Written for Architects, operations leaders and sponsors who are being asked to “replace RPA with agents,” or to put an agent on a process that already has a reliable path.
What's in this white paper
This 22-page white paper is a technical point of view: what an agent is and is not, a mechanism test already set out in the Workflow Automation Playbook, a firm position that human involvement is architecture rather than a prompt, and six illustrative architecture patterns, labeled as such. They are not case studies, and they are not a sample of “what leading organizations are doing.”
The library title around this paper has sometimes promised six real deployment architectures, a migration roadmap off RPA, and “adaptive, self-healing workflows.” This document does not contain those. It will not present six client deployments it cannot show, it will not publish a roadmap that moves every reliable workflow onto an agent, and it will not promise that a workflow will heal itself.
It is written for architects, operations leaders and sponsors who are being asked to “replace RPA with agents,” or to put an agent on a process that already has a reliable path. The financial method behind any case you then fund is the method in The Complete Guide to AI Automation ROI. The failure-analysis framework in this library is the companion for the patterns that appear when an agent is funded without a gate, an operator or a baseline.
The mechanism test, and the gate
An agent, in this paper, is a goal-directed mechanism with a bounded action set, not a synonym for “automation” and not a replacement for every reliable path. The same continuum as the Workflow Automation Playbook applies: a rule, a reliable workflow, a document process, a model, an agent, or a person — chosen, in writing, because a simpler mechanism cannot do the job to the required standard. “We have the platform” is not a reason.
Human involvement is architecture, not a prompt. A gate the model cannot bypass, a closed list of permitted actions, and a switch-off are designed into the system. A prompt that says “be careful” is not a gate. The paper will not describe an unsupervised high-consequence agent as a first case, and it will not recommend a program whose aim is to “replace all RPA with agents.”
After go-live, the same discipline as the rest of this library applies: an operator who is not only the builder, a log, a sample, a value review, a switch-off, and a date to look again. The companion workbook holds a mechanism test, a pattern register and a gates sheet. A completed workbook is not a certificate, and it is not a finding that a pattern has been deployed.
Six illustrative architecture patterns
The six patterns are designs, labeled illustrative. They are not six real deployments, not a migration roadmap, and not client cases. Each pattern has a counterpart the paper will not recommend as a first case.
P1
After-hours inquiry agent, with a CRM hand-off. Questions by phone, chat or email when the usual team is not there, or when volume spikes. A look-up, a closed set of inquiry types, and a hand-off.
P2
Document intake: rules first, an agent on the exceptions. Inbound documents classified and checked; an agent on what the rules cannot decide, with a person on the decision.
P3
An exception layer on existing automation. A reliable workflow already runs on the known work. Exceptions are the agent’s job, not a program to replace the workflow.
P4
An internal, read-only research assistant. An employee asks a question of a corpus the institution already holds, and receives a draft with sources. Not a customer-facing or write-back use.
P5
A multi-step operations agent, with an approval gate. Several steps — look-ups, a draft, a permitted change — where a named person approves anything that writes back.
P6
The same patterns, on your hardware, with an optional air gap. Not a seventh kind of work. If data cannot leave, P6 is decided before the case is funded. The offline guide is the companion.
What is not a first case: an unsupervised assistant that can change an account; an agent that “handles documents end-to-end” with no rule layer and no person; a program to replace all RPA; a research assistant allowed onto a customer channel; a goal-directed agent with an open action set and a prompt in place of a gate.
Human involvement as architecture
Three things the paper treats as design, not as a courtesy or a prompt.
A gate the model cannot bypass
A person on the action that matters — a send, a write-back, a decision that affects someone — in the system, not only in the instructions.
A closed permitted-action list
What the agent may do is written down. An open action set, with a prompt saying “be careful,” is not this design.
A switch-off
A way to stop the agent today, without a project, and a named operator who is not only the builder.
Get the guide and workbook
Enter your name, work email and company. The files appear on this page immediately. A backup download link is also emailed when email delivery is available.