That is not a coincidence. It is a thesis about where the value in enterprise AI actually lives — and it is not where most executives have been looking.

The model was never the hard part

OpenAI’s new unit deploys what it calls Forward Deployed Engineers — practitioners embedded inside client organisations to redesign workflows around AI systems. The $4 billion in backing came from 19 investment firms including TPG, Bain Capital, Brookfield, and SoftBank. To seed the unit with immediate capacity, OpenAI simultaneously acquired Tomoro, an applied AI consulting firm, adding approximately 150 deployment specialists from day one.

Anthropic’s parallel move targets mid-sized businesses. Its applied AI engineers will work alongside the new company’s team to identify use cases, build custom systems, and support customers over time — a long-cycle engagement model, not a product sale.

Both labs are making the same bet: that getting AI to work inside a real organisation is a professional services problem, not a software problem. The implication is direct. The model — the thing both labs have spent years and billions developing — is increasingly a commodity input. The configuration, integration, change management, and governance work is where the actual value concentrates.

Traditional system integrators have held that position for decades. Accenture, Deloitte, and their peers built multi-billion-dollar practices reselling enterprise software and managing the implementation complexity on behalf of clients. What OpenAI and Anthropic are signalling is that they intend to own that layer themselves — or at least take a significant share of it.

The ROI data enterprises don’t want to publish

While the labs were announcing their services arms, Gartner published a finding that should sit uncomfortable alongside most AI strategy decks currently circulating in boardrooms.

Approximately 80% of enterprises piloting autonomous business capabilities have reduced headcount. The workforce reduction rates among companies reporting high AI ROI and those with modest or negative outcomes were nearly identical. The cuts are happening. The returns are not following.

The companies with the highest AI gains used the technology for what Gartner calls people amplification — making workers more productive rather than replacing them. A second Gartner prediction sharpens the stakes further: 50% of enterprises without a people-centric AI strategy will lose their top AI talent by 2027. The practitioners who know how to make these systems work in production will leave for organisations that give them room to grow alongside the technology.

Across the technology sector, more than 92,000 workers have been laid off in 2026 through mid-May. Meta and Microsoft cut 20,000 jobs in April. Coinbase cited AI workflow consolidation for a 14% workforce reduction. PayPal plans to cut 20% of staff over two to three years. In most cases, the public attribution points to AI-driven efficiency. The Gartner data suggests the efficiency gains are not materialising at the rate the announcements imply.

What the labs understand that most enterprises don’t

The labs are not launching professional services businesses because they have spare engineering capacity. They are doing it because they can see, at scale, where their own products fail inside enterprise environments — and they understand that failure is structural, not technical.

Enterprise AI deployment fails at the integration layer. It fails when automated workflows don’t connect to core systems of record. It fails when governance is absent and agents proliferate without oversight — a problem Microsoft’s Agent 365, which went generally available in May, is explicitly designed to address. It fails when the workforce reduction narrative runs ahead of the change management required to make the new model work.

The Forward Deployed Engineer model is a direct response to this. You embed practitioners with the client, you redesign the workflow from the inside, and you carry accountability for the outcome rather than handing over a licence and a user guide. It is expensive. It does not scale like software. But it works in a way that self-service AI deployment, for complex enterprise environments, demonstrably does not.

What this means for enterprise IT leadership

Three things follow from this for anyone managing AI strategy in a large organisation.

The buy vs. build decision has a new variable. If the frontier labs are prepared to embed engineers inside your organisation, the question is no longer just whether to buy a platform or build on an API. It is whether to engage the lab directly as an implementation partner — and what that means for data governance, model dependency, and negotiating leverage over time.

The headcount-reduction narrative deserves more scrutiny than it is getting. Gartner’s data is not arguing that AI cannot generate efficiency gains. It is arguing that cutting people to fund AI, without a clear account of where the productivity improvement is going, is not a strategy — it is cost accounting dressed as transformation. The organisations generating real AI ROI are treating it as a capability multiplier, not a replacement programme.

Governance is no longer optional. Colorado’s AI Act takes effect June 30, 2026 — the first major US state law imposing requirements on algorithmic employment decisions, including impact assessments and employee notification obligations. Illinois has been in effect since January. The regulatory surface area for enterprise AI is expanding in real time, and compliance is now a core operational requirement, not a future consideration.

The signal in the timing

It is worth sitting with the fact that OpenAI and Anthropic made the same strategic move in the same week. Both read the same enterprise feedback. Both concluded that the deployment and integration problem is large enough, and sticky enough, to justify building a services capability from scratch — or acquiring one outright.

That is a clear message about where enterprise AI value is concentrating. The model is a commodity input. The configuration, integration, and governance work is the moat. Most enterprise AI strategies are still organised around the former. The labs just told you, with $4 billion and a press release, that the latter is what matters.

If this is relevant to where your organisation is in the AI journey, I share observations from the enterprise technology frontline regularly on LinkedIn. Connections and follows welcome.

Based on what they announced this week, I don’t think they’re being ambitious enough.

What Microsoft Did

Copilot Cowork shifts Microsoft’s AI ambition from generative to agentic. Until now, Copilot has been good at producing — drafting emails, summarising documents, generating slides. Useful, but fundamentally assistive. You still had to do the work.

Copilot Cowork is different. You describe the outcome. The AI creates a plan and executes it across your M365 environment — autonomously, across multiple applications, with you monitoring rather than doing.

That’s a meaningful line to cross. Microsoft is calling it “delegation.”

It’s currently in the Frontier programme — enterprise beta, not GA. The rollout will be careful. That’s very much by design.

Microsoft’s carefulness isn’t irrational. Their core enterprise value proposition is the security boundary. They’re trusted because they move deliberately, because they’ve earned the right to sit inside the compliance perimeter of some of the world’s most regulated organisations. Stability over pace is a defensible position — as long as their enterprise customers can afford the same.

That assumption is getting harder to hold. Agentic AI is not evolving on a slow enterprise adoption curve. Results are real, timelines are compressing, and competitors — both inside and outside the Microsoft ecosystem — are moving faster. Copilot Cowork is a meaningful step, but it’s structurally a one-shot: you define the task, the agent executes it, and the engagement ends. There’s no iteration loop, no mechanism for the agent to reflect on outcomes and sharpen its approach over time. For enterprise customers who can still afford patience, that’s fine. The question is how long that describes most of them.

What ServiceNow Did

ServiceNow’s announcement is, in some ways, more ambitious in its framing. They’re not calling these “AI assistants” or “copilots.” They’re calling them AI specialists — entities that own a job, end to end, the way a new team member would.

The stress test they cite is compelling: when Moveworks joined ServiceNow, their IT helpdesk load doubled overnight. AI absorbed 90% of L1 tickets without missing an SLA. They didn’t just survive the surge — they productised it.

That’s a real result, not a demo. And the language around it is unusually direct for a vendor announcement. “AI that finally clocks in.” “They own a job, not just a task.”

ServiceNow’s advantage is structural. They already own the workflow layer across HR, IT, procurement and finance in a huge chunk of enterprise. They’re not trying to build a new beachhead — they’re deepening something they already have. Adding intelligence to a layer that already touches every employee in every transaction.

That’s a strong position. The risk is that it’s also a constraining one.

ServiceNow’s frame is: make existing work smarter. Automate the ticket. Accelerate the process. Remove the friction from the workflow that’s already there. It’s a compelling efficiency argument, and the ROI is measurable. But a specific process is a narrow goal — and a narrow goal leaves no room for the agent to evolve, to find paths that look different from what you defined at the start. Like Microsoft, it’s a conservative deployment model. And it carries the same underlying assumption: that their enterprise customers have the luxury of thinking incrementally.

The Gap Neither Is Talking About

Both announcements are fundamentally about automation — replacing human effort inside existing processes with AI effort. That’s valuable. It’s also the conservative version of this opportunity.

What neither addresses is what happens when you give an agent a goal instead of a process.

A useful illustration: Oliver Henry gave an AI agent called Larry a single brief — grow his app’s TikTok presence. Not a content calendar. Not a process. An outcome. Larry executed Henry’s initial ideas, then started generating strategies Henry hadn’t considered — and they outperformed anything he would have done himself. Half a million views in five days, converting into paying subscribers.

The key was that Larry wasn’t constrained to Henry’s methods. Given a goal rather than a process, it found better work to do.

Apply that to enterprise support functions. HR, finance, legal, procurement — these exist in their current form as adaptations to human cognitive limits. Those constraints are shifting. If you hand an agent a goal rather than a workflow, you create conditions for it to find paths that look nothing like the current process. That’s a different kind of value than automation delivers.

What Microsoft and ServiceNow announced this week doesn’t go there — at least not publicly. The goal-oriented frame, where the agent challenges the process rather than executing it, isn’t part of the story yet.

Why It Matters Now

Microsoft and ServiceNow are setting the visible frontier of what enterprise AI looks like. That frontier will shape how executives think, how budgets get allocated, and how IT strategies get written for the next 18 months.

Both vendors are, rationally, optimising for stability. They’re large, their customers are large, and the cost of getting it wrong inside a regulated enterprise is high. Conservative deployment — narrow goals, one-shot execution — is a reasonable position for an organisation with that much to protect.

The risk isn’t that they’re wrong about their own constraints. It’s that they’re building platforms that impose those constraints on their customers too, without leaving a path to evolve inside the walls. As agentic AI matures faster than these platforms can absorb — delivering broader goals, iterating on outcomes, compounding over time — enterprise customers who want to keep pace will find themselves with one option: go outside the ecosystem.

That’s the disconnect. Not a competitive threat to Microsoft or ServiceNow, at least not yet. A strategic trap for their customers — built with the best of intentions.

The smart move is to ask both questions simultaneously: how do we make today’s work more efficient, and what does work look like if we give agents room to evolve it? The second question is harder. It’s also the one your platform may not be designed to answer.

Anders L. Munck is an IT executive and founder of IT Leadership Services, working with organisations navigating complexity and transformation.

I love that type of games myself, but every time I saw one it ended up being some weird coin collection madness with cards, in-app purchases, and nothing like the advertised game.

So I decided to make one myself with nothing but the fun stuff. It starts with three games and a global high-score system that will allow me to do more when I get the time.

Before you read on, if you have a real interest in this area, please read the original papers and share observations. Any misinterpretations or omissions here are entirely my own, and we all get smarter faster by comparing notes.

Opportunities

In a recent working paper form Harvard Business School called “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality”, they compared the performance of 758 BCG consultants on 18 specific tasks to provide a baseline for comparing before and after. 

The results showed that consultants using GPT-4 finished 12.2 % more tasks, 25.1% more quickly, and increased quality by 40%.

In another study of 95 developers by Microsoft and MIT, the results showed a 55.8% increase in productivity, and a third online study of 444 professionals, indicate a 37% faster completion rate on common tasks like minutes and memos, while also improving quality.

One of the conclusions of the BCG study, which seems to have been confirmed elsewhere as well, was that one of the biggest impacts of generative AI, was as a skill leveler. It allows lower performing employees to catch up. This is of course very tempting to employees, and as illustrated in this X (formerly known as Twitter) poll, the majority of employees may already be using it without necessarily telling anyone.

In short, the potential is very real, most of your employees probably already know and use it, but they may still be too shy to tell anyone.

Risks

The short-term risks may be obvious to most. Things like intellectual property rights, content fact-checking, or compliance concerns cause most to have a rather conservative approach to formal deployments. But with the easy access to tools, and the promise of concrete personal performance improvements outlined above, it is hard to see how anyone can truly prevent it

The problem is that the value to the individual seems obvious, the limitations of generative AI are not obvious – for example, the ability to generate ideas is much better than calculating basic math – and although these gaps are rapidly evolving they are still large enough that any untrained employee can easily fall in.

In the BCG study, they call this the “jagged frontier” and to measure its impact, they split consultants into three groups, the first didn’t use GPT-4 at all, the second used it without training, and the third received training. While there were significant improvements in both groups using GPT-4, it was clear that variation in quality of work delivered by people without training was larger.

In this paper from Princeton University, they analyzed the impact of Generative AI on occupations, and conclude that among the most impacted will be the highly skilled and highly paid knowledge workers. Without training or policies in place, the risk is that the parts of your workforce most crucial to creating knowledge, will rapidly increase volume without necessarily increasing quality.

Longer term, this may limit opportunities. The content created today will become the source of the content we train our AI’s with tomorrow, and having to filter vast amounts of poor-quality AI-generated content could put a big dampener on things.

Business tactics

I’m sure everyone will have unique challenges and opportunities, and that the current rapid disruption will only make predictions harder. However, a few obvious tactics for dealing with the impact of generative AI stand out:

1. Allow multispeed adoption – Like with most disruptive technologies, businesses will have to adopt a multispeed adoption strategy. First, front-runners in low-risk areas will get to play, then a combination of larger groups or more risky areas will be added as experience grows.
2. Increase master-data scope – Employee-generated content is becoming a strategic resource, and each business will have to identify what content is critical to their core functions and start tightening the reigns to ensure quality at scale.
3. Let Workplace IT move in with HR – This has been true since Excel apps were the rage, but the need only increases every year: It really is time for Workplace IT, internal Communications and HR to become department buddies. The days of preventing employees from doing stupid things are long gone. To stay compliant you need to nudge, push, and train each individual to be able to take accountability for their own content, and to maximise opportunity you need to create a safe space for your front-runners to play.
4. Do nothing – I call this the “McKinsey option” as it is always in their slides. The go-to tactic for any CIO who feels they have enough on their plate, is to simply try to contain usage until the hype is over and the more certain benefits materialise. This is how most CIOs deal with technology innovation that doesn’t seem crucial to their core business.

Those are my five cents on this but let me know what I missed. As I said, we all get smarter faster when we share, and I’m sure I have a lot to learn.

By the way, when pasting the title of this article into Midjourney, this is what I got. Make of that what you will. Until then, see you out there.