AGENT OR SIMPLER WORKFLOW? — BUSINESS DECISION BRIEF
MLAI editorial aid, 2026-09-11
Not a security certification, customer case study, supplier quote or savings forecast. All figures and records are invented. Independent owner/technical review pending.
No AI system, approval gateway or customer workflow was evaluated. All business acceptance and approval demonstrations are NOT RUN.

SUPPLIED FICTIONAL CASE — M-104
M-104 — Invented incoming request: The storeroom light is flickering. Please arrange someone next week. Site: Example Works. Unit and on-site contact are not supplied.
ACCESS-A / v1 — Invented permitted access note: Reception handles access for Example Works.
ACCESS-B / v2 — Invented permitted access note: A nominated unit contact must arrange access. No contact name is recorded.
CURRENT STATUS — Unresolved source state: No effective dates or confirmed replacement relationship were supplied. A larger version label alone does not resolve the conflict. Actual source lookup, model output and safety assessment: NOT RUN.

THREE DESIGNS — SAME MONTHLY WORKLOAD
Baseline: 50 requests x 8 active minutes = 400 minutes. Case time includes checking, correction and fallback; extra oversight is separate, not the same review counted twice. A$60/hour values internal capacity, not avoidable payroll.

Fixed automation
M-104 handling: Required-field rules can flag the missing unit and route M-104 to a person. Do not guess which access note applies.
Error consequence: A weak rule may accept an incomplete request. A scheduler checks completeness; missed details and recovery costs are not measured.
Failure and maintenance owner: Office manager owns the rule list, checks rejected requests and fixes changed field requirements.
Planned monthly work: 50 x 6 = 300 case minutes + 60 extra oversight minutes. Potential capacity release: 40 minutes.
Setup: A$150 external cash + 3 internal hours (A$180) = A$330 combined economic setup.
Monthly cash: subscription A$10, usage A$0, external support A$10, extra labour A$0; total A$20.
Monthly cash change: −A$20. No positive cash payback.
Decision: Prepare an improved-form and manual-queue trial first; no live trial is approved.

AI-assisted workflow
M-104 handling: A fixed extraction-and-review sequence could draft the question from the request and both notes. Conflicting access information still needs a person.
Error consequence: A draft may omit a conflict or invent a booking. Office staff must inspect the original and every proposed commitment; sending remains unavailable.
Failure and maintenance owner: Office manager reviews drafts; the implementer maintains extraction versions, source references and the failure queue.
Planned monthly work: 50 x 5 = 250 case minutes + 100 extra oversight minutes. Potential capacity release: 50 minutes.
Setup: A$400 external cash + 6 internal hours (A$360) = A$760 combined economic setup.
Monthly cash: subscription A$20, usage A$10, external support A$20, extra labour A$0; total A$50.
Monthly cash change: −A$50. No positive cash payback.
Decision: Consider only if the improved form leaves a material free-text burden. The planned 50-minute release is not an observed gain.

Bounded agent
M-104 handling: Choosing which permitted note to read cannot establish which conflicting note is current. M-104 does not demonstrate a need for dynamic tool choice.
Error consequence: Wrong lookups, looping and cross-record access add failure paths. Read-only permissions, limits and review still require verification.
Failure and maintenance owner: Office manager handles unresolved work and can request a pause; the implementer owns access revocation, limit enforcement and incident diagnosis.
Planned monthly work: 50 x 4 = 200 case minutes + 180 extra oversight minutes. Potential capacity release: 20 minutes.
Setup: A$900 external cash + 12 internal hours (A$720) = A$1,620 combined economic setup.
Monthly cash: subscription A$20, usage A$30, external support A$40, extra labour A$0; total A$90.
Monthly cash change: −A$90. No positive cash payback.
Decision: No agent justified yet. Require evidence of useful variable decisions that the simpler workflow cannot handle, within tested boundaries.
Sensitivity: seven assisted case minutes rather than five gives 50 x 7 + 100 = 450 minutes, or −50 net minutes. It uses more time than the baseline; this is not an observed outcome.
Actual prices, support terms, error losses, revenue, taxes, financing and workload variability are not established. Obtain quotes and consistent cost/tax treatment; these exclusions prevent a complete ROI claim.

PROPOSED APPROVAL AND RECOVERY DEMONSTRATION — ALL NOT RUN
1. Prepare D-1
Proposed clarification to maintenance-contact@example.invalid: “Please confirm the unit and access contact. No visit time is confirmed.”
Expected: Internal draft only. The current design has no send capability. The address and wording are synthetic.
Inspect: Original M-104, both access-note versions, draft D-1 and recipient identity.
Actual: NOT RUN.
2. Approval A-1
In a separately proposed future send test, the office manager would approve only recipient maintenance-contact@example.invalid, action send-clarification and exact content D-1.
Expected: A-1 is not a general permission to book work or approve a different message. This step has not occurred.
Inspect: Approver authority, allowed action, immutable approved payload/version and approval status at execution.
Actual: NOT RUN.
3. Changed draft D-2
The test changes the text to “A technician is booked for 9 am.” and retries using A-1.
Expected: Reject the stale approval; no send. Booking commitments also remain prohibited, even if a model asks for approval.
Inspect: D-1 versus D-2, attempted approval reference, downstream rejection and absence of a send in the delivery system.
Actual: NOT RUN.
4. Recover the work
Expected recovery: retain M-104 and the failed attempt; discard D-2 and return the request to the office manager’s unresolved queue.
Expected: Clarify the source conflict manually before a new draft. An already sent message cannot be undone merely by restoring records.
Inspect: Queue item, responsible owner, preserved source and retry history. All four demonstration steps: NOT RUN.
Actual: NOT RUN.

COMPLETED FICTIONAL EXAMPLE — NOT A DEPLOYED SYSTEM
1. Workflow owner and business outcome:
Fictional maintenance-office manager; give a scheduler a complete administrative request without promising a price, attendance time or safety assessment. No client has approved this brief.
2. Current steps and non-AI alternative:
Staff read the request, inspect required fields and resolve omissions. Improve the form and retain a visible manual queue before adding AI.
3. Decision that cannot use a fixed rule, with an example:
M-104 has two access notes with no established current authority. Dynamic lookup cannot decide which is true. No evidence yet that an agent is necessary.
4. Permitted information and access scope:
Only the supplied invented M-104 and ACCESS-A/v1, ACCESS-B/v2 records for preparation. Live data permission and vendor retention remain unresolved; no customer credentials.
5. Read-only tools and allowed draft outputs:
If later built, selected authorised lookups and an internal proposed clarification with source references. These tools do not exist in this worksheet.
6. Actions requiring approval; approver and exact approved version:
Office manager checks each internal draft. In a separate future-send demonstration, A-1 would bind one recipient, action and D-1 payload; D-2 invalidates it. All steps NOT RUN; no actual approval exists.
7. Prohibited actions and technical enforcement:
No sends in the initial design; no booking, safety dispatch, pricing, invoice, deletion or cross-customer access. Require unavailable capabilities and downstream permission checks, not only prompt instructions.
8. Step, time and spending limits:
Call, elapsed-time and cash caps must be chosen with the owner and implementer before any test. Unchosen limits mean test not ready. No universal safe cap or approved budget is asserted.
9. Failure queue, responsible person and manual fallback:
One unresolved work item per original request ID. Office manager owns missing/conflicting details; implementer investigates access, duplicate and outage defects. Handoff acceptance and pause/revoke procedures still need demonstration.
10. Recovery limits and irreversible consequences:
Preserve M-104, discard D-2, retain failure/retry evidence and return to manual clarification. Restoring a database cannot undo an already delivered message.
11. Representative acceptance cases and evidence:
Missing unit; conflicting notes; attempted other-customer export; unavailable lookup; repeated M-104; changed D-2 after A-1. Expected/actual results, source/model/software versions, review time and downstream evidence must be recorded. Actual outcomes: NOT RUN.
12. Baseline, ongoing costs and continue/stop decision:
Invented 50 requests × 8 minutes = 400 monthly minutes. Fixed, assisted and agent plans release 40, 50 and 20 minutes after extra oversight. Monthly cash changes are −A$20, −A$50 and −A$90; no avoidable payroll or revenue is established. Prepare the improved-form comparison; no agent justified yet and no purchase/live trial approved.

BLANK BRIEF
Workflow owner and business outcome:
Current steps and non-AI alternative:
Decision that cannot use a fixed rule, with an example:
Permitted information and access scope:
Read-only tools and allowed draft outputs:
Actions requiring approval; approver and exact approved version:
Prohibited actions and technical enforcement:
Step, time and spending limits:
Failure queue, responsible person and manual fallback:
Recovery limits and irreversible consequences:
Representative acceptance cases and evidence:
Baseline, ongoing costs and continue/stop decision:

BLANK COMPARISON — COPY FOR FIXED, ASSISTED AND AGENT OPTIONS
Same task / representative period / cases / baseline time and rework:
Required output / permitted sources / variable decision / simpler alternative:
Case review, correction and fallback minutes / extra oversight:
Setup cash / internal time / subscriptions / usage / external support / extra labour:
Error consequences / access controls / failure owner / unknown costs and quotes:
Evidence of benefit over simpler option / reasons to stop / owner decision and date:

ACTUAL ACCEPTANCE RECORD — COPY PER CASE
Case and source IDs / versions / expected outcome / actual outcome or NOT RUN:
Model/software/configuration / reviewer / review time / costs / retained failures:
Action / recipient / exact content version / approval identity and current authority:
Downstream evidence / duplicate count / denied access / failed or irreversible action:
Pass, fail, unresolved or NOT RUN / manual queue owner / decision / retest:

BEFORE SHARING
Remove private customer records, credentials and confidential details. The file is not automatically attached to an enquiry. Acceptance, availability and savings are not guaranteed; confirm market eligibility with MLAI.
Optional client enquiry: https://mlai.au/mlai-studio/start-project
Full guide: https://mlai.au/articles/featured/what-is-an-agent-in-artificial-intelligence

SOURCE ROLES — checked 11 September 2026
https://www.anthropic.com/engineering/building-effective-agents — workflow/agent distinction, simpler designs and stopping conditions; dated architecture discussion, not current vendor ranking.
https://genai.owasp.org/llmrisk/llm062025-excessive-agency/ — limited capabilities, permissions, approval and downstream authorisation; not certification of the proposed design.
https://aws.amazon.com/what-is/ai-agents/ — memory, tools and learning/reflection terminology; not proof that a particular agent learns or improves.
Codex assisted this fictional record and local arithmetic/resource checks. Independent review, live permission, actual caps and delivery validation remain pending.