AWS is putting $1 billion behind a new Forward Deployed Engineering organisation that will send AI engineers directly into customer teams. For a company selling increasingly autonomous software as a way to compress months of work into days, the decision to spend that much money putting highly skilled humans closer to the problem says quite a lot about where enterprise AI is still getting stuck.
The AWS Forward Deployed Engineering programme will eventually involve thousands of experts working alongside customers to build and deploy agentic AI systems using their own data, governance and business processes. AWS says these engagements can reduce deployment timelines from months to days, while leaving customers capable of operating and extending what has been built after its engineers leave. There’s an obvious commercial logic to this: AWS wants more companies to move from experimenting with generative AI to running production workloads on AWS, and helping customers through the difficult stretch between buying access to AI technology and actually changing a business process is one way to get them there.
The investment also exposes a less convenient reality. Models have become far more capable, cloud providers have spent the past few years building platforms around them, and businesses have been encouraged to find AI use cases almost everywhere. Yet AWS is now committing $1 billion to engineers who can sit inside those businesses and help make the pieces work together. Better models haven’t removed the deployment problem; they’ve made the gap between what the technology can do in theory and what organisations can make it do reliably much harder to ignore.
AWS says customers have moved beyond asking what AI can do and now want to make it part of how they operate. Its own description of the FDE model suggests that doing so involves far more than giving employees access to a chatbot or adding an agent to an existing workflow. AWS engineers will work with customers’ business, engineering and security teams, connecting AI systems to organisational data while dealing with governance, internal processes and the realities of how work actually gets done. Those are precisely the places where ambitious AI strategies tend to become difficult, because organisations are usually messier than product demonstrations allow for.
Enterprise AI is becoming an implementation problem
None of this makes Forward Deployed Engineering a new idea. Palantir has used forward deployed engineers for years, with technical teams working closely with customers to understand problems, configure software and build applications around actual operating environments. Palantir has described the model as a way of keeping engineers close enough to customers to understand what needs to be built and feeding those lessons back into the product.
What AWS is doing differently is taking a model associated with highly involved enterprise deployments and putting the weight of a hyperscaler behind it. That tells us something useful about where the AI market has reached because enterprise software has always had an implementation layer. Consultants, systems integrators, solution architects and internal IT teams exist because a product that works beautifully in a demonstration can become considerably less elegant when it encounters a 20-year-old database, an approval process nobody has documented properly, three departments that describe the same customer differently and a regulator that would quite reasonably like to know where the data is going.
AI adds another set of complications because the system may now be expected to interpret information and take actions rather than simply move data from one predetermined place to another. AWS’s response is unusually hands-on for a cloud provider: its engineers aren’t supposed to arrive, write a strategy document and disappear. According to the company, customer engineers will move from observing the work to building alongside AWS and eventually operating the systems themselves. AWS says engagements will leave behind architectural documentation, runbooks, knowledge graphs and trained internal staff, while a semantic layer deployed inside the customer’s AWS account connects enterprise data and gives agents governed information to reason over.
That promise of self-sufficiency deserves some scrutiny because there’s a natural tension between teaching a customer to need your engineers less and building a much deeper relationship with that customer’s infrastructure, data and workflows. The AI system may eventually be operated by the customer, but if it has been designed around AWS infrastructure and services, independence from the engineering team doesn’t necessarily mean independence from AWS. The customer may gain capability while AWS gains a more deeply embedded workload, and both of those things can be true at the same time.
“Months to days” needs more evidence
AWS says Forward Deployed Engineering can compress AI deployments from months into days and points to work with organisations including the Allen Institute, Cox Automotive, the NBA, the NFL, Ricoh and Southwest Airlines. The NFL example is the most developed in AWS’s announcement: NFL CIO Gary Brantley says AWS engineers worked alongside its team to bring products including NFL Fantasy AI and NFL IQ into production within weeks, with measurable engagement from fans and broadcasters. AWS also cites earlier engineering work with BMW, Jabil and Lyft, including a claim that its work with Lyft helped resolve driver-support issues 87% faster.
Those examples show what AWS says the approach can achieve, but they don’t yet establish that months-long enterprise AI projects can routinely be reduced to days. The announcement doesn’t provide enough detail about project complexity, baselines, costs or deployments that didn’t work as expected to make the comparison especially meaningful. A three-day deployment may be impressive, but without knowing what was deployed, what systems it had to connect to and how much preparation had already happened, the number can only tell us so much.
Speed is also an odd thing to emphasise too heavily for AI systems that may end up touching consequential business processes. AWS talks about human oversight, security, encryption and keeping customer data inside its governance framework, all of which are necessary, but an organisation can get a system into production quickly and still spend months finding out whether people trust it, whether it produces the intended economic benefit and whether it behaves sensibly around situations nobody anticipated during implementation.
The customers AWS is targeting make that tension clearer. The company says FDE is aimed particularly at organisations moving beyond experimentation into production, including regulated industries, financial services and government, where governance becomes harder rather than less important once an AI system starts doing useful work. The quickest route into production isn’t always the best measure of whether an AI deployment was handled well.
South Africa’s problem isn’t access to a chatbot
There’s a useful South African dimension to this because local AI adoption statistics can make the market look further along than many organisations probably are operationally. South Africa may be ahead of much of the continent in AI usage, but AI adoption still trails the country’s mch broader internet usage, which is one reminder that access and meaningful integration aren’t the same thing.”
A company can have thousands of employees using ChatGPT, Copilot, Gemini or Amazon Q and still have legacy systems that don’t speak to one another, poor data quality, scarce specialist skills and governance processes written for software that behaves predictably. Using AI and rebuilding a business process around AI are different thresholds, and Forward Deployed Engineering is effectively AWS betting that the distance between them has become commercially important enough to justify $1 billion and thousands of engineers.
In South Africa, that distance can be widened further by skills constraints, infrastructure differences, uneven digital maturity and the cost of integrating systems that may have been built at different times for different purposes. Whether South African customers will benefit directly from the investment is less clear because AWS’s announcement doesn’t provide country-level deployment plans, pricing or details about how many FDE engineers will be available in individual markets, so there isn’t enough in the supplied material to turn this into a specific local rollout story.
The South African relevance lies more in the problem AWS is trying to solve: many organisations don’t lack access to AI tools so much as the engineering capacity, usable data and internal coordination needed to make those tools part of how the business actually runs.
The model also complicates one of the simpler narratives surrounding enterprise AI, which is that increasingly capable agents will steadily reduce the amount of specialist human expertise required to build software and change organisations. AWS is certainly using agents inside the development process. Its new AI-Driven Development Lifecycle is supposed to use agents throughout software development while engineers verify and guide their work. That follows the direction AWS has already been taking with frontier agents designed to work with greater autonomy across software development and operations.
At the same time, AWS is investing heavily in putting more engineers closer to customers, not fewer. Better agents may eventually make more of this implementation work repeatable, and AWS says each engagement should contribute knowledge that can make subsequent projects faster. For now, the company’s $1 billion bet suggests that enterprise AI still depends heavily on understanding companies in all their messy specificity. The useful test of Forward Deployed Engineering will be whether customers are genuinely more capable once the engineers leave, and whether AWS can make such a labour-intensive model scale without quietly replacing one form of dependence with another.