Enterprise AI · Deployment infrastructure
The problem is not the model. It is not the budget. Every CIO I talk to has approved the investment, run the demos, and watched the board nod. What they have not done is ship. The gap between pilot and production is a deployment model problem — and the five pieces of infrastructure your Forward Deployed Engineer needs to close it are probably not in place yet.
Spend enough time in rooms with CIOs, Chief AI Officers, and the founders building AI-native companies, and a pattern becomes uncomfortable to ignore. The organizations with the most sophisticated AI strategies — the ones with dedicated AI councils, rigorous vendor evaluations, and boards that have seen the McKinsey slide on $4.4 trillion in potential value — are, in many cases, running below 20% production deployment rates. The pilots are elegant. The production workflows are not there.
The standard diagnosis is that this is a change management problem, or a data quality problem, or a trust problem. Those are real. But they are downstream of something more structural: the enterprise does not have the right person doing the right thing with the right infrastructure in place. What OpenAI recognized when it built its Forward Deployed Engineering function — and what the $1.5B joint venture between Anthropic, Blackstone, and Goldman Sachs named directly — is that the bottleneck is not model capability. It is the human operator embedded inside the business, with the tools to translate AI capability into auditable, production-grade enterprise workflows.
That person is the Forward Deployed Engineer. And before they can do anything that matters inside your environment, five things have to exist. Not as roadmap items. As working systems.
▪️<20% Enterprise AI production deployment rate
▪️$4.4T Annual value AI could add, per McKinsey
▪️$1.5B Anthropic–Blackstone–Goldman FDE deployment JV
The five infrastructure requirements
01 A workflow engine
An agent without a workflow engine is a demo. It completes a task in a controlled environment, once, with someone watching. A workflow engine is the spine that defines the sequence of steps, the conditions that trigger each one, and the branching logic when reality — as it always does — deviates from the plan. The FDE's core skill is not writing prompts. It is designing a workflow narrow enough that a language model can evaluate whether each step succeeded. The narrower the goal, the better the evaluation. That equation — narrow goal plus LLM evaluator plus FDE as governance layer — is what makes autonomous enterprise workflow possible. It did not exist before 2023. The workflow engine is where the goal definition lives, and goal definition is the highest-leverage work in the entire deployment.
02 An audit trail
Every CIO I talk to who operates in a regulated environment eventually asks the same question: if the AI made this decision, who is accountable for it? The audit trail is not a compliance checkbox. It is the mechanism that makes the FDE model defensible — to your General Counsel, to your CFO, to OSFI, to a board that needs to assess model risk without relying on vendor assurances. Every agent action logged. Every input documented. Every output traceable to the decision logic that produced it. RBC Borealis, now ranked third globally for AI maturity and targeting $700M–$1B in enterprise AI value by 2027, built 950 embedded engineers and data scientists partly because the governance posture of that model is one the bank's regulators can evaluate. The FDE who ships a workflow without an audit trail has not shipped a production system. They have created a liability that will surface at the worst possible moment.
03 Exception handling
Here is the scenario that breaks most AI deployments: a purchase order is written for 1,000 units. The bill of lading says 600 arrived. The ERP stalls — it was designed for a clean three-way match. The agent encounters a reality the workflow designer did not anticipate, and it either fails silently or fails loudly. Neither is acceptable in a system that is supposed to replace a human specialist who would have resolved this before lunch. Exception handling is the defined logic for what happens when the environment does not cooperate. It is also the asset that compounds across FDE deployments. The pattern library — every exception condition encountered, every resolution encoded — reduces the cost and time of every subsequent deployment. The traditional SI starts each engagement from scratch. The FDE operation that has built its exception library does not. That compounding is the structural economic advantage of the model.
04 Escalation paths
Not every exception should be resolved autonomously. Some decisions — a payment above a materiality threshold, an action in a regulated workflow, a situation where the agent's confidence is low — require a human. Escalation paths are the defined routes by which the agent stops, surfaces the situation to the right person in the right format, and waits for instruction. This is also the contractual and governance answer to the question your General Counsel will ask about a $50M payment that misfires: the escalation architecture is what creates an accountability structure that a regulator can evaluate. An FDE who designs escalation well is not just deploying software. They are building the organizational trust deposit that makes the next workflow faster to approve, and the one after that faster still. Without it, you have a single well-performing deployment that no one will authorize expanding into adjacent functions.
05 Integration into ERP, CRM, and ITSM systems
Enterprise value does not live in standalone agents. It lives in the systems of record that your organization has spent decades and hundreds of millions building. An agent that cannot read from and write back to your ERP is not doing enterprise work — it is performing a demonstration. Integration is also where the hardest deployment problems are concentrated: authentication models that were not designed for programmatic access, data schemas that do not match the workflow assumption, rate limits that surface only under production load. The FDE's advantage here is not technical novelty. It is operational proximity — being inside the environment long enough to know the sharp edges before they become production incidents. The Accenture-Microsoft joint FDE practice, announced this year, is targeting exactly this layer: the deployment gap between what the software can theoretically do and what it is actually doing inside a live enterprise environment.
"The FDE's returns compound. Each deployment produces a reusable agent template, a pattern library of exception resolutions, and an organizational trust deposit. The traditional SI's returns do not compound — because each engagement starts from scratch."From the TruNorth AI Leadership Summit session proposal, Vancouver 2026
These five elements are not a maturity model you progress through sequentially. They are simultaneous requirements for a production-grade agentic system. An agent with a workflow engine but no audit trail is not FDE-ready — it is a pilot that cannot be defended to a regulator. An agent with exception handling but no escalation paths has an accountability gap that will surface on the worst possible transaction. An agent with a clean integration layer but no workflow governance is a script with ambitions.The organizations getting this right — RBC Borealis internally, the FDE-native firms building on Palantir's original model, the frontier labs now deploying FDE as their primary enterprise go-to-market — are not distinguished by better models. They are distinguished by how quickly they can get all five in place and how well those assets compound across deployments.The question worth sitting with: which of these five does your organization not yet have working? The answer tells you exactly what to build next — and how far away production actually is.
FDE readiness scan — four diagnostic questions
Q1 Can you define the goal of your target workflow in one sentence?
Q2 Is someone with production coding ability sitting inside the relevant business unit — not in IT, not in a centre of excellence, but inside the function?
Q3 Do you have an audit infrastructure that can log and explain every agent decision to a regulator or a board?
Q4 Does your governance posture hold an AI system accountable — not just monitor it?
Four yes answers means you are FDE-ready today. Two or three means you know exactly what to build next. One or fewer means the gap between your AI slide and your AI deployment is wider than the board currently understands — and closing it starts with naming the infrastructure problem, not acquiring a better model.
The FDE is not defined by the model they use. They are defined by the infrastructure they build around it. The organizations that understand that distinction — and act on it before the category stabilizes — are the ones that will not be paying the platform tax when someone else builds the infrastructure layer first.
Adapted from a session proposal for the TruNorth AI Leadership Summit North America — Vancouver, September 2026. Written by Rajeev Kapoor, Chairman of the Board at Alpha.AI and CEO at NAP AI The full proposal covers five beats, three core theses, and five open questions the audience will not stop thinking about.