An automation prioritization framework ranks the AI and automation projects competing for the same budget, so the first one funded is the one most likely to return measured hours. This one scores each candidate on six weighted criteria, starting from the measured baseline of the work it would change: addressable hours, payback, process stability, service risk, measurability, and time to a first measured result. Three rules override the total, and every project leaves the sheet with the baseline it will be measured against after go-live.
A scoring sheet for the AI strategy team, the CFO, and the operations leader to fill in together, before any project starts.
The list arrives from every direction. A document extraction tool for intake. A copilot for the people who write file notes. A model that would take decisions off your most senior staff. Each one has a champion, a vendor deck, and a projected saving, and the projections were all written by people who want the project funded.
The framework rests on one rule: no baseline, no rank. A candidate is scored on the work it would change, measured as that work runs today. A candidate whose work nobody has measured does not get a score. It gets a measurement.
How to Run the Framework
Run it once per budget round, on one operation at a time, with the operations leader, someone from the AI strategy team, and someone from finance in the room. Five steps take you from a list of ideas to a funded sequence.
- List the candidates: Write down every proposed AI or automation project for the operation, one row each, and name the activities each one would change. A project described as “AI for claims” is not a candidate yet; “extract data from first notice of loss documents” is.
- Baseline each candidate: For every activity a candidate touches, record the measured hours per week, the volume, the minutes per item, the rework rate, and the spread in handle time across people. Candidates without measured activity data go straight to the Measure first wave.
- Agree the weights: Set the six weights before anyone sees a score, and have the operations leader and finance both sign them. Weights chosen after the scores are known are a way of picking a winner.
- Score and apply the rules: Score each candidate 1 to 5 on the six criteria, calculate the weighted total, then apply the three override rules. The rules beat the total every time.
- Sequence the waves and write the after-read: Assign each candidate to a wave, and for each funded project fill in the before half of the after sheet, so the measure it will be judged by is fixed before it starts.
The Six Criteria and Their Weights
Every candidate is scored on the same six criteria. The weights add up to 100, so a candidate scoring 5 on every criterion totals 100 and one scoring 1 on every criterion totals 20.
| Criterion | Default weight | What it asks | Where the number comes from |
|---|---|---|---|
| 1. Addressable hours | 25 | How much measured work could the project take on, as a share of the operation’s hours? | Measured hours on the activities it touches × their automatable share |
| 2. Payback | 20 | How many months until the hours it releases have paid for it? | Addressable hours, a realization factor, loaded hourly cost, and the project’s first-year cost |
| 3. Process stability | 15 | Does the work run the same way across people, or with rework and wide variation? | Rework rate and handle-time spread from the baseline |
| 4. Service risk | 15 | What happens to customers, Service Level Agreements (SLAs), and regulated decisions when the project gets it wrong? | The SLA and compliance map for the activities in scope |
| 5. Measurability | 15 | Can the result be read against the baseline, with the same measure, after go-live? | Whether the baseline exists at the activity level and can be rerun |
| 6. Time to first measured result | 10 | How many weeks until there is a first result to read against the baseline? | The project plan, counted to the first read, not to go-live |
The automatable share in criterion 1 comes from scoring each activity, not the project. Use the 0, 25, 50, 75, and 100% scale in the Workforce Capacity Planning Model, or translate a workforce automation analysis band: automate now is about 75%, augment is 25 to 50%, keep human is 0%.
Scoring each criterion from 1 to 5
| Criterion | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|
| Addressable hours (share of the operation’s measured hours) | Under 2% | 2 to under 4% | 4 to under 7% | 7 to under 12% | 12% or more |
| Payback | Over 36 months | 24 to 36 months | 12 to under 24 months | 6 to under 12 months | Under 6 months |
| Process stability (score the lower of the two signals) | No standard way of doing the work, or rework over 20% | Spread over 2.5x, or rework 12 to 20% | Spread 2 to 2.5x, or rework 8 to under 12% | Spread 1.5 to under 2x, or rework 3 to under 8% | Spread under 1.5x and rework under 3% |
| Service risk | The project makes or changes a regulated decision itself | Errors can feed a regulated decision or breach a contractual SLA | Errors can reach a customer or an SLA, and are reversible | Errors are caught at an existing review step before anything leaves the team | No customer, SLA, or compliance exposure |
| Measurability | No agreed way to tell whether it worked | Only output volume can be compared before and after | Hours are measurable, but part of the value is quality nobody measures today | Baseline measured; the after-read needs one extra metric, such as a quality sample | Baseline measured at the activity level, and the same measure can be rerun after go-live |
| Time to first measured result | Over 52 weeks | 27 to 52 weeks | 14 to 26 weeks | 6 to 13 weeks | Under 6 weeks |
Spread is the median minutes per item of the slowest quarter of the people doing the activity, divided by the median of the fastest quarter. Rework is the share of items touched a second time to correct or complete them. Both come from the baseline, which is why criterion 3 cannot be scored from a process map.
The weighted total: multiply each score by its weight, add the six results, and divide by 5. The total runs from 20 to 100.
The Three Rules That Override the Score
A weighted total treats every point as interchangeable. Three situations are not, so they override the total wherever they apply.
Rule 1: No baseline, no rank. If the activities a candidate touches have not been measured, the candidate is not scored. It goes to the Measure first wave, and the action is to measure the work. A score built on estimated hours ranks the confidence of the person who estimated them.
Rule 2: The worst place to start. If a candidate’s savings case depends on taking work away from your most seasoned, most expensive people, it cannot go in the first wave, whatever it scores. That work looks like the biggest saving per head, and it is the work that holds your SLAs and your customer service. Prove the realization factor on lower-risk work first.
Rule 3: The service floor. A candidate scoring 1 on service risk cannot go in the first wave either. A project that makes regulated decisions itself needs a human decision point designed into it before it is scored again.
The Prioritization Worksheet
The worksheet has three parts: a baseline block per candidate, the scoring sheet, and the weights. Copy them into a spreadsheet or a document. Fill in the baseline block first and leave the scores blank until every candidate has one, because scoring before measuring is how the project with the loudest champion ends up first.
Baseline block (one per candidate)
| Activity the candidate touches | Measured hours per week | Volume per week | Minutes per item | Rework rate | Handle-time spread | Automatable share |
|---|---|---|---|---|---|---|
| Total |
Scoring sheet
| Candidate | Addressable hours | Share of measured hours | Payback (months) | 1. Hours (×25) | 2. Payback (×20) | 3. Stability (×15) | 4. Service risk (×15) | 5. Measurability (×15) | 6. Time to result (×10) | Weighted total | Rule applied | Wave |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
Weights and assumptions
| Item | Your figure |
|---|---|
| Total measured hours per week, whole operation | |
| Realization factor (share of addressable hours that actually comes back) | |
| Working weeks per year | |
| Weights: hours / payback / stability / service risk / measurability / time (total 100) | |
| Signed by (operations) | |
| Signed by (finance) |
Rebuilding the worksheet in a spreadsheet
Enter up to 10 candidates in rows 2 to 11 and the assumptions in column T. The formulas score the two calculated criteria, total the six, and assign the wave, so changing a weight or the realization factor re-ranks the whole list.
ASSUMPTIONS AND WEIGHTS (column T)
T2 Total measured hours per week, whole operation
T3 Realization factor, as a decimal
T4 Working weeks per year
T6 Weight: addressable hours (default 25)
T7 Weight: payback (default 20)
T8 Weight: process stability (default 15)
T9 Weight: service risk (default 15)
T10 Weight: measurability (default 15)
T11 Weight: time to first result (default 10)
T12 =SUM(T6:T11) must equal 100
CANDIDATE ROWS (row 2 shown, copy down to row 11)
A2 Candidate
B2 Measured hours per week on the activities it touches
C2 Automatable share: 0, 0.25, 0.5, 0.75 or 1
D2 =B2*C2 Addressable hours
E2 =D2/$T$2 Share of measured hours
F2 Loaded cost per hour
G2 First-year project cost
H2 =IF(D2>0,G2/(D2*$T$3*$T$4*F2)*12,"") Payback in months
I2 =IF(E2>=0.12,5,IF(E2>=0.07,4,IF(E2>=0.04,3,IF(E2>=0.02,2,1))))
J2 =IF(H2<6,5,IF(H2<12,4,IF(H2<24,3,IF(H2<=36,2,1))))
K2 Process stability score, 1 to 5
L2 Service risk score, 1 to 5
M2 Measurability score, 1 to 5
N2 Time to first result score, 1 to 5
O2 =(I2*$T$6+J2*$T$7+K2*$T$8+L2*$T$9+M2*$T$10+N2*$T$11)/5
P2 Rule: M for no baseline, W for worst place to start, blank if none
Q2 =IF(P2="M","Measure first",IF(O2<40,"Redesign or drop",IF(O2<60,"Hold",IF(OR(O2<80,P2="W",L2=1),"Second wave","First wave"))))For a blended automatable share across several activities, use the hours-weighted average from the baseline block rather than the share of the biggest activity.
Reading the Score: Five Waves
The weighted total sets the wave, unless one of the three rules has already set it.
| Wave | When | What to do |
|---|---|---|
| Measure first | Rule 1 applies | Measure the activities before anyone scores the project. Re-enter it in the next round with a baseline |
| First wave | 80 to 100, and no rule applies | Fund it, write the after sheet, and read it at the first measured result |
| Second wave | 60 to 79, or 80+ with Rule 2 or Rule 3 applied | Hold the budget until a first-wave project has a measured realization factor, then re-score with it |
| Hold | 40 to 59 | Fix the weakest criterion first: standardize the process, add a review step, or define the after-read. Re-score next round |
| Redesign or drop | Under 40 | The project as described does not earn its budget. Narrow its scope to the activities that score, or take it off the list |
A Worked Example: Five Candidates in One Claims Operation
Here is the framework filled in for one hypothetical operation: a property and casualty claims operation of 120 FTE, measured at 4,080 hours a week. The numbers are illustrative, not client data. They exist to show how the sheet behaves, including how far a ranking moves when one input is an estimate.
Assumptions. Realization factor 75%. Fifty working weeks a year. Loaded cost $50 an hour for claims staff and $85 an hour for senior adjusters. Default weights.
The five candidates on the AI strategy team’s list:
- A. Intake document extraction: Pulls data from first notice of loss documents and indexes attachments. Touches 620 measured hours a week, 50% automatable.
- B. Status request automation: Answers claim status requests from agents and policyholders by email and portal. Touches 410 hours a week, 75% automatable.
- C. Adjuster copilot for file notes: Drafts file notes and claim summaries. Touches 900 hours a week, 25% automatable.
- D. Complex claims decision model: Takes coverage decisions on large losses off the senior adjusters. Touches 520 hours a week, 25% automatable.
- E. Subrogation flagging: Identifies claims with recovery potential. The work is spread across adjusters and nobody has measured it.
The calculated criteria.
| Candidate | Addressable hours | Share of 4,080 | First-year cost | Hours released a year at 75% | Payback |
|---|---|---|---|---|---|
| A. Intake extraction | 310 | 7.6% | $350,000 | 11,625 ($581,250) | 7.2 months |
| B. Status requests | 307.5 | 7.5% | $150,000 | 11,531 ($576,563) | 3.1 months |
| C. Adjuster copilot | 225 | 5.5% | $600,000 | 8,438 ($421,875) | 17.1 months |
| D. Complex claims model | 130 | 3.2% | $900,000 | 4,875 ($414,375 at $85) | 26.1 months |
| E. Subrogation flagging | not measured |
The scored sheet.
| Candidate | Hours | Payback | Stability | Service risk | Measurability | Time to result | Weighted total | Rule | Wave |
|---|---|---|---|---|---|---|---|---|---|
| B. Status requests | 4 | 5 | 5 (spread 1.3x, rework 2%) | 3 (a wrong status reaches a customer, and is corrected) | 5 | 5 (5 weeks) | 89 | First wave | |
| A. Intake extraction | 4 | 4 | 4 (spread 1.6x, rework 6%) | 4 (caught at adjuster review) | 5 | 4 (10 weeks) | 83 | First wave | |
| C. Adjuster copilot | 3 | 3 | 2 (spread 2.4x, rework 14%) | 2 (a summary feeds a coverage decision) | 3 (note quality is not measured today) | 3 (20 weeks) | 54 | Hold | |
| D. Complex claims model | 2 | 2 | 2 (spread 2.8x) | 1 (it makes the coverage decision) | 3 | 2 (40 weeks) | 40 | Rules 2 and 3 | Hold, redesign before re-scoring |
| E. Subrogation flagging | Rule 1 | Measure first |
What the sheet says. The least glamorous project on the list goes first. Status requests carry no AI headline, and they are 7.5% of the operation’s measured hours, stable, measurable within five weeks, and paid back inside a quarter. Intake extraction follows it.
The decision model arrived with the biggest projected saving per head, because senior adjusters cost the most per hour. Measured, it addresses 3.2% of the hours, pays back in more than two years, and makes regulated decisions itself. Rules 2 and 3 would hold it out of the first wave even if it had scored 90.
File notes are written at least three different ways across the team, with a 14% rework rate, so a copilot would automate the variation. Its first project is standardizing how notes are written, which the same baseline can measure.
How far an estimate moves the ranking. Suppose the copilot’s hours had come from a survey instead of measurement, and the survey put file notes at 1,300 hours a week instead of the measured 900. Addressable hours rise to 325, or 8.0% of the operation, which scores 4. Payback falls to 11.8 months, which also scores 4. The total rises from 54 to 63, and the copilot moves from Hold to Second wave on a number nobody measured. That is Rule 1 at work: the estimate changes which project gets the budget.
Bring your list. If your candidates are scored on estimates today, we can score them from measured activity instead. Book a 30-minute strategy call, bring the list, and we will walk through which rows would move once the work is measured.
The After Sheet: Measuring Each Project Against Its Baseline
Every funded project gets an after sheet. The first half is written before the project starts, from the baseline block, and nobody edits it afterwards. The second half is filled in at the read date, with the same measure.
Written before the project starts
| Project | Baseline period | Measured hours per week | Volume per week | Minutes per item | Rework rate | Addressable hours | Projected hours released | Read date |
|---|---|---|---|---|---|---|---|---|
Filled in at the read
| Project | Measured hours per week | Volume per week | Hours expected at this volume | Hours released | Measured realization factor | Rework rate | Payback, re-read |
|---|---|---|---|---|---|---|---|
Volume-adjust the read. Volume rarely holds still for 12 weeks, so compare against what the baseline would have needed at the new volume: hours expected = baseline minutes per item × volume after ÷ 60, then hours released = hours expected – measured hours after. The measured realization factor is hours released ÷ (hours expected × automatable share).
The example, read at week 12. Status requests were baselined at 410 hours a week, 4,920 requests, 5.0 minutes per request, and 2% rework. At week 12 the same measure shows 220 hours a week on 5,010 requests. At the baseline’s 5.0 minutes, 5,010 requests would have needed 417.5 hours, so the project released 197.5 hours a week. Against 313.1 addressable hours at the new volume, that is a measured realization factor of 63%, not the 75% the sheet assumed. Payback, re-read at the measured figure, is 3.6 months instead of 3.1.
The project still earned its place. More useful is the 63%: it replaces the assumed 75% for similar work in the next round of scoring, so the second wave is ranked on a realization factor measured in your own operation instead of one written into a vendor’s projection.
Why the Framework Starts From the Baseline
Most AI prioritization frameworks score a project on an impact estimate written by the person proposing it.
When an AI strategy team looks at a baseline, the fair question is whether it is just one more item on the list of AI projects they are evaluating. Our answer is that it comes first. It measures the specialized tools on the list, before they start and after they go live, which is what lets the CFO see what each project actually returned.
The baseline also shows something a grid never does. Captured across a team, the same activity runs differently depending on where the person works, how experienced they are, and how much rework they do. That variation is where much of the opportunity sits, and it is also the strongest warning that a project should standardize the work before it automates it. That is why process stability is a criterion here, scored from measured spread and rework rather than from a process map.
How This Fits With the Other Tools
This framework ranks projects. It leans on the pages that score the work underneath it, and does not repeat them.
- Whether a process can be automated at all is the job of the workforce automation analysis, which scores one process on seven factors. Its result feeds the automatable share here.
- What the freed hours mean for headcount is the job of the Workforce Capacity Planning Model, which turns automation scores into FTE scenarios.
- Which tasks are candidates in the first place is covered in how to know what to automate in operations, with five tests for a good candidate.
- The prioritization method in prose, and why impact-against-feasibility grids fail, is in how to prioritize AI projects in operations.
- Presenting the waves to a board that has already set a number is what the Mandate Response Framework is for.
- What finance should measure once projects are live is in AI ROI in operations.
Where Prioritization Scores Usually Go Wrong
The hours are estimated. Every downstream number inherits the hours: addressable hours, payback, and two of the six scores. The worked example shows a survey moving a project up a wave. If the baseline block is filled in from interviews, mark it as estimated and treat the whole row as provisional.
The tool gets scored instead of the work. A vendor’s capability says what the software can do, not how much of your work it would touch. Score the activities in your operation, then ask what share of them the tool would take on.
The weights are set after the scores. Once a favoured project is on the sheet, every weight change has a winner. Agree and sign the weights first, and change them only between rounds.
Nothing is read after go-live. A framework that ranks projects and never measures them repeats the same assumptions every round. The after sheet is what turns the second round into a better one than the first.
Frequently Asked Questions
What is an automation prioritization framework? An automation prioritization framework is a fixed set of weighted criteria that every proposed automation or AI project is scored against, so the projects competing for one budget are ranked the same way. This one scores six criteria from the measured baseline of the work, applies three override rules, and sets the measure each project is read against after go-live.
What criteria should you use to prioritize automation projects? Six cover most operations: addressable hours, payback, process stability, service risk, measurability, and time to a first measured result. The default weights are 25, 20, 15, 15, 15, and 10. Agree and sign the weights before anyone sees a score.
Can this be used as an AI use case prioritization framework? Yes. The criteria apply the same way to an AI use case as to rules-based automation, because both are scored on the work they would change. Score the activities the use case touches, not the model’s capability.
Who should score the candidates? The operations leader, the AI strategy team, and finance together. Operations knows the work and the SLAs, the AI team knows what each tool can take on, and finance owns the payback and the after-read. Both operations and finance sign the weights.
What if we do not have baseline data for a candidate? It goes to the Measure first wave and is not scored. Estimated hours change which project wins, as the worked example shows. Measure the activities it would touch and score it in the next round.
How do you measure an automation project after it goes live? Write the baseline hours, volume, minutes per item, and rework rate into the after sheet before the project starts. At the read date, rerun the same measure, volume-adjust it, and calculate the hours released and the measured realization factor. Use that factor in the next round.
Is there an automation prioritization template in Excel? Yes. The worksheet on this page rebuilds in any spreadsheet: a baseline block per candidate, the scoring sheet, the weights, and the cell formulas listed above. It is free and needs no email address.
The framework runs on one input most operations do not have: measured hours per activity, by person. The Summit Trails 90-day Phase 1 assessment produces it: activity captured and classified across the team, each activity scored for automation potential, and a prioritized deployment roadmap built from the result. Book a 30-minute strategy call and bring your list of candidates.
To see how activity is captured and classified, read our approach, and the results page shows what the output looks like.
Use this framework with: the AI Readiness Scorecard to check whether your data can support the baseline, the Workforce Capacity Planning Model, and the other free tools for operations leaders.
Rather talk it through? Wendy runs a live session for operations leaders facing an AI mandate, with upcoming dates on the webinar page.