← Back to Insights

Insight

Fifteen Deployments Deep

Ariel Agor
Updated
Experience is counted in deployments, and vendors get more of them. Their engineers get more repetitions.

Key facts

FactValueSourceAs of
Anthropic commitment to train Frontier Deployed Engineers$100 millionanthropic.com
AWS commitment to its forward deployed engineering initiative$1 billioncnbc.com
U.S. forward deployed engineer job postings, first nine months of 2026 (Lightcast)more than 7,300cnbc.com
Median advertised salary for forward deployed engineers (Lightcast)$195,456cnbc.com

Listen · Read by Ariel

0:00 / n/a

On October 2, 2026, Anthropic said it would spend $100 million to train 10,000 Frontier Deployed Engineers by the end of 2027. On October 15, CNBC reported that AWS had committed $1 billion to its own forward deployed engineering initiative, that OpenAI's new deployment business was backed by more than $4 billion in initial investment, and that Accenture and Google Cloud were building a 1,000-person forward deployed engineering workforce. On November 10, IBM and Microsoft launched Imagine Enterprise, an agent platform that comes with forward-deployed engineering support from both companies for qualifying engagements. Two days later Xero, the accounting software company in Wellington, announced Xforce, a team of forward-deployed engineers and partner consultants for accounting firms.

So the product now ships with a person attached, and that person is the same person you were planning to hire. If you are building an internal AI team this quarter, the forward deployed engineer wave matters more to your plan than any model release this autumn. The vendors are hiring, training and lending out the exact profile you need. They also hold an advantage you cannot match with salary. Their engineers get more repetitions.

In this trade, experience is counted in deployments, and a vendor's engineer collects them across dozens of clients while yours collects them one at a time. The companies that build lasting internal capability will be the ones that redesign their AI team to produce repetitions on purpose. Headcount comes second.

The market put a price on the person

The Lightcast figures in the CNBC piece are worth reading slowly. U.S. job postings for forward deployed engineers went from fewer than 200 in 2023 to about 1,200 in 2025, and then to more than 7,300 in the first nine months of 2026. The median advertised salary was $195,456, up 40% from 2023 and more than three times the median across all U.S. postings. "Demand for forward deployed engineers is rising fast," Elena Magrini of Lightcast told CNBC.

The list of who is posting those jobs matters more than the totals. The top recruiters in the first nine months of 2026 were Deloitte with 872 postings, Accenture with 688, Google with 499 and KPMG with 447. Those are advertised roles, not completed hires, as CNBC is careful to say. Even so, the shape is plain. The biggest buyers of this talent are consultancies and cloud vendors who plan to rent it back to you.

Jeff Christian of the executive search firm Christian & Timbers gave CNBC the reason for the hurry. "The Street is starting to become impatient," he said. Investors have paid for the data centers. They want to see the revenue, and the revenue depends on models turning into working systems inside companies that buy them. The forward deployed engineer is the vendor's answer to a capital markets problem. It happens to be a person you also need, for your own reasons, and the two of you are now in the same auction.

Ram Ramalingam of Accenture said the quiet part out loud. "The reality is that experienced FDEs cannot simply be recruited at the scale clients need," he told CNBC. If Accenture, with 688 open postings, cannot recruit them at scale, a mid-market manufacturer with a two-person AI budget line will not either. The supply has to be grown. Everyone in the auction has figured this out, which is why the money in October went to academies and residencies as much as to recruiters.

Experience comes in deployments

Two lines from the CNBC article explain why the vendor's engineer is so hard to match. Christian suggested that forward deployed engineers with 10 to 15 successful deployments could supervise teams of five or six. Val Henderson, the CEO of the AWS partner Caylent, put the same idea in a sentence I keep turning over. "There is no compression algorithm for experience," she said.

Put those two together and the unit of seniority in applied AI becomes visible. Years on the job count for little. Tenure in a particular stack counts for little too. What counts is the number of times a person has taken a model from a demo to something a business runs on, watched it break, and fixed it. Christian's number, 10 to 15, is a rough threshold for when someone has seen enough failure modes to lead others through them.

Now look at how that experience accumulates. An engineer at AWS or the OpenAI deployment business moves from client to client. Each engagement is a repetition, with a new data mess, a new security review and a new set of people who do not want their jobs changed. After two years that engineer has a mental catalogue of what goes wrong in claims processing, in loan origination, in warehouse slotting. They carry it to the next client.

An engineer you hire into a typical internal AI team gets assigned to one big initiative. It takes nine months, or a year, or it gets reorganized halfway. Suppose two years pass that way. They have one or two deployments to their name. They may be brilliant. They will still be junior by the only measure that matters here, and they will be junior next to the vendor engineer sitting across the table from them in your own conference room.

This is the structural problem with the default way companies build an internal AI team. They hire for talent and then starve that talent of repetitions. The vendor does the opposite. It hires for talent and then feeds it repetitions at the client's expense.

The half you already own

There is a second fact buried in the reporting, and it cuts in your favor. Steve Corfield of Anthropic described the gap this way: "A skill in short supply is an engineer who can understand how the business works." Henderson said something close. "The hardest thing to teach is the customer EQ, without question."

Magrini told CNBC that forward deployed engineering roles share roughly half their skills with ordinary software engineering. The other half is understanding a business well enough to know which workflow to change, who will resist, which data is lying, and what "done" looks like to the person who signs the invoice. A vendor engineer has to learn all of that about your company from scratch, every time, in a few weeks. Your own people already know it. They know that the returns desk in Ohio runs on a spreadsheet nobody admits exists. They know which vice president will quietly kill a project that makes her team look slow.

The technical half turns out to be teachable faster than most executives assume. Accenture told CNBC its forward deployed engineer training and onboarding takes 30 to 90 days, depending on background. Anthropic's residency includes 12 weeks of mentorship. In May 2026, The Asian Banker reported that Techcombank in Vietnam, which had about five dedicated AI engineers and around ten data specialists, had launched a two-month upskilling programme, and that engineers with Java or Python experience could move into AI roles within one to two months.

So the scarce profile splits into two halves with very different supply curves. Business understanding takes years and lives inside your walls. Model and agent engineering takes weeks to a few months and can be taught to people you already employ. The vendor has to buy the slow half every engagement. You start with it.

Most companies throw that advantage away. They post a job for an "AI engineer," hire a stranger at a price set by Deloitte's recruiting budget, and then spend six months teaching the stranger the business. That is the slow half, taught to a person who may leave for a vendor that pays more, once they finally understand your company well enough to be dangerous.

Building an internal AI team that produces repetitions

If experience is counted in deployments, then the design question for an internal AI team is how to produce more of them per engineer per year without producing junk. Here is how I would build it.

Make the unit of work small

The Techcombank account contains a number that should change how you scope projects. According to The Asian Banker, the time to build a fully functional agent at the bank dropped from about eight months to under a week once it had a reusable agent platform in place. The bank's business analyst agent started as a pilot in mid-June 2025 with about 20 business analysts and grew to around 30 active users across retail and corporate banking. The initiative was reported to have saved about 1,000 hours, estimated at 5.6 virtual full-time equivalents.

Those are modest numbers, and that is the point. A one-week agent with 30 users is a complete repetition. Someone scoped it, wired it to real data, got it through security, put it in front of users, saw what broke and fixed it. A twelve-month transformation program contains maybe one such loop, and the lessons from it arrive too late to help the next project.

Say your team has four engineers and each spends the year on a single large initiative. Say instead the same four ship one small deployment each per month, each one used by a real team. For illustration, that second team ends the year with 48 deployments and four people who are each near Christian's threshold of 10 to 15. The first team ends with four. The salary bill is identical.

Small units also make failure cheap enough to study. A team that ships a deployment every month can afford to kill a third of them and write down why. Those write-ups become the internal catalogue of failure modes that the vendor engineer carries in their head, except yours stays in the building.

Rotate across the business

The vendor engineer's catalogue comes from variety. They see the same pattern break in different contexts, and that is how they learn which parts of a solution are general and which are local. An internal team can get the same variety by moving its engineers across functions on purpose.

Assign each engineer a sequence of short engagements in different parts of the company, the way a vendor would, with finance one month and operations or customer support after that. Keep a small platform group stable underneath them, so the shared pieces get built once. The rotating engineers carry patterns between departments. Within a year, they know the business more broadly than any single department head does, which is a strange and useful position to put a 30-year-old engineer in.

Use the vendor's school and keep the graduate

The design of the Anthropic residency is the most interesting thing announced in October, and few executives have read the fine print. According to Anthropic's announcement, participation is by nomination only, and each nominee arrives with a named Claude project to lead on return. After a multi-day in-person program and a graded practical, the engineer spends 12 weeks leading a real Claude use case at their own organization, supported by Anthropic engineers and their cohort. The final credential is a Claude Frontier Deployed Engineer badge, with the first ones expected in early 2027. Anthropic says prior experience building AI agents is not required.

Read that structure from the buyer's side. A lab is offering to coach one of your employees through a real deployment inside your company, on your problem, with its own engineers on call. The first cohorts included engineers from Commonwealth Bank of Australia, Morgan Stanley and Novo Nordisk, sitting next to people from Accenture, Bain, Capgemini, Deloitte and McKinsey. Corfield said in the announcement, "No AI company has invested in developing that talent inside its customers and partners at this depth."

The consultancies understood immediately what that offer was worth. Their nominees come home with a badge they can bill. Your nominee comes home with a deployment, a cohort of peers at other firms, and a direct line to the lab's engineers. The repetition happens on your premises. The person who earned it is on your payroll.

Two cautions. The credential is vendor-branded, and it teaches one vendor's tools, so treat it as a source of repetitions and keep it out of your architecture decisions. And the nomination route runs through your Anthropic account team or partner manager, so if nobody at your company has asked, nobody at your company is in a cohort. That is a decision someone made by default.

Count deployments and stop counting hires

The metric most boards see for an internal AI team is headcount, sometimes with a budget line next to it. Headcount tells you how much you are spending. It says nothing about how much capability you are building.

Replace it with two numbers. The first is deployments shipped to real users per engineer per quarter. The second is the number of engineers who have crossed something like Christian's threshold, people with ten or more deployments who can now lead a small team of their own. That second number is your internal capacity to absorb AI work without renting it. When it grows, your dependence on vendor engineers shrinks. When it sits still, you are paying salaries for an expensive waiting room.

What the vendor's engineer takes home

None of this is an argument against using forward deployed engineers. On April 27, 2026, American Banker reported that Customers Bank, a regional bank in Pennsylvania, had signed a multiyear partnership under which OpenAI technical staff would work onsite to build custom AI capabilities. Customers Bank CEO Sam Sidhu said the deal brought frontier models, engineering expertise, and the ability to co-create a roadmap. Commonwealth Bank's Rodrigo Castillo said in the Anthropic announcement that "our teams have produced up to 3x more code changes in the past year." Vendor engineers can move fast, and on a deadline that can be worth a great deal.

The question is where the repetition lands. When an OpenAI engineer ships a deployment at your company, the code may belong to you. The experience belongs to the engineer and, through them, to OpenAI. Their fourteenth deployment teaches them something about your industry that they will bring to your competitor's fifteenth. The pattern library walks out the door at the end of the engagement, and it was partly built on your data, your edge cases and your mistakes.

The evidence on outcomes is also thinner than the spending suggests. CNBC noted that the gains it cited, including PennyMac completing some engineering tasks up to six times faster in a Caylent engagement, were company-reported and not independently verified returns. Anthropic told CNBC it has no outcome data yet for the residency and plans to track how quickly participants' projects reach production. Everyone is buying a bet. The difference is who keeps the winnings if it pays.

Xero's chief executive, Sukhinder Singh Cassidy, framed Xforce this way: "AI can help, but to scale it firms need hands-on support to build it into the way they work." She is right about the need. The open question for any company is whether that hands-on support stays a rental forever or becomes a bridge to a team of its own. A firm that uses vendor engineers without pairing them with internal people has chosen the rental, permanently, whether it meant to or not.

So pair them. Every vendor engineer on site should have an internal engineer attached, shipping alongside them and owning the deployment after handover. Write that into the statement of work. The vendor will agree, because it wants the logo and the expansion revenue. You get a repetition for your own person at the vendor's speed, which is the cheapest training you will ever buy.

The auction you are already in

Go back to the list of recruiters. Deloitte, Accenture, Google and KPMG are not hiring forward deployed engineers to work on Deloitte, Accenture, Google and KPMG. They are hiring them to work on you. Each one has done the arithmetic on repetitions and concluded that a pool of experienced deployers, moved quickly from client to client, is worth $195,456 a head and more. They are probably right.

You cannot win that auction on price, and you do not need to. You hold the slow half of the skill already, sitting in your operations staff, your analysts and the Java developers who maintain systems nobody else understands. What you lack is a structure that turns those people into deployers by giving them repetitions faster than a single big program ever will.

That structure is a design decision. A tool purchase will not make it for you, and a vendor partnership will make the opposite decision for you if you let it. It covers how projects are scoped and who rotates where, which training programs you nominate people into, how vendor engineers are paired, and what number the board sees each quarter. Get it wrong and in two years you will own a large AI budget, a few excellent engineers with one deployment each, and a standing invoice from someone else's team. Get it right and you will have people inside your company who have done this fifteen times and can teach the next ten.

Agor AI Advisory works with leadership teams on exactly this design: scoping the first wave of small deployments, building the rotation model, choosing which vendor programs feed your bench, and writing the pairing terms into contracts before the next engagement starts. We are not selling you engineers. We help you build the team that will make you need fewer of them. The vendors began their own hiring push in October, so your plan needs to be in place before your next budget cycle.

Sources

Want this kind of automation working for your business?

Agor AI designs and ships the systems these posts describe, scoped in weeks, not quarters.

Book a Free Strategy Call