Written for Architects, operations leaders and sponsors who are being asked to replace automation with agents, or to fund a multi-agent system that would complete the act.
What's in this playbook
This playbook is the operational document behind AsymiLink Meta's Agentic AI Systems consulting page. It is a method for deciding when a goal-directed agent is the right first case, and for designing that case with a human gate as architecture — a gate the model cannot bypass, not a prompt that asks the model to be careful. It is complementary to Agentic Workflows: Beyond Traditional RPA, already in this library, and is not a replacement for it. That white paper remains the mechanism test, the position on human involvement as architecture, the six illustrative patterns, and four refusals: no self-healing promise, no replacement of all reliable automation, not six real deployments, and not unsupervised customer-facing work without gates. This playbook will not reprint those six patterns as if they were six real deployments, and it will not publish a migration roadmap off reliable automation.
An agent earns its place where judgment is needed at run time. A reliable known path should stay. The honest default first case is a recommend-and-review agent on a named exception: a proposed next step, with evidence, for a named person who can accept, override or refuse, with the override recorded. An unsupervised multi-agent system that completes the act, or a program to "replace our automation with agents," is not a first case this playbook will recommend. A person sits on every irreversible act, every write, and every customer-facing message. Hosting is decided before the case is funded. Where data cannot leave, AsymiLink AI runs on your own hardware, with an optional air gap, and the gates travel with the workload.
How the tests and the method are meant to be used
The financial method is The Complete Guide to AI Automation ROI: a named process, a baseline agreed before work starts, and a calculation you can defend. This playbook will not invent an agentic-sector return or a self-healing saving, and it will not promise a twelve-week go-live. The 3.8x average first-year ROI across client engagements is an average, not a forecast or a commitment for any individual project. ROI is measured across client engagements against a baseline agreed before work starts. It is an average, not a forecast for any agentic project, and is not used as a sector benchmark. Built to SOC 2 standards is the exact wording we use; it is not a certification of the reader's system, and not a finding that an agent is safe to run unattended.
An illustrative mid-size operations organization, introduced in this playbook and labeled hypothetical, sits at 19, Developing, on the eight-dimension readiness model. Process, Infrastructure and Adoption are relative strengths; Data Foundation and Governance are constraints. Its honest first work is the mechanism test, the inventory, and a bounded recommend-and-review agent, with a next-step target of 24, still Developing — a next step, not a timeline, and not a promise. A companion workbook, asymilink-agentic-ai-playbook.xlsx, holds a family register, the fourteen tests, questions for counsel, a baseline sheet, and the illustrative organization.
Six families of agentic work
each with an honest first case, a case that is too wide, and a gate. Each item below is the shape of the map; the full wording, honest first cases and “too wide” warnings are in the PDF.
F1
A recommend-and-review agent on a named exception
F2
Intake with rules first and an agent on the residue
F3
An internal, read-only research assist
F4
A multi-step operations agent with an approval gate (later)
F5
A customer-facing inquiry with a closed action list and a hand-off (later)
F6
Hosting, decided before the case is funded
The PDF gives, for each family, an honest first case, a case that is too wide, and a defined gate. The summaries above are the shape of the map, not a substitute for that wording.
Fourteen design tests (excerpt)
Fourteen design tests are applied to every case. The answers will differ. The standard does not. The full wording is in the PDF and the companion workbook.
T1
including the mechanism test, a gate as architecture rather than a prompt, a permitted-action list, a named operator and a stop, and a refusal to promise self-healing.
A completed test sheet is not a certificate, not a compliance score, and not a finding that a system is safe or aligned with a named regime.
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.