AI's so smart it needs teams to deploy it.
The rise of Forward Deployed Engineers, the AI image problem, and the end of cheap tokens.
By Victor Sowers — 15 years scaling B2B SaaS GTM
The Signal
The Shift
If AI is the future three things need to happen. First, we need to sell the vision of this future as something positive for people. Second we need to make sure that AI actually gets deployed in ways that solve important problems for businesses. Third, it has to stay affordable.
On selling a vision of AI unlocking some glorious future? That's simply not happening:
!Share who say AI is moving too fast, by age group — Economist/YouGov, May 2026
And the industry's response - or lack thereof - reveals an enormous amount of hubris. The consensus thinking seems to be that AI is inevitable. Get on or get out of the way.
But of course, it's not that simple.
For starters, AI's gains could reflect and exacerbate the socioeconomic disparities we're already grappling with. We've talked about this before when we covered how new cohorts of displaced white collar workers are entering the underemployed, and the junior-senior divide and historically bad job market for new grads. No wonder only 22% of Gen-Z have a favorable view on AI.
The other crazy part of this hubris? AI literally requires physical infrastructure and approvals for that building and energy consumption require political will and voter agreement. The polling data is damning.
The second thing that has to happen for AI to "be the future" is that it has to work for enterprises. And "work" is not as simple as giving everyone the latest frontier model and getting out of the way. Everything from security, to cost, to context has to be managed.
So who does the managing? What does it take to succeed? That's the main topic for this week both because it matters and because there's a lot of industry activity happening here.
The headline news is probably that OpenAI launched DeployCo this week. $4 billion of capital, 150 Forward Deployed Engineers from the Tomoro acquisition, majority-controlled by OpenAI, backed by Blackstone. The frontier model provider looked at enterprise adoption and decided selling model access wasn't enough. They need to be in the room deploying it to make it work and make sure those tokens get spent.
At the same time ServiceNow and Accenture launched an FDE program. EY started hiring FDEs as embedded senior engineers. Google Cloud posted 59 open FDE roles across four countries, with senior OTE listed at an eyewatering $700K.
Less pure consulting but in the same vein SAP built a workflow orchestration platform at Sapphire and invested in n8n at $5.2 billion to run it.
No wonder the Financial Times called the forward deployed AI engineer the most in-demand job in tech with interest up 800% since January 2025. Other than Palantir this role feels like it basically didn't exist in most org charts eighteen months ago.
So what's driving the momentum?
Aaron Levie, the CEO of Box, highlights how complex the work is: "Deploying agents is far more technical of a task than most people realize." At Box, Aaron is hiring AI automation engineers whose job is to build the harness and operating system surrounding the models. Think wiring systems to agents safely, designing human-in-loop workflows, and solving security and access questions.
Then there's the glaring lack of AI expertise basically anywhere.
Arvin Jain, the CEO at Glean put it well when he said that "most enterprise users are not builders. Engineers can often work around gaps in the product. Most other functions cannot be expected to." Nor should they really. As Leif Kothe put it: "For every AI operator or GTM engineer or Claude Code savant, there's a hundred great knowledge workers who have roughly zero interest in personal knowledge management systems, agent-building 'substrates,' ContextOps and intel pipelines." They want it to just work.
It's interesting. The forward deployed engineer motion is exploding *because* the "everyone's a builder" narrative is both right and wrong. The reality is most functions need someone else to do the deployment work for them.
More data and intel on the story below.
The Deployment Industry
Key takeaway: Two years of "just deploy AI" has produced a $4B deployment industry because demos work and rollouts stall at seventeen handoffs across four systems.
Two years of "just deploy AI" has led to a lot of broken pilots and early growing pains:
- BCG: only 5% of companies achieve AI value at scale. 60% see little or no material value.
- PwC CEO survey: 56% report no significant financial benefit.
- McKinsey: workflow redesign had the single largest effect on EBIT impact among 25 tested attributes. Not model selection. Not data quality. Not budget. Does this help McKinsey sell a bunch more overpriced consulting work? Yeah, probably. Doesn't mean it isn't true.
!The Deployment Gap — BCG, PwC, McKinsey failure stats
### Why Vendors Moved Downstream
The vendor moves make sense against that data. If your customers can't capture value from your model, or your model is viewed as commoditized against competitors, then your renewal is at risk regardless of how good the model is.
This isn't really new.
Palantir figured this out years ago. Their FDE model pairs "Delta engineers" (production-grade builders who navigate broken data environments) with "Echo strategists" who translate operational reality into technical requirements. They've since made AI FDE generally available and partnered with Accenture to deploy 2,000+ Palantir-skilled professionals.
### AI So Smart. Needs Babysitting.
But why exactly do we need so much help deploying a technology that is supposedly this smart?
According to an NBER paper "Chaining Tasks" the answer is that AI's impact comes from deploying it against contiguous chains of tasks, not isolated task automation. Fragment the chain and realized value drops. That's why a demo works and a rollout stalls. The demo chains three tasks in a clean environment. The rollout hits seventeen handoffs across four systems with different permission models and dozens of potential edge cases.
Of course, when we get into contiguous chains of tasks and non-deterministic tools things get messy. Context matters and so does clear thinking. Ryan Eade put it in well when he said: "Agents make it easier to move fast, but they also punish vague thinking faster than traditional workflows."
The result is often that individuals develop "local superpowers without a shared operating model." In other words, one person gets absurdly productive with Claude Code but their methods don't scale leading to only 32% of leaders reporting sustained enterprise-wide AI impact.
It'll be interesting to see how this all plays out. For now the gravy train shows no signs of slowing down with BCG reporting that AI spending is set to rise from 0.8% to 1.7% of revenue, with 94% of executives saying they'll keep investing even if returns don't materialize in 2026.
In the meantime, maybe go apply to that $700k OTE job? Just don't be surprised if it doesn't come with great benefits anymore because after all, someone has to pay for all those tokens.
The Subsidy Era Is Over
Key takeaway: Anthropic users consume roughly $8 in compute for every $1 of subscription revenue. The Uber-subsidy playbook is ending and consumption-based pricing is replacing flat rates across the industry.
AI has been absurdly cheap relative to its costs thus far. It's the Uber-subsidy, burn that VC money playbook all over again.
We talked about this previously when talking about how OpenAI lost $5 billion on $3.7 billion revenue, with some users generating $35K in compute on $200/month plans.
Cracks are starting to show in the good times.
On May 14, Anthropic split Claude billing into two pools. Pool 1 contains first-party usage (Claude chat, Claude Code) whereas Pool 2 governs third-party agent usage, metered at API rates. Monthly programmatic credits: $20 for Pro, $100 for Max 5x, $200 for Max 20x. In a continued run of crap messaging Anthropic tried to spin this as a positive for users. No one was buying it.
And to be fair, prices probably do need to come up.
One analysis found Anthropic users consume roughly $8 in compute for every $1 of subscription revenue. A team of 50 on Claude Pro costs $1,000/month. Equivalent API usage: $15,000 to $40,000. On the agent side a single autonomous agent running through OpenClaw could consume $1,000 to $5,000 per month in API-equivalent costs on a $200 subscription. Roughly 135,000 OpenClaw instances were affected. Even 300MW of datacenter and 220,000+ GPUs can't sustain unmetered agent demand.
The big question is, are we locked in? How quickly do OpenSource/local models or super complex routers to drive down inference costs evolve (a topic for another day)?
### This Isn't Just Anthropic
This isn't just about Anthropic or OpenAI when they're done trying to cynically poach users with two free months of Codex. The story is also about all the SaaS or AI-native products with token/consumption/outcome based pricing creeping in.
Kyle Poyar's State of B2B Monetization report (230 companies, April-May 2026) highlights how 37% of companies now use hybrid pricing combining subscriptions with AI consumption fees, up from 25% twelve months ago. Also fun-fact for the SaaS apocalypse debate: The median gross margin target for AI capabilities is around 50%, versus 70-80% for traditional SaaS. 75% of companies above $50M ARR modified their pricing in the past year.
What's interesting about this bucket of consumption fees is that it means 70% of AI spending comes from existing technology budgets. What happens when those budgets suddenly get bigger and less predictable? That's already happening. Goldman Sachs found companies overrunning AI budgets "by orders of magnitude" while GitHub is moving to usage-based billing on June 1.
Closing Thoughts
Key takeaway: The resolution isn't "make everyone a builder." It's the FDE — someone builds the invisible infrastructure, everyone else uses it.
"Everyone's a builder now" is the rallying cry of the AI operator community. In reality it's probably more of a bumper sticker for a subculture, not a description of the workforce (and I say that as someone who screams its value to everyone in my life). Most people just want it to work. Hence forward deployed engineers trying to make the infrastructure invisible.
Reading Corner
- Emily Kramer / MKT1: How to Research in Claude Code — A good tactical breakdown on using Claude Code for market research at scale.
- HBR: AI increases productivity but reduces motivation — AI increases productivity but simultaneously reduces motivation for non-AI tasks. The builder class is energized. The other 99% are somewhere between indifference and quiet dread.
- AI-generated papers flood scientific journals — The pattern is broader than academia. I'm seeing it flood Wikipedia edits, stall PRs inside engineering orgs, and in general it suggests that filtering plus AI/human visibility is the next category to build.
- AI isn't failing, your enterprise systems are — A CDO's blunt take: "AI will not fix bad data. If anything, it does the exact opposite."
- The Agent Operator: The New Emerging GTM Role — What that Claude Code power user in your org's future title will be.
Tool Watch
- Marketing Skills by Corey Haines — 38+ marketing skills with a foundational context layer every skill references. The CRO coverage is the standout: 6 skills split by conversion moment (page, signup, onboarding, form, popup, paywall). AI-SEO for LLM citation is forward-looking. (source)
Get the verdict every Wednesday.
The AI x GTM briefing for operators. Free forever.
One email per week. Unsubscribe anytime. No spam, ever.
More Issues
Everybody's building. What aren't they asking?
The AI wave is exciting and unsettling at the same time. Unemployment for recent grads hit 5.6% — worst in 37 years. Meanwhile Anthropic went from $9B to $19B ARR in 90 days. The build momentum is real. So are the costs nobody modeled: maintenance economics, token subsidies with expiration dates, and the apprenticeship disappearing underneath us.
Build momentum is real. So are the costs nobody's modeling.
The build wave is legitimate. Operators are shipping faster than vendors can sell. But three hidden trade-offs are catching teams off guard: maintenance economics, token subsidies with expiration dates, and model-version fragility that just got concrete.