inside customers. Two moves, one direction: the AI business is shifting from selling models to selling deployment, because 95% of enterprise pilots still deliver no measurable return.
|
|
The Lead
OpenAI ships its biggest work update today. Microsoft just committed $2.5B to put 6,000 engineers inside customers. The product stopped being the model.
Two different moves, one direction. The leading AI vendors have stopped selling models and started selling deployment. The product is no longer the token or the model card. It is the outcome, and they want to be accountable for it inside your building.
The reason is a number that has haunted every board deck for a year: MIT found 95% of enterprise AI pilots deliver no measurable return. The vendors have concluded they cannot fix that from the API. They have to come inside.
|
|
Your AI vendor wants to move in.
| F |
or two years the AI business was a model business. You bought access to a capability, wired it into a workflow, and the vendor's job ended at the API. This week that model quietly changed. OpenAI is streaming its largest ChatGPT for Work update today, positioning the product as the place work happens rather than a tool bolted onto it, and it arrives alongside the wider rollout of the GPT-5.6 family (Sol, Terra, and Luna) tuned for long-horizon agentic tasks. A week earlier, Microsoft committed $2.5 billion to a new operating company, Frontier, whose entire purpose is to put 6,000 engineers and industry specialists inside customer organizations to design, deploy, and keep improving AI systems against measured business outcomes. |
The two announcements point the same way. The frontier labs and the hyperscalers have concluded that selling capability is not the same as selling results, and that the gap between the two is where their growth now lives. Microsoft was explicit about why: it cited the MIT finding that roughly 95% of enterprise generative-AI pilots produce no measurable profit-and-loss impact, and framed Frontier as the answer. Not a better model. More people, sitting closer to the work.
|
The model was the product for two years. The outcome is the product now, and the vendors want to own the last mile of it.
|
For enterprise leaders, this is a procurement shift dressed as a product launch. What you are being sold is moving from a license to an engagement: embedded engineers, shared workspaces, outcome-based scopes, and a vendor with a direct stake in whether your deployment works. That is genuinely useful if your pilots keep stalling at the integration step, which most do. It is also a dependency you are choosing on purpose. The forward-deployed-engineer model, pioneered by Palantir and now copied by Microsoft, AWS, and the labs, works because the vendor learns your operation intimately. That is exactly what makes it hard to leave.
|
The Spearhead Take
Take the help, keep the thinking. Embedded vendor engineers are the fastest way to get a stalled pilot into production, and refusing them on principle is a luxury most AI programs cannot afford this year. But the deployment gap is an organizational problem before it is a technical one, most pilots fail because nobody could specify the workflow or owned the outcome, and no amount of rented engineers fixes a company that cannot do that for itself. Use the vendor to ship. Build the internal muscle to specify, govern, and eventually run the system yourself, or you will be renting the outcome forever.
|
|
|
The Obvious & The Overlooked
The launch is the headline. The procurement shift underneath it is the story.
|
The Obvious
The vendors are selling deployment now, not just models.
OpenAI and Microsoft both moved this week to own the outcome inside the customer. TechCrunch
OpenAI is making ChatGPT the workplace, not a tool in it.
Today's update pushes shared workspaces, admin controls, and custom assistants. Crypto Briefing
GPT-5.6 is priced for agentic work.
Sol at $5/$30, Terra at $2.50/$15, Luna at $1/$6 per million tokens. AI Pricing Guru
|
The Overlooked
The whole pivot is a response to one MIT stat.
95% of enterprise AI pilots show zero P&L impact, and vendors have stopped pretending a better model fixes it. Fortune
The EU compliance cliff is 24 days out.
Annex III high-risk obligations hit August 2, with fines up to 7% of global turnover. ComplianceHub
Most companies are not ready for it.
As of April, 78% of organizations had taken no meaningful compliance steps and over half lacked an AI inventory. Cloud Security Alliance
Embedded engineers are a moat, not a service.
The forward-deployed model works by learning your operation so well you cannot easily switch. GeekWire
|
|
|
Moving Pieces
Five developments worth a CIO's attention.
Product
OpenAI ships its biggest work update and widens GPT-5.6
OpenAI is streaming what it calls its largest update to ChatGPT for Work today, with expected features spanning team collaboration, shared and resettable workspaces, and custom workplace assistants. It lands alongside the wider rollout of the GPT-5.6 family, Sol for frontier reasoning and long-horizon agentic work, Terra as the balanced everyday model, and Luna as the cheapest and fastest, with Sol also reaching customers on Cerebras hardware at up to 750 tokens per second. The enterprise read: OpenAI is repositioning ChatGPT from an assistant you visit to the surface where work is done, which puts it in direct competition with Microsoft 365 Copilot and Google Workspace for the seat, not just the query. Availability is staged and some access remains gated, so treat the capability as arriving in waves rather than switched on everywhere.
Strategy
Microsoft commits $2.5B and 6,000 engineers to fix enterprise AI
Microsoft launched Frontier, a new operating company backed by a $2.5 billion investment, to embed 6,000 industry and engineering experts inside customer organizations to co-design, deploy, and continuously improve AI systems against measured outcomes. Microsoft framed it directly against the MIT finding that 95% of enterprise AI pilots deliver no P&L impact, and said it will work alongside partners including Accenture and KPMG. The enterprise read: this is the forward-deployed-engineer model, long associated with Palantir, going mainstream at hyperscaler scale, and it signals that Microsoft believes the bottleneck to AI revenue is deployment labor, not model capability. It also reframes the sale from software licensing to an outcome-based engagement, which changes how you scope, price, and govern the relationship.
Policy
The EU AI Act's high-risk deadline is 24 days out
The EU AI Act's most consequential obligations, the Annex III high-risk system requirements plus transparency duties, conformity assessments, CE marking, and AI Office enforcement, take effect on August 2, 2026. Penalties reach 7% of global annual turnover, above GDPR's 4% ceiling, and large-enterprise compliance costs run $8 million to $15 million. A separate Digital Omnibus proposal could defer stand-alone Annex III obligations to December 2027, but only if formally adopted and published before the August date, so the deadline stands until it does not. The enterprise read: if you deploy AI in hiring, credit, education, or other listed domains for the EU market, this is a hard date with real money attached, and as of spring most companies had not started. Build the AI inventory now, because you cannot assess what you cannot see.
Deals
Together AI raises $800M to grow public cloud capacity 50x
Together AI closed an $800 million Series C led by Aramco Ventures, with NVIDIA, Vista Equity Partners, and General Catalyst participating, and disclosed that annual bookings crossed $1.15 billion in the second quarter. The company plans to use the capital to expand its public cloud capacity by a factor of 50 over five years. The enterprise read: the market for hosting and serving open-weight models, the layer beneath the cost-driven model switching we covered yesterday, is now a multi-billion-dollar bookings business in its own right. For buyers routing workloads to cheaper open models, the neutral inference provider is becoming a strategic vendor category, not a commodity, and the biggest ones are raising to lock in capacity ahead of demand.
Capital
Abu Dhabi's MGX closes a $49B fund for AI infrastructure
MGX, the Abu Dhabi state-backed investor, closed its first fund at $49 billion, above a $45 billion target, and has already deployed across 14 companies spanning semiconductors, AI infrastructure, and AI platforms. It is developing what it calls Europe's largest AI campus near Paris, with 3 gigawatts of planned compute capacity, and took part in a $40 billion consortium deal to acquire Aligned Data Centres. The enterprise read: sovereign capital is now a first-order force in where AI compute gets built and who controls it, which matters for enterprises negotiating capacity, data residency, and geopolitical exposure in multi-year cloud commitments. When the landlord of your future compute is a Gulf sovereign fund building in France, vendor diligence extends past the logo on the invoice.
|
|
On the Radar
Eight signals, sharpened.
| Product |
OpenAI upgraded ChatGPT voice mode with a redesigned experience and CarPlay support, rolling out ahead of today's work event. The assistant is pushing into more of the workday, including the commute. 9to5Mac |
| Deals |
Salesforce says Agentforce annual recurring revenue crossed $1 billion, up triple digits year over year. Agentic revenue is now material at the incumbents, not just the startups. PYMNTS |
| Product |
ServiceNow opened its Action Fabric to any external agent via an MCP server, with Anthropic's Claude as launch partner. The enterprise platforms are competing to be the place agents act, regardless of who built them. ServiceNow |
| Research |
Gartner projects 40% of enterprise applications will have embedded agents by year-end, up from under 5% in 2025. The agent is moving from a feature you buy to a default you inherit. To The New |
| Governance |
Over 40% of agentic AI projects are at risk of cancellation by 2027, per Gartner, with governance gaps a leading cause. Deployment speed is outrunning control, and the bill comes due at audit time. Agentic AI Institute |
| Deals |
Kling AI closed about $2 billion at an $18 billion valuation, backed by General Atlantic, for its video generation models. Chinese generative-media capital is scaling even as US enterprises weigh model provenance. Crescendo AI |
| Infrastructure |
OpenAI is serving GPT-5.6 Sol on Cerebras hardware at up to 750 tokens per second. Inference speed is becoming a product differentiator for agentic workloads, not just a spec. Explain X |
| Compliance |
As of April, 78% of organizations had taken no meaningful EU AI Act compliance steps and more than half lacked a basic AI inventory. The readiness gap is the story, not the deadline. Cloud Security Alliance |
|
|
Quick Hits
The wider field, one line each.
| Microsoft's Frontier will work alongside Accenture and KPMG on enterprise deployments. The Decoder |
| OpenAI's ChatGPT Work stream is billed as its biggest update for work in ChatGPT. Testing Catalog |
| GPT-5.6 Terra is pitched as GPT-5.5-competitive at roughly half the cost. Explain X |
| MIT's NANDA report analyzed 300 public AI deployments to reach the 95% zero-return figure. Fortune |
| MIT found buying from specialized vendors succeeds about 67% of the time, roughly triple the rate of internal builds. Healthcare IT News |
| Together AI's annual bookings crossed $1.15 billion in Q2 2026. Crescendo AI |
| MGX closed Fund I at $49 billion, above its $45 billion target. Crunchbase News |
| MGX is building what it calls Europe's largest AI campus near Paris at 3GW. StartupHub.ai |
| SAP's Autonomous Enterprise embeds 200-plus agents and uses Anthropic's Claude as a primary reasoning engine. The Next Web |
| Tripo AI raised roughly $150 million to advance interactive 3D foundation and world models. Crescendo AI |
| EU AI Act fines for high-risk violations can reach 7% of global annual turnover. Legal Nodes |
| Microsoft called Frontier the largest outcome-driven engineering organization in the industry. GeekWire |
|
|
The Number
6,000
Engineers Microsoft will put inside customers
The people Microsoft says it will embed inside customer organizations through its new $2.5 billion Frontier company.
Set it against the model race. For two years the competition was measured in parameters, benchmarks, and tokens per second. Microsoft's answer to the enterprise AI problem this week was not a number about the model at all. It was a headcount, thousands of people it will station inside other companies to make the technology actually pay off. The most telling metric in AI right now is not how smart the model is. It is how many humans the vendor thinks it takes to make yours work.
|
|
Counter-Signal
Governance
Rented engineers do not build a capability.
The deployment pivot reads as a gift: the vendor finally owns the hard part, and your stalled pilot ships. The catch is that shipping and capability are not the same thing. When Microsoft's or OpenAI's engineers design, deploy, and continuously improve the system, the institutional knowledge of how it works lives with them, not with you. The forward-deployed model is a moat precisely because it makes leaving expensive, and an outcome you cannot reproduce internally is an outcome you rent indefinitely.
The MIT research the vendors keep citing points somewhere more uncomfortable for them. The pilots that failed did not fail for lack of vendor engineers. They failed because organizations invested in visibility projects over operational ones, could not specify the workflow, and did not empower the line managers who actually run the work. Those are governance and organizational-design problems, and outsourcing them to a vendor's staff can paper over the symptom while leaving the disease. The workable version is deliberate: use embedded engineers to ship the first production system, and treat internal capability transfer, documentation, and the right to operate it yourself as non-negotiable terms of the engagement. Take the deployment help. Do not let it become the reason you never learn to deploy.
|
|
From the Field
The instinct, when 95% of pilots stall, is to reach for more capability. This week two of the biggest vendors conceded that was the wrong reach.
A better model. A faster one. A cheaper one. The vendors spent two years selling into that instinct. The problem was never that the model could not do the work. It was that the organization could not absorb it.
That is why the offer changed from a model to a team. Microsoft is betting $2.5 billion that the missing ingredient is people who sit inside your operation and stay until it works. OpenAI is betting that if it owns the surface where work happens, adoption stops being something you have to engineer. Both are probably right that deployment, not capability, is the binding constraint. Both are also selling you a dependency, and the honest read is that you should take it with your eyes open.
The vendor can close your deployment gap. Whether you own what gets built on the other side of it is a decision, not a deliverable.
Let's get to production,
AK
|
|
|
|
Anthropic is a Spearhead technology partner, and its Claude model produced this edition under human editorial direction. Anthropic is not the subject of this edition; it appears in supporting items as an ecosystem player (Claude as ServiceNow's Action Fabric launch partner and SAP's reasoning engine). The Big Story and The Number center on OpenAI and Microsoft, and no favorable or unfavorable framing was applied on account of the partnership. Today's OpenAI ChatGPT Work livestream (July 9) is described from pre-event reporting; specific features are expected rather than confirmed, and GPT-5.6 availability is staged, with sources conflicting on the breadth of general access. Microsoft Frontier ($2.5B, 6,000 experts) is multi-sourced (TechCrunch, CNBC, Microsoft, GeekWire, The Decoder). The MIT NANDA statistic (95% of pilots show no P&L impact) is from the widely cited 2025 "State of AI in Business" report and is presented as backdrop, not fresh news. GPT-5.6 pricing and the Cerebras throughput figure trace to secondary trackers; Together AI, MGX, Kling AI, and Tripo AI details come from funding roundups; Salesforce, ServiceNow, and SAP items reflect earlier-2026 announcements cited as ecosystem context. The EU AI Act August 2 deadline is a scheduled date a pending Digital Omnibus could defer to December 2027 if adopted first. No emoj | |
