Automating Workday applications

Workday is deployed per employer, with its own hostname, its own configuration and its own selection of steps and questions. Two employers on the same underlying product can present noticeably different application flows.

So a handler written against one employer’s instance will not simply transfer. The platform-level structure is shared, but the specific steps, required fields and custom questions are configured by each employer, and an agent has to treat that configuration as something it discovers per site.

The right architecture follows directly: a platform-level handler that knows the general shape, plus a per-tenant configuration learned on first encounter and cached. The first application to an employer is expensive and every subsequent one is cheap, which is a perfectly good trade as long as the caching actually happens.

Applications typically require a candidate account on that employer’s instance, which means a separate account per employer — not one login that carries across. Some flows also involve email verification.

An agent should not be silently creating accounts and inventing passwords on someone’s behalf. Detect the requirement, stop, and let the candidate decide and complete it. Anything involving a code sent to their inbox is theirs, and the agent’s job is to resume cleanly afterwards.

Resuming cleanly is the part that makes this acceptable rather than annoying. If stopping means the candidate re-does everything, they will stop using the tool; if it means clicking a link, completing a verification and having the agent pick up at the exact step it paused, the interruption costs a minute.

• One account per employer instance, not a shared login
• Verification steps that need the candidate inbox
• Credentials stored by the user, encrypted and scoped, never invented
• A resumable handoff so stopping is cheap

These flows commonly run to several screens — profile, experience, education, questions, disclosures, review. Each screen is a chance to mis-fill, to trip a validation rule, or to lose a session.

Verify after every step: did the page advance, did a validation message appear, is the field populated with what was intended. Without that, an agent can proceed for four screens on a state it got wrong on the first.

Sessions expire mid-flow more often than anywhere else, simply because the flow takes long enough for it to happen. An agent that detects the expiry, re-authenticates and resumes from a recorded step is doing something a candidate manually filling the form usually cannot — which is one of the few places this automation is strictly better than a person.

Uploading a CV often pre-populates the work history, and the parse is rarely perfect — dates shifted, an employer name mangled, a title truncated. Accepting it wholesale produces an application containing errors the candidate never wrote.

Reconcile the pre-filled values against the structured profile you already hold, field by field, and correct rather than trust. A recruiter reads those fields as the candidate’s own account of their career.

Constrained inputs are their own category of problem. Where a field is a dropdown — country, institution, degree type — the candidate’s true value may simply not be in the list, and the correct behaviour is to flag it rather than to select the nearest plausible option on their behalf.

It is worth deciding in advance what makes an application not worth automating, because a long flow will otherwise absorb unbounded effort. A verification wall, an unrecognised required field, a validation error the agent cannot resolve in two attempts, or a screen whose purpose it cannot identify are all reasonable stopping points.

Stopping should produce something useful rather than an error. A saved partial application, a note about which screen it reached and what blocked it, and a direct link back is a handoff the candidate can finish in three minutes.

The alternative — retrying, guessing, or abandoning silently — produces the two worst outcomes available: a wrong application submitted, or a candidate who believes an application exists when it does not.

Voluntary disclosure sections are the candidate’s to answer or decline, and an agent completing them is making a personal declaration for someone else. Employer-specific screening questions with legal weight — sponsorship, notice, compensation — belong to stored answers the candidate confirmed, not to inference.

And review before submitting. This is a long flow with many opportunities for a wrong value to survive to the end; the confirmation screen is the last place to catch it, and it is worth showing the candidate in full.

Compensation expectations deserve a line of their own. A number entered by an agent on a candidate’s behalf anchors a negotiation that has not started yet, and no amount of inference from market data substitutes for the candidate deciding what to put in that box.