A Salary Formula for Small Startups Where Every Developer Uses AI

Abstract

Founders of small software companies usually set pay at hiring and move it by negotiation. When every developer works with AI coding tools, hours at the desk and visible effort say little about what a person delivers. We propose a transparent salary formula for startups of 5 to 10 employees. Pay is anchored to the employer's market for the role, blended with the local market through published cost-of-living tiers and a rule-based weight. A six-month delivery score, taken from our earlier developer scorecard and smoothed for small samples, sets the person's position in a pay band of 0.85 to 1.15 of that anchor. Pay moves toward its target in steps of at most 6 percent per half-year, never down, inside a merit pool the founder sets. We state the formula's properties, test it in a synthetic simulation of an eight-person team over ten reviews, and list its threats to validity. This is a proposal with a simulation, not a validated result.

1 Introduction

Two developers work on the same team. One leaves at five. The other stays until midnight, most days. Over six months the first ships more, with fewer defects. Who gets the raise?

Most founders know the answer they would like to give. Fewer have a rule that gives it, and the rule they have is often hours, presence or whoever argues best. When every developer works with AI, none of those tracks output: an agent can run while someone sleeps.

Pay set by negotiation also has a cost we described in the leave model: when nobody writes down what earns a raise, people fill the gap with hours. A salary formula is the written version. It tells the employee what the number depends on and lets them check it.

This article gives no salary figures, only a way to compute one from your own market data, your own tiers and the six-month score from the developer scorecard. The reader is a founder with 5 to 10 employees, some of whom may hold shares, options or phantom shares. Contractors are out of scope.

This is the third article in a series on running a small team that works with AI. The first, the developer scorecard, defines how to measure delivery. The second, the leave model, defines when people work, rest and are away. This one turns the score into pay and keeps hours out of the formula, as the leave model keeps them out of the rules for time off. Every number in the worked example and the simulation is invented or synthetic and is labelled where it appears, and nothing here is legal or tax advice.

1.1 Design requirements

1.2 Research questions and hypotheses

• RQ1. Can a score-driven formula meet R1 to R7 at the same time, and where do the requirements conflict?
• RQ2. How does the formula behave over several reviews under score noise and a moving market?

We state three hypotheses and test them only on synthetic data in Section 6:

• H1. Smoothing the score toward its previous value reduces review-to-review movement in target pay and reduces the upward drift that a never-down rule creates under noise.
• H2. A merit pool keeps payroll growth within budget, at the cost of slower convergence when the pool is small relative to market growth.
• H3. With a 6 percent cap and a stable market, a person well below target reaches it in a predictable number of reviews.

2 Related work

2.1 Published salary formulas

Two companies are the best-known publishers of formula-based pay. GitLab's compensation calculator multiplied a San Francisco benchmark by a location factor, a level factor, a compa-ratio and an exchange rate [24]. GitLab retired the calculator on 30 May 2025, as part of a job architecture project, and its handbook now marks the earlier framework as historical [24]. GitLab argued for local rates on competitive grounds: uniform pay would either raise costs enough to limit hiring or lower pay enough to hurt retention [25].

Buffer published its first formula in December 2013: a job-type base times seniority times an experience multiplier from 1.0 (junior) to 1.3 (master), plus a location amount [26]. Its 2017 version multiplied a San Francisco median benchmark by a cost-of-living multiplier in three bands (100, 85 and 75 percent), a role multiplier and an experience factor [27]. Buffer's current page lists two bands, Global at 90 percent and High at 100 percent, in use since February 2022 [28].

Both start from a market benchmark and a level, and adjust for location. Neither includes a direct measure of delivered output. We did not find a published formula that does, though our search was not exhaustive.

2.2 Market data on AI skills

Market data suggest a premium for AI skills, with wide disagreement on size. Lightcast reported in July 2025, from more than 1.3 billion job postings, that postings asking for AI skills advertised salaries 28 percent higher [29]. PwC's 2025 AI Jobs Barometer reported a 56 percent average wage premium for workers with AI skills in the same occupation [30]. Levels.fyi reported in May 2023 that AI engineers earned 8 to 12.5 percent more than non-AI software engineers at the same company and level, without published sample sizes [31]. These are posting-based or self-reported comparisons, and none shows that AI skill causes higher pay. They are inputs to M, not a method.

2.3 The merit matrix

A merit matrix crosses a performance rating with the compa-ratio (pay divided by the band midpoint), so a strong performer low in the band gets a larger raise than one near the top [33]. Our source is a vendor glossary. The formula is a continuous version of this, with the rating replaced by a score.

2.4 Incentive effects of pay for performance

The evidence is mixed and depends on the task. In a field setting, Lazear found that moving windshield installers from hourly pay to piece rates raised output per worker by about 44 percent, through both effort and selection [2]. That is a single, countable task. Holmstrom and Milgrom showed that when an agent has several tasks and only some are measurable, strong incentives on the measured ones pull effort away from the rest, and a fixed wage can be optimal [3]. Kerr's older essay makes the same point in management language [6]. Software work is multi-task, and review, mentoring and maintenance are poorly measured.

On motivation, Deci, Koestner and Ryan's meta-analysis of 128 experiments found that performance-contingent rewards undermined free-choice intrinsic motivation [5]. Huffman and Bognanno found, in a within-subject experiment with paid workers in a real work setting, that crowding out extends to high-powered incentives, with a minority of workers showing the opposite effect [1]. Gneezy, Meier and Rey-Biel review when incentives backfire [4]. Together these argue for a bounded incentive and slow-moving base pay.

2.5 Fairness and transparency

Colquitt separates distributive, procedural, interpersonal and informational justice [8]. A formula addresses mainly the procedural and informational parts; whether outcomes are fair is a separate question.

Pay transparency cuts both ways. When Card, Mas, Moretti and Saez told University of California employees how to look up colleagues' pay, those below the median for their unit and occupation reported lower satisfaction and more job search; those above were unaffected [9]. Cullen and Pakzad-Hurson found that US state laws protecting workers' right to discuss pay were followed by wages about 2 percent lower, consistent with employers bargaining harder [10]. Cullen's review distinguishes horizontal transparency (seeing peers' pay) from vertical and cross-firm transparency, which have more positive effects [11]. Our design publishes the formula, tiers and bands, not individual salaries.

2.6 Why nominal pay rarely falls

In Kahneman, Knetsch and Thaler's telephone surveys, people judged a wage cut that exploits a shift in demand as unfair, though they accepted cuts when profits were threatened [12]. Bewley's interviews with more than 300 managers, labour leaders and advisers in the early 1990s recession found that managers avoided pay cuts mainly to protect morale [13]. Grigsby, Hurst and Yildirmaz, using administrative payroll data, found that only about 2 percent of job-stayers had a nominal base wage cut in a given year [14]. The never-down rule follows common practice. Section 6 measures its cost.

2.7 Hours, crunch and health

A WHO and ILO analysis estimated 745,000 deaths from stroke and ischemic heart disease attributable to long working hours in 2016, and estimated that working 55 hours or more a week was associated with a 35 percent higher risk of stroke and a 17 percent higher risk of dying from ischemic heart disease, compared with 35 to 40 hours [17] [18]. Pencavel's study of British munitions workers in the First World War found output proportional to hours below about 49 hours a week and rising at a decreasing rate above that [16]. That is factory work a century ago: read the shape, not the numbers.

Software-specific evidence is thin. The Game Outcomes Project surveyed game developers and found crunch negatively correlated with project outcomes; its authors note the results were not peer-reviewed [34]. They considered whether crunch is a response to troubled projects rather than a cause, and report no sign that it rescued those projects, but the data are correlational. The nonprofit Take This published a white paper in 2016 on the harm of long-term, cultural crunch [35]. Neither is a controlled study.

2.8 Credibility weighting

Actuarial credibility theory weights a group's own experience against a prior according to how much data it has [15]. We use that form of weight to shrink a six-month score toward the previous one.

3 Model and notation

All pay figures are annual base pay in one reference currency: the employer's, or the euro or US dollar where the employer's currency is volatile (Section 9.2). A review happens every six months, indexed t.

Assumptions. A1: the scorecard is run as its article describes, with scope and capacity written down before work starts. A2: M comes from one named source and is revised once a year. A3: the person is an employee, not a contractor. A4: the team is small enough that the founder can see every review.

4 The formula

The formula has five parts: anchor, score, target, step and bonus. All coefficients are starting values that we chose, not values derived from research.

4.1 Geography: the weight w and the tiers L

The anchor blends the employer's market with the local one:

w = 0.25 + 0.25 x (scarcity + overlap + customer)
A = M x (w + (1 - w) x L)

Each of the three inputs takes the value 0, 0.5 or 1, set once per role and published:

So w runs from 0.25 to 1.0. A scarce, high-overlap, customer-facing role is paid at the employer's market rate wherever the person lives; a role the local market could fill is weighted toward it. A published rule, not a case-by-case judgement, sets w, as R6 requires.

Tiers. L comes from a published price index, not a negotiation. We suggest the World Bank's price level index (GDP) from the World Development Indicators [23]. Take the ratio of the person's country to the employer's country and place it in a tier:

L is capped at 1.00, so a person in a more expensive country than the employer's is anchored to the employer's market. Tiers are revised once a year, at the same review as M. A price index measures prices, not developer wages. If you have a trustworthy local salary source for the role, you may use the ratio of local median pay to M instead, but say which you used and use the same source for everyone.

4.2 Score smoothing and the new-hire prior

The scorecard's own article warns that three deliverables make a noisy trend. We shrink each new score toward the previous smoothed score, more strongly when there are few deliverables:

z     = n / (n + 3)
S_t   = z x Perf_t + (1 - z) x S_(t-1)       when n is 3 or more
S_t   = S_(t-1)                              when n is under 3: no valid score
S_0   = 0.4 + 0.6 x p_offer                  for a new hire
p_offer = clamp((P_offer / A - 0.85) / 0.30, 0, 1)

With 3 deliverables the new score gets half the weight; with 12, 80 percent. The constant 3 is what the prior is worth in deliverables, a choice tested in Section 6.

A new hire's prior is the score their offer position implies: hiring at the middle of the band assumes a score of 0.7, and the formula says so. The first review comes at the first scheduled date after six months of service.

With fewer than three closed deliverables, the score carries over and base pay holds. A flat or lower score never cuts pay either: the target stays or falls, and once it is at or below current pay, pay holds until a higher score or a higher anchor lifts the target again. A person still below their target keeps moving toward it by the capped step in Section 4.4. The lead cuts work into deliverables, so two invalid windows in a row trigger a review of the lead's scoping, not a sanction on the person.

4.3 Target

p = clamp((S_t - 0.4) / 0.6, 0, 1)
T = A x (0.85 + 0.30 x p)

A smoothed score at or below 0.4 places the person at the bottom of the band. A score of 1.0, which means filling the agreed scope at baseline cost and quality, places them at the top. The band is 0.85 to 1.15 of the anchor. Chapter 6 of this playbook, on growth without titles, adds a cap on p by skills-ladder stage; the cap limits the target and never cuts pay.

4.4 Capped step and merit pool

r_i  = max(0, min(T_i, 1.06 x P_i) - P_i)     zero if no valid score
Pool = beta x (sum of P_i over the team)
f    = min(1, Pool / sum of r_i)              f = 1 if no raises are proposed
P_i' = P_i + f x r_i

Each raise is the gap to target, capped at 6 percent. If the raises exceed the pool, all are scaled by the same factor. The pool is the budget control; the cap is the per-person control.

Convergence. With a fixed target and a pool that does not bind, a person below target reaches it in the smallest whole number of reviews k such that 1.06 to the power k is at least T / P. A 16 percent gap takes three reviews and a 35 percent gap takes six. If the market moves the target up by g per review, the remaining gap shrinks by a factor of (1 + g) / 1.06 each review, so the person converges only if g is below 6 percent and the pool is large enough. A fixed number of reviews can be promised only under these conditions.

Re-anchoring. M and the tiers are revised once a year at one review. A, the band and every target are recomputed. Base pay is never cut. A person whose pay is above the new band's top keeps it, with no raise until the band catches up; compensation practitioners call this a red-circle rate.

4.5 Bonus

Bonus = 0.5 x max(0, min(Perf_t, 1.2) - 1) x P x 0.5     valid window only

The bonus pays half the excess of the raw score over 1.0, on six months of base pay. The scorecard's score is not bounded above, because delivered scope can exceed agreed capacity, so an uncapped bonus would have no ceiling. Capping the input at 1.2 limits the bonus to 10 percent of six months' pay. It uses the raw score because it rewards one window. It comes from a separate budget line and does not touch base pay, the band, the step or the pool.

4.6 Human sign-off

The scorecard article says the score should not be the sole basis for pay. Here the score sets the default outcome, inside a band, a 6 percent cap and a pool. The lead and a founder sign every review. Either may change the outcome in either direction, within the band and never below current pay, only with a written reason that the employee sees. Overrides are counted and reported to the team, without names. The scorecard article keeps the score between the developer and the lead; using it for pay adds one founder to that circle and nobody else. This narrows the tension with the scorecard article without removing it (Section 10).

4.7 What the employee sees

The founder publishes M's source, the tiers, the rule for w and each role's inputs, the band, the cap, the pool rate for the review and the smoothing rule. Each employee sees their own score inputs and can compute their target and the reviews needed to reach it. No one sees anyone else's pay. We make no promise about titles; a pay position can be explained without one.

5 Worked example

Every number in this section is invented to show how the formula behaves. The countries are named only to make the example concrete, and the tiers are illustrative assignments, not values looked up from the price index. Pay is an index where M = 100.

We reuse the four developers from the scorecard article with their invented scores. That article's example scored each developer on a single service, which would be an invalid window under the three-deliverable minimum, so we add an invented deliverable count n and a previous smoothed score for each. The pool rate is an invented 3 percent of the team's payroll for this review.

How to read it:

• B scores 1.07 raw but 0.996 smoothed, because a 0.80 history keeps 27 percent weight. The target is 80.4 and the cap allows 76.3. Proposed raises total 8.76 against a pool of 8.55, so all are scaled by 0.976 and B moves to 76.2, a 5.9 percent raise, with the target now 5.5 percent above pay: one more review if nothing changes and the pool allows. B's bonus is 0.5 x 0.07 x 36 = 1.26, 3.5 percent of six months' pay.
• C's target is 6.7 percent above C's pay. The cap binds slightly and the pool scales the raise to 5.9 percent, leaving C less than 1 percent short.
• A and D are above their targets, so their pay holds, not falls. Their raw scores are at or below 0.4, which starts the notice process in Section 7.3.
• Geography. A and D share w and L, so they share an anchor despite different employers. B's anchor is lower than A's only because of the tier.

6 Properties and simulation

6.1 Properties

The following hold by construction. Our script asserts the never-down rule and the 6 percent bound in every simulated review, and spot-checks monotonicity and continuity.

• Never down (R3). New pay is current pay plus a scaled raise that is zero or more.
• Bounded (R4). A raise is at most 6 percent per review. Total raises are at most the pool. The bonus is at most 10 percent of six months' pay.
• Monotone in the score. All else fixed, a higher score never gives a lower target or new pay, and under pool scaling a person's own raise still rises with their proposed raise. Others' raises can fall when one person scores higher, as with any fixed budget.
• Continuous in the score. Clamp, min, max and the pool factor are continuous. Two discontinuities are deliberate: the three-deliverable rule, and tier boundaries, where a small index move shifts L by a whole tier. The never-down rule cushions a downward tier change.
• Equal treatment across tiers. Two people with the same smoothed score get the same position p. Their targets differ only by the published ratio of their anchors, and their speed of convergence does not depend on the tier.

6.2 Simulation method

We wrote a Node.js script that simulates a synthetic team of eight people over ten half-yearly reviews, five years, with 2,000 seeded runs per scenario. All inputs are invented:

• Two people in each of the four tiers (L of 1.00, 0.75, 0.55, 0.40), with w from the rule in Section 4.1, between 0.375 and 0.75.
• Each person has a true ability, from 0.45 to 1.10, that drifts by a random walk with a standard deviation of 0.03 per review.
• Each window has 2 to 12 deliverables, drawn evenly, so about 9 percent of windows are invalid. Each deliverable scores true ability plus normal noise (standard deviation 0.25), clamped to 0 to 1.2; the raw score is their mean.
• M grows a normally distributed 1.5 percent per half-year (standard deviation 1.5 percent). Starting pay is drawn evenly from 0.85 to 1.05 of the anchor, which sets the prior. The default pool is 4 percent of payroll per review.

"True-score target" means the target computed from true ability instead of the measured score. The script is salary-sim.mjs and its raw output is salary-sim-output.txt; both are kept with the article's working files.

6.3 Results

All figures below are synthetic. They describe the formula's behaviour under our invented assumptions, not real people or real payrolls.

H1, smoothing. Smoothing cut the review-to-review movement of the target by about a third, from 5.4 to 3.5 percent of the anchor, and the never-down drift from about 1.0 to 0.6 percent above the true-score target. The drift effect is small because the cap already limits what one lucky window can do. It did not slow a genuine improver: a synthetic person whose ability jumped from 0.6 to 1.0 at review 3 reached the new target a median 5 reviews later either way, and about 15 percent of runs had not reached it by review 10 in both cases. The cap, not the smoothing, sets the speed. H1 is supported for volatility and weakly supported for drift.

H2, pool. With a 4 percent pool, payroll grew a median 1.29 times over five years against market growth of 1.16 times, ending within 0.5 percent of the sum of true-score targets; the excess is catch-up, since about three quarters of the synthetic people started below target. The pool bound in 18 percent of reviews. A 2 percent pool covers the 1.5 percent market drift but little of the catch-up on top of it. Median payroll growth then equalled the budget path, 1.22 times, but pay ended a mean 4.8 percent below target and 63 percent of those starting below target never reached it. H2 is supported, with a warning: the pool has to fund market growth plus catch-up, and one that covers little more than market growth leaves most people behind, visibly.

H3, convergence. With a still target and no pool limit, gaps of 10, 16, 25 and 35 percent take 2, 3, 4 and 6 reviews, as the convergence formula predicts. Under noise, market drift and the default pool, people starting below target reached 99 percent of it in a median 5 reviews (90th percentile 8); about 5 percent had not within 10. The cap bound in about 31 percent of valid person-reviews. H3 is supported.

Market fall and bonus. When the synthetic market fell 8 percent at review 4, 52 percent of person-reviews had no raise and pay ended a mean 1.6 percent above target: held, not cut. The bonus cost a mean 0.8 percent of six months' payroll per review, at most 3.8 percent in any single review of the base scenario.

Notice process. A synthetic person with true ability 0.60 had three consecutive valid windows at or below 0.4 in 0.6 percent of runs. A person with true ability 0.45, just above the trigger, did so in 26 percent of runs. Under these assumptions the rule rarely catches a mid-band performer by noise. For someone near the trigger, noise decides much of the outcome, which is one reason the process ends in a human decision.

Answers to the research questions. RQ1: the formula meets R1 to R7 by construction (R1 in Section 4.7, R2 because hours appear nowhere in Section 4, R3 and R4 in Section 6.1, R5 through smoothing, R6 in Section 4.1, R7 because every step is spreadsheet arithmetic). The requirements still pull against each other. R3 costs accuracy: pay drifts 0.6 percent above the true-score target, and 1.6 percent after a market fall. R4 costs speed: a tight pool leaves people below target. R1 is weakened by the override in Section 4.6, which an employee cannot compute in advance. RQ2: H1 to H3 above, on synthetic data only.

7 Hours, exceptional effort and sanctions

7.1 Hours are not in the formula

Go back to the two developers. The one who leaves at five and ships more gets the higher position. The one who stays until midnight gets no premium for staying.

Hours still matter for health. Given the WHO and ILO estimate [17] [18], treat a sustained week of 55 hours or more as a risk to fix, not as commitment. Overtime law comes first. In the United States, the Fair Labor Standards Act requires non-exempt employees to be paid at least one and a half times their regular rate for hours over 40 in a week, and compensatory time off in place of that pay is a public-sector provision [21]. In the European Union, the Working Time Directive limits average weekly working time, including overtime, to 48 hours [22].

7.2 Exceptional effort

Every company has moments when it needs its people. One of our team has worked with more than a dozen companies, and in at least three of them spent a night in the office and worked into the middle of the next day. Someone who steps in at 3 a.m. for an incident deserves thanks. We keep it outside the formula: base pay rewards what is repeatable, nobody can be required to work beyond what the law and contract allow, and nobody should be penalized for declining, because people have limits and a life. We propose a separate, after-the-fact thank-you:

• The lead logs the event: who, when and why.
• The default thanks is time off, which in the leave model is Off and is rarely refused, at one and a half hours for each hour worked beyond the agreed day, taken within the next month. A person who prefers cash can have it, where the law allows the choice. Where overtime pay is mandatory, as for non-exempt US employees, the legal minimum applies and time off is not a substitute [21].
• We pay for the act and do not try to read motives. We publish no price list in advance, because a pre-announced premium risks turning a favour into a transaction [1] [5]. That literature does not test this policy, so this is a reasoned choice, not a finding.
• The count is a signal about the company, not the person. A team called in two or three days a month is in crunch, and needs a review of scope and staffing, not a larger thank-you.

7.3 Pay never goes down, and the notice process

A pay cut for low performance damages trust and teaches people to hide problems [12] [13]. The sanction in this formula is that pay holds.

When holding is not enough, there is a process. A notice is triggered by a valid six-month window with a raw score at or below 0.4, which is the bottom of the band. The person gets a written notice with the score and its inputs, plus training time, a mentor and a plan with a date. A second such window brings a second notice. After a third, separation is a fair outcome: the first two reviews are the company's chance to help, the third its right to move on.

The trigger uses the raw score, not the smoothed one, so that a person who recovers is not held back by their history. The simulation in Section 6 counts consecutive windows, and we would reset the count after a valid window above 0.4. Dismissal law differs by country and often requires notice periods, procedures and severance, so adapt this with local counsel.

8 Top of band and equity

A strong person eventually reaches the top of the band, and base pay stops until the market moves the anchor. This is where companies lose people or break the band. The options, in the order we would try them:

1. Equity. Shares or share options if you can issue them, phantom shares if you cannot. In the US, an option stays outside the deferred-compensation rules of Section 409A only if, among other conditions, its exercise price can never be below the stock's fair market value at grant. An independent appraisal no more than 12 months old is one valuation method presumed reasonable [19]. Phantom shares are a promise to pay a cash bonus equal to the value of a number of shares, without transferring ownership, and are taxed as ordinary income when paid [36]. They can fall under Section 409A, where non-compliance brings income inclusion at vesting plus a 20 percent additional tax; many plans pay within the short-term deferral period, generally by 15 March of the year after vesting for a calendar-year taxpayer [19]. The company must also plan the cash owed at an exit. Rules outside the US need local advice.
2. The bonus. For a raw score above 1.0, as in Section 4.5.
3. Reduced hours at the same pay. Because hours are not in the formula, a four-day week at the same salary is a raise in hourly terms. The scorecard shows whether output holds.
4. A second job, if there is no conflict. The person moves to a part-time contract with us and takes other work, which raises their income without breaking our band. Pay is pro-rated to the contracted hours: a 60 percent contract pays 60 percent of the full-time rate. Band position, raises and the never-down rule apply to that full-time rate, so the move is not a pay cut, and it happens only at the person's request. This needs clear terms on confidentiality, intellectual property and conflicts, and some countries restrict exclusivity clauses.
5. A fair goodbye. If none of this is enough, do not block the person; help them find the next place.

A bonus costs cash now; equity costs dilution, and phantom shares cost cash later. A founder granting equity at the top of the band should still pay the computed bonus, or say in advance that equity replaces it.

9 Mobility, tax and legal exposure

9.1 When someone moves country

A remote employee can move. The agreement should say what happens, with one of three clauses chosen at hiring:

• Fixed. Pay and tier do not change when the person moves.
• Repriced. The tier follows the new location from the next review. Pay is not cut; after a move to a lower tier it holds until the target catches up.
• By approval. A move needs the company's agreement first, and the agreement says what changes.

9.2 Currency

The formula is computed in the reference currency of Section 3: the employer's, or the euro or US dollar where the employer's own currency is volatile. The never-down rule is tested in that currency. If you pay in a local currency, convert at each review at a published rate averaged over the previous three months, fix the local amount until the next review, and test the guarantee on the reference amount. Paying in the reference currency leaves exchange-rate risk with the employee when they spend locally; paying locally against a reference-currency guarantee leaves it with the employer. A guarantee set only in a high-inflation local currency would protect very little, which is why we do not use one.

9.3 Tax and legal exposure

An employee working from another country can create obligations for the company: payroll registration, social contributions and, in some cases, a permanent establishment that brings corporate tax. The OECD's 2025 update to the Commentary on Article 5 of its Model Tax Convention, adopted in November 2025, sets out when a home office can be a fixed place of business [20]. As summarised by advisers, home working below 50 percent of total working time over 12 months is generally not a place of business, and above that a commercial reason, such as meeting local customers, is needed; convenience, talent retention or saving office costs are not commercial reasons [37]. A role with local customer contact may therefore be both better paid under our w rule and more exposed. Most treaties follow older versions of the Model, so the effect varies by country.

One adviser lists four common routes for paying people abroad: a local entity, non-resident employer registration, an employer of record, and contractors. It also notes that a single employee can create a permanent establishment and that an employer of record does not eliminate that risk [38]. That source sells those services. Take cross-border tax advice before hiring in a new country.

10 Threats to validity and limitations

Construct validity. The score measures scoped delivery against an agreed capacity, adjusted for quality, cost, time and simplicity. It does not measure value to customers, and it ignores review, mentoring and maintenance. Someone who spends a half-year on that work scores low unless the lead reduces their capacity in writing beforehand. M depends on one data source, and sources disagree. The tier index measures prices, not wages.

Internal validity. The simulation's conclusions follow from its assumptions: normal, independent noise, a fixed team, a random-walk market with one shock, and people who do not react to the formula, so gaming, effort responses and attrition are absent by construction. Correlated errors, such as a lead who consistently under-scopes one person, would not be removed by smoothing; they would be smoothed in. The results show how the rules interact, not that they are good for a real team.

External validity. The design targets 5 to 10 employees. The score exists only for developers, so other roles need their own delivery measure before the formula applies to them. Above that size, a founder cannot see every review, and the sign-off rule weakens. Tax, overtime and dismissal rules differ by country and change. The equity discussion is mostly US-specific.

Goodhart effects. When a measure becomes a target, it stops being a good measure, in Strathern's generalisation of Goodhart [7]. The moment the score sets pay, people will push on it. The likely routes are inflating scope estimates before work starts, negotiating capacity down, and neglecting unmeasured work [3] [6]. The band, cap and pool limit what gaming can earn, and sign-off gives a human a place to say no. They do not remove the incentive: a gamed scorecard is paid for, up to the limits.

The tension with the scorecard article. That article says the score should not be the sole basis for pay and that, if money is attached, the amount should stay small and a human should decide. This article lets the score set the default outcome. We reconcile the two only partly: pay moves at most 6 percent per half-year inside a 30 percent band, the founder can override with a written reason, and overrides are reported. Whether that is small is a judgement; a founder who disagrees should widen the human role.

Fairness across geography. Two people with the same score will earn different amounts because they live in different places. That is defensible by labour market [25], but it is not the same as fair. Publish the tiers and the rule for w, and review them once a year.

Ratchet and drift. The never-down rule plus noise pushes pay slightly above what perfect measurement would give: about 0.6 percent in our synthetic base case with smoothing. Budget for it.

Parameters are starting values. The band, the 0.4 to 1.0 mapping, the 6 percent cap, the 0.5 bonus rate, the 1.2 bonus cap, the smoothing constant 3, the w rule and the tier cut-offs are our choices. Vary each by 25 percent, rerun the example on your own team, and see whether the order of people changes. If it does, treat that part as noise.

11 Implementation checklist

1. Name one market data source and record M for each role and level, with the date.
2. Score each role on scarcity, overlap and customer contact; publish w and the tier table.
3. Choose the reference currency and the move-country clause for every offer letter.
4. Run the scorecard for six months; record n and Perf, and set new hires' priors from their offer.
5. Set the pool rate first, then compute S, p, T and the raises in a spreadsheet, and scale them to the pool.
6. Have the lead and a founder sign, with written reasons for any override. Pay bonuses from a separate line and send notices where due.
7. Revise M and the tiers once a year, and book the next review six months out.

If you want help setting this up, our Fractional CTO service covers pay structures for small engineering teams.

Frequently Asked Questions

Should we pay the same salary to everyone in the same role, wherever they live? You can, but then you pay the most expensive market's rate everywhere. This formula pays a published blend of your market's rate and the local rate, weighted by role scarcity, required overlap and customer contact.

Should an employee who works fewer hours be paid less? Not under this formula. Pay follows the delivery score, so delivering more in fewer hours is not penalized. Watch for unsustainable hours instead, and follow local overtime law.

How often should a startup review salaries? Annual review is the more common practice; one vendor survey of 140 companies reports it as the most frequent pattern [32]. We use two, because the scorecard runs on six-month windows and, with 5 to 10 people, a review costs little.

Are share options better than phantom shares? Neither is better in general. Options give real ownership and, in the US, need an exercise price at or above fair market value to stay outside Section 409A. Phantom shares pay cash tied to share value without issuing shares, but need cash planned for an exit and care with deferred-compensation rules. Get local advice before granting either.

References

Peer-reviewed and scholarly

1. Huffman, D., and Bognanno, M. (2018). High-powered performance pay and crowding out of nonmonetary motives. Management Science, 64(10), 4669-4680. https://doi.org/10.1287/mnsc.2017.2846
2. Lazear, E. P. (2000). Performance pay and productivity. American Economic Review, 90(5), 1346-1361. https://doi.org/10.1257/aer.90.5.1346
3. Holmstrom, B., and Milgrom, P. (1991). Multitask principal-agent analyses: Incentive contracts, asset ownership, and job design. Journal of Law, Economics, and Organization, 7, 24-52. https://doi.org/10.1093/jleo/7.special_issue.24
4. Gneezy, U., Meier, S., and Rey-Biel, P. (2011). When and why incentives (don't) work to modify behavior. Journal of Economic Perspectives, 25(4), 191-210. https://doi.org/10.1257/jep.25.4.191
5. Deci, E. L., Koestner, R., and Ryan, R. M. (1999). A meta-analytic review of experiments examining the effects of extrinsic rewards on intrinsic motivation. Psychological Bulletin, 125(6), 627-668. https://doi.org/10.1037/0033-2909.125.6.627
6. Kerr, S. (1975). On the folly of rewarding A, while hoping for B. Academy of Management Journal, 18(4), 769-783. https://doi.org/10.2307/255378
7. Strathern, M. (1997). 'Improving ratings': Audit in the British university system. European Review, 5(3), 305-321. https://doi.org/10.1002/(SICI)1234-981X(199707)5:3%3C305::AID-EURO184%3E3.0.CO;2-4
8. Colquitt, J. A. (2001). On the dimensionality of organizational justice: A construct validation of a measure. Journal of Applied Psychology, 86(3), 386-400. https://doi.org/10.1037/0021-9010.86.3.386
9. Card, D., Mas, A., Moretti, E., and Saez, E. (2012). Inequality at work: The effect of peer salaries on job satisfaction. American Economic Review, 102(6), 2981-3003. https://doi.org/10.1257/aer.102.6.2981
10. Cullen, Z., and Pakzad-Hurson, B. (2023). Equilibrium effects of pay transparency. Econometrica, 91(3), 765-802. https://doi.org/10.3982/ECTA19788
11. Cullen, Z. (2024). Is pay transparency good? Journal of Economic Perspectives, 38(1), 153-180. https://doi.org/10.1257/jep.38.1.153
12. Kahneman, D., Knetsch, J. L., and Thaler, R. (1986). Fairness as a constraint on profit seeking: Entitlements in the market. American Economic Review, 76(4), 728-741. https://ideas.repec.org/a/aea/aecrev/v76y1986i4p728-41.html
13. Bewley, T. F. (1999). Why Wages Don't Fall during a Recession. Harvard University Press. https://doi.org/10.4159/9780674020900
14. Grigsby, J., Hurst, E., and Yildirmaz, A. (2021). Aggregate nominal wage adjustments: New evidence from administrative payroll data. American Economic Review, 111(2), 428-471. https://doi.org/10.1257/aer.20190318
15. Bühlmann, H. (1967). Experience rating and credibility. ASTIN Bulletin, 4(3), 199-207. https://doi.org/10.1017/S0515036100008989
16. Pencavel, J. (2015). The productivity of working hours. Economic Journal, 125(589), 2052-2076. https://doi.org/10.1111/ecoj.12166
17. Pega, F., Náfrádi, B., Momen, N. C., et al. (2021). Global, regional, and national burdens of ischemic heart disease and stroke attributable to exposure to long working hours for 194 countries, 2000-2016. Environment International, 154, 106595. https://doi.org/10.1016/j.envint.2021.106595

Official, legal and statistical sources

18. World Health Organization (2021, 17 May). Long working hours increasing deaths from heart disease and stroke: WHO, ILO. News release. https://www.who.int/news/item/17-05-2021-long-working-hours-increasing-deaths-from-heart-disease-and-stroke-who-ilo
19. US Treasury Regulations, 26 CFR 1.409A-1, Definitions and covered plans, paragraphs (b)(4) and (b)(5), https://www.law.cornell.edu/cfr/text/26/1.409A-1; and Internal Revenue Code, 26 U.S.C. 409A(a)(1), for the income inclusion and the 20 percent additional tax, https://www.law.cornell.edu/uscode/text/26/409A
20. OECD (2025). The 2025 Update to the OECD Model Tax Convention. OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/11/the-2025-update-to-the-oecd-model-tax-convention_c7031e1b/5798080f-en.pdf
21. Fair Labor Standards Act, 29 U.S.C. 207, sections (a)(1) and (o). https://www.law.cornell.edu/uscode/text/29/207
22. Directive 2003/88/EC of the European Parliament and of the Council concerning certain aspects of the organisation of working time, Article 6. https://eur-lex.europa.eu/eli/dir/2003/88/oj
23. World Bank. Price level index (GDP), World Development Indicators. https://data.worldbank.org/indicator/PA.NUS.GDP.PLI

Company, industry and adviser sources (not peer-reviewed)

24. GitLab. The Compensation Calculator, and Compensation at GitLab. GitLab Handbook, retirement noted as 30 May 2025. https://handbook.gitlab.com/handbook/total-rewards/compensation/
25. Flowers, A. (2019, 28 February). Why GitLab pays local rates. GitLab blog. https://about.gitlab.com/blog/why-we-pay-local-rates/
26. Buffer (2013, 19 December). Introducing open salaries at Buffer: Our transparent formula and all individual salaries. https://buffer.com/resources/introducing-open-salaries-at-buffer-including-our-transparent-formula-and-all-individual-salaries
27. Buffer (2017, 6 December). Introducing the new Buffer salary formula. https://buffer.com/resources/salary-formula/
28. Buffer. Transparent salaries. Accessed November 2026. https://buffer.com/salaries
29. Lightcast (2025, 23 July). New Lightcast report: AI skills command 28% salary premium as demand shifts beyond tech industry. https://lightcast.io/resources/blog/beyond-the-buzz-press-release-2025-07-23
30. PwC (2025). The Fearless Future: 2025 Global AI Jobs Barometer. Press release. https://www.pwc.com/gx/en/news-room/press-releases/2025/ai-linked-to-a-fourfold-increase-in-productivity-growth.html
31. Kolesnikova, A. (2023, 2 May). Here's how much AI engineers are getting paid. Levels.fyi blog. https://www.levels.fyi/blog/ai-engineer-compensation.html
32. Ravio. Pay review survey: how 140 companies do comp reviews. Vendor report, undated. https://ravio.com/reports/pay-review-survey
33. Deel. Merit matrix. Glossary entry, vendor. https://www.deel.com/glossary/merit-matrix/
34. Tozour, P., et al. (2015, 20 January). The Game Outcomes Project, part 4: Crunch makes games worse. Game Developer. https://gamedeveloper.com/business/the-game-outcomes-project-part-4-crunch-makes-games-worse
35. Take This (2016). Crunch hurts: How unmitigated overwork harms employee health, productivity, and your studio's bottom line. White paper. https://www.takethis.org/wp-content/uploads/2016/08/CrunchHurts-TakeThis.pdf
36. National Center for Employee Ownership. Phantom stock and stock appreciation rights (SARs). https://nceo.org/articles/phantom-stock-appreciation-rights-sars
37. EY (2025). OECD 2025 Update: New rules on permanent establishment for remote work. Tax alert. https://www.ey.com/en_ch/technical/tax-alerts/oecd-2025-update-new-rules-on-permanent-establishment-for-remote-work
38. Vistra. Hiring and paying remote workers abroad: Understanding your risks and options. https://www.vistra.com/insights/hiring-and-paying-remote-workers-abroad-understanding-your-risks-and-options

---

Need a pay structure your whole team can read? Book a diagnostic call or see how our Fractional CTO service can help.