Asking Questions and Defining Metrics

A useful analysis begins with a decision someone needs to make. In the Ask phase, the analyst defines the problem and learns what stakeholders expect. This happens before collecting or analyzing data: the question determines which evidence the later phases need.

Find the gap behind the request

A business problem is a gap between the current situation and the desired situation that matters to the organization. It may cost time, money, customers or quality. “We need more customer data” names an activity. “Repeat orders fell from 40% to 31% of first-time customers in six months” identifies a measurable gap.

Requests often arrive as proposed outputs: “Build a dashboard of app downloads.” Ask, “What decision will this help you make?” The answer may be that the marketing lead must decide by June whether to continue an advertising campaign. That decision determines what the analysis needs to show. A dashboard is one possible deliverable, not the business problem itself.

Describe the symptom without guessing the cause

A symptom is what you can observe. A cause produces it. Falling sales are visible; late deliveries might explain them. Offering discounts before checking the cause can spend money while leaving the delivery problem untouched.

Write the problem statement around five details: who is affected, what is happening, how large the gap is, since when, and why it matters. Then name the decision and its deadline. For example:

“Since January 2026, the share of first-time customers who place a second order within 30 days has fallen from 40% to 31%. The growth team must decide by June 30 whether to fund a welcome offer, and needs to know which customers and regions account for the drop.”

This statement identifies an observable change without claiming that price, delivery or the welcome experience caused it. Each explanation remains open to investigation.

Learn enough business context

The problem domain includes the people, processes, products and systems that affect the problem. Repeat orders may depend on delivery, pricing, ordering and customer support. Office rent and payroll processing are unlikely to help explain this decision.

Ask what changed during the period. A new price, app release, holiday or recording system may change the meaning of a number. A rise in support tickets after switching ticketing systems may partly reflect how tickets are counted. Write down these changes so the analysis can examine them.

Data-driven decision-making uses evidence to guide a specific decision, alongside costs, experience, ethics and strategy. Data-inspired decision-making draws more broadly on data and judgment to explore ideas. In either case, people remain responsible for deciding what to do.

Starting with an available table can lead to interesting findings with no use. Begin with the decision, identify evidence that could change it, and then look for the data that can provide that evidence.

Naming the kind of problem helps you plan the answer. Six common types describe different jobs an analyst may do. A project can combine them: the labels guide the work, while the business decision gives it a purpose.

Match the question to the type

A prediction is an informed estimate, so report the assumptions and uncertainty that affect it. Finding a past pattern is a different task: it describes what has tended to happen. A future estimate can then use that pattern along with other relevant information.

Categories and themes do different work

Categories need clear sorting rules so two people can classify the same item consistently. An “other” category can catch cases outside the named groups. Categories such as “price shown late,” “forgot password” and “too many form fields” describe separate kinds of comment. A theme such as “checkout feels like hard work” connects several categories into a broader idea. Themes are especially useful with interview notes and open-ended survey answers.

Define normal before flagging unusual

Suppose a room's temperature normally changes by less than 1 °C per hour, then rises by 6 °C in one hour. The rise stands out because a usual range has been established. Without a baseline, a large-looking number may be normal for that place, time or group. Choose the reference that fits the question, such as a comparable day or similar location.

Compare connections using the right base

A shop links delivery records to reviews. Of 2,000 late deliveries, 160 received a one-star review: 8%. Of 18,000 on-time deliveries, 180 did: 1%. More one-star reviews followed on-time deliveries in raw counts, yet the rate was much higher after late deliveries. Compare the share within each delivery group to examine the connection. Establishing a cause requires stronger evidence than this comparison alone.

A fraud project may first find patterns of normal spending and then spot payments that depart from them. When planning, identify each task separately: what counts as normal, which payments differ, and what decision a flagged payment will trigger. Forcing the whole project into one type can hide a needed step.

A broad problem becomes manageable when its reasoning is visible. Structured thinking means recognizing the situation, organizing information, revealing gaps and opportunities, and identifying options. These steps help you choose what to investigate first.

Organize what is known and unknown

A gym chain's monthly cancellation rate rose from 3.0% to 4.2% after a February price increase. Membership and visit records exist, but the cancellation form's reason field is optional. That leaves a gap: most leavers' reasons are unknown. Options include comparing visits before cancellation, comparing clubs, and surveying recent leavers. Writing down the gap prevents optional comments from being treated as a full account of why members left.

Build an issue tree

An issue tree splits a main question into smaller questions until each branch can be answered with data or a direct check. For revenue, start with:

Orders can be split again into visitors multiplied by orders per visitor. Follow the arithmetic one step at a time:

• Last month, 50,000 visitors at 0.020 orders per visitor produced 1,000 orders. At $60 per order, revenue was $60,000.
• This month, 50,000 visitors at 0.018 orders per visitor produced 900 orders. At $60 per order, revenue was $54,000.

Visitors and average order value stayed the same. The revenue drop sits in the orders-per-visitor branch. That branch can be investigated further: did fewer visitors buy, did buyers place fewer orders, or did more people stop during checkout?

Check that branches cover the problem cleanly

A MECE breakdown is mutually exclusive and collectively exhaustive: its groups do not overlap and together cover the relevant whole. “New” and “returning” customers can meet this rule when each customer belongs to exactly one group. “Mobile users” and “customers under 30” overlap, and they miss older desktop users. Use one consistent basis for a split so results are not double-counted or missed.

Turn possible causes into checks

The five whys asks why repeatedly to work back from a symptom. Late deliveries may follow late van departures, caused by late packing, caused by a supplier moving its arrival from 6:00 to 8:45. Each link needs evidence. Stop when you have a useful, supported explanation; the chain may need fewer or more than five questions and may branch.

A hypothesis states an explanation that evidence could contradict: “The fall in orders per visitor is concentrated among mobile users after the April 1 update.” Compare mobile and desktop results before and after that date. Decide what finding would count against it before looking. This reduces confirmation bias, the habit of favoring evidence that fits an existing belief.

Rank branches by how much they could explain and how easily the evidence can be obtained. Investigate promising, feasible branches first instead of jumping straight to a discount or trying to examine everything.

“Is our delivery app working?” sounds useful but leaves the analysis open. Working could mean faster ordering, fewer errors or more on-time deliveries. The SMART question checklist turns that broad concern into a question with a usable answer.

Apply all five checks

Specific means the question names a clear subject and enough context: which customers, product, place or process? “Are customers happy?” leaves the population open. A question about people who contacted app support about a late delivery gives a clear focus.

Measurable means the answer can be assessed in a defined way. “What percentage of app orders arrived within the promised 30 minutes?” has a calculable answer. A number in a sentence is not enough: the thing counted still needs a definition.

Action-oriented means the answer helps someone change something or make a decision. Identifying delivery zones with low on-time rates can guide staffing or route decisions. Ask what the stakeholder would do differently after seeing the result.

Relevant means the question serves the business problem. A question about the app's color scheme can be easy to measure yet do little to resolve late deliveries. Check the link to the decision before spending time on it.

Time-bound means the question states the period being studied. “Recently” lets different analysts choose different dates. Name dates or a precise period anchored to an event.

Rewrite a broad question

Start with “Is our new delivery app working?” If the problem is late deliveries, a sharper version is:

“In the 12 weeks after the app update on April 1, 2026, what percentage of app orders each week arrived within the promised 30-minute window, compared with the 12 weeks before the update, and which delivery zones stayed below 90%?”

Read the revision against the checklist:

• App orders and delivery zones make it specific.
• Weekly on-time percentages make it measurable.
• Zones below 90% point toward where attention is needed.
• On-time delivery connects it to the business problem.
• The two 12-week windows fix the comparison periods.

The question can reveal where and how much performance changed. “Did delivery improve?” may be too limited when the decision needs that detail. A yes/no question still has a place when a factual confirmation is all that is needed.

Keep question checks and target checks distinct

SMART has several meanings in goal setting. A common goal-setting version uses achievable for A. In analytics questions, A means action-oriented. Achievable asks whether a target can be reached; action-oriented asks whether an answer can guide action. A question can be useful even when the answer shows that a proposed target is unrealistic.

After wording the question, check its definitions and data requirements. A clear question about “active users” still needs a shared meaning of active before anyone calculates an answer.

People can give full, honest information only when the question lets them. Wording can suggest a preferred answer, hide an assumption or force one response to cover two different experiences. Read each question from the respondent's point of view.

Remove pressure and hidden assumptions

A leading question signals the answer the asker wants. “Don't you think the new checkout is much faster?” encourages agreement. “How would you describe the time it takes to pay with the new checkout?” lets the person describe what happened.

An unfair question assumes something that may be false or makes an honest response difficult. “Why do you prefer our brand?” assumes a preference. First establish whether the person has a preference. “What car do you drive to the store?” assumes car ownership and travel by car; a question about how the person usually travels leaves those possibilities open.

A double-barreled question combines two issues. Someone can like a price and dislike delivery speed, so “How satisfied are you with the price and delivery speed?” has no single clear answer. Ask about price and delivery separately.

Choose the response format for the purpose

A yes/no question, a list of categories and a 1-to-5 rating scale are all closed-ended. “Does this table include online orders?” is a useful factual check. A monthly satisfaction scale supports comparison across months. But “Are you satisfied?” gives little detail about what needs changing. Follow it with an open question when the reason matters.

Supply context and usable choices

A vague question leaves key details open. “Does the tool work?” could refer to any user or task. A clearer question asks how much time managers spend preparing the weekly schedule since changing tools, compared with the spreadsheet used before.

Answer choices need nonoverlapping boundaries and enough coverage. Age bands “18–25” and “25–35” both include 25. For an adult survey, “18–24,” “25–34” and “35 or older” avoid that overlap. Where relevant, “other” or “prefer not to say” gives people a truthful response route.

Use follow-up questions well

At a first stakeholder meeting, ask what decision depends on the analysis, what success looks like, who will use the answer, when it is needed and what has already been tried. “Can you give me an example?” can turn “the service is slow” into a specific process to investigate. Restate the request in your own words to check your understanding.

Try survey questions with a few people before sending them widely. Watch where they hesitate or interpret a word differently. Good intentions do not remove leading wording or unclear choices.

A well-worded question still needs a route from records to an answer. Define the terms, the population, the unit of analysis, the period and the comparison. Then check that the necessary data exists and can be used.

Make everyday words measurable

An operational definition states exactly how a concept will be measured. “Active customer” might mean a customer with at least one order in the last 90 days, or in the last 30 days. Among 10,000 customers, 6,200 qualify under the first rule and 4,100 under the second. The active share is therefore 62% or 41%. Neither figure can be interpreted until the rule is known.

Agree definitions before reporting a result. Writing “active” in a SMART question does not settle what activity counts, who is eligible or which records will be checked.

Choose what each observation represents

The unit of analysis is the thing described by each data point: a customer, order, visit, store or day. Changing it changes the question. Suppose 800 customers placed 1,200 orders worth $72,000:

The answers are $60 per order and $90 per customer. Multiple orders from one person explain why the values differ. Choose the unit that matches the decision, and label it in the result.

Name the comparison and the kind of claim

A cancellation rate of 4.5% becomes useful when compared with a target, another region, an earlier period or a group without a change. Name the comparison in the question so it is selected for a reason before the results are known.

A causal question needs stronger evidence, such as a fair comparison group or a planned experiment. A rise after a price increase could also reflect seasonal changes. Separate an observed sequence from the explanation you want to test.

Check feasibility before promising an answer

Ask whether records exist at the needed level of detail, cover the whole population and period, and are usable under the organization's privacy rules. Monthly totals cannot answer every question about individual orders. If the necessary detail is missing, revise the question or agree a way to collect it.

A gym project might use membership records for cancellation rates, gate records for visits before leaving, and optional cancellation comments for themes. Report how many leavers supplied a reason; those comments may describe only a small part of the group.

Early findings can justify sharper questions, such as focusing on two clubs. If that change affects outputs or dates, agree it with stakeholders and update the plan.

An analysis affects more people than the person who requests it. A stakeholder may use the findings, supply data, approve resources, carry out the resulting change or experience its effects. Identifying these people early helps reveal requirements that would otherwise arrive at the end.

Find roles, not just job titles

Executives set strategy and approve budgets. Project managers coordinate the work. Customer-facing teams understand daily problems and often implement changes. Data teams know where records are stored and how they are created. Customers and employees can also be affected by the final decision.

Titles differ between organizations. Ask what each person contributes, controls or needs. A data engineer may have little interest in a marketing decision yet be essential for granting access. A store manager may have no budget authority but know that a proposed process cannot work during busy hours.

Primary stakeholders are directly responsible for the project's goals and usually own the decision. Secondary stakeholders contribute or are affected without owning that outcome. In a repeat-order project, the head of growth deciding on a welcome offer is primary. Finance, support and the data engineer contribute different requirements as secondary stakeholders.

Map influence and interest

The power–interest grid helps choose how to involve each person:

Power means the ability to influence the project. Interest means how much attention someone has in its work or results. Low power does not mean a person's experience is unimportant. The grid guides communication effort; people with practical knowledge can still shape the analysis. Revisit it as interests and responsibilities change.

Discover different meanings of success

For a delivery project, sales may want more orders, finance lower cost per delivery, and support fewer complaints. These measures can pull in different directions. Record them before choosing the final metric set so the decision owner can agree priorities.

For each key stakeholder, ask about the decision and deadline, what they already know, what they need to learn, and the level of detail and certainty required. Ask how they prefer to receive information and who can help with data access or blocked work. A short note for each person can prevent a long chain of missed assumptions.

Finish the first conversation by asking who else will use the findings or be affected. Consulting only the requester risks overlooking the people who must carry out the recommendation. Add them to the stakeholder list before finalizing the plan.

Expectations include the answers stakeholders believe they will receive, their format, the deadline and the level of certainty. State these explicitly. A careful analysis can still disappoint when one person expects a quick estimate and another expects a complete investigation.

Confirm the request in writing

First restate the request and ask what decision sits behind it. Then send a short confirmation, such as:

“By Friday, May 22, you need weekly on-time delivery rates for each zone since January, to decide which zones receive extra drivers in June. Records from the old ordering system before January are out of scope.”

This gives both sides a concrete agreement to correct. Check access, preparation needs, approvals and competing commitments before estimating dates. Include time for problems that are likely to occur. A deadline based only on the final calculation ignores the work needed to obtain usable inputs.

Handle pressure for an early answer

Speed and accuracy often involve a trade-off. A useful response has three parts:

1. Reframe the question around the decision that matters now.
2. Outline the problem and plausible options so the first proposed solution is not the only one considered.
3. Agree a timeline for an initial answer and any fuller follow-up.

A marketing director wants to cancel a month-long 50% discount campaign after one flat sales day. “Is the campaign failing?” may be premature. “Do customers know about the discount?” suggests a first check of campaign reach. Low visibility and unappealing discounted products are different possibilities. Agree a reach update in two days and a product sales review in a week.

The first answer should state its limits. Missing periods, small groups and optional fields can affect the conclusions. Explaining these early lets stakeholders decide whether the preliminary answer is sufficient for their immediate choice.

Keep new requests visible

Scope creep is gradual growth through extra requests that have not been formally agreed. “While you're at it, add returns” may sound small but require different records, definitions and reviews.

Record the request and estimate its effect on work and dates. Ask the decision owner to choose: extend the deadline, replace a lower-priority task or schedule the request later. Update the written scope after the choice. This routine lets priorities change without silently promising more work in the same time.

Report progress in useful pieces

A short status update states what is done, what comes next, risks or blockers, and any decision needed. If access is delayed, explain which milestone is affected and when a decision is needed. Report a changed estimate while the stakeholder still has options. Agreeing to every deadline to please people makes later trade-offs harder to manage.

Good communication helps people use the same evidence for different responsibilities. Decide what the audience needs before choosing the detail, vocabulary and format. Four questions help: Who is the audience? What do they know? What do they need to know? How can you best communicate it?

Put the needed answer first

An executive deciding where to act needs the result and its consequence first. For example: “Repeat orders fell mainly in two regions; the welcome offer should start there.” Keep supporting detail available for questions. A data colleague checking the result needs definitions, filters and sources. The underlying finding stays the same, while the explanation changes with the task.

Give numbers a reference point, such as a target, prior period or comparable group. “The on-time rate is 88%” becomes easier to act on when the audience knows the target is 90%. Explain unfamiliar technical terms where they first matter.

A clear written message opens with its purpose, gives the essential information and ends with a specific request and date. For example:

“Decision needed by Wednesday: zones for extra drivers. The on-time rate is below 90% in zones 3 and 7. Please confirm by Wednesday, June 3, whether both zones should get extra drivers from June 15.”

The reader can tell what happened, why the message was sent and what response is required.

Make meetings produce decisions

Read the agenda, bring the relevant evidence and prepare for likely questions. Arrive on time, pay attention and ask when something is unclear. Afterward, send a short record of decisions and action items. Each action needs an owner and a date; “someone will check the figures” leaves the responsibility unresolved.

Resolve conflicting requests through the shared goal

Sales may ask for more orders while finance asks for lower delivery cost per order. Ask what each request is trying to achieve. Both may serve profitable growth. Propose questions and metrics that show both order growth and cost, then let the decision owner set priorities where the goals genuinely compete.

Reframing moves the discussion from who is right to the problem being solved. If a large analysis is requested overnight but processing needs three days, explain the dependency and arrange a brief conversation about the decision. Agree which answer can come first and which requires more time.

Keep the evidence independent of preferences

Objectivity means reporting what the data shows even when the result disappoints someone. Do not choose periods or filters simply because they make a favored campaign look successful. Explain the agreed comparison and its limitations. A disagreement about priorities belongs with the decision owner; a disagreement about a calculation calls for checking the evidence and definitions.

A row of facts, a calculated measure and a goal for that measure play different roles. Keeping them separate makes a report easier to interpret and a business objective easier to track.

Move from records to measures

Data is a collection of facts, such as one row per order recording its date, customer, items and amount. A metric is a defined numerical measure calculated from data: monthly revenue, average delivery time or the share of orders arriving on time.

A key performance indicator (KPI) is a metric selected to show progress toward an important objective. Every KPI is a metric, but a team also tracks supporting metrics that are not central indicators. The choice depends on the objective. Page load time can help investigate a checkout problem without becoming the organization's main performance measure.

A target adds a desired value and time frame to a metric. “Monthly cancellation rate at or below 4.0% by September 2026” contains a measure, a threshold and a deadline. “Increase sales” leaves the calculation, amount of improvement and timing open.

Combine quantities and experiences

Quantitative data consists of numerical counts, amounts, durations or ratings. It answers how many, how much or how often. Qualitative data describes qualities and experiences, often through words in interviews, comments or support tickets.

Suppose numerical records show that 18% of trial users stop at the payment page. Interviews suggest that a delivery fee appearing at that step surprises people. The numbers show the size and location of the problem; the words suggest an explanation to test. Comments can also be coded into categories and counted, such as the share mentioning price. The original comments and the resulting counts are different forms of evidence.

Choose a useful metric form

Counts and totals describe volume: orders this week or revenue this month. Ratios, rates and percentages put one quantity against another: on-time deliveries as a share of all eligible deliveries. Averages summarize quantities such as order value. Durations measure elapsed time, such as days from sign-up to first purchase.

Totals tend to grow with business size. A larger store may have more complaints simply because it serves more orders. A rate supplies a base for comparison. The metric form should match the question, not just the fields easiest to total.

Calling every reported number a KPI spreads attention too thin. Start with the objective, select a short set of indicators that show progress and keep explanatory measures as supporting metrics.

A metric name alone is not a definition. Two teams may both report “conversion” while counting different things. A complete definition lets people reproduce the same calculation and understand what a change means.

Write the measurement rule

Record the name, the question it answers, its formula, included and excluded records, time window, unit, data source, update frequency, owner and direction of improvement. For a rate, state the numerator and denominator explicitly. Both must refer to the same eligible population.

For the 30-day repeat-purchase rate of March's first-time customers, count customers whose first-ever order was in March and who ordered again within 30 days of that first order. Divide by all customers whose first-ever order was in March. With 775 repeat buyers among 2,500 eligible customers, the rate is 31%. Earlier customers belong in neither count. Allow every eligible customer the full observation window before treating the cohort's result as complete.

Match the base to the thing counted

A site records 40,000 sessions from 25,000 distinct visitors. Orders occur in 800 sessions belonging to 700 visitors.

Both results are valid, but they answer different questions. Counting orders and dividing by visitors gives orders per visitor. Because one person can order several times, that is not the share of people purchasing.

Fix period and edge-case rules

One monthly churn definition counts customers present at the start who cancel during the month, divided by customers present at the start. If 180 of 4,000 opening customers cancel in May, churn is 4.5%. Customers joining during May are excluded from both parts of this particular measure.

Write rules for refunded orders, test accounts, paused subscriptions and the time zone used to assign dates. A record at 23:30 can fall on different dates under different time zones. These choices belong in the shared definition, not in an analyst's private habits.

State the kind of average

Nine orders of $40 and one of $640 total $1,000. The mean is $100 per order, while the median is $40. The mean relates the total to the number of orders; the median describes the middle value after sorting. Specify which average the question needs.

Keep definitions in a shared metric glossary. If “active user” means opening an app to one team and completing an action to another, agree the rule before comparing their results. Date definition changes. Recalculate earlier periods where possible; otherwise mark the break so a recording change is not mistaken for a change in behavior.

Business calculations become easier when you identify the base before calculating. A share uses a whole, a change uses an old value, and a pooled rate uses all eligible records. State the unit and keep full precision until the final result.

Shares, changes and percentage points

Revenue falling from $60,000 to $54,000 is a 10% decrease. The fall is $6,000, divided by the old $60,000. Reversing the comparison gives a rise of about 11.1%, because the base is now $54,000.

Percentage points describe the difference between two percentages. Conversion moving from 2.0% to 2.5% rises by 0.5 percentage points. Relative to the original 2.0%, that is a 25% increase. Write the unit explicitly: “0.5% improvement” could be read as a much smaller relative change.

Successive percentage changes use successive bases. Starting at 1,000, a 10% rise gives 1,100. A 10% fall from 1,100 gives 990, which is 1% below the start. Adding +10% and −10% overlooks the changed base.

Compare and combine rates

Store A has 120 orders with at least one complaint out of 8,000 orders: 1.5%. Store B has 90 out of 3,000: 3.0%. Store B has fewer affected orders but twice the complaint rate. For a percentage of orders, count each affected order once, even if it produces several complaints.

Pool the counts to describe both stores together:

A simple average of 1.5% and 3.0% gives 2.25%, which assigns the small and large stores equal weight. The pooled rate represents every order equally.

Calculate return and cost measures

A campaign costs $20,000 and brings $26,000 in additional gross profit before the campaign cost. Subtracting $20,000 leaves a net gain of $6,000; dividing by the investment gives 30% ROI. State which costs and gain period the definition includes. Additional sales revenue still has production and delivery costs, so using it as the gain while subtracting only campaign cost overstates return.

Average order value is revenue divided by orders. Customer acquisition cost is sales and marketing spending divided by new customers gained in the period: $30,000 for 600 new customers is $50 each. For the same opening group and period, retention is 100% minus churn; 4.5% churn gives 95.5% retention. Label each result with its unit, such as “$50 per new customer.”

A metric can be accurate and still be a poor guide to action. Choose a measure because it helps steer the objective, and check what happens when people try to improve it.

Build a useful set

A useful steering metric is linked to an objective, actionable, clearly defined, comparable over time, available soon enough and based on reliable data. If a number changes, ask what decision that change could support. A cumulative download total may look impressive while giving little information about whether new customers receive value.

Lagging indicators measure outcomes after they happen, such as quarterly revenue or completed cancellations. Leading indicators change earlier and help anticipate an outcome. A fall in weekly subscriber logins may give an earlier warning than churn. An early signal remains something to check against the later outcome, not a guarantee.

Vanity metrics attract attention without guiding a decision. Actionable metrics show where action may help. The share of new downloads leading to a first purchase within seven days, broken down by channel, gives more direction than total downloads since launch.

Trace the drivers

A metric tree connects a top-level KPI with the measures that produce it. Revenue can be broken into visitors, orders per visitor and average order value. If revenue falls, inspect the drivers before proposing a fix. Each driver may then be split by channel, device or another relevant group.

A North Star metric is a single measure selected to represent the value customers receive, such as weekly customers placing an order for a grocery service. Supporting team measures connect their work to that value. The metric still needs a precise definition and suitable checks.

Check indirect measures and side effects

A proxy measures something indirectly. Saving a payment card might be used as a proxy for trust, but the link needs a reason and periodic checking. A rise in saved cards could reflect a changed interface instead of greater trust.

A guardrail metric tracks something that could worsen as a team pushes its KPI. If a support team cuts average call time, repeat calls within seven days and satisfaction scores can show whether customers leave with unresolved problems.

This matters because rewards can change behavior. Goodhart's law describes the risk that people improve a targeted number without improving what it was meant to represent. Judging agents only on calls closed per hour can encourage early endings. Use balanced measures and examine the process behind sudden improvements.

Apply the set to retention

For meal-kit subscriptions, monthly churn can be the lagging KPI, with a baseline of 4.5%. The share skipping two or more deliveries in the last month can be a leading indicator. Churn by plan and region helps locate the problem. Discount cost per retained subscriber is a guardrail against buying retention at excessive cost. Each measure has a different job in the same decision.

A clear boundary makes an analysis finishable. Scope states what the project covers and excludes. A scope of work (SOW) records that boundary together with the outputs, schedule and responsibilities that stakeholders agree before the main work begins.

Describe what will be handed over

A deliverable is an item or task completed for handover, such as a two-page findings summary and a 20-minute presentation. A milestone marks significant progress, such as access being granted or first findings being reviewed. The timeline gives the dates and order. Progress reports state how often updates will be sent and in what form.

A deadline alone leaves these parts unstated. Agree what the recipient will receive and what must happen first.

Name the boundaries explicitly

For falling repeat orders, include first-time customers from January to April 2026, all regions, and linked delivery, pricing and support data. Exclude corporate accounts, loyalty-program redesign and website changes. A finding may eventually suggest redesign work, but delivering that redesign is a different commitment.

Record the background, objective and data needed alongside these boundaries. Scope should be narrow enough for the decision and wide enough to answer it. “Analyze all customer data” gives no practical stopping point.

Separate assumptions from constraints

An assumption is something the plan takes as true before it has been confirmed, such as records being complete from January onward. Check important assumptions early. If one fails, explain which question, date or deliverable is affected.

A constraint is a fixed limit, such as a deadline, available staff, budget or a rule that personal data stays in a secure system. Plan the work within these limits. Labeling an access hope as a certainty hides a risk; treating a fixed deadline as flexible can create an impossible promise.

Put the sequence on dates

A project can agree questions at a May 4 kickoff, gain data access by May 8, prepare and check records by May 15, review first findings on May 22 and hand over the final presentation and summary on June 1. Friday updates let stakeholders see delays before the final deadline. The access milestone matters because preparation depends on it.

Change the agreement deliberately

Record each new request, estimate its effect and ask the primary stakeholder to add it, swap it for existing work or defer it. Then update the SOW. Small additions can otherwise grow into scope creep.

SOW can also mean statement of work in a formal supplier contract, including payment terms. For an internal analytics project, a scope of work is a smaller agreement about the work. Share it and get written confirmation so each stakeholder is working from the same boundaries.

Success needs an agreed test before results arrive. Success criteria state measurable conditions for deciding whether the analysis or a resulting business change has achieved its purpose. Choosing them afterward allows the goal to move to whatever the numbers happen to show.

Separate two kinds of success

Project success concerns the analysis: answering the agreed questions, delivering on time and meeting acceptance conditions. Business success concerns the effect of a later decision on the organization. An analysis that shows a proposed change would be unhelpful can be a successful project. The project is not required to produce a positive business finding.

A deliverable can have acceptance criteria of its own. For a weekly report, these might require refreshing every Monday by 9:00 and matching the finance total for the same week within 1%. These conditions say whether the output is usable, not whether revenue has improved.

Write a measurable criterion

Include the metric, its baseline, a target, a time frame and the measurement method. The baseline is the starting value over a named period. For example:

“Monthly churn rate, baseline 4.5% averaged over January to March 2026, at or below 4.0% in each month from July to September 2026, measured with the standard subscription-database definition.”

The baseline makes the starting point clear. The target says what counts as enough. Requiring each month specifies how long the result must hold. The shared definition prevents a favorable result created by changing who counts.

Set targets using the baseline, recent trend, comparable teams and the expected effect of the action. “Reach 5% churn” is a weak improvement target if the current rate is already 4.5% and lower is better.

Agree the action for each result

A decision rule connects the result to a next step. A team might agree to roll out a welcome offer if its 30-day repeat-purchase rate is at least 3 percentage points above a comparable group without the offer, and stop if the difference is smaller. Write the rule before seeing the outcome.

Include relevant guardrails in success criteria. A retention offer should not count as a success simply because fewer customers leave while discounts become too costly. State the allowed side effect and how it will be checked.

Check the pattern against normal variation

If churn usually ranges from 4.2% to 4.8%, one month at 4.4% sits inside that usual range. Look at several months, the size of normal changes and a comparison group where available before attributing improvement to a new action. Agree how long a target must be sustained. A single favorable month can satisfy a poorly worded target while leaving the business question unresolved.

An analysis plan, or project brief, brings the Ask phase into one shared page. Writing the parts together exposes gaps: a question may lack a metric, a metric may lack records, or a deadline may depend on access that has not been arranged.

Start with the problem and decision

A chain of 12 fitness clubs has 18,000 members. After a February 2026 price increase, monthly cancellations rose from 3.0% to 4.2%: about 540 to 756 members at that membership level. The operations director must decide by September 30 whether to introduce a cheaper off-peak membership.

The director is the primary stakeholder. Club managers, finance, member services and the data team contribute knowledge, cost checks and access. Write these roles beside the decision so approval and consultation are clear.

Connect questions to evidence

The project combines finding patterns across months and clubs, spotting unusually large increases and identifying themes in cancellation reasons. Two questions are:

• “From January to August 2026, what was the monthly cancellation rate at each club, and which clubs rose by more than 1.5 percentage points after the price increase?”
• “Among members who canceled from March 1 to August 31, what share visited fewer than four times in their final month, compared with members who stayed?”

Define cancellation rate using opening members who cancel divided by opening members, excluding staff memberships. Measure visits as recorded gate entries per member per calendar month. Use cancellation-form responses for themes and report the share of leavers who answered the optional field.

Each question now has a measure and a source. The themes can suggest reasons, while the response share shows how much of the leaver group supplied that evidence.

Fix scope, timing and success

Include all 12 clubs, January to August, and membership, visit and cancellation-form records. Exclude personal-training sales and setting the price of the proposed plan. These boundaries keep the work focused on the director's decision.

Project success means presenting findings and a recommendation by September 15. If the off-peak plan is introduced, business success means cancellation at or below 3.2% within six months. The main evidence risk is that only a minority of leavers supply a reason. Record that risk alongside the plan for checking it, so themes are not presented as everyone's experience.

Use the plan through the project

A practical one-page layout has fields for the problem and decision, stakeholders, problem types, questions, metric definitions, data, scope, success criteria, timeline and main risks. Each field should support the same decision. Remove unrelated requests or give them an agreed later phase.

Share the plan with the primary stakeholder before collecting data. At milestones, use it to check whether the work still answers the agreed questions. Record changes with a reason and date. The final presentation should answer those questions and show how the recommendation follows from the evidence.