THE AI LAB GOAL LADDER
Our method for setting the right goals for AI in a business. Five rungs, climbed in order: business outcome, objective, use case, target and guardrail. You end up with one sentence per goal that says what will change, by how much, by when, and what you will not trade away to get there.
FOR AUSTRALIAN BUSINESSES OF 10 TO 200 PEOPLE · BUILDS ON THE READINESS FRAMEWORK
- 1OUTCOMETake on more BAS clients this year without hiring ahead of revenue.
- 2OBJECTIVECut the time bookkeepers spend keying supplier invoices by the end of next quarter.
- 3USE CASEAI reads each invoice and drafts the coded entry in the ledger. A bookkeeper reviews and approves it.
- 4TARGETFrom 6 minutes per invoice (measured over two weeks) to 2 minutes, review included, by week 8.
- 5GUARDRAILCoding errors no higher than today. A bookkeeper approves every entry. Client data stays in Australia.
A 12-person accounting practice in Sydney. Example only. Not a client, and the figures are illustrative. Use your own baseline.
001 / THE LADDER
Most AI projects that disappoint were never given a goal anyone could check. The usual cause is starting on the third rung, with a tool or a use case, and never going back for the other four. Each rung below answers one question. Skip none.
RUNG 01 · OUTCOME
Start with a business outcome
Name the result the owners care about, in words that would make sense with no AI in the room: more capacity, faster turnaround, fewer errors, lower cost to serve, less risk. If the outcome only makes sense when you mention a tool, it is not an outcome yet.
WEAK
We need to be using AI this year.
STRONG
Take on more clients this year without hiring ahead of revenue.
QUESTIONS TO ASK
- ?What would be different in the business in 12 months if this worked?
- ?Who in leadership owns this outcome and will answer for it?
- ?Would we still want this if AI turned out not to be the way to get it?
RUNG 02 · OBJECTIVE
Turn it into one objective
Pick the single process where a change would move that outcome most, and say which direction it needs to move and by when. One objective per outcome to begin with. This is the O in an OKR: short, specific and owned by one person.
WEAK
Improve efficiency across the business.
STRONG
Cut the time bookkeepers spend keying supplier invoices by the end of next quarter.
QUESTIONS TO ASK
- ?Which one process, if it changed, would move the outcome most?
- ?Who runs that process day to day, and do they agree it is the problem?
- ?What is the deadline, and what happens if we miss it?
RUNG 03 · USE CASE
Choose the use case
Only now pick where AI helps. List the candidate tasks inside that process and rank each on value (how much time, money or risk it touches) against feasibility (is the data usable, is the task bounded, can a person check the output). Start in the high-value, high-feasibility corner. The Readiness Framework's six evidence checks are how you test feasibility.
WEAK
Roll out an AI assistant to everyone and see what sticks.
STRONG
AI reads each invoice and drafts the coded entry. A bookkeeper reviews and approves it.
QUESTIONS TO ASK
- ?Is the task bounded, repeated often and easy for a person to check?
- ?Is the data legally usable, reachable and good enough?
- ?If the AI gets one wrong, who notices and what does it cost?
RUNG 04 · TARGET
Set a target from a baseline
Measure how the process performs today before you build anything. Then set one primary number, a date and the share of work it applies to. A target without a baseline is a guess, and nobody can tell afterwards whether the project worked.
WEAK
Save lots of time. Aim for 10x productivity.
STRONG
From 6 minutes per invoice (measured over two weeks) to 2 minutes, review included, by week 8.
QUESTIONS TO ASK
- ?What does the process take today, measured rather than estimated?
- ?What is the floor we would accept, the target we expect and the stretch?
- ?On what date do we check, and who brings the numbers?
RUNG 05 · GUARDRAIL
Add the guardrails
Write down what must not get worse while you chase the target, and what would make you stop. Cover quality, privacy and where data is processed, who reviews the output, running cost and how to switch it off. A goal with no guardrail tends to get hit by quietly breaking something else.
WEAK
We will sort out the risk side once it is live.
STRONG
Coding errors no higher than today. A bookkeeper approves every entry. Client data stays in Australia.
QUESTIONS TO ASK
- ?What must not get worse: accuracy, client experience, compliance, cost?
- ?Who is accountable for the system, and who reviews its output?
- ?What result would make us pause or switch it off, and how do we do that?
002 / THE TEMPLATE
Fill in the blanks from the top of the ladder down. If you can't fill one in, that is the rung to work on before anything gets built.
FILL IN THE BLANKS
To business outcome we will change in one process by using AI to bounded task, with who checks it moving measure from baseline today to target by date without what must not get worse Owned by one named person We stop or rethink if stop rule
BEFORE YOU SIGN IT OFF
- ✓It still makes sense if you delete the word AI.
- ✓It names one process, one measure and one owner.
- ✓The baseline was measured, not remembered.
- ✓The date is close enough that the people who set it will be the ones who check it.
- ✓The guardrail is something you can count or observe.
The PDF has this as a blank worksheet you can print and fill in by hand. Get the PDF
003 / TARGETS AND BASELINES
We won't give you an industry-average number to aim for, because your process is not the average. Measure first, then set the number. Our companion method, the AI Lab Value Scorecard, covers which measures to track once the system is running.
- 01
Pick one primary measure
Time per item, turnaround, error rate, cost per item or volume handled. One number carries the goal. Others can be tracked, but only one decides it.
- 02
Measure the baseline first
Two to four weeks of real records before any build: sample cases, timings, error counts. If the records do not exist, collecting them is the first task of the project.
- 03
Set a range, not a single number
A floor (below this it was not worth doing), a target (what you expect) and a stretch. Early AI projects are uncertain, and a range lets you make an honest call at the gate.
- 04
Pair it with a counter-measure
For every speed or cost target, name the quality measure that must hold: error rate, rework, complaints, review time. This is where the guardrail becomes a number.
- 05
Say how much of the work it covers
A target on 80 per cent of invoices is different from a target on the easy 20 per cent. State the share of volume, and what happens to the exceptions.
- 06
Book the review date now
Put the gate reviews in the calendar when the goal is set, with the person who will bring the numbers. Goals drift when the check has no date.
004 / PILOT TO SCALE
A pilot with no exit criteria neither scales nor stops. Agree the conditions before it starts, book the review dates, and make a call at each gate. Gate 1 uses the same evidence checks as the Readiness Framework.
GATE 1 · BEFORE ANY BUILD
Ready to start
- ✓The goal sentence is complete, with a named owner
- ✓The baseline has been measured from real records
- ✓The data is legally usable and you know where it will be processed
- ✓Real test cases with known correct answers are set aside
- ✓A named person will review outputs, and people know where problems go
If not: Close the gap first, or pick a different use case. Do not start the build to find out.
GATE 2 · HALFWAY THROUGH THE PILOT
Worth continuing
- ✓The people it was built for are using it in real work
- ✓The primary measure is moving towards the target
- ✓The counter-measure is holding at or better than baseline
- ✓No guardrail has been breached
- ✓Problems found so far have an owner and a fix date
If not: Revise once, with a clear change and a new check date. If the same gap is still there at the next check, stop.
GATE 3 · END OF THE PILOT
Ready to scale
- ✓The result is at or above the floor of the target range on the agreed share of work
- ✓Guardrails held for the whole pilot, with evidence
- ✓Running cost per month is known and the value is worth it
- ✓Monitoring, support and a way to switch it off are in place
- ✓The team that will run it has been trained and has agreed to own it
If not: Stop, write down what you learned and release the budget to the next idea on the list.
SCALE
Every pass condition at the gate is met. Harden it, roll it out and set the next goal.
REVISE
The use is real but a measure or guardrail is short. One change, one new date, then decide again.
STOP
The floor was missed, a guardrail was breached and cannot be fixed, or people will not use it. Stopping on evidence is a result, not a failure.
005 / GOALS BY HORIZON
The same three horizons as the Readiness Framework. At each one, set a business goal, a people goal and a governance goal. Leave one out and the other two tend to stall.
H1
Foundation pilots
0–3 MONTHS
BUSINESS GOALS
One to three goals, each on a single bounded process, each measured against its own baseline.
PEOPLE AND ADOPTION
The pilot team uses it in real work each week, and every user knows how to report a wrong answer.
GOVERNANCE AND RISK
A named owner per system, a person reviewing outputs, a simple register of what is in use and a written rule for what staff may put into AI tools.
H2
Scale and hardening
3–9 MONTHS
BUSINESS GOALS
Goals for the pilots that passed Gate 3: coverage across the whole process, value measured monthly, plus a second wave of use cases.
PEOPLE AND ADOPTION
Adoption goals by role, training completed, and time saved actually moved to agreed higher-value work.
GOVERNANCE AND RISK
Monitoring, cost caps, incident handling and compliance evidence, including the automated decision-making transparency duty from 10 December 2026 where it applies.
H3
Enterprise capability
9–18+ MONTHS
BUSINESS GOALS
Portfolio goals: total value across use cases, speed from idea to decision, and reuse of what has already been built.
PEOPLE AND ADOPTION
The business sets and reviews its own AI goals each quarter without outside help.
GOVERNANCE AND RISK
A standing review rhythm: goals, risks and the register checked together, with systems retired when they stop paying their way.
006 / WORKED EXAMPLE
EXAMPLE · A 12-person accounting practice in Sydney. Example only. Not a client, and the figures are illustrative. Use your own baseline.
THE FIVE RUNGS
- OUTCOME
- Take on more BAS clients this year without hiring ahead of revenue.
- OBJECTIVE
- Cut the time bookkeepers spend keying supplier invoices by the end of next quarter.
- USE CASE
- AI reads each invoice and drafts the coded entry in the ledger. A bookkeeper reviews and approves it.
- TARGET
- From 6 minutes per invoice (measured over two weeks) to 2 minutes, review included, by week 8.
- GUARDRAIL
- Coding errors no higher than today. A bookkeeper approves every entry. Client data stays in Australia.
THE GOAL IN ONE SENTENCE
To take on more BAS clients without hiring ahead of revenue, we will cut invoice keying time by using AI to draft coded entries for bookkeeper review, moving handling time from 6 minutes to 2 minutes per invoice by week 8, without raising the coding error rate or sending client data offshore. Owned by the practice manager. We stop or rethink if errors run above today's rate for two weeks in a row.
AT THE GATES
- Gate 1. Two weeks of timing and error records collected. Client engagement terms checked for data use. Practice manager named as owner. 200 past invoices set aside as test cases.
- Gate 2. Week 4: bookkeepers are using it on most invoices, handling time is trending down, error rate is at or below baseline. Decision: continue.
- Gate 3. Week 8: target met on the agreed share of invoices, guardrails held, monthly running cost known. Decision: scale to all BAS clients, then set the next goal.
H1
Invoice entry pilot for one client group, with every entry reviewed.
H2
Roll out to all BAS clients, add monitoring and a cost cap, then start a second use case (for example, chasing missing documents).
H3
A quarterly rhythm: the partners review goals, retire what is not paying its way and pick the next two ideas.
007 / COMMON MISTAKES
- 01
Starting from the tool
The goal is to use a product, so nobody can say what success looks like. Start from the outcome and let the tool be the last decision.
- 02
No baseline
The project finishes and the team argues about whether it helped. Two weeks of measuring first settles it.
- 03
Too many goals at once
Ten pilots, none finished. One to three goals in the first horizon is enough for a business of 10 to 200 people.
- 04
Targets with no guardrail
Speed goes up and accuracy quietly drops. Every target needs the measure that must not get worse.
- 05
A pilot with no exit criteria
Without gates, a pilot neither scales nor stops. It just keeps running. Decide the conditions before you start.
- 06
Counting licences as adoption
Seats bought is not work changed. Set a goal for real use in real tasks, and for where the saved time goes.
- 07
Nobody owns it
A goal owned by a committee is owned by no one. One named person per goal, and one per system.
008 / FREE PDF
The whole method in a 13-page A4 guide: the five rungs, a blank goal worksheet to fill in by hand, the six steps for setting a target, the three gates, goals by horizon and the worked example.
- ✓Printable goal worksheet, one sheet per goal
- ✓Gate checklists you can tick off in the room
- ✓No statistics we can't source, no vendor pitch
010 / QUESTIONS
Anything else is quicker to answer on a call. Thirty minutes, free, with the people who would build it.
It is our method for setting goals for AI in a business. You climb five rungs in order: business outcome, objective, use case, target and guardrail. The result is one sentence per goal that names the process, the measure, the baseline, the date, the limits and the owner.
Bring one process that annoys you. In thirty minutes we will walk it down the ladder and tell you honestly whether it is worth building.
Book a free call→