By Jason Hreha

Guide

Employee Incentives: Designing for Quality, Quantity, and Cooperation

By

Published 18 min read

Your customer support team is slow. Customers wait about nine hours for a first answer, longer than a working day, and leadership has given you a budget to fix it. The obvious idea is a bonus: pay each agent a little extra for every ticket they close above a weekly target.

Before you announce it, ask two questions. What behavior will this scheme actually reward? And what useful work might it push aside? This guide answers both with one worked example and a worksheet.

Employee incentives are rewards or penalties tied to something an employee does or achieves: a bonus, a commission, a prize, public recognition. The incentives glossary entry gives the general definition. This guide is about one decision: designing an incentive scheme, or incentive program, for a particular team and job.

Start with the behavior, not the reward

Here is the running example. It is hypothetical.

Dana manages 16 support agents at a software company, in four pods of four. Customers wait a median of nine hours for a first reply, and many problems take several exchanges to fix. Leadership has approved $30,000 a year for an incentive. Dana’s first draft pays $3 for every ticket an agent closes above 60 in a week.

If every agent closed ten extra tickets a week for 50 weeks, that would cost about $24,000. Nothing caps it, though. If agents found ways to close more tickets, the cost would rise with them.

The draft rewards one number: tickets marked closed. But slow first replies are a symptom. Dana wants customers’ problems solved quickly and correctly. Write three layers down separately:

  • The goal: what the organization needs. Problems solved quickly and correctly.
  • The target behaviors: what people must do to reach it. Diagnose accurately, fix or escalate well, write clear answers, document fixes.
  • The measured outcome: the number you will pay on. Tickets closed.

People will attend to the layer you pay on. Steven Kerr made this point in a 1975 paper, “On the Folly of Rewarding A, While Hoping for B”. People look for what is rewarded and then do it, or “at least pretend to do” it. In one of his cases, an insurance claims office tracked complaints from policyholders. Underpayments drew complaints and overpayments often did not, so new hires learned an informal rule: “When in doubt, pay it out!”

I have argued for the same order in behavior change generally: define the problem, choose the behavior that solves it, and only then design the intervention (Behavior Market Fit; my 2025 notes on human behavior). An incentive is an intervention. It cannot rescue a scheme that rewards the wrong behavior.

What the evidence says about incentives and motivation

One popular claim says money is the main lever. Another says rewards destroy people’s real motivation. The research supports neither in its simple form.

Incentives tend to move quantity more than quality

In a 1998 meta-analysis of 39 studies, G. Douglas Jenkins and colleagues found that financial incentives had a corrected correlation of .34 with performance quantity and were not related to quality.

Christopher Cerasoli, Jessica Nicklin, and Michael Ford (2014) combined 183 samples and more than 212,000 people, most of them students. Intrinsic motivation, doing a task because it is interesting or satisfying, predicted performance whether or not incentives were present, and the link was if anything stronger in the 42 workplace samples. It was more closely linked with quality (.35) than with quantity (.26). Combined with the Jenkins estimates, incentives explained more of the variation in quantity, and intrinsic motivation much more of the variation in quality.

Most of the studies in Cerasoli’s review were correlational, and not every review agrees. Yvonne Garbers and Udo Konradt’s 2014 meta-analysis of 146 studies found a larger effect of individual incentives on qualitative measures (g = 0.39) than on quantitative ones (g = 0.28). Treat the pattern as a reason to measure quality separately, not as a law. For Dana, a per-ticket bonus will probably raise the count. Better answers depend on things it does not pay for: knowledge, care, and pride in solving hard problems.

Do rewards destroy intrinsic motivation?

Not as a general rule. Edward Deci, Richard Koestner, and Richard Ryan pooled 128 experiments in 1999. Expected tangible rewards reduced later free-choice engagement with a task (d = −0.36), mainly for tasks people already found interesting. Positive feedback increased it (d = 0.33).

Judy Cameron, Katherine Banko, and W. David Pierce reanalyzed 145 studies in 2001 and found no reliable overall effect on free-choice engagement. Negative effects appeared on high-interest tasks when rewards were tangible, expected, and only loosely tied to performance.

The two teams disagree about how to group reward conditions and which studies to include. Their findings overlap on where the risk sits: a predictable, tangible payment for doing work people already enjoy. In both analyses, praise tended to help. Most of these experiments measured free-time engagement, not job performance.

The motivational crowding out entry explains the proposed mechanisms. Also check your assumptions: people tend to overweight external rewards when explaining others’ motivation, the extrinsic incentives bias. Some of Dana’s agents may care more about cracking a hard problem than about $3.

Do bigger bonuses buy better work?

In 2022 I argued, in an essay for HR leaders, that large cash bonuses can hurt work that requires thought. Its sharpest line was this: “If your task requires even a modicum of thought, the bigger the monetary bonus the lower the performance.” The essay drew on Dan Ariely’s research.

I no longer rely on that research. In 2021, the journal PNAS retracted a 2012 paper he co-authored. The researchers behind the blog Data Colada had reported what they called very strong evidence that the data in its field experiment were fabricated. Ariely said that the insurance company running the study had collected and prepared those data before sending them to him.

In September 2026, Psychological Science retracted a 2002 paper he co-authored, saying the editor could no longer attest to the reliability of its findings. Both authors agreed to that retraction. Neither retraction concerns the bonus study my essay cited. I set that research aside because of the wider record, not because that study was withdrawn.

A larger test published in 2023, by a team that did not include Ariely, points to a narrower claim. Benjamin Enke, Uri Gneezy, and colleagues (2023) offered 1,236 university students in Nairobi a bonus worth more than a month’s income for solving reasoning problems. They spent about 40% longer on the problems, but their accuracy barely improved on most tasks, and it did not reliably fall. The reviews by Jenkins and by Cerasoli, above, fit the same pattern: incentives move how much people produce more reliably than how well they do it.

So my advice now is narrower. Do not expect a bigger cash prize to buy better thinking. I can no longer say, on evidence I trust, that it makes thinking worse.

The case for noncash rewards needs the same care. In a field experiment by Sebastian Kube, Michel André Maréchal, and Clemens Puppe (2012), people recruited on a German university campus for a one-time, three-hour job cataloging library books received a surprise thank-you gift. A thermos bottle worth €7, about a 20% addition to their pay, raised output by about 25%. The same €7 in cash raised it by about 5%, a difference too small to be reliable. The same €7, with the note folded into an origami shirt, raised output by about 30%.

These were small groups and a one-off job, and the gifts were thanks, not rewards for performance. The lesson is narrow: how a reward is given can matter as much as what it is worth.

For Dana, the bigger risk is not the size of the bonus but what it pays for.

Check who controls the measured outcome

A ticket count depends on more than effort: the queue an agent works, how hard its tickets are, whether a product bug floods the team with duplicates, and the season. An agent on large enterprise accounts may close fewer tickets because the problems are harder.

Paying on outcomes people only partly control turns rewards into partly a lottery, which people may reasonably see as unfair. Economists describe this tradeoff in the principal-agent problem. For each measure, list what besides effort moves it. If those factors are large, adjust the measure (compare agents within the same ticket type), measure the team, where routing evens out, or use the measure only to spot problems.

Map how people could hit the number without doing the work

Any measure you pay on attracts effort toward the number itself. Goodhart’s law describes the result: a good indicator can stop being one once people are rewarded for moving it. A perverse incentive is the worst case, where the reward pays people to make the problem worse. The Goodhart entry describes how sales goals at Wells Fargo led to millions of accounts or products provided under false pretenses or without customers’ consent.

In Dana’s draft, agents could:

  • close tickets before the customer confirms the fix
  • split one problem into several tickets
  • take easy tickets and leave hard ones in the queue
  • escalate difficult tickets early so they leave the agent’s count
  • ask customers to open a new ticket for a follow-up question

Each route raises the count while leaving customers no better off. List your own routes before launch, and add a check for each. A reopen rate within 14 days catches premature closing. A weekly audit of randomly chosen tickets catches poor answers. The share of tickets closed within minutes of the first reply flags suspicious speed.

List the work the scheme might displace

Some of the team’s most valuable work never becomes a closed ticket. Agents write knowledge-base articles, coach new hires, handle angry escalations, and report recurring bugs so whole categories of tickets disappear.

Bengt Holmström and Paul Milgrom explained the risk in their 1991 analysis of jobs with several tasks. Paying for tasks that are easy to measure pulls attention away from tasks that are hard to measure. Their example was the debate over paying teachers for test scores: critics warned that teachers would neglect curiosity and communication skills. In that situation, they showed, a fixed wage can be the better contract even when good output measures exist. If the displaced work matters, reward it too, measure the team, or weaken the incentive.

Look for costs that arrive later

A ticket closed too early returns as a repeat contact next week. A frustrated customer cancels at renewal months later. An agent who chased the target for a quarter burns out. Kerr noted that many reward systems “pay off for short run sales and earnings only.” So measure over a window long enough to catch delayed failures, such as counting a ticket as resolved only if it stays closed for 14 days, and pay after that window closes.

Decide between individual and team rewards

The right choice depends on how interdependent the work is and whether individual contributions can be seen.

Individual rewards give each person a clear line from effort to payoff, but they discourage helping.

Pay relative to coworkers changes the social cost of effort. On a UK fruit farm studied by Oriana Bandiera, Iwan Barankay, and Imran Rasul (2005), pay depended on output relative to the field average, so working harder lowered coworkers’ pay. Productivity was at least 50% higher after a switch to simple piece rates. Under the relative scheme, workers held back most beside friends who could see them.

Team rewards encourage helping and even out luck in routing. Garbers and Konradt found a positive average effect across 30 studies of team rewards, with rewards divided by contribution outperforming equal splits. The known risk is social loafing, when individual effort cannot be seen or seems unnecessary.

Support work is interdependent: one agent’s article speeds up everyone. That points Dana toward a team reward, plus a way to make individual contributions visible.

Keep pay adequacy separate

Is total pay fair and competitive? Does the structure of pay reward the right behavior? This guide addresses the second question. An incentive cannot fix pay that is too low, and recognition is no substitute for money people need. If good agents are leaving for better-paid jobs, settle the pay level first.

Choose the reward type, size, and timing

  • Type. Cash suits outcomes the whole team can affect and that can be verified. Recognition suits contributions that are visible but hard to price. Dana uses a cash team pool, plus monthly recognition naming knowledge-base articles other agents used.
  • Size. Start from the budget and the outcome. Dana’s $30,000 becomes a $7,500 quarterly pool, about $470 per agent, kept modest so it does not stand in for base pay.
  • Timing. Pay after delayed costs can appear. Dana pays quarterly, after the 14-day window for the quarter’s last tickets.

Then state exactly what earns the reward. Dana’s hypothetical rule has one target and two guardrails. The target is a median time to a confirmed resolution at least 10% below the four-week baseline. The guardrails are a 14-day reopen rate within two percentage points of baseline and an average audit score at or above baseline.

Meet all three and the team earns the pool. Miss the target and the pool is withheld. Hit the target but breach a guardrail and the pool is withheld and the scheme reviewed.

Redesign Dana’s scheme

Question First draft Redesign (hypothetical)
What is paid for? Tickets closed above 60 a week, per agent Team median time to a confirmed resolution at least 10% below baseline
Who controls it? Partly the agent, partly routing and ticket mix The team, across all queues
Gaming routes Premature closing, splitting, cherry-picking Reduced, and watched through guardrails
Guardrails None Reopen rate within two percentage points of baseline; audit score at or above baseline; very fast closures tracked
Displaced work Knowledge base, coaching, bug reports Recognized monthly
Timing Weekly, at closing Quarterly, after the 14-day window
Cost About $24,000 a year if each agent closes ten extra tickets a week; no cap $7,500 a quarter, split by hours worked; $30,000 a year at most

Splitting by hours rather than measured contribution departs from Garbers and Konradt’s finding, but it avoids reintroducing individual counting. That judgment is worth testing.

Pilot it with stopping rules

Test a new scheme in part of the team, compare, and decide in advance what would make you stop. For Dana:

  1. Choose the pilot group fairly. Pair the four pods by ticket mix and flip a coin within each pair. The two chosen pods (eight agents) get the new scheme; the other two continue as before.
  2. Record a baseline. Collect four weeks of data on every measure first.
  3. Run one full payout cycle. The pilot lasts one quarter (13 weeks), with a $3,750 pool, half the quarterly pool for half the team. The last tickets need 14 days to count as confirmed, so results are complete in week 15.
  4. Track the guardrails weekly in both groups. Reopen rate, audit scores (at least ten tickets per pod each week), very fast closures, and complaints. For a month afterward, watch whether documentation and resolution time hold up.
  5. Set stopping rules in advance. Pause and review if the pilot pods’ reopen rate rises three or more percentage points more than the comparison pods’, each against its own baseline, for two weeks running. Also pause if their two-week average audit score falls below baseline. Stop immediately on any evidence of falsified closures.
  6. Set the decision rule. In week 15, extend the scheme only if the pilot pods’ median time to a confirmed resolution improved more than the comparison pods’, relative to baseline, and no guardrail was breached. Otherwise, revise or drop it.

With four pods, the pilot cannot prove the scheme works; pod differences could explain a modest gap. Its job is to catch gaming cheaply and show whether the direction is promising. If more rides on the decision, the research methods guide explains how to design a randomized comparison with enough power and why to record the plan first.

When a simple per-unit incentive fits

Change the conditions and the answer changes. When Safelite Glass moved its windshield installers from hourly pay to piece rates with a guaranteed minimum, Edward Lazear (2000) found a 44% increase in output per worker, and profits rose too. In his own summary, only about half of that gain came from the incentive effect itself. The rest came from changes in the workforce, mostly in who was hired, which matters whenever you read a before-and-after figure. Lazear described the work as a simple process with an easily monitored product of a single standard.

Cerasoli and colleagues suggest that straightforward, repetitive tasks suit close links between pay and output. Garbers and Konradt found the opposite pattern: individual incentive effects were smaller on less complex tasks (g = 0.19). So the better test is measurability, not simplicity. A per-unit incentive is a reasonable candidate when:

  • output is easy to count
  • quality can be checked reliably
  • the person controls most of what moves the number
  • little valuable work sits outside the measure
  • coworkers do not depend heavily on one another

Dana’s job fails most of these tests; Lazear’s description of windshield work suggests it met at least the first two. As I put it in my 2025 notes, strong situations with clear incentives, constraints, and feedback can produce large effects when they match people’s traits and skills. The match is the hard part.

Try it: a sales commission

A hypothetical software company pays its sales team 10% of a new customer’s first-year contract value when the contract is signed. What would you check?

An answer. Gaming: deep discounts to close quickly, poor-fit customers who will leave, deals pulled into the current period. Displaced work: handoff notes, help for colleagues, attention to existing customers who earn no commission. Control: lead quality and pricing rules sit partly outside the representative’s hands. Delayed costs: a poor-fit customer may cancel six months later, long after the commission was paid.

A redesign might pay part of the commission after the customer has paid and stayed six months, limit discounts without approval, and add a small team component tied to renewals. The reasoning is the same as Dana’s, though the job differs.

Use the worksheet

The incentive design worksheet follows this guide’s sections in order. The full worksheet below includes a filled example of Dana’s scheme.

Incentive design worksheet

Use this worksheet with the guide Employee Incentives: Designing for Quality, Quantity, and Cooperation. Its twelve steps follow the guide’s sections in order. Fill it in before you announce a scheme. A step you cannot answer is the part of the scheme to work on next.

Part 1 is blank for your own scheme. Part 2 shows the guide’s hypothetical support-team example, filled in.

Part 1: Your scheme

Team and job: ______________________________

Scheme being considered: ______________________________

Budget: ______________________________

1. Goal and target behaviors

What does the organization need? ______________________________

What must people do to get there? List the target behaviors.

2. Measured outcome

What number would the scheme pay on? ______________________________

Which target behaviors does this number capture, and which does it miss?

  • Captures: ______________________________
  • Misses: ______________________________

Is the cost capped? ☐ Yes ☐ No. If no, what happens to cost if people find ways to move the number? ______________________________

3. Quantity and quality

Does the measure count output, judge its quality, or both? ______________________________

How will you measure quality separately? ______________________________

4. Who controls the measure

What moves this number besides the person’s own effort? (Routing, case mix, other teams, seasons, products, customers, luck.)

How large are these outside factors? ☐ Small ☐ Moderate ☐ Large

If moderate or large, choose one: ☐ Adjust the measure ☐ Measure at team level ☐ Use the measure only to spot problems

How: ______________________________

5. Gaming routes and checks

List every way someone could raise the number without doing the target behaviors. Add a check for each.

Gaming route Check that would catch it

Could anyone earn the reward by making the underlying problem worse? ☐ No ☐ Yes: ______________________________

6. Displaced work

Which valuable activities will compete with the paid one for time and attention?

How will you protect them? ☐ Reward them too ☐ Measure the team ☐ Weaken the incentive ☐ Other: ____________

7. Delayed costs

What could go wrong after the payout? (Repeat work, cancellations, burnout, neglected maintenance.)

How long must you wait to see these costs? ____________

Will the measurement window and payout timing cover that period? ☐ Yes ☐ No: change to ____________

8. Individual, team, or both

How much do people depend on one another’s work? ☐ Little ☐ Some ☐ A lot

Can individual contributions be seen? ☐ Yes ☐ Partly ☐ No

Choice: ☐ Individual ☐ Team ☐ Both

Will anyone’s pay depend on output relative to coworkers? ☐ No ☐ Yes: ______________________________

If team: how will the reward be divided? ______________________________

Social loafing risk and how you will handle it: ______________________________

9. Pay adequacy check

Is total pay fair and competitive for this job? ☐ Yes ☐ No ☐ Unknown

If no or unknown, resolve the pay level before designing the incentive structure. Notes: ______________________________

10. Reward type, size, and timing

Type: ☐ Cash ☐ Noncash (voucher, gift, time off) ☐ Recognition ☐ Mix. Why this type? ______________________________

Size, and how it was set from the budget: ______________________________

Paid when (after the delayed costs in step 7 can appear): ______________________________

Earning rule. What exact result earns the reward, against what baseline, and which guardrail limits must hold?

  • Target: ______________________________
  • Guardrail limits: ______________________________
  • If the target is missed: ______________________________
  • If the target is met but a guardrail is breached: ______________________________

Is anyone being paid a predictable, tangible reward for work they already enjoy? ______________________________

11. Pilot design

Pilot group, and how it was chosen: ______________________________

Comparison group: ______________________________

Baseline period: ______________________________

Pilot length (at least one full payout cycle plus any confirmation window): ______________________________

Pilot reward pool: ______________________________

Primary outcome (decide now; do not swap later): ______________________________

Guardrails tracked weekly in both groups, with sample sizes:

What will you watch after the pilot ends? ______________________________

12. Stopping and decision rules

Pause and review if (measured against each group’s own baseline): ______________________________

Stop immediately if: ______________________________

Decision point (week): ____________. Extend the scheme only if: ______________________________

Otherwise: ☐ Revise ☐ Drop

Part 2: Filled example (hypothetical)

This example is the support team from the guide. Dana, the manager, and every number below are hypothetical illustrations, not results.

Team and job: 16 customer support agents at a software company, in four pods of four.

Scheme being considered: $3 per ticket closed above 60 per agent per week.

Budget: $30,000 a year.

1. Goal and target behaviors

What does the organization need? Customers whose problems are solved quickly and correctly. Today the median wait for a first reply is nine hours, a symptom of the larger problem.

Target behaviors:

  • Diagnose problems accurately.
  • Fix them, or escalate them well.
  • Write clear answers.
  • Document fixes so other agents can find them.
2. Measured outcome

What number would the scheme pay on? Tickets marked closed, per agent.

  • Captures: speed and volume of closing.
  • Misses: whether the fix worked, answer quality, documentation, help to teammates.

Is the cost capped? No. About $24,000 a year if each agent closes ten extra tickets a week for 50 weeks; more if agents find ways to close more.

3. Quantity and quality

The draft counts output only. Measure quality separately with a weekly audit of randomly chosen tickets (at least ten per pod), scored against a written rubric, and with the 14-day reopen rate.

4. Who controls the measure

Outside factors:

  • Queue assignment. Enterprise tickets are harder.
  • Product bugs that create waves of duplicate tickets.
  • Seasonal volume.

How large? Large.

Choice: Measure at team level, where routing and ticket mix even out.

5. Gaming routes and checks
Gaming route Check that would catch it
Closing before the customer confirms the fix 14-day reopen rate
Splitting one problem into several tickets Audit sample; duplicate-ticket report
Taking easy tickets, leaving hard ones Queue age of unassigned tickets
Escalating hard tickets early Escalation rate and engineering feedback
Asking customers to open new tickets for follow-ups Repeat-contact rate within 14 days
Closing suspiciously fast Share of tickets closed within five minutes of the first reply

Could anyone earn the reward by making the problem worse? Yes. Premature closing creates repeat contacts, which add more tickets to close.

6. Displaced work
  • Writing and updating knowledge-base articles.
  • Coaching new hires.
  • Handling angry escalations.
  • Reporting recurring bugs to engineering.

Protection: Measure the team, and recognize documentation. Each month, the manager names knowledge-base articles other agents used.

7. Delayed costs
  • Repeat contacts the following week.
  • Cancellations at renewal, months later.
  • Burnout after a quarter of chasing a count.

Wait: at least 14 days for repeat contacts; renewal effects take longer and need separate tracking.

Window and timing: Count a ticket as resolved only if it stays closed for 14 days. Pay quarterly, after that window.

8. Individual, team, or both

Dependence: A lot. One agent’s article speeds up everyone.

Individual contributions visible? Partly.

Choice: Team pool, plus individual recognition. No pay relative to coworkers.

Division: Split by hours worked. (Splitting by measured contribution would bring back individual counting; worth testing.)

Social loafing: Make documentation and audit scores visible by person, without paying per unit.

9. Pay adequacy check

Is total pay fair and competitive? Assumed yes for this example. If agents were leaving for better-paid jobs, Dana would raise base pay first.

10. Reward type, size, and timing

Type: Cash team pool, because the whole team affects and can verify the outcome; recognition for documentation, which is visible but hard to price.

Size: $7,500 a quarter ($30,000 a year), about $470 per agent per quarter, kept modest so it does not stand in for base pay.

Paid when: Quarterly, after the 14-day window for the quarter’s last tickets.

Earning rule:

  • Target: median time to a confirmed resolution (the ticket stays closed for 14 days) at least 10% below the four-week baseline.
  • Guardrail limits: 14-day reopen rate within two percentage points of baseline; average audit score at or above baseline.
  • If the target is missed: pool withheld; nothing else changes.
  • If the target is met but a guardrail is breached: pool withheld and the scheme reviewed.

Interest: Several agents enjoy solving hard problems. The redesign avoids a per-unit payment for that work.

11. Pilot design

Pilot group: Two of four pods. Pair the pods by ticket mix, then flip a coin within each pair.

Comparison group: The other two pods, on the current arrangement.

Baseline period: Four weeks before the pilot, on every measure.

Pilot length: One quarter (13 weeks). Results are complete in week 15, after the 14-day window for the last tickets.

Pilot reward pool: $3,750, half the quarterly pool for the eight pilot agents.

Primary outcome: Median time to a confirmed resolution.

Guardrails tracked weekly in both groups:

  • 14-day reopen rate.
  • Audit score, at least ten tickets per pod per week, compared as two-week averages.
  • Share of tickets closed within five minutes of the first reply.
  • Complaints.

After the pilot: Watch documentation and resolution time for the following month.

Limit: With four pods, pod differences could explain a modest gap. The pilot is for catching problems and checking direction, not proof.

12. Stopping and decision rules

Pause and review if: The pilot pods’ reopen rate rises three or more percentage points more than the comparison pods’ rate, each against its own baseline, for two weeks in a row. Also pause if the pilot pods’ two-week average audit score falls below their baseline.

Stop immediately if: There is any evidence of falsified closures.

Decision point: Week 15. Extend the scheme only if the pilot pods’ median time to a confirmed resolution improved more than the comparison pods’, relative to baseline, and no guardrail was breached.

Otherwise: Revise or drop.

Step Answer before launch
1. Goal and target behaviors What does the organization need, and what must people do?
2. Measured outcome What number will you pay on, and what does it miss?
3. Quantity and quality How will you measure quality separately?
4. Who controls the measure What besides effort moves it?
5. Gaming routes and checks How could people hit it without the work, and what catches each route?
6. Displaced work What valuable work could it push aside?
7. Delayed costs What could go wrong after the payout?
8. Individual, team, or both How interdependent is the work?
9. Pay adequacy check Is total pay fair before you design the structure?
10. Reward type, size, and timing What reward, how large, when, and earned by what rule?
11. Pilot design Which group, baseline, length, outcome, and guardrails?
12. Stopping and decision rules What makes you pause, stop, or extend?

So, what will your scheme reward, and what might it push aside? You will not know until you write down the behavior you want, the number you will pay on, and every route between them. Start with step 1, and test the result in part of your team before you scale it.

References