Every Civil Service recruitment campaign begins with the same quiet arithmetic. A vacancy attracts two hundred applications, each one containing a CV and a personal statement written against a published framework, and each one legally entitled to the same standard of consideration as every other. Automating Civil Service Success Profiles screening is not an efficiency luxury for departments and arm's length bodies working at this volume. It is the difference between a sift that finishes on schedule and a sift that quietly compresses the last eighty applications into an afternoon nobody wants to talk about.
This article sets out a technical blueprint for scoring CV and statement text against the Success Profiles framework, what a transparent data output should contain, and where automation has to stop and hand the decision back to a human panel. It is written for resourcing leads, recruitment business partners and the agencies supporting public sector campaigns.
What Success Profiles Actually Asks a Sifter to Do
Success Profiles replaced the older Civil Service Competency Framework and deliberately widened what a campaign can assess. Rather than one lens, it defines five elements, and a vacancy chooses which of them to assess and at which level.
- Behaviours: the actions and activities people demonstrate at work. There are nine, and each is defined across levels one to seven, where level one maps to the most junior grades and level seven to senior leadership.
- Strengths: the things a person does regularly, well, and with genuine motivation.
- Ability: the aptitude to perform a task, usually measured through tests rather than narrative.
- Experience: the knowledge and achievements gained through work and life to date.
- Technical: specific professional, occupational or craft skills and qualifications.
The nine behaviours are Seeing the Big Picture, Changing and Improving, Making Effective Decisions, Leadership, Communicating and Influencing, Working Together, Developing Self and Others, Managing a Quality Service, and Delivering at Pace. A campaign might sift on two behaviours at level three plus one technical requirement. Another might sift on experience and technical only, holding behaviours back for interview.
The Part That Makes It Slow
A sifter is not asked whether the candidate seems good. They are asked, for each assessed element, whether the written evidence meets the standard for that level, and to record a score that a panel chair can later defend. Two hundred applications assessed against three elements is six hundred discrete evidence judgements, each needing a rationale solid enough to survive a feedback request from an unsuccessful applicant who is entitled to ask why.
That is why Success Profiles sifting resists the shortcuts that work elsewhere in recruitment. You cannot skim for a job title. The framework explicitly rewards a candidate who evidences Delivering at Pace from a non-obvious background over one whose job title merely sounds relevant.
Why the Manual Sift Breaks at Volume
Three failure patterns show up repeatedly in public sector campaigns, and none of them is about effort.
Drift across the pile. Assessors are consistent with themselves over twenty applications and measurably less consistent over two hundred. The standard applied at application 190 is rarely identical to the standard applied at application 10, and the drift is usually in the direction of severity as fatigue accumulates.
Position effects. A merely adequate application reads as strong immediately after three weak ones. Manual sifting in submission order imports that ordering into the outcome, which is precisely the sort of arbitrary factor the Recruitment Principles exist to exclude.
Rationale decay. Under deadline pressure, recorded rationales get shorter. By the end of a large sift, the record often reads "does not meet required level" with no reference to what evidence was considered. That note is adequate for a spreadsheet and inadequate for a complaint.
What Automating the Sift Should and Should Not Mean
This is the point where most public sector conversations about screening automation go wrong, so it is worth being precise.
Automation should mean evidence extraction, structuring and scoring proposals: reading the application text, locating the passages that speak to each assessed element, mapping them to the level descriptors, and presenting that mapping to the panel with the source text attached.
Automation should not mean an unreviewed machine rejection. Under UK data protection law a decision with a significant effect on a person, made solely by automated means and without meaningful human involvement, is tightly restricted. A recruitment rejection sits comfortably inside the category of decisions people will argue about. Beyond the legal position, the Civil Service Commission's Recruitment Principles require selection on merit through fair and open competition, and a panel that cannot explain its own sift has not satisfied that duty regardless of how the score was produced.
The practical model is therefore a proposal layer, not a decision layer. The system does the reading and the structuring. A named human confirms, adjusts or overrides, and that confirmation is what actually moves a candidate.
A Technical Blueprint for Scoring Against the Five Elements
Behaviours: Extract the Incident, Not the Adjective
The signal for a behaviour is a specific incident with a context, an action attributable to the candidate, and an outcome. Most applicants have been coached toward a STAR structure, which helps, but plenty submit prose that buries a strong example inside a paragraph about their team.
A screening layer should extract candidate examples as structured objects: the situation described, the action the applicant claims personally, the stated result, and the scale implied. Scale is what separates levels. "Coordinated a review of a team process" and "led a review affecting four directorates and a statutory reporting deadline" are the same behaviour at very different levels, and the difference is legible in the text.
The critical design rule is attribution. Sentences that describe what a team or a department did are not evidence of what the applicant did. A system that cannot tell "we migrated" from "I designed and delivered the migration" will systematically over-score collective language, which is exactly the weakness that well-drilled applicants exploit.
Experience: Map to Requirements, Not to Titles
Experience assessment should resolve to the stated requirement. If the requirement is experience of managing a devolved budget, the system should look for budget ownership, quantum, and duration, and it should find it whether the applicant held it in a local authority, a charity or a private contractor. Title matching does the opposite: it privileges people who have already worked in the sector, which narrows the field in exactly the way public sector recruitment is meant to avoid.
Technical: Qualifications Are a Verification Problem, Not a Scoring Problem
Technical elements frequently rest on a named qualification, professional registration or accreditation. A screening layer can reliably identify a claim, normalise its variants, and flag whether the claim is present, absent or ambiguous. What it must never do is treat the claim as verified. Verification is a separate step against the awarding body or register, and the output should say "claimed, unverified" so that no downstream reader mistakes parsing for proof.
Strengths and Ability: Be Honest That a CV Cannot Carry These
Strengths are about motivation and natural preference, and Ability is measured through tests. Neither is safely inferable from written application text, and a vendor claiming to score them from a CV is overselling. The correct engineering decision is to decline: return no score, state that the element is not assessable from the submitted evidence, and leave it to the assessment instrument designed for it. A screening tool that returns a confident number for something it cannot see is worse than one that returns nothing, because the number will be believed.
What a Transparent Data Output Looks Like
The output is the part that determines whether the campaign is defensible later, and it should be boring, complete and exportable. For each applicant and each assessed element, the record should carry:
- The element and the level assessed against.
- The proposed rating, and the descriptor language it was matched to.
- The verbatim source passage the proposal drew on, with its location in the application.
- An explicit note where required evidence was absent, distinct from evidence that was present but weak.
- The confirming assessor, the timestamp, and any override with its stated reason.
That last line is the one that matters in a dispute. A record showing that a human assessor reviewed a proposal, considered identified evidence and reached a recorded conclusion is a defensible sift. A score with no provenance is a liability, whether a person or a model produced it.
Fitting This Alongside an Existing Campaign Process
Most departments and agencies are not going to replace the system their campaigns already run through, and they do not need to. The screening layer sits between application export and panel sift.
- Applications are exported from the campaign system in bulk once the vacancy closes.
- The batch is uploaded into the screening layer and mapped to the vacancy's assessed elements and levels.
- Evidence extraction and scoring proposals are generated for the whole cohort under one consistent standard, with no ordering effect.
- Assessors review proposals element by element, which is faster and more consistent than reading whole applications cold, and record confirmations or overrides.
- The confirmed outcome and full evidence record are exported back for the panel record and for feedback to applicants.
Reviewing by element rather than by applicant is a small change with a disproportionate effect. Assessing one behaviour across two hundred applications holds the standard still, because the assessor is comparing like with like rather than rebuilding the standard from scratch on every new application.
Guarding Against the Bias You Are Trying to Remove
Automated screening can reduce inconsistency and can also entrench a pattern if nobody is watching. Two safeguards are worth building in from day one.
First, monitor outcomes by protected characteristic at cohort level, not to set quotas, but to detect a scoring pattern that disadvantages a group without job-related justification. Indirect discrimination under the Equality Act 2010 does not require intent, and public bodies carry an additional equality duty on top.
Second, keep the evidence visible rather than reducing an applicant to a number. A panel that can see the passage behind a proposal can spot when the model has latched onto phrasing fluency rather than substance. An interface that shows only a score invites the panel to defer to it, which converts a proposal layer back into a decision layer through the back door.
Frequently Asked Questions
Can Success Profiles Behaviours Be Scored Automatically from a CV?
Behaviour evidence can be extracted and mapped to level descriptors automatically, and a proposed rating can be generated with the source text attached. The rating should be confirmed by a human assessor before it affects a candidate's progress. Behaviours are assessable from written evidence; Strengths and Ability are not, and should be left to the appropriate assessment method.
Does Automated Sifting Comply with the Civil Service Recruitment Principles?
It can, provided selection remains on merit through fair and open competition and the panel can explain its decisions. The Principles do not prescribe a tool. They require a defensible, merit-based process, which in practice means a documented standard applied consistently and a recorded human judgement behind each outcome.
How Do We Handle the Disability Confident and Guaranteed Interview Commitments?
Those commitments operate on the outcome of the minimum criteria assessment, so the screening layer must expose whether minimum criteria were met as a distinct field rather than folding it into a single overall score. If your output cannot answer "did this applicant meet the minimum criteria" independently, it cannot support a guaranteed interview scheme correctly.
What About Name-Blind Sifting?
Anonymisation should happen before evidence extraction, not after scoring. Remove names, contact details, institution names where policy requires it, and any other identifying field from the text the screening layer sees. Redacting the display while scoring the full text achieves nothing.
How Much Time Does This Realistically Save?
The saving comes from removing the first read, not from removing judgement. Assessors still make every decision, but they make it against extracted, structured evidence instead of hunting through free text. The larger the cohort, the greater the effect, and the consistency gain across a big pile is usually valued more highly than the hours recovered.
The Standard to Hold a Vendor to
If you are evaluating screening technology for a regulated public sector campaign, the questions worth asking are narrow and awkward on purpose. Can it show the source evidence behind every score? Does it decline to score what it cannot see? Does it record who confirmed each outcome? Can it export a complete audit record that a panel chair, an internal auditor or a tribunal could read without a demonstration? Where does the candidate data rest, and who else can reach it?
CVSense was built around evidence-based validation rather than keyword frequency for precisely these reasons: a public sector sift stands or falls on whether the reasoning behind it can be produced on demand. If you run high-volume Success Profiles campaigns and want to see what an element-by-element evidence record looks like against your own vacancy, that is a short and specific conversation worth having.
Sources
Civil Service. Success Profiles.
https://www.gov.uk/government/publications/success-profiles
Civil Service Commission. Recruitment Principles.
https://civilservicecommission.independent.gov.uk/recruitment/recruitment-principles/
Information Commissioner's Office. Rights Related to Automated Decision Making Including Profiling.
https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/individual-rights/rights-related-to-automated-decision-making-including-profiling/
Equality Act 2010.
https://www.legislation.gov.uk/ukpga/2010/15/contents
InsightCircle
