A compensation committee approves a 30% increase to the senior band and files it under “investing in engineering quality.” Nine months later, the people hired at the new band ship at about the same rate as the people hired at the old one. The link between what an engineer costs and how well that engineer performs is much weaker than most hiring plans assume, and the weakness has a fixable structure.
TL;DR
- A salary is a price, not a score. It records where someone lives, who employed them, how well they negotiated, and when they signed.
- Prior salary anchors everything. Two engineers of equal ability can drift $90K apart in seven years purely from where their first offer landed.
- The market prices stack scarcity before engineer quality, and the premium decays as supply catches up.
- Raising the band 30% mostly buys more competition for the candidate, not more capability from the hire.
- A US engineer at $190K and a LATAM engineer at $80K can be the same engineer. Low price is not evidence of low quality, and it is not evidence of high quality either.
- Money does buy performance when it goes into onboarding, retention, and tooling.
What You Are Actually Paying For When You Pay $220K
Take a backend role you would fill at $150K in Denver and fill instead at $220K in San Francisco. That $70K gap is real money, and almost none of it measures engineering ability. It is the sum of several markets that happen to settle on one number: where the person sleeps, which logo sat on their last badge, how well they run a negotiation, which comp cycle they last signed in, and how urgent you looked while you were doing it.
A working decomposition of that $70K premium, estimated from how offers get built rather than measured:
| Component | What it actually reflects | Rough share of the premium |
|---|---|---|
| Metro cost of labor | Rent and which employers are within reach | 30-40% |
| Previous employer brand | A hiring bar cleared three or four years ago | 15-25% |
| Negotiation and competing offers | Skill at running a job search | 15-25% |
| Timing of the last signature | A 2021 base that never reset downward | 10-20% |
| Technical ability seen in your loop | Skill at the job | 10-15% |
Accept even the generous end of that last row and roughly 85% of what you pay above baseline is priced on things unrelated to how the person will perform on your codebase.
Your Bands Are Repeating Someone Else’s Guess
Most US states now bar asking for salary history, and it barely matters. Candidates state expectations, expectations come from current base, and current base comes from an offer someone made years ago with less information than you have today.
The compounding does the damage. Two engineers write comparable code in 2019. One starts at a well-funded company at $165K, the other at a regional healthcare payer at $110K. Both change jobs twice with standard 12-18% bumps. By 2026 the first asks $265K and the second asks $175K, and nothing in that gap came from writing better software. Your committee then reads $265K as seniority and $175K as a weaker candidate, which is anchoring bias doing your evaluation for you, using data generated by other companies’ recruiters.
The Market Prices Scarcity, Not Quality
Bands get built from stack labels because that is what salary surveys can count, so the price moves with the supply of a keyword. An engineer with four years of production Go and Kubernetes prices 20-30% above the same engineer with four years of production Java, even when the underlying work is the same distributed systems problem.
When a stack gets hot the premium appears within a quarter. When bootcamps, cloud vendors, and internal migrations flood the supply two years later, it flattens. Scarcity pricing rewards label acquisition over judgment, and engineers respond rationally by chasing whichever stack is currently expensive.
What Does Predict Performance
Four signals correlate with how an engineer performs in month three, and none of them appear in a salary number.
- Evidence of prior work you can inspect. Not tutorial repos. A system they owned, with constraints named: traffic, team size, deadline, what they cut. Ask what broke.
- Decision quality under constraints. Give a real backlog scenario with a hard limit, like two weeks and no new infrastructure. Strong engineers ask what the failure mode costs. Weak ones start describing an architecture.
- Ramp-up speed in code they did not write. The highest-value signal and the least tested. A paid 3-4 hour task in a small unfamiliar repository with one deliberately confusing module tells you more than any resume line.
- Behavior in code review. Hand them a pull request with a subtle concurrency bug, an unnecessary abstraction, and a naming problem. Whoever leads with the naming nit and misses the race condition has told you a lot for $0.
That loop costs six to eight hours of team time per finalist, which is cheaper than one bad hire at any band.
The Same Engineer, Two Prices
A senior backend engineer in a US tech metro prices at roughly $175K-$210K base. A comparable senior in Sao Paulo, Buenos Aires, or Medellin prices at roughly $65K-$95K all-in. The reflex is to assume the second engineer is 60% as good, which is not what those numbers describe.
Price is set by the best alternative offer available in the person’s local market. A senior engineer in Colombia competes against local banks, regional startups, and a few foreign employers, while a senior engineer in Seattle competes against a dozen companies that can pay $200K plus. Both may sit at the top of their local distribution. Only one distribution has a bidding war attached.
This cuts both ways, and it needs saying plainly: a lower price is also not evidence of quality. LATAM markets contain excellent engineers and weak ones at similar prices, because the same signal noise applies locally. Salary tells you almost nothing in either direction, which makes the evaluation work above mandatory in both markets. There is still a floor: pay meaningfully below a market’s real senior range and you select for people no one else wanted.
What Money Actually Does Buy
Spending does move performance. It has to go somewhere other than the offer letter.
- Onboarding. A structured first 30 days with a named buddy, a scoped starter project, and a documented setup cuts time to first meaningful pull request from 5-6 weeks to 2-3. On a $90K engineer that is roughly $5K of recovered output per hire.
- Retention. Replacing a mid-level engineer costs four to six months of loaded salary once you count the open seat, the search, and the ramp. Paying $8K-$12K above your own band to hold someone for a third and fourth year beats paying $40K more for a stranger.
- Tools and leveling. CI that runs in 4 minutes instead of 22 gives a team of eight several hours back every week. A mid-level engineer who reaches senior in eighteen months because a manager gave them scope is the cheapest performance gain available.
Set Bands by Impact Level, Not Market Percentile
Percentile targeting (“we pay at the 75th”) locks you into repeating whatever the market did last year, including its mistakes. Price the work instead.
- Write four impact levels defined by decisions owned, not years. L2 executes a defined task; L3 owns a service and its tradeoffs; L4 owns a domain across services; L5 changes what the team builds.
- Assign the level for every open role before anyone sees a resume or a rate.
- Pull the local range for that level in each geography you hire in. Different bands per geography is not unfairness, it is pricing the alternative-offer reality that produced those numbers.
- Set the band at the local median for the level plus a 10-15% retention premium, with no adjustment for the candidate’s last salary. Delete that field from your process. Equity is a separate conversation.
- Ring-fence 5-8% of total comp spend for onboarding, tooling, and training, and defend it in the meeting where you defend the bands.
- Twice a year, plot every engineer’s total cash against their last performance rating. Teams doing this for the first time usually find their highest-paid quartile spread evenly across ratings, which is what a weak correlation looks like once you graph it.
Frequently Asked Questions
Does this mean we should lowball candidates?
No. Underpaying against a market's real range selects for people without alternatives and raises turnover, which erases the savings within a year. Price and performance correlate weakly, so paying above your band buys little while paying below it costs a lot.
What do we do with a candidate asking well above our band?
Evaluate them like everyone else, then decide against the band rather than the ask. If they clear the bar and the band does not stretch, say so directly and move on. A meaningful share take the level-appropriate number when the scope beats what they have.
How do we defend paying less for LATAM engineers without it feeling unfair?
Pay the top of the local band for the impact level, not the bottom, and be explicit that the level and scope match a US counterpart. Engineers accept geographic bands when leveling is transparent. What they reject is senior scope at a mid-level title because of where they live.
Isn't a top-tier previous employer a legitimate signal?
It is a weak positive signal that a bar was cleared once, and it decays fast. It says nothing about what the person did there, whether they owned decisions or executed someone else's, or how they work without a large platform team behind them. Use it to open a conversation, never to set a price.
What single change gives the fastest improvement?
Run the comp-versus-performance plot for your current team. It takes an afternoon with data you already have, and it turns an argument about compensation philosophy into a chart the committee can read in ten seconds.
