RESOURCES · THE FORWARD-DEPLOYED METHOD
Palantir's method, brought downstream.
Every company can now buy the same AI. What separates them is how it is put to work inside their own business. This page explains the method we use, where it comes from, and why it fits companies of 50 to 500 people.
8 October 2026 · Agent Forge Works is not affiliated with Palantir
WATCH · TWO MINUTES
One quoting desk, sorted step by step.
The tools are no longer the edge
A new model ships almost every week, and most companies now run the same assistants and coding tools. When everyone has the same intelligence, having it stops being an advantage.
The advantage moves to deployment: where AI goes to work, on which steps, and how it is checked. Handing everyone a tool and hoping is how a budget gets spent with nothing in the business running differently.
Where the method comes from
Palantir coined the term forward-deployed engineer. It paired one platform with engineers embedded at each client, who learned how the work actually ran and fitted the software to it. The value was in the fitting. We run the same method for companies that cannot hire a forward-deployed engineer of their own.
| Palantir | Agent Forge Works | |
|---|---|---|
| Clients | Governments and the largest enterprises | Owner-run Canadian companies of 50 to 500 people |
| Software | Its own platform | The ERP, CRM and inbox you already run. Nothing to buy. |
| Engineers | Embedded on site | Embedded on site. The person who scopes the work ships it. |
| Method | Learn the workflow, then fit the software | The same: assess, evaluate, deploy, repeat |
Two kinds of judgment, in one team
The work needs two kinds of judgment that rarely sit in one person.
Business judgment
Workflows, cost, incentives, risk, adoption, and who owns which step. Consultancies are strong here.
Technical judgment
Models, systems, data, reliability, evaluations and guardrails. Engineers are strong here.
A forward-deployed team holds both, and turns what it learns about the business into software that runs in it.
The process on paper is not the process
The procedure says an email arrives. On site, it comes from forty senders in no common format, the data sits in PDFs and forwarded threads, half the cases are exceptions, and the routing rules live in one person's head. An interview gets the job as people describe it. A day beside them gets the job as it runs.
A model only where judgment is needed
Every step in the mapped process is sorted into one of four, in this order:
- 01
Delete
Steps that exist only because of another step.
- 02
Code
Rules and integrations: cheap, fast, predictable.
- 03
Model
Reading and classifying what a rule cannot.
- 04
Person
Approvals, exceptions and anything risky.
Assess, evaluate, deploy, repeat
01
Assess
Map the work as it runs, sort every step, and measure the baseline before anything is built.
02
Evaluate
Test the design on your own history before it touches live work.
03
Deploy
Shadow mode, then a person approving exceptions, then live, with every action logged.
04
Repeat
Each workflow fixed shows the next bottleneck. The loop runs again.
Start with where the work piles up.
Thirty minutes. You leave knowing which workflows are worth handing to agents first, what the loops cost today, and what it would take.
30 minutes · no deck · no cost. Bring the work that piles up.