You lose a candidate at offer stage and the feedback comes back as money. So the band gets reviewed, someone pulls a salary guide, and the number moves up by a few thousand. Six weeks later the same thing happens with the next hire.
The band probably was wrong. But most benchmarking exercises are built on data that can’t support the decision being made with it.
Why the published numbers mislead
Start with coverage. According to APSCo’s Q2 2026 analysis with CV-Library and Lightcast, only 34% of UK technology vacancies publish a salary at all. That third isn’t a random sample. Employers advertise a number when the number is competitive, so advertised data skews high by construction.
Then look at the premiums. On AI skills, PwC has reported a wage premium of around 62%, while payroll data drawn from more than 1,600 companies puts it closer to 12%. Both are real measurements. They’re measuring different things, and a salary guide quoting either without context will send you somewhere expensive.
Part of the gap is definitional. A senior platform engineering role that mentions a machine learning pipeline gets counted as an AI role, and its salary, driven mostly by seniority, gets counted as an AI premium.
What to benchmark against instead
Your own payroll, first.
What you already pay a senior engineer who is performing well is the most reliable number you have access to. If a new offer sits materially above it, you have an internal equity problem arriving alongside your hiring problem.
Accepted offers, not advertised ranges.
Ask your recruitment partner what candidates at this level have actually accepted in the last quarter, in your sector, at your stage. That number is harder to obtain and far more useful.
Total cost, not base.
Employer National Insurance sits at 15% above the £5,000 secondary threshold. A £5,000 base increase is not a £5,000 decision.
Tenure, not just the hire.
Mercer puts voluntary turnover in AI and machine learning roles at around 18%, against 12% for software engineering. A premium that buys 18 months is a different investment from one that buys five years, and should be appraised differently.
Money is rarely the cheapest lever
Harvey Nash’s 2026 survey of more than 3,600 technologists found that 41% would accept a lower salary in exchange for more home working.
That doesn’t mean you can underpay. It means that when you’re losing offers, pay may be the stated reason without being the binding one. Before raising the band, check the things candidates are actually weighing:
- How many interview stages, and over how many weeks
- Whether the work is greenfield or maintenance, and whether you were honest about it
- Flexibility, stated specifically rather than as “hybrid”
- Who they’d report to, and whether they met them
One thing to know about next year
The EU Pay Transparency Directive’s transposition deadline passed in June 2026, and employers across the EU will increasingly publish pay ranges before interview. As mandatory ranges spread, the distance between advertised and paid narrows. Expect the advertised premium to fall, not because specialist talent gets cheaper, but because the advertising gets more honest.
If your bands are currently set against advertised data, that correction will make them look wrong in a different direction.
Get a benchmark you can act on
We place technology talent across the UK, which means we see accepted offers rather than advertised ones, at role level and in your sector.
If you’re setting bands for 2027, or you’ve lost a hire recently and want to know whether the number was really the problem, send us the role and we’ll tell you what it’s actually going for.
