What responsible recruitment AI looks like in practice

Generative AI has weakened the traditional application. Candidates can produce polished CVs and cover letters quickly, application volume rises, more applicants appear convincing on paper, and recruiters can no longer discern suitability from this traditional route.

The obvious solution is automation, and indeed, the majority of products out there use AI to automate the traditional pipeline. However, as I discussed in Parts 1 & 2 of this series, there are legal, not to mention ethical, restrictions on this and for good reason - AI in/AI out does a disservice to both candidates & recruiters.

We at Decision Agent believe the promise of AI lies in allowing us to get more out of the candidate, and interrogate that output rigorously. Give candidates a level playing field and a fair chance, and recruiters the information they need to make good, informed decisions.

Principle 1: Test candidates in the work

AI is excellent at content generation, and we believe this is one of the best uses for it.

An intelligent LLM with the right research harness can write realistic, expert-level scenarios for any role, a level of individualization and content generation previously only possible for headhunted roles.

Decision Agent turns the role into realistic scenarios & conversations that simulate the decisions that candidates will need to make, and test the level of knowledge they should have to succeed in the job.

Principle 2: Define the standard first

The hiring team defines the relevant behaviours and scoring framework before candidates respond, both positive & negative. Applying consistent, role-relevant criteria makes the assessment possible to justify, test and audit as opposed to an opaque model that invents its own idea of suitability. Obscuring these standards through a scenario narrative allows us to reveal how candidates would act, rather than how they think they should answer.

Principle 3: Make every recommendation traceable

Recruiters can see what the candidate faced, what they said, how each behaviour was judged and which evidence supports the finding. The result is focused interview questions, allowing recruiters or even hiring managers directly to start an interview with relevant contextual questions for each candidate. We believe recruitment specialists should be able to devote their time to shaping the pipeline, rather than acting as a screen ahead of hiring managers.

Principle 4: Let recruiters correct the system

Recruiters can challenge individual judgements, make corrections and rerun the report using their feedback. Human oversight is only legally present when the reviewer can understand and change the result, and we are not looking to replace human expertise, only leverage it.

Principle 5: Make only the claim the evidence supports

We assess observable behaviour in a defined interaction. This makes validation much more achievable - can our platform reliably assess the presence of this behaviour? As mentioned in Part 2, the broader the claim, the harder it is to validate, and without validation responsible automation is not possible. We do not infer emotion or personality, or claim that one response proves suitability or future performance.

Principle 6: Set the conditions for automation

Automation of judgemental tasks is not what we offer right now, nor should any other product on the market. What we provide is the framework to collect the data to achieve automation in the future. Through collation of candidate, recruiter and employee performance data over time, we aim to achieve validation - can we predict successful hires from how they act in our assessments. That is the pot of gold at the end of the rainbow, and it's not about saving money - it is instead about fairness. Candidates should not be subject to inconsistent evaluation if at all possible, and it is this consistency that only a machine can provide, not subject to the vagaries of the human condition explained so clearly in Thinking, Fast & Slow.

A worked example

For a Customer Success Manager role, the candidate could face an unhappy customer whose usage is falling and who wants a discount. The hiring team would define behaviours in advance: identify the commercial risk, seek the missing evidence, and propose a clear next step. They would also define negative behaviours they would want exposed, for example over-promising, dismissal of concerns or lack of technical understanding. Decision Agent assesses the transcript against those behaviours, shows the relevant evidence and lets the recruiter correct any finding. It makes no claim about the candidate’s emotion, personality or future performance.