OpenAI launched Presence on 22 July, an enterprise platform for voice and chat agents with policy controls, guardrails that can interrupt an interaction while it is running, and escalation to a human. There is no self-serve tier. Every installation runs through OpenAI’s forward deployed engineers or an approved systems integrator, and access goes through an account team. Pricing is not published. OpenAI says deployments “are scoped individually based on each customer’s use case and implementation needs.”
The company that meters intelligence by the token has shipped a product you cannot buy without also buying a service. For two years every frontier lab argued that its competitive advantage was the capability of the model itself. It is now adding people to the model.
In May, OpenAI launched the OpenAI Deployment Company with more than $4bn from nineteen outside investors, led by TPG with Advent, Bain Capital and Brookfield as co-leads, and bought Tomoro, a British AI consultancy. The acquisition brought roughly 150 forward deployed engineers and deployment specialists on day one. The job title is Palantir’s. It describes staff who sit inside a customer’s operation for months, selecting workflows, connecting internal systems, setting permissions and moving agents into production.
Anthropic did the same thing in the same month. Ode, a joint venture with Blackstone valued at $1.5bn, took $300m each from Anthropic, Blackstone and Hellman & Friedman, with Goldman Sachs, Apollo, General Atlantic, GIC, Leonard Green and Sequoia alongside. Google is hiring forward deployed engineers into Google Cloud and lending them to the customers of firms such as Accenture, Deloitte and McKinsey. AWS committed $1bn to the same kind of engineer working on client sites. Meta is reported to be building an Enterprise Solutions unit on the same model, with product managers running client engagements and engineers wiring Meta’s systems into the customer’s own.
Five firms announced the same move in ten weeks. None of them announced a new model alongside it. All of them added people who know a particular customer’s processes, systems and politics from the inside.
I should say straight away that I am not neutral here. I run an independent AI advisory, so the argument that implementation is the hard part suits me. Take the rest as a report on what the labs did with their own money, and judge it on that basis.
What it cost them
The price OpenAI and Anthropic were willing to pay is the part worth reading.
Both put the work in a separate company rather than an internal division. Google, Meta and Amazon did the opposite and kept their engineers on their own payroll. The difference between the two groups is what their share price rests on. Google, Meta and Amazon already run enormous services and advertising operations, and an advisory line changes nothing about how the market values them. OpenAI and Anthropic are valued purely as software businesses.
The precedent is in Palantir’s accounts. Palantir runs an adjusted gross margin of about 84%, higher than Snowflake and Databricks, while providing more integration services than either of them. Professional services run at 18 to 20% of revenue. The margin holds at that level because a share of forward deployed engineering cost is booked in R&D and in sales and marketing rather than in cost of goods sold. A services firm would book those people in COGS. Reclassify them and the gross margin falls materially.
This is not about net profit. Neither OpenAI nor Anthropic is profitable, and neither is valued on a multiple of earnings. It is about what an investor pays per dollar of revenue. Services revenue is valued at roughly one to three times annual sales; software and AI revenue at twenty or thirty times. Advisory work sitting in the same income statement dilutes that ratio across the whole company. In a separate company financed by outside capital, it does not. OpenAI guaranteed its private equity backers a 17.5% annual return over five years to get the structure signed. Those investors are paid on that schedule whether or not a single deployment works.
That is the measure of how seriously to take this. Two companies whose valuation depends on the market reading them as software accepted a permanent financing cost and a guaranteed payout to outside capital in order to own implementation capacity. Nobody pays that for something peripheral.
The consultancies were not displaced
The nineteen investors in OpenAI’s Deployment Company include Bain & Company, Capgemini and McKinsey.
TechCrunch described the arrangement when both ventures were announced. They “get preferred sales access to their investors’ portfolio companies, while the investors will capture more value from any resulting contracts.” Blackstone, TPG, Advent, Brookfield and Apollo hold hundreds of companies across healthcare, manufacturing, financial services and logistics. Those are the customers.
KPMG, named an OpenAI Elite Partner on 21 July, explained the reasoning in public. Chad Seiler, its US industry leader for technology and media, argues that what consultancies hold is client-specific institutional knowledge accumulated over decades: a customer’s business model, people, culture, systems, data and politics. KPMG also maintains parallel partnerships with other model providers rather than betting on one.
None of that is written into the model, and that is where the difficulty sits. The knowledge belongs to one company, it takes months to acquire, and it does not move anywhere else.
Gartner’s 2026 forecast puts worldwide IT services spending at $1.87tn against $1.44tn for software. Services is still the largest line in enterprise IT by more than $400bn, in the third year of the generative AI build-out. The thesis sold since 2023 was that agents would automate implementation labour and the budget would shift from services into software and tokens. The firms building the agents are buying implementation labour themselves.
The limits of this
The valuations do not reconcile. OpenAI’s Deployment Company was reported at $10bn in May and $14bn later, and neither figure is worth citing. No revenue has been disclosed for any of these ventures. The Enterprise Solutions unit at Meta comes from an internal memo reported by The Information; Meta has confirmed no customers, no headcount, no pricing and no timing, so treat it as a plan rather than a business.
It is also unclear where the money for that guarantee comes from. OpenAI is deeply lossmaking and is sustained primarily by successive investor rounds rather than by a surplus earned from customers. I could not establish whether the 17.5% obligation sits on OpenAI itself or on the joint venture’s balance sheet. That difference matters and nobody has explained it publicly.
The reference number for Presence, 75% of inbound calls resolved without a human, comes from OpenAI’s own English-language phone support line. Pareekh Jain of Pareekh Consulting has said that CIOs should treat it “as evidence that the technology can work, rather than as a benchmark that every enterprise can expect to reach.” His reasoning is the ordinary one: fragmented legacy systems, uneven knowledge bases, heavier compliance load. The same article carries the line that matters more than any of the launch numbers. “Often the biggest cost of enterprise AI is not tokens but integration and governance.”
Impact on the Polish market
No Polish enterprise will engage an OpenAI forward deployed engineer this year or next. The arithmetic rules it out at 150 people, and the commercial structure rules it out before the arithmetic does, because the channel is already allocated to the backers’ portfolio companies. What reaches Warsaw reaches it through partners, and mostly through your own staff.
That changes how the budget is built. If the firms with the best models in the world concluded they had to place their own engineers at a customer site for months, then the main cost of your AI programme sits in integration, data preparation and process redesign. In traditional IT the split between services and licences runs at roughly half and half, and sometimes seventy-thirty in favour of services. With AI there is a third line that was not in that calculation: the cost of tokens, which grows with use and is hard to estimate at business-case stage. It is worth checking whether your last AI business case separated those three, or treated implementation as something the team would absorb along the way.
Issue 56 covered the contract clauses that stop an agent programme before it starts. Issue 57 covered who reads the meter once it runs. This is about what has to happen inside your own organisation for either of those questions to matter at all.
Three questions for your board:
- Across our last three AI initiatives, what share of the effort went to model selection and what share went to integration, data preparation and process redesign? Did the original budget reflect that ratio correctly?
- Who holds the institutional knowledge our AI programme depends on, our own people or a supplier’s?
- When a pilot succeeds, what specifically has to change in our systems and processes before it reaches production, and who is accountable for making those changes?
Briefing
Microsoft is funding Mistral’s European compute build-out. The expanded partnership, announced 21 July, puts billions into Mistral data centres in France running Nvidia’s Vera Rubin GPUs. Mistral’s models go into Microsoft Foundry and Copilot Studio. Through Azure Local it offers three deployment modes: cloud, connected, and fully air-gapped. Brad Smith, Microsoft’s vice chair and president, confirmed the deal carries no new equity stake. So what: air-gapped deployment is not itself new. Open-weight models from Llama to Mistral have been self-hostable for years and a substantial share of AI infrastructure in financial services already runs that way. What is new is a frontier model running in that mode with commercial support and a contract behind it, rather than open source you maintain yourself.
Two dozen firms signed an open letter backing open-weight models on 24 July. Meta, Microsoft, Nvidia, IBM, Dell, CrowdStrike, Palantir, ServiceNow, Hugging Face, Perplexity, Mistral, Andreessen Horowitz, Y Combinator, the Linux Foundation and Mozilla asked US lawmakers to expand compute access, fund shared training data and evaluation frameworks, and avoid premature restrictions on open models. (AI News) So what: whether open weights stay freely deployable decides if a self-hosted model is a real fallback in your architecture or just a slide.
Canada set a date for explainable AI in finance. OSFI’s revised Guideline E-23 takes effect 1 May 2027 and requires federally regulated banks and insurers to inventory every model including vendor-supplied AI, classify it by autonomy and potential damage, and hold audit trails showing the reasoning behind a decision rather than a log of the action. Quebec’s AMF has a parallel guideline on the same date. So what: OSFI is not first. New York’s NYDFS has required audit trails for AI touching customer data since 2023, and Colorado’s rules take effect on 1 January 2027, four months before the Canadian ones. All three are worth watching: they are the closest working templates for what a regulator actually expects, ahead of the EU’s Annex III duties.
Summary
OpenAI shipped an enterprise agent platform you cannot buy without hiring its engineers. Behind it is a $4bn joint venture built on a consultancy acquisition. Anthropic made the same move with Blackstone in Ode, and Google, Meta and AWS built in-house equivalents. Five firms announced the same thing in ten weeks and none of them announced a new model alongside it. Two of them, valued purely as software companies, went further and moved the work off their own books, with OpenAI guaranteeing its investors a 17.5% annual return to do it, because services revenue is valued at a few times sales while software revenue is valued at a few dozen times. That is what the concession was worth to them. For a Polish enterprise the conclusion is a budgeting one: the money and the difficulty sit in integration, data and process, and the bill now carries a token cost that classic IT never had.
Stay balanced, Krzysztof
Krzysztof Goworek is founder of Quintant — AI advisory that gets enterprises from experiment to production value.