FDE Field Notes
Analysis for leaders deciding how to qualify, build, govern, measure, and transfer consequential AI work.
When remembered context can shape a future answer or action, it becomes production data. Govern agent memory with explicit scope, provenance, retention, correction, deletion, and acceptance tests.
Read field note →Reverting an agent version can restore configuration while leaving in-flight work and external effects behind. Production recovery needs separate controls for execution, business state, evidence, and resumption.
Read field note →Keeping agent activity in customer-controlled storage resolves one custody question. Production still requires named responsibility from signal detection through containment, investigation, and resumption.
Read field note →An agent evaluation is decision-ready only when the system under test cannot rewrite the baseline, success oracle, or retained evidence used to approve it.
Read field note →A connected-tool inventory shows where an agent can reach. Production requires an authority ledger that makes each consequential action attributable, conditional, time-bound, testable, and revocable.
Read field note →Buying AI creates access. Deployment begins when a named workflow has an owner, evidence standard, operating boundary, and path to adoption.
Read field note →OpenAI, Anthropic, and Google disclose different ways to allocate field delivery. Compare the responsibilities and handoffs, not the labels.
Read field note →Build, buy, and partner are not mutually exclusive bets. They are choices about who should own each layer of the first production workflow.
Read field note →OpenAI’s public FDE role connects discovery and production adoption to eval-driven product and model feedback. The return path is where field work can compound.
Read field note →Anthropic’s public FDE roles expose qualification, prioritization, partner routing, commercial coordination, and handoff as one delivery system.
Read field note →Google's public partner design shows how frontier delivery can distribute capacity without losing accountability—if context, evidence, and escalation have named owners.
Read field note →A full FDE queue can hide weak allocation. Govern capacity as a portfolio of qualified deployments with a workflow, evidence plan, owner, route, and stop condition.
Read field note →Field delivery can increase adoption and improve the product—or build a shadow product. A four-ledger review shows whether services are compounding or consuming software economics.
Read field note →Usage shows that an AI system was touched. A decision-ready evidence plan shows whether a bounded workflow changed enough to expand, hold, reshape, or stop.
Read field note →Traditional vendors distribute forward-deployed work across guidance, services, success, specialists, and partners. Buyers should map ownership and handoffs, not search for one title.
Read field note →Microsoft's explicit FastTrack scope can become a practical ownership map for the custom deployment work that remains.
Read field note →Salesforce's FDE Partner category is a real market signal, but buyers still need to inspect accountability, evidence, operation, and transfer.
Read field note →An installed base brings enterprise context and distribution, but also every unresolved data, authority, integration, and operating decision.
Read field note →A scalable FDE organization starts with explicit responsibilities and artifact handoffs, not an ever-rarer all-in-one hire.
Read field note →Field work compounds only when each custom asset becomes supported product, reusable delivery capability, a visible exception, or a documented stop.
Read field note →An enterprise agent is defined as much by what it must escalate as by what it can automate. Put authority, handoffs, and recovery into the specification before integration.
Read field note →Customer feedback can mean data, patterns, evals, requirements, or code. A durable field-to-product loop gives each one an owner and a disposition.
Read field note →The way out of pilot limbo is not a larger transformation promise. It is a consequential workflow with visible evidence, owners, dependencies, and a stop rule.
Read field note →A runbook written at the end cannot repair ownership decisions deferred at the beginning. Transfer must shape architecture, access, evidence, and delivery from intake.
Read field note →Senior judgment is necessary, but it does not make a repeatable practice. The practice needs explicit systems for qualification, service design, delivery control, evidence, and ecosystem transfer.
Read field note →