AI Operations Assessment: A Workflow Scoring Checklist
The wrong first automation can look excellent in a spreadsheet. High volume, expensive handling, obvious repetition—and a decision that should never leave a person's hands. My AI operations assessment checklist starts by separating those questions rather than blending them into one attractive score.
TL;DR: Inventory the work, establish a baseline, and score value, feasibility, and exposure separately. Remove unnecessary work before automating it. Use the worksheet below to choose a bounded pilot with a named owner, a human-review gate, and a measurable outcome—not a tool shopping list.
What to gather before an assessment
I start with the work, not a vendor demo. Gather a representative period of workflow data, the instructions people actually use, and a few completed cases—including exceptions. Ask the people doing the work where the documented process differs from reality. A missing handoff can make an apparently simple automation impossible to evaluate.
Record the measurement period and source beside each number. Handling time means active labor; cycle time includes waiting. Keep them separate. Finance should supply a consistent loaded hourly labor rate, including the costs it normally allocates to labor. Do not quietly substitute salary for loaded cost halfway through the comparison.
Use de-identified examples where possible. Establish who may access the source data and which systems may process it before uploading anything. If access or baseline quality is unknown, mark it unknown. That is a readiness gap, not permission to assume the best case.
Build a workflow inventory
Describe one workflow per row: its trigger, inputs, action, output, and exception path. 'Customer support' is too broad; 'draft a shipping-status response using an authenticated order lookup' is testable. Name the owner who can approve process changes, not just the person who answers questions.
The worksheet keeps economic inputs beside operational constraints. Copy it into a spreadsheet, add your sources, and replace the illustrative rows. Rework should be measured on completed cases, not guessed from complaints. An error caught before sending and an error delivered to a customer need different severity labels.
Copyable assessment worksheet
All three populated workflows below are illustrative, not descriptions of clients. Monthly figures assume the same measurement period. Rates are modeling inputs, not SPS prices.
| Workflow | Volume/month | Handling min/case | Loaded labor USD/hour | Errors/rework baseline | Systems | Data sensitivity | Owner | Baseline | Human-review gate | Success metric | Value 1–5 | Feasibility 1–5 | Exposure 1–5 | Decision | Baseline window | Data source | Cycle time | Error severity | Evidence links | Missing inputs |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Illustrative: shipping-status reply drafts | 1200 | 5 | 36 | 60 reopened cases/month (5%) | Helpdesk + order lookup | Customer contact and order data | Support lead | 100 active hours/month; log end-to-end resolution time | Support agent verifies authenticated lookup and approves every reply; missing/disputed status escalates | Active min/completed case with reopen rate no worse than baseline | 3 | 4 | 2 | Automate drafts only; shadow test first | ||||||
| Illustrative: duplicate weekly report reconciliation | 400 | 8 | 42 | 20 corrected entries/month (5%) | CRM export + spreadsheet | Internal commercial reporting | Operations manager | 53.3 active hours/month; 20 corrections | Report owner confirms required fields and approves removal of duplicate step | Hours removed while required report completeness is maintained | 2 | 5 | 1 | Eliminate duplicate step before considering AI. | ||||||
| Illustrative: credit eligibility decisions | 600 | 15 | 48 | 12 reviewed corrections/month (2%) | Application system + policy records | Sensitive financial and identity data | Risk lead | 150 active hours/month; 12 corrections | Authorized reviewer retains every eligibility decision; no autonomous approval or denial | Reviewer minutes/case plus decision-quality checks set by risk owner | 5 | 2 | 5 | Retain human judgment; assess bounded assistance separately. | ||||||
| Your workflow |
Select and copy worksheet text
For each row, add the baseline window, data source, cycle time, severity of errors, evidence links, and any missing inputs before making a decision. These are illustrative starting scores; your organization must calibrate the scale.
Score value, feasibility, and exposure separately
For this AI operations assessment checklist, I use three ordinal scores. They organize discussion; they do not predict ROI.
- Value: 1 means little avoidable effort or delay; 3 means meaningful, measured recurring work; 5 means a major operational burden supported by volume, time, and quality evidence. Use 2 and 4 for intermediate cases. Calibrate those anchors against your own workflow inventory before scoring.
- Feasibility: 1 means unreliable inputs or unclear rules; 3 means usable inputs with material integration or exception work; 5 means stable rules, accessible data, and a reversible test. Intermediate cases score 2 or 4.
- Exposure: 1 means limited, reversible internal impact; 3 means consequential customer or business impact requiring review; 5 means sensitive, high-stakes decisions with serious harm potential. Intermediate cases score 2 or 4. Legal or policy restrictions apply regardless of score.
Do not sum these into a leaderboard. First exclude prohibited actions and workflows without a workable review path. Then compare value and feasibility among eligible candidates. Review load can erase an apparent advantage; record it explicitly.
In the examples, credit decisions consume 150 hours per month, versus 100 for shipping-status replies. Yet their exposure score of 5 and feasibility score of 2 rule out autonomous decision-making. Even a high-value workflow can lose. A separate document-preparation assist might be worth assessing, but it is a different scope with its own gate.
Monthly baseline labor value equals volume × handling minutes ÷ 60 × loaded hourly rate. That gives $3,600 for shipping-status replies, approximately $2,240 for duplicate reporting, and $7,200 for eligibility work. Those are baseline labor values—not achievable savings. Tool costs, implementation, retained review, and rework still matter.
Eliminate, automate, or retain human judgment
The Eliminate, Automate, Too Late framework asks a useful question before any technical one: should this work exist?
Duplicate reporting is an elimination candidate only after its owner confirms it serves no unique control or decision. Shipping-status replies are an automation candidate with verified source data and agent approval. Eligibility decisions retain human judgment because the consequences and rules matter more than the apparent repetition.
'Too Late' is not a universal instruction to ignore a complaint. Keep people on cases where judgment or exposure is high. A scoring sheet should record that boundary rather than turn a category into a shortcut.
Released capacity is not realized cash savings. If an agent spends less time on replies but payroll remains unchanged, you have capacity to redeploy. Cash savings require an actual expense reduction; document how it happens. Avoid counting both that reduction and the same redeployed hours as separate benefits.
Set human-review gates
Write the gate as an operational rule: who checks what evidence, before which action, and what happens on failure. 'Human in the loop' is not specific enough.
For reply drafts, the support agent checks the order source and customer context before sending. Missing records, contradictory statuses, disputes, and sensitive exceptions return to the normal human queue. Record the input, draft, decision, and correction so a reviewer can reconstruct the case.
Agree stop conditions before testing. An unauthorized disclosure or an unapproved high-stakes action should trigger an immediate stop and investigation. Other thresholds should follow your baseline and risk owner's judgment, not a universal accuracy percentage. Name a backup reviewer and a rollback owner; otherwise a gate can become a queue nobody owns.
Turn the shortlist into a measured 90-day plan
Treat 90 days as a sequence of decisions, not a promise to automate everything. An illustrative plan starts with baseline validation and scope approval, moves to shadow testing, then allows limited production only after the owner accepts the evidence. Expand, revise, or stop based on completed-case outcomes.
For shipping-status drafts, compare active minutes per completed case, review time, reopened cases, and total operating cost. Keep the case mix comparable. Report uncertainty if the sample is small or the mix changed; a shorter handle time is not a win if customers need another contact.
Each phase needs an owner, evidence required to proceed, a stop condition, and a budget range agreed during scoping. A readiness scorecard can start the conversation, but it cannot replace workflow-level evidence or authorize a production rollout.
What an assessment engagement delivers
My AI Operations Assessment turns this inventory into a scored shortlist, build/buy/wait recommendations, and a sequenced plan with owners, metrics, and a budget range. It starts with workflow mapping and data rather than a generic strategy deck. The service page explains the two-to-three-week scope and what to prepare.
Bring the process that eats the most time—and the evidence behind it. Read the assessment scope and book its existing scoping call to decide whether the right next step is elimination, a bounded automation pilot, or better support for human judgment.