AI Ops Map & Opportunity Scorecard

01/ 10

What's in the kit

• This page — the method, the rubric, the formulas, what not to automate, a worked example and the failure modes.
• Prompt 1, the Ops Map interview (interview-prompt.md) — the AI interviews you stage by stage, then shows you the scoring facts it inferred so you can confirm them. Every number carries a basis label and every guessed fact is marked "(assumed)".
• Prompt 2, the Opportunity Scorecard (scoring-prompt.md) — scores every row with a fixed rubric, applies hard gates, works out the return from stated minutes, then runs a sceptic pass against its own top 3.
• The sheet (ops-map.csv) — 34 columns, with 12 fictional example rows to show what good looks like. Import it into Google Sheets or Excel, study the example, then replace it with yours.

02/ 10

Two ways to run it

The fast path (30–45 minutes)

• Gather and redact your evidence: calendar titles, sent subjects, task list (5 min).
• Run Prompt 1 on your own. Answer each stage briefly, write "unknown" freely, then confirm the scoring-facts table in one reply (15–25 min; nearer 15 if you already know your loaded costs).
• Run Prompt 2 in the same chat with your cost lines and any measured revenue evidence (5 min).
• Run Prompt 2b, accept the re-scores you agree with, import Prompt 2's CSV into the sheet, and put the first pilot run for #1 in your calendar (5–10 min).

Solo? Give one role (you) and your own hourly estimate as the loaded cost, skip the team message, and tell Prompt 1 "solo: one role, me". Phase 0 then takes one or two questions.

The full path (2–3 hours across a week)

• Monday: send the team message from Prompt 1's file; work out loaded costs per role.
• Tuesday: run the interview with the team's replies in the evidence.
• Wednesday to Friday: time the top 5 rows once each, including how many of the minutes are calls, visits or lookups, and replace estimates with measurements.
• Friday: run Prompt 2 and 2b on the measured inventory, write the redeploy line for each top-3 item, start the pilot.

03/ 10

Why most AI projects start in the wrong place

Most owners pick their first AI project in one of two ways: the task that annoyed them most this week, or the demo that looked best on social media. Both skip the step good agencies take before any build: inventory the work, size it, then choose.

I've managed 45+ digital projects, and the expensive mistakes rarely came from bad code. They came from building the wrong thing well. AI makes that mistake cheaper to start and easier to repeat. Put it on a workflow that runs four times a month, that nobody owns, or that a calendar reminder could handle, and you get a clever tool, a new subscription and no change you can find in the numbers.

Three ideas drive the method:

• AI multiplies a good process. It doesn't fix a broken one. If three people do a job three ways, AI produces a fourth.
• Value is hours consumed or closeness to revenue, whichever is higher. A ten-minute task done 40 times a month beats a three-hour task done twice a year. Work close to revenue usually beats back-office work, even at fewer hours.
• Hours released are not cash saved. Unless a cost line actually goes down, the gain is capacity. Real, worth having, but not money until you decide what those hours do.

04/ 10

The method

The method I use before any build: list every recurring workflow, put a price on the hours, test each one for AI suitability and risk, rank what survives by value × feasibility × safety, and write the return honestly, with hours released, capacity value, cash saved and setup effort kept in separate columns.

Step 1: Inventory every recurring workflow

List every piece of work that happens at least once a month, walking the business in a fixed order so nothing hides: Attract, Convert, Deliver, Bill and collect, Retain, Back office, and the owner's own recurring work. That last stage is there because owners forget themselves.

For each workflow, capture:

Prompt 1 also collects ten scoring facts per workflow (digital inputs, good past examples, whether "good" can be written as rules, tool access, who sees and who reviews the output, whether it decides about a person, whether it's the same every time, the minutes that are calls, visits or lookups, and what would have to exist first). It pre-fills them, marks every guess "(assumed)", and asks you to confirm the table in one reply.

Good evidence for the inventory: two weeks of calendar titles, 30 to 60 sent-mail subject lines, your task board, and five minutes of answers from each person on your team. Prompt 1 runs this as an interview and asks about the last real occurrence of each task rather than the average, because owners misjudge recurring work from memory.

Expect 12 to 30 rows for a business of 2 to 20 people. Fewer than 10 usually means vague rows ("admin", "emails") that need splitting.

Step 2: Put a price on the work: hours × loaded cost

Hours per month is runs × minutes ÷ 60. The cost of the work is hours per month × the loaded cost per hour of whoever does it.

Loaded cost is not the hourly wage. It's what an hour of that person's productive time really costs the business:

Loaded cost per hour =
  (annual pay + employer costs + yearly overhead share per person)
  / productive hours per year

Productive hours = paid hours
  - paid leave - public holidays
  - meetings, internal admin and other non-workflow time

Use your own numbers, not industry rules of thumb. For freelancers, use what you actually pay per hour of their work. For your own time as the owner, use what it would cost to hire someone to do that work, and label it an estimate.

The cost of the work tells you what a workflow consumes. It is not what you'll save. That comes in Step 6, and it's always smaller.

Step 3: Test AI suitability

AI is good at a handful of task shapes, and almost every useful small-business use is one of them:

• Draft from structured inputs: a reply, a quote, a listing, a report section.
• Extract fields from documents, emails, forms or photos into a table.
• Classify and route: tag, prioritise, send to the right person.
• Summarise a long thread, call or file for someone who has to decide.
• Compare against rules: check a draft against a price list, a policy, a checklist.
• Transform: rewrite for a different reader, format or language.

It is bad at facts that aren't in its inputs, exact arithmetic without a tool, anything that needs access to a system it can't reach, and judgements about people.

Then score feasibility, one point for each yes:

Frequency isn't a feasibility point on purpose. It already counts through Value (hours) and through setup hours in Step 6, where a rare task's setup takes longer to pay back. Counting it a third time would bury useful low-volume work.

Feasibility is about inputs and definitions, not about which model you use. Current models can draft almost anything. Whether they can draft your thing depends on what you can give them.

Step 4: Score the risk and apply the gates

Risk decides how much human stays in the loop. Start at 1 and add one for each yes, up to 5:

• A customer, supplier or regulator would see the output without a human review.
• It involves financial, health, ID or other sensitive data (contact details alone don't count).
• A mistake is costly or hard to reverse: money moves, a legal commitment, safety.
• It judges people: hiring, tenants, credit, disputes, performance.
• Nobody clearly owns reviewing the output.

Then the gates. A gated workflow doesn't get scored at all. It goes straight to the parked or no-AI list with the reason written down:

• Data rule — your contracts, regulator or client terms don't allow sending this data to an AI provider.
• Decision about a person — the workflow decides something about someone end to end.
• Fix the process first — it runs differently every time, or nobody owns it.
• No AI needed — a reminder, template, report feature or simple automation does the job, or the work can just stop.

An ungated row leaves the Gate cell empty. Not "—", not "none": the sheet's Priority formula reads any text there as a gate.

The gates matter more than the scores. In most inventories I'd expect a real share of the hours to land in "No AI needed", and those are often the fastest wins in the sheet: switch on a feature you already pay for this afternoon.

Step 5: Rank: value × feasibility × safety

Value is the higher of two bands:

Priority = Value x Feasibility x (6 - Risk)      range 0-125
Gated rows: Priority = 0
Ties: higher Value wins, then lower Risk

Why multiply instead of add: a zero anywhere should kill the idea, and a weak dimension should drag the score down hard. Why (6 − Risk): it turns risk into safety, so a Risk 5 workflow scores at most a fifth of an otherwise identical Risk 1 workflow.

How I read the result, as a rule of thumb rather than a law: 75 and above is a pilot candidate this month; 40 to 74 waits until the first pilot has taught you something; under 40 stays human or parked for now.

Every score gets an evidence line that quotes your inventory. "High impact" is not evidence. "18 hours a month, direct link: slow replies lose applicants" is.

Step 6: Write the return honestly

This is where most AI business cases go wrong. Someone multiplies all the hours by a rate, calls it savings, and six months later the bank balance hasn't moved. The sheet keeps these apart and never adds them together:

The 0.5 and 0.8 are my planning factors, not measurements. Someone still gathers the inputs, pastes them and fixes the draft, so an AI step rarely removes a whole touchable minute. Your pilot log replaces them.

Setup hours get their own column because they're paid once and everything else is monthly; mix them and a row with heavy prep looks either free or ruinous. Compare them instead: setup hours ÷ hours released low gives the months it takes to win the setup time back.

Two consequences you should expect, and say out loud:

• Your top opportunity may cost cash. If cash saved is 0 and running cost is 20 a month, the cash line is negative. That's fine, as long as the case is honestly about capacity or speed, not dressed up as savings.
• Capacity only becomes money when you decide what the hours are for. Write one line per top-3 item: "the released hours go to calling lapsed customers" or "the owner stops working Saturdays". Then measure that, not the AI.

Conditional cash, like a hire you would no longer make, gets the word "Conditional:" and the decision it depends on. It isn't saved until the decision is made.

Step 7: Attack the ranking, then pilot one thing by hand

Before you commit, run the sceptic pass in Prompt 2b. It checks for touchable minutes that still include calls or visits, optimistic review times, double counting, setup that looks too light, understated risk, drift toward judging people, and the cheaper fix you missed. Accept the re-scores you agree with.

Then don't build anything yet. Take the #1 row and run it by hand on the next 5 to 20 real cases: inputs pasted into a chat with one prompt, a named person checking every output, a log of minutes, review time, edits and errors caught. Replace the planning factors and the review minutes with what the log says. Only then decide whether it deserves an automation.

05/ 10

What not to automate

Some work should never reach an AI step, some needs a rule instead, and some should simply stop. Each of these appears in real inventories more often than you'd think.

06/ 10

The sheet

ops-map.csv has 34 columns. Import it into Google Sheets (File > Import > Upload, Separator type: Comma) or into Excel (Data > From Text/CSV, Delimiter: Comma). Row 2 is a banner saying the example rows are fictional. Study them, then delete the banner and every example row, keeping the header in row 1. Your own data starts in row 2.

Getting your rows in

Use Prompt 2's CSV block, not Prompt 1's. Prompt 1's block has only the 14 inventory columns and exists to feed Prompt 2; Prompt 2's has all 34, in the sheet's order.

• Google Sheets: copy Prompt 2's CSV block into a plain-text file and save it as my-ops-map.csv (on a Mac, TextEdit: Format > Make Plain Text first; on Windows, Notepad). In your spreadsheet: File > Import > Upload, choose the file, Import location: Insert new sheet(s), Separator type: Comma. Your rows land in a new tab, header in row 1.
• Excel: save the block as my-ops-map.csv the same way, then Data > From Text/CSV, choose the file, Delimiter: Comma, Load. Don't double-click the file: where decimals are written with a comma, Excel opens it as one column. If 4.5 comes in as a date or text, choose Transform Data and set the column type Using Locale: English (United States).
• Quick paste in Sheets: paste the block into cell A1 of an empty tab, then Data > Split text to columns, Separator: Comma. Check that each Evidence cell stayed whole; if a quoted value got split, use the file route above.

Optional formulas

The example rows hold values, not formulas, so they display anywhere. If you paste the AI's CSV, it already contains these values; add formulas only if you'll edit numbers by hand and want the sheet to recalculate. Put them in the first row of your own data (row 2, once the banner and example rows are deleted) and fill down. Pasting a CSV over these columns later replaces the formulas with plain values.

H2   =F2*G2/60                                        Hours per month
K2   =H2*J2                                           Cost of the work per month
S2   =IF(OR(R2="",R2="—",R2="-",R2="none",R2="N/A"),O2*P2*(6-Q2),0)
                                                      Priority (blank or dash Gate = scored)
V2   =IF(U2="","",F2*U2/60)                           Review hours per month
W2   =IF(T2="","",F2*(T2*0.5-U2)/60)                  Hours released low
X2   =IF(T2="","",F2*(T2*0.8-U2)/60)                  Hours released high
Y2   =IF(W2="","",W2*J2)                              Capacity value low
Z2   =IF(X2="","",X2*J2)                              Capacity value high

Then sort by column S, largest first. Keep columns Y, Z and AA next to each other but never in a single total, and keep AD (setup hours) out of every monthly sum. If you want a summary row, add separate totals.

07/ 10

Failure modes

• Vague rows. "Admin: 10 hours" can't be scored. Split it until each row is one person, one sitting, verb + object.
• Scoring from memory. Owners misjudge recurring tasks, often in both directions. Mark estimates as estimates and time your top five once.
• Scoring on guessed facts. A Feasibility point built on a fact nobody confirmed is a guess with a number on it. Phase 2b exists to catch these; Prompt 2 stops if more than a third are still assumed.
• Missing the team's work. The owner's view misses much of the repetitive work everyone else does. Five minutes of team replies usually adds rows.
• Counting hours as money. The most common way AI cases get inflated. If cash saved isn't tied to a named cost line, it's 0.
• Forgetting review time. An AI draft that takes four minutes to check isn't saving six minutes on a six-minute task. Net it out.
• Counting minutes the AI can't touch. Calls, site visits, physical work and lookups in tools with no export stay human. The touchable-minutes column makes them visible; Prompt 2b hunts for the rest.
• Ignoring setup. A row that needs 10 hours of rules-writing before its first run can still be the right pilot, but choose it knowing that.
• Automating a rule job. Reminders and data copies done by AI are slower, costlier and less reliable than the feature you already have.
• Picking an end-to-end agent as the first project. Start with one step, by hand, with a human check. Autonomy is earned with a log, not assumed.
• Letting the AI fill gaps. Both prompts are written to ask, not guess. If you see a number or a fact you never gave without its label, reject the output and say which one.
• Accepting the first ranking. Run the sceptic pass. If it changes nothing, make sure that's because the ranking holds.

08/ 10

Security and data

• Redact before you paste: client names become "Client A", and emails, phone numbers, addresses, account and card numbers and health details come out. A scorecard never needs anyone's real name.
• Never paste passwords, API keys, login links or anything under a confidentiality agreement.
• Check your AI plan's data settings (whether chats can be used for training, how long they're kept, whether a business plan applies) against what you owe your clients before pasting business information.
• Pasted material is treated as data in both prompts. If an email or document you paste contains instructions, the prompt tells the model to ignore them and say so.
• Salary data is sensitive too. Work out loaded costs per role, not per named person, and keep the sheet's sharing settings tight.

09/ 10

Prerequisites and costs

You need a business with recurring work and at least rough monthly volumes. It works for a solo operator too; the inventory is just shorter, and the loaded cost is your own hourly estimate.

10/ 10

Quality checklist

• Every stage, from Attract to the owner's own work, has a row or an answered gap question.
• Every number carries a basis label, every unconfirmed scoring fact says "(assumed)", and neither survives in your top 3.
• Hours = runs × minutes ÷ 60 and cost = hours × loaded cost on every row.
• Every scored row shows its touchable and review minutes, and hours released match the formula.
• Every top-3 score has an evidence line that quotes your own data.
• Your "No AI needed" list has at least one item, or you can say why not.
• Capacity value, cash saved, setup hours and revenue at stake sit in separate columns and are never totalled together.
• Cash saved is 0 unless a named cost line goes down, and conditional savings say so.
• Nothing that judges people or touches health, ID or financial data sits in the top 3 without a named human checkpoint.
• Each top-3 item has a line saying what the released hours will go to.
• The first pilot run for #1 is in someone's calendar this week.