Playbook: Agent or Not?
AI Agent Hype is here, leaders want movement, vendors want deployment, teams want to show progress, but what can Operators do to understand and benefit from the hype cycle.
Vieews operator layer
Agent or Not?
How to decide whether a workflow needs an agent, simpler automation, human-led AI assist, or nothing at all.
In 60 seconds
The agent hype is here: Leaders want movement, vendors want deploymentv teams want to show progress and a lot of people are being asked to use AI without being trained to decide what kind of AI intervention actually fits the work.
In this series, Operator means anyone being asked to make the work run: a business lead, ERP lead, service lead, project lead, process owner, product owner, IT partner, finance lead, team manager, trainer, analyst, or frontline support lead.
Most AI workflows do not need an agent, some need automation, some need human-led AI assist, some need documentation, data, or ownership fixed first. Most importantly, some workflows should be left alone.
Use this playbook to triage selected workflow candidates before you add another digital layer.
Start here
Your vendor probably called it an agent but that does not mean it should act like one.
If you have read the Digital Work Portfolio playbook, you have already started mapping what digital work exists in your operation: tools, reports, bots, spreadsheets, workflows, training hubs, local workarounds, and the other digital pieces people rely on to get work done.
That first playbook helps you ask: keep, fix, or stop?
This playbook asks the next question: which selected workflows deserve AI and which ones deserve an agent?
If you are arriving here first because someone handed you an AI mandate, a Copilot/Agent rollout, a vendor agent pitch, or a pilot list with no clear decision logic, you are still in the right place. Use this playbook now, then go back to the Digital Work Portfolio if you need to see the wider estate.
Why this exists now
Enterprise agent tooling is getting easier to buy, easier to connect, and easier to deploy. Ease is useful, but it also creates a new sprawl risk within your digitization portfolio. The easier it gets to launch agents, the easier it gets to create agent sprawl.
Recent market moves point in the same direction: frontier AI providers and major platforms are building enterprise deployment motions, partner networks, services layers, and control planes. Anthropic announced a new enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs. Gartner has warned that more than 40% of agentic AI projects may be cancelled by the end of 2027 because of cost, unclear value, or weak controls. The signal is clear: the deployment layer is accelerating, but the operator layer is still thin.
This playbook serves as a decision layer support before the deployment layer. It does not ask, "Can the platform build this?" It asks, "Should this workflow become an agent at all?"
From Digital Work Portfolio to Agent or Not?
Digital Work Portfolio gives you visibility. Agent or Not? gives you triage.
Do not take every item in your digital inventory and turn it into an AI idea. That is how existing tool sprawl/chaos becomes agent sprawl/chaos.
| Portfolio call | How Agent or Not? applies | Likely lane |
|---|---|---|
| Keep | Run triage only if AI is proposed, already exists, or could clearly improve work that matters. | Often Assist or Automation. Agent only if complexity earns its keep. |
| Fix | Use triage to avoid adding AI before the real blocker is clear: process, data, documentation, owner, control, or value. | Often Leave Alone until the fix is done. Assist may help expose gaps. |
| Stop | Do not agentify it. Confirm it should not become the next AI pilot by accident. | Usually Leave Alone / Kill. |
| Unknown | Go back and finish basic visibility first. | No verdict yet. |
The practical handoff is simple: choose 5 to 10 pressure points from the portfolio, not the whole estate.
The four lanes
Agent
A system that can move a multi-step workflow forward across tools with some autonomy. Example: a service entitlement workflow that checks contract status, asset history, inventory, field availability, and routes an approval-ready recommendation.
Automation
A deterministic workflow with known rules and predictable steps. Example: standard user access requests, fixed approval routing, invoice reminders, or RPA-style data moves.
Assist
A human stays accountable, but AI reduces effort. Example: a SharePoint-style training assistant that answers questions from approved ERP training content without updating the ERP.
Leave Alone
No AI change yet. Example: a duplicated local tracker with no owner, no baseline, and unclear value. Fix, merge, or stop the workflow first.
Signal corner
Before asking if AI removes jobs, map the work.
Most people hear "agent" and jump straight to the jobs question. That is understandable. But the first operator move is not to argue over job titles. It is to ask what part of the workflow changes.
| Lane | Human work signal |
|---|---|
| Automation | Repetitive steps may be removed or routed. People may shift toward exception handling. |
| Assist | The person stays accountable, but AI reduces searching, drafting, summarising, comparing, or preparation work. |
| Agent | The workflow may need redesign because the system can move work forward across steps or tools. |
| Leave Alone | No AI change yet. The better move may be cleanup, documentation, training, ownership clarity, or stopping a weak idea. |
This is not a layoff model. It is an early work-design signal. The jobs question is real, but job panic starts too late. Map the workflow first.
Read the linked Signal: Before asking if AI removes jobs, map the work.
The seven triage tests
1. Is this a workflow or just a task?
If it is one task, it probably does not need an agent. Summarising a document is Assist. A multi-step issue resolution process may be a workflow.
2. Does the work involve ambiguity that rules cannot handle well?
If the rules are stable and the decision tree is known, use Automation. If exceptions and context drive the work, Agent or Assist may be worth considering.
3. Does it need to act across systems?
If it mostly prepares information for a human, it is probably Assist. If it needs to update, route, trigger, or resolve across tools, Agent may be relevant. Connection is technical capability. Permission is an operating decision.
4. Would failure be reversible and catchable?
If a wrong action could affect customers, payments, safety, compliance, production, warranty, or critical operations, autonomy needs tighter controls.
5. Is there a named owner?
No owner, no agent. Autonomy needs accountability.
6. Is there a measurable outcome?
No baseline means no credible scale decision. Track cycle time, cost per case, first-time fix, quote turnaround, backlog, rework, claim leakage, or another meaningful metric.
7. Is the process stable enough?
Broken workflows do not become better because AI is added. They usually become harder to see.
Verdicts
| Verdict | Use when | Decision options |
|---|---|---|
| Agent | Multi-step, ambiguous, valuable, tool-connected, measurable, owned, safe enough to test with controls. | Scale / Fix / Pause / Kill |
| Automation | Repeatable, rules-based, structured, predictable, stable, not judgment-heavy. | Scale / Fix / Pause / Kill |
| Human + AI Assist | Human judgment stays in control; AI supports drafting, summarising, comparing, extracting, or preparing. | Scale / Fix / Pause / Kill |
| Leave Alone | Low-value, duplicated, unstable, unowned, risky, immature, or missing a baseline. | Fix later / Stop / Merge / Document |
Common failure modes
Agent washing
A chatbot, assistant, RPA, or workflow automation gets relabelled as agentic.
Pilot theatre
The pilot creates meetings and demos, but no operating decision changes.
Broken-process automation
A team tries to agentify a workflow nobody has cleaned up.
Ownership fog
Everyone uses it. Nobody owns the outcome, exceptions, or stop/scale decision.
Permissions first
The team connects the agent to tools before deciding what it should be allowed to do.
No kill criteria
The pilot does not scale, but nobody stops it.
The first move
Start with 5 to 10 live AI ideas, pilots, or workflows. Take them from your Digital Work Portfolio if you have one. If not, grab the most visible AI ideas currently floating around your team.
| Field | What to capture |
|---|---|
| Workflow name | What work is being changed? |
| Owner | Who owns the outcome? |
| Current state | Idea / pilot / live |
| Workflow or task? | Is this real workflow or just one task? |
| Human work impact | No change / less manual effort / exception shift / human approval / role redesign likely / training needed |
| Verdict | Agent / Automation / Assist / Leave Alone |
| Decision | Scale / Fix / Pause / Kill |
Then hold a 45-minute review, not to admire the list but to make decisions.
What good looks like
A good triage session does not produce 20 more ideas, it produces a shorter, sharper portfolio.
You should leave with a few use cases that may deserve agents, a few that should be automation, a few that should stay human-led with AI assist, and a few that should be paused or killed.
The point is not more AI activity. The point is better AI decisions.
What comes next
This playbook answers: should this be an agent at all?
It does not answer: is this ready to go live? That comes next in the Readiness Gate, where the questions get sharper: permissions, source systems, human approval, fallback paths, logging, monitoring, audit, support ownership, and rollback.
Related pieces
Free member downloads
Get the editable Agent or Not? pack
Use the working files to run 5 to 10 AI ideas through the triage in one session. Keep the public insight open. Gate the practical implementation assets.
