AI Salary Benchmarking for HR Consultants: Sources, Bands and Checks

Use AI for the clerical work of benchmarking, not for market data: give it survey rows and job adverts you have collected, let it match jobs, clean pay figures and draft band formulas, then recalculate every number in a spreadsheet. Check each band by tracing figures to a source row, comparing independent sources and testing for invented numbers.

Pay-check tools answer a different question: where does one person's salary sit? A client has to explain why a band is what it is to an employee, an auditor or a tribunal. Ask an assistant cold for the median pay of a prepress designer and you get a tidy number whether it was read from a table or assembled from patterns, and nothing tells you which. A band needs the opposite: every figure carries a source, a date and a row you can point to.

Follow me on Instagram@sagnikteaches

What AI may touch in a benchmarking job, and what it may not

Divide the work before you open a chat window. Assistants are good at three clerical jobs here: comparing a client's job description with survey job summaries, cleaning pay figures pasted from adverts, and writing the formulas that turn evidence into bands. They are the wrong tool for supplying the market figure itself, because a figure from memory has no source, date or sample behind it.

Connect on LinkedInSagnik Bhattacharya

Privacy comes before any of it. Pay for named employees is personal data, so swap names and employee numbers for codes before anything reaches a chat. The main providers' business plans do not train on business content by default, while consumer plans carry a training switch in their privacy settings, and my notes on whether free AI plans are safe with staff data go through the settings. Licensed survey data needs the same care: read the licence before you upload it, because some licences limit who may process the data.

Subscribe on YouTube@sagnikteaches

Build the evidence file before you write a single prompt

Open a spreadsheet and give every data point its own row and an ID. The IDs matter more than anything else in this method, because every check later in this tutorial works by matching a figure back to a row. Put the row ID in column A, the source and the job as the source defines it in B, the data date in C, the statistic (median, 25th percentile, advert minimum and maximum) in D and the value in E. Then add pay basis (base or total cash), unit (annual or hourly), sample size and a notes cell.

Three kinds of evidence go in, and they are not equal.

• Licensed survey data is the strongest: percentile pay by job, an effective date and a sample size. Its weakness is job definitions, which is why matching comes first.
• Job adverts are timely and increasingly common as pay-range rules spread, but they record what an employer hopes to pay, not what it paid. Collect them by hand. LinkedIn's user agreement, for one, bans scripts, robots and other automated means of copying its data, and other sites have comparable terms.
• The client's own payroll never counts as market evidence. It comes in after the bands exist, to show where current pay sits against them.

The evidence file for the worked example later in this tutorial has three entries. Every figure is illustrative.

The nine adverts, in thousands, run 40 to 48, 42 to 50, 38 to 45, 44 to 52, 45 to 55, 36 to 44, 41 to 49, 43 to 47 and 39 to 46. Their midpoints are 44, 46, 41.5, 48, 50, 40, 45, 45 and 42.5, and the median of those midpoints is $45,000.

Match client jobs to survey jobs on duties, not titles

Matching is where AI saves the most time and misleads most quietly, because a job title is the cheapest thing for a model to compare. Force it to compare duties, scope and level, rate each match on a scale you define, and say what does not fit. Paste in the survey provider's own job summaries so it has nothing else to draw on. The quality of the client's side matters too, so start from a proper job description; my tutorial on writing job descriptions with AI covers that.

You are helping an HR consultant match one client job to benchmark jobs.
Use ONLY the text below. Do not use outside knowledge of pay.

CLIENT JOB (anonymised):
[duties, reporting line, number of people managed, budget responsibility, qualifications]

BENCHMARK JOBS (the survey provider's own summaries):
B1: [paste summary]
B2: [paste summary]
B3: [paste summary]

Return a table: Benchmark ID | Rating (A = core duties and level match,
B = partial match, C = title similarity only) | Duties that match (max 3) |
Duties that do not match (max 3) | Scope gap (people, budget).
Finish with one line: the best match, or "no adequate match".
Do not mention any salary figure.

An illustrative answer for a tutoring centre's centre manager, who runs the day-to-day, rosters twelve tutors, handles enrolment and parent contact, and holds no budget beyond petty cash, looks like the table below. The first three columns are the assistant's; the last is the ruling.

The rating said A while the note beside it listed duties the centre manager does not have. When a rating and its reasoning disagree, trust the reasoning and overrule the rating. Then confirm the ruling against the survey's own definition, which takes about a minute per job and is the step that keeps a keyword match from setting a client's pay.

Clean advertised ranges into market points

Adverts arrive messy, and cleaning them is the most tedious step, so it is the best one to hand over. Give the assistant the pasted text and the rules below, and ask for one row per advert plus a column for anything it excluded and why.

1. Keep the unit the advert uses, then convert on the client's contracted hours (weekly hours times 52), not on a default full-time year.
2. Record the minimum, maximum and midpoint. The midpoint is the data point, and the raw minimum and maximum stay in the file.
3. Exclude "up to", "from", "negotiable" and "competitive" adverts, and any range whose maximum is more than 50% above its minimum, because that usually spans several levels. The 50% is my cut-off, so set your own.
4. Count duplicates once: the same role reposted by an agency or on a second board is one data point.
5. Summarise with the median of the midpoints, never the mean, because one wide advert can drag a mean a long way.

One slip in this step deserves an illustration of its own. Suppose a wedding planner hires a seasonal assistant coordinator for 30 hours a week over 46 weeks, and the adverts show $18 to $22 an hour. If the assistant converts them on 2,080 hours (40 hours times 52 weeks), the $20 midpoint becomes $41,600. This client's actual annual cost at $20 is 20 × 30 × 46 = $27,600. Part-time and seasonal roles should be benchmarked as hourly rates, with the band kept hourly and converted to an annual cost only in the budget sheet, using the client's real hours.

The prepress designer band at a 40-person print shop, step by step

Consider a 40-person print shop that wants production bands from 1 January 2027, starting with the prepress designer. The 3% annual pay movement below is an assumption the consultant sets from the client's pay budget or from the movement the survey provider publishes. It is not a market fact, and every figure here is illustrative.

Age each source to the effective date

A survey figure dated last January understates today's pay if pay has moved, so each figure is carried forward to the date the bands will apply: raw figure times 1.03 to the power of the years between data date and effective date.

The weights are a judgement the client's pay policy can change. Survey A gets the most because it has the larger sample and the closest job match; the adverts get the least because they record hopes, not payments. Write the weights in the method note.

Test the spread, then blend

The highest aged figure ($47,895) sits 6.0% above the lowest ($45,191). I use 15% as the trigger for investigating job matches, dates and samples before blending, and this set is well inside it, so the sources agree. The blend is 0.4 × 47,895 + 0.3 × 45,191 + 0.3 × 45,334 = $46,315, rounded to the nearest $250 for a midpoint of $46,250.

Draw the range

With a 30% range spread, the minimum is the midpoint divided by 1.15, which is $40,217, rounded to $40,250. The maximum is twice the midpoint minus the minimum, which gives $52,250, so the midpoint stays the true middle of the band. The band is $40,250 to $52,250.

Hold it against payroll

Three prepress designers, anonymised as A, B and C, earn $41,000, $45,500 and $51,000. Their compa-ratios (pay divided by midpoint) are 0.89, 0.98 and 1.10, and their range penetration (how far through the band each sits) is 6%, 44% and 90%. Nobody is outside the band. A sits near the bottom, so check tenure and performance before treating that as underpayment.

Spreadsheet formulas do all of this, and the assistant can write them if you describe the sheet in plain English and ask for formulas only. My guide to ChatGPT for Excel covers the habit of testing each formula on one row worked by hand. This is the set, using the evidence columns above, a summary row that holds the median of the advert midpoints, and three named cells for the settings: AnnualMove (3%), EffDate (1 January 2027) and Spread (30%).

Aged value    =E2*(1+AnnualMove)^YEARFRAC(C2,EffDate)
Blend         =SUMPRODUCT(AgedRange,WeightRange)
Midpoint      =MROUND(Blend,250)
Minimum       =MROUND(Midpoint/(1+Spread/2),250)
Maximum       =2*Midpoint-Minimum
Compa-ratio   =Salary/Midpoint
Penetration   =(Salary-Minimum)/(Maximum-Minimum)

From one band to a structure: progression and overlap

One band is easy to defend alone. A structure gets challenged when neighbouring bands sit too close together or overlap too much, so build the other grades in the production family the same way and test them as a set.

Overlap here means the share of the lower band that sits above the next band's minimum, so G1's 43% is (37,250 − 34,000) ÷ (37,250 − 29,750). I set a 30% to 50% window for this client's policy, and all three pairs fall inside it. Midpoint steps rise from 16% to 19%, a smooth progression. Two things should send you back upstream: a step that shrinks or reverses, and an overlap above 60%, which makes grades hard to tell apart. Both usually mean a matching error or a weighting slip, not a pay-policy decision.

Seven checks that catch invented or drifting figures

Run these in order. The first two catch most invented numbers, and the full set takes under half an hour for a five-job band pack. For the habit behind them, see my tutorial on catching made-up figures in AI-drafted proposals.

Check 1: look every quoted figure up in the evidence sheet

Make the assistant return each figure it uses together with the row ID it took the figure from. Paste that table into a check sheet and let a lookup compare it with the evidence, where column E holds the cited row ID and column F the quoted value:

=IF(ISNA(MATCH(E2,Evidence!A:A,0)),"NO SUCH ROW",
 IF(ABS(F2-INDEX(Evidence!E:E,MATCH(E2,Evidence!A:A,0)))<=0.5,"match","MISMATCH"))

Suppose the first answer on the print shop file quotes "S1: $47,800" when row S1 says $46,500, and cites a "Survey C" at $45,900 that does not exist. One cell returns MISMATCH and the other NO SUCH ROW. Both figures look plausible, which is the problem: reading the answer would not catch either.

Check 2: plant a canary row the evidence cannot answer

Add a job to the request that has no evidence rows at all. Take the print shop file again, with a bindery technician as the canary. A loose first prompt can return a band such as $38,000 to $49,000 with the note "typical market range". One added instruction closes the gap: "If a job has fewer than three evidence rows, return INSUFFICIENT DATA and no figures." The rerun should then answer INSUFFICIENT DATA for the canary. If your canary gets a number, the prompt is not holding the model to the file, and no other figure should be trusted until it does.

The other five checks

• Check 3, recalculate outside the chat. Rebuild each midpoint with the spreadsheet formulas and compare it with the assistant's quoted midpoint, allowing one rounding step ($250 here) and no more. If you do let an assistant calculate, use a mode that runs code on an uploaded file: ChatGPT's data analysis writes and runs Python code (availability varies by plan), and Claude's code execution handles CSV files up to 30MB on every plan. Read the code it shows. A model doing arithmetic in prose is predicting digits, not computing them.
• Check 4, run the spread test. Aged figures from independent sources should sit within your tolerance (15% in the example). When they do not, look at the job match, the data dates and the sample sizes before you blend anything.
• Check 5, repeat the run. Send the same prompt in a fresh chat with the evidence rows in a different order. A figure that changes between runs was not read from the file.
• Check 6, test the structure. Progression and overlap from the previous section: no reversals, no overlap above your ceiling.
• Check 7, test against payroll. Work out compa-ratios for the client's staff by code. A median far from 1.00, say below 0.90 or above 1.10, does not prove the bands are wrong, but it means finding out why before the pack goes to the client.

A thin-data miss: three adverts for a yoga studio manager

Take a yoga studio that needs a band for its studio manager, an illustrative case in which only three usable adverts exist: $30,000 to $36,000, $28,000 to $60,000 and $32,000 to $38,000. An assistant left to itself might average the midpoints (33,000, 44,000 and 35,000) to $37,333 and label the result high confidence. Two things would be wrong. The 28 to 60 advert spans several levels and breaks the 50% rule in the cleaning list, and three points are too few for any confidence rating above low.

Corrected, the run drops the wide advert and leaves two points, a median of $34,000. Two points are a hint, not a market. The fix is to mark the band provisional and low confidence, add front-of-house manager roles from fitness and hospitality adverts as proxies (tagged as proxies in the notes), and revisit once there are five usable points. The report to the studio should say so in words: provisional, two adverts and one proxy source.

Pay-transparency rules your bands will be read against

Bands matter more as pay-transparency rules spread, because a band is what an employer publishes, defends or reports. Rules differ by where a client's staff and applicants are, so record the jurisdiction beside each band and treat the table below as a prompt for the client's employment adviser, not as legal advice. Status checked on 1 November 2026.

Two practical consequences follow. First, "objective, gender-neutral criteria" makes the method matter as much as the number: a band lifted from an AI answer with no sources is hard to defend against that test, while a dated evidence file with stated weights is easy to. Second, one band sits behind several published figures, so changing it later changes every advert, which is a reason to freeze the evidence file with the band. The posting side has its own tutorial in the pay-transparency checklist for AI-written job ads, and the client questions an HR consultant should answer on AI compliance cover the wider conversation.

The one-page note that travels with the bands

Every band pack should carry a single page the client can file. Filled in for the prepress designer band, it reads like this (illustrative):

• Job and match: prepress designer, matched to survey A "graphic production artist" (rating A) and survey B "prepress technician" (rating B), with scope differences recorded.
• Evidence: rows S1, S2 and J1 to J9, dated 1 January, 1 April and 1 October 2026. Nine adverts collected by hand, none excluded.
• Method: aged to 1 January 2027 at 3% a year, weighted 40% survey A, 30% survey B and 30% adverts, midpoint rounded to $250, 30% range spread.
• Result: $40,250 to $52,250, midpoint $46,250.
• Confidence: high, meaning two independent sources and at least five data points in each; source spread 6.0%.
• Checks run: row lookup (one mismatch and one missing row corrected), canary row (passed after a prompt change), recalculation, repeat run, structure test, payroll test.
• Exceptions and review: one designer at 0.89 compa-ratio flagged for a tenure and performance look; refresh the evidence in twelve months.

A client who asks next year why a band moved can be handed the old evidence file and this page. That trail is the product. AI makes it quicker to produce, and the checks keep it honest. If you are pricing the work as a fixed service, the HR consultant proposal template shows how to state scope and exclusions.

Questions HR consultants ask about AI benchmarking

Can I put a client's payroll file into ChatGPT or Claude?

Only after removing names and employee numbers, and only on a plan whose terms suit the client's contract. Business plans from the main providers do not train on business content by default, while consumer plans have a training switch in privacy settings. Safer still, keep the payroll in your spreadsheet and ask the assistant only for formulas, so no individual's pay ever enters the chat.

How many data points does a pay band need before I publish it?

There is no universal number, so write your own rule into the method note. A workable one is at least two independent sources and five usable data points per job, with anything thinner labelled provisional and low confidence. A job with fewer than three evidence rows should return insufficient data rather than a band, whatever the assistant is tempted to offer.

Are advertised pay ranges good enough to replace survey data?

They make a useful second source and a timely cross-check, not a replacement. Advertised ranges show what employers hope to pay rather than what they paid, and they vary in width and honesty. Use them for roles a survey does not cover, to test survey figures with the spread check, and label any band that rests partly on advertised pay.

How often should pay bands be refreshed?

Review them at least once a year, and age the evidence to the new effective date each time instead of reusing last year's blended figure. Scarce or fast-moving roles may need a mid-year look. Keep every old evidence file, because a client asked to justify a pay decision may need to show what the data said on the day.

Should the AI decide which grade a job belongs in?

No. Grading rests on job evaluation against agreed criteria, and the EU directive expects those criteria to be objective and gender-neutral. An assistant can draft factor scores and flag inconsistencies, but a person or a panel should agree the grade and record the reason, because the reason is what you will be asked to produce later.

Further reads

• AI Survey Analysis for HR Consultants: Coding Employee Comments Safely — Code staff comments safely before pairing them with pay findings.
• HR Analytics and AI: Analyse Staff Turnover From a Spreadsheet Export — Test whether pay position explains who leaves.
• ChatGPT for Payroll Teams: Safe Prompts That Never Include Pay Data — Prompt habits that keep individual pay out of chats.
• How HR Consultants Use AI for Policies, Letters and Cases — The wider set of HR consulting jobs AI can take on.
• A Five-Minute Fact-Check Routine for AI Output Before It Goes Out — A five-minute routine for any AI output before it leaves.
• HR Audit Checklist: Run a Client Audit With AI and Check the Gaps — Fold a pay review into a client HR audit.
• Can AI Analyse My Sales Spreadsheet? What to Upload and Check — What to upload and check when AI analyses a spreadsheet.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the text of Directive (EU) 2023/970 on EUR-Lex; the government's gender pay gap reporting guidance; the consultation page on equal pay and pay discrimination; transposition and payroll-vendor trackers of pay-transparency rules (rechecked 4 November 2026); OpenAI and Anthropic help-centre pages on file analysis; LinkedIn's User Agreement.