Early in 2027 a small group of engineers will receive a digital badge. It will say "Claude Frontier Deployed Engineer," and Anthropic will have issued it. Anthropic announced the program on October 2, 2026. It calls the program the Claude Frontier Academy, it puts $100 million behind it, and it plans to train 10,000 engineers by the end of 2027. The first cohorts come from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. Some of those firms are the people you pay to give you independent advice about which AI to buy.
I keep thinking about that list. For twenty years the deal with a systems integrator was simple. You paid them to know every vendor and to owe none of them. The engineer who walked into your building was supposed to be a translator between your mess and the market's products. In 2026 forward deployed AI engineering became the hottest job in enterprise software, and the people doing it are now being trained, credentialed and in some cases funded by the companies whose models they install. The translator went to school at one of the publishers.
This is a piece about what that does to your architecture, and what you should do about it before the badges arrive.
What forward deployed AI engineering actually decides
The title sounds like a field sales job with a laptop. It carries more weight than that. A forward deployed engineer sits inside your company, finds a workflow an agent can take over, and builds the thing that takes it over. Anthropic's announcement describes them as engineers "who can take Claude from an idea to a system in production." The residency it designed has engineers lead "a real Claude use case at their own organization" for 12 weeks, supported by Anthropic engineers.
Look at the decisions that sit inside those 12 weeks. Somebody picks the model. Somebody decides how your data gets chunked, cleaned and exposed to it. Somebody writes the tool definitions, so the agent knows which of your systems it may touch and how. Somebody builds the test set that tells you whether the agent is good enough to ship, and that same person decides what "good enough" means. Somebody chooses where the logs go and which cloud the whole thing runs on.
In most enterprises nobody on the executive team makes any of those decisions on purpose. They get made by the engineer in the room, in week three, because the demo is on Thursday. Each one is small. Together they are your AI architecture, and they compound. An agent built around one vendor's tool-calling conventions, evaluated with that vendor's harness, logging to that vendor's console, can be moved. Moving it costs a quarter.
Executives tend to think they buy a model and then hire people to install it. In practice, the installer picks the model and everything that grows around it. The purchase order follows the person.
Five schools, one bench
Accenture makes the pattern easy to see, because it has enrolled everywhere.
According to TechCrunch's reporting on September 8, 2026, Accenture this year launched forward deployed engineering programs with Microsoft in March, ServiceNow in May and SAP in June. In September it formed the Accenture Gemini Enterprise Business Group with Google Cloud, and Google said it would train up to 1,000 of Accenture's FDEs to build on Gemini Enterprise. Earlier in the year, the same report notes, Google committed $750 million to a partner program that embedded its FDEs at Capgemini, Cognizant and Deloitte. Then in October Accenture sent engineers to Anthropic's academy. Ram Ramalingam of Accenture, quoted in Anthropic's announcement: "Anthropic built the residency the same way we think about the role."
Read one way, that looks like healthy neutrality. Accenture can serve any client on any stack. Read from inside a single client engagement, it looks different. The integrator's bench is now partitioned into vendor-shaped cohorts. When you hire them for an agent project, you are staffed from one of those cohorts. The Gemini group gets the Gemini work. The person with the Claude badge gets the Claude work. The partner who sells you the engagement may be neutral. The engineer who shows up has a curriculum.
Deloitte appears twice in the record above, once in Google's embedded program and once in Anthropic's first cohort. Capgemini also appears twice. McKinsey and Bain, firms that sell judgment about strategy, are sending people to learn hands-on deployment from a single lab. None of that is scandalous. It is how training markets work. It does change what you are buying when you buy advice.
Why the labs pay for the classroom
The money explains the behavior.
On October 15, 2026, CNBC reported that AWS has committed $1 billion to its forward deployed engineering initiative and is starting with hundreds of engineers. Asa Kalavade, AWS's vice president of forward deployed engineering, gave the reason plainly: "We are investing $1 billion now because customers need this hands-on help turning their AI investments into results at speed." The same report says OpenAI launched a deployment business backed by more than $4 billion in initial investment.
The talent pool those billions are chasing is tiny. Lightcast data cited by CNBC counts fewer than 200 US FDE job postings in 2023, about 1,200 in 2025, and more than 7,300 in the first nine months of 2026, with a median advertised salary of $195,456, up 40% from 2023. In July, TechCrunch covered a study by the executive search firm Christian & Timbers that estimated about 2,000 US engineers have the mix of sector knowledge and applied AI experience to deliver enterprise returns consistently. The study's phrasing is the best line anyone has written on the subject: "Not 2,000 available," it reads, "2,000 total."
So the labs face a supply problem they cannot hire their way out of. AWS's answer is to send its own people. Anthropic's answer is cheaper and cleverer. It trains engineers who already work for customers and partners, on Anthropic's dime, using Anthropic's standards. Steve Corfield, Anthropic's global head of business development and partnerships, told CNBC: "The AI skills gap affects every customer and every partner, so it is a market problem."
He is right that it is a market problem. The lab that solves it also gets something valuable. Every engineer it trains carries its habits into an organization that pays that engineer's salary. Anthropic does not have to win your procurement process if the person writing your agent harness already thinks in its idioms.
Market share data shows why the labs care so much. Ramp's August figures, cited by TechCrunch, put Anthropic at 43.5% of enterprise AI spending among Ramp's US customers, OpenAI at 39.7% and Google at about 6%. Google disputes how representative that sample is, since Ramp's customers skew away from the large enterprises that sign big Google Cloud deals. Either way, the gap between those numbers is why Google needs Accenture's bench and why Anthropic wants McKinsey's. Model quality now moves in months. A trained engineer who stays inside a client for years moves share much more slowly, and more durably.
What a vendor-built agent looks like
On November 9, 2026, Northern Trust and Microsoft announced an AI agent for client onboarding on Northern Trust Asset Servicing's Front Office Solutions platform. Northern Trust's own team and Microsoft's Frontier Company Forward Deployed Engineering team built it together. It converts years of a new client's data and transaction history into the standard format Northern Trust's reporting runs on. Data preparation time fell 83%, from about two and a half weeks to three days, according to the announcement.
It is a good result, and I take it at face value. The detail worth studying is the stack. The agent runs on Microsoft Azure, uses Microsoft Foundry, and hands its work to analysts for review inside Microsoft Excel and SharePoint. The quality score of 96.5% came, per the release's footnote, from controlled testing on seven representative test cases using anonymized historical workbooks. Microsoft improved that score with what it calls an "eyes-off" approach, using only anonymized data so that client data stayed under Northern Trust's control.
Every one of those choices is reasonable. Northern Trust is a Microsoft shop, and building on what analysts already open every morning is smart design. My point is narrower. When the forward deployed team comes from the platform vendor, the platform is settled before the first workshop. The evaluation set (seven cases, in this instance) was built inside the partnership. If Northern Trust someday wants to test whether a different model does the job better, it will need a test harness that sits outside the partnership. It may already have one. Most companies would not.
Run the same analysis on any of the vendor programs above and the shape repeats. AWS says its FDEs work beside customer teams so those customers can eventually maintain and extend the systems themselves. That is a sincere goal, and it is also a description of a customer who inherits an AWS-native agent and the skills to run it. Anthropic's residents build "a real Claude use case." Google's Accenture group builds "on the Gemini Enterprise platform." The deliverable comes pre-labeled.
The evidence is still thin, which raises the stakes
You might expect all this spending to come with proof. It mostly does not, yet. CNBC's own reporting says available evidence does not yet confirm a financial return from FDE projects. The case studies it cites come from the vendors: an AWS-provided study showing some PennyMac engineering tasks done up to six times faster, and a Smarsh study estimating $5 million to $10 million in freed engineering and product capacity, which CNBC notes is neither an isolated FDE contribution nor cash savings. Corfield told CNBC it was too early for outcome data on the academy and that Anthropic plans to track how fast participants' projects reach production.
That gap matters for a specific reason. When results are hard to measure, whoever designs the measurement controls the story. If the vendor-trained engineer builds the evaluation, picks the baseline and writes the success report, then the one party with the clearest incentive to declare success is also grading the work. No bad faith is required for this to go wrong. Engineers trained on one stack will reach for that stack's tools to measure things, and those tools will make that stack look natural.
Suppose a mid-sized insurer runs two agent pilots in 2027, one staffed by a Gemini-cohort integrator team and one by an engineer fresh from the Claude residency. Suppose each team builds its own test set. The insurer will end up with two success stories that cannot be compared, and a CFO who has to pick a platform based on which team presented better. I have watched versions of that meeting. Nobody leaves knowing anything.
The independence you now have to build
Neutrality used to come bundled with the integrator invoice. You now have to build it yourself. That takes three moves, and none of them needs a large team.
Own the test before anyone owns the build
Write the evaluation set for every agent workflow before a forward deployed engineer starts. Pull real cases from your own history, including the ugly ones: the misfiled claim, the client spreadsheet with merged cells, the contract with a handwritten amendment. Decide in writing what accuracy you need and what a failure costs. Keep the harness in your own repository, runnable against any model through a thin interface.
Say your onboarding team pulls 200 historical cases and grades them by hand over two weeks. That set becomes the one durable asset of the whole program. Models will change three times before your contract renews. The test set keeps its value through every change and lets you ask a question no vendor-trained engineer will volunteer: what would the other model have scored?
Decide the model on paper first
Make the platform decision an explicit executive choice, recorded with its reasons, before the deployment team arrives. If you choose the vendor's stack because your data already lives there, good. Write that down, along with the conditions under which you would revisit it. The danger is a stack that nobody chose, assembled one sprint at a time by someone whose training made one answer feel obvious.
Ask every integrator which vendor program staffed the engineers on your account. Ask it in the statement of work. Accenture, Deloitte and Capgemini are all in several programs at once, so the answer will vary by team, and you are entitled to know which team you got. Ask also who owns the prompts, tool definitions and evaluation code at the end of the engagement. If the answer is vague, the answer is the vendor.
Grow engineers who hold two badges
Anthropic's academy works by nomination, and it trains your own employees. Use that. Nominate people. Then send a different engineer through Google's or Microsoft's program, or AWS's. The goal is an internal bench where at least two people can build production agents on two different stacks, and who argue with each other about which one to use.
Christian & Timbers says there are about 2,000 truly qualified people in the US. You will not hire your way into that pool at a $195,456 median when AWS is spending $1 billion to staff its own. You can, however, let the labs pay to train your people, as long as you train them on more than one lab's ideas. The labs have made that offer because they expect the training to pull you toward them. Take the training and keep the pull in check.
What the badge actually certifies
The vendors are doing something sensible. AI that sits in a pilot earns them nothing, and the shortage of people who can get it into production is real. Anthropic, AWS, Google, Microsoft and OpenAI are spending billions to fix that shortage, and enterprises will benefit. The Northern Trust agent cut more than two weeks out of a process that has frustrated asset servicing clients for years. That gain is real.
Still, the badge certifies skill on one product. It cannot certify judgment about which product you should have bought. Those were once the same job, done by the integrator you trusted. In 2026 the labs split them apart, funded the first half, and left the second half with you.
Architect this before the residents arrive
Buying another tool will not fix this. A tool cannot tell you whether the people building on it were the right people to choose it. What fixes it is architecture: an evaluation harness you own, a platform decision you made on purpose, a contract that names who trained your engineers, and an internal bench taught by more than one teacher. Companies that set this up in the next two quarters will get the benefit of every vendor's academy while staying free to leave any of them. Companies that skip it will discover in 2028 that their agent architecture was chosen by the curriculum of whichever residency their integrator attended in the autumn of 2026.
Agor AI Advisory takes no vendor's training money and holds no lab's badge. We build the evaluation sets, write the platform decision with you, and structure your integrator contracts so the forward deployed engineers you bring in build for your company first and their school second. The first badges ship in early 2027, so the time to set up your side of the room is now. Schedule a strategic consultation with us today.
Sources
- Anthropic, "Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gap," October 2, 2026
- CNBC, "AWS, Anthropic race to deploy army of specialized engineers key to proving AI spending will pay off," October 15, 2026
- Microsoft Source, "Northern Trust and Microsoft Frontier Company collaborate to transform client onboarding through AI," November 9, 2026
- TechCrunch, "Google Cloud races to catch up in the AI deployment wars with Accenture deal," September 8, 2026
- TechCrunch, "Forward-deployed engineers are the AI industry's latest talent obsession," July 30, 2026