Workable

Named systems cover every category the product touches, all listed as available on every plan: job board distribution (200+), Indeed Easy Apply, Apply with LinkedIn, LinkedIn Recruiter System Connect, two way calendar sync (Gmail and Microsoft 365), Zoom, Teams and Meet, background check integration, native e-signatures and integrations, and payroll integrations, with a partner estate at partners.workable.com.

These are authenticated writes into systems Workable does not own, and the direction is named in several places. Job distribution pushes a posting out to more than two hundred external boards. Two way calendar sync writes interview events into the customer's Google or Microsoft calendar and reads availability back. Self scheduled interviews create meetings in Zoom, Teams or Meet. LinkedIn Recruiter System Connect is an authenticated two way link into the customer's LinkedIn Recruiter seat.

The agent acts through these rather than beside them: it sends personalized emails that drive applications, chats with candidates to capture facts, and invites them to a screening call, so outbound mail and calendar bookings land in systems outside Workable, unattended and at volume.

The Chrome LinkedIn profile extension is a capture path the recruiter operates, ingestion rather than browser use. Workable also offers its own API and MCP server for outside callers. The job board figure is the vendor's own, and the partner catalog is not enumerated.

Sourceworkable.com/pricing feature matrix and /workable-ai agent stepsread 2026-09-12

The pipeline is published as eight numbered steps spanning intake to handover: understand the role and the ideal candidate profile; fine tune that profile, marking must haves, disqualifiers and what appears in the job post; source from 400M+ profiles; send personalized emails that drive applications; chat with candidates to fill gaps and capture basic facts; score responses, CVs and profiles against the job criteria; move strong candidates to the shortlist or invite them to a screening call; and hand qualified candidates to the team. That is eight stages with a branch at step seven and a terminal handover.

The spec produced at step two governs every step after it: the ideal candidate profile determines what is sourced, what is asked in the chat and how the scoring weights apply. The launch material describes this as forcing clear articulation of what is needed before the first candidate is reviewed.

The flow loops back rather than running once: "As candidates respond to messages, take assessments, or progress through your pipeline, the agent re-evaluates them in real time. Your shortlist dynamically updates." A candidate can return to scoring long after first contact, so the process holds state across days and external events.

Participants are distinct and the handoffs are defined: the recruiter who sets the spec and can override at any point, the agent, the candidate who is messaged and chatted with directly, and the hiring team who receive the shortlist. The agent yields automatically the moment a human engages a candidate, a defined transfer of control. Obvious rejections are filtered so they do not slow the queue.

Sourceworkable.com/workable-ai eight-step breakdown and FAQread 2026-09-12

The agent is grounded in two maintained corpora, both persistent and both queried. The external one is a standing index of more than 400 million candidate profiles, listed in the pricing feature matrix as "400M+ candidate profiles" alongside passive candidate sourcing, and the agent sources candidates from it as step three of its pipeline. The internal one is the customer's own estate: a talent CRM database with past candidate resurfacing, so prior applicants can be searched again rather than archived. Both stand between runs and far exceed any context window.

The spec is a structured retrieval object, not a prompt. The agent builds an ideal candidate profile through a guided intake conversation, structured into matching categories the vendor enumerates (experience level, tools and technologies, certifications, languages, domain knowledge and others), with the customer marking each as a must have, a nice to have or a disqualifier. Candidates are then evaluated against "up to 14 criteria specific to your job," each scored independently and normalized, so retrieval is matched against a customer authored specification rather than a free text query.

Grounding is continuous: "As candidates respond to messages, take assessments, or progress through your pipeline, the agent re-evaluates them in real time. Your shortlist dynamically updates." Resume parsing is automated on intake and the agent chats with candidates to fill gaps, so the corpus is actively completed rather than passively read.

No document ingestion for the customer's own policy content, citation to source documents or customer facing search over the vendor's training data is claimed.

Sourceworkable.com/workable-ai agent steps and FAQ, with /pricing sourcing suite featuresread 2026-09-12

Workable publishes its oversight model as its own section, headed "You stay in control," with three mechanisms named separately.

Configurable autonomy: "Configure the agent per account, per job, and per candidate. Set sourcing parameters, screening thresholds, engagement rules, and scoring weights. Change any setting in seconds." The buyer decides what the agent may do, at three levels of granularity.

Reversibility with a stop condition: "Every action, from sourcing to engagement, is logged and reversible. The agent stops the moment you engage a candidate. Review, override or reject any decision instantly."

A global off switch: "Disable the agent across your entire account at any time. No downside. Your existing candidate data, scorecards, and pipelines stay intact."

The stop condition is the most interesting of the three: the agent yields automatically when a human enters the conversation, so the handover is a property of the system rather than a discipline the recruiter has to maintain. A confidence threshold also routes to a person, since low confidence inferences pass to a human, so uncertainty escalates by design. The scope of autonomy is itself elective; the launch material states "you choose what the agent does: sourcing, outreach, scoring, auto-advance."

The decision boundary is stated too: recruiters still own the relationship and the final call, and the agent hands over an interview ready shortlist rather than making hiring decisions. Workable also states that you maintain human oversight at every step as part of its EU AI Act alignment, so oversight underpins its own regulatory position.

Sourceworkable.com/workable-ai control and compliance sectionsread 2026-09-12

Workable holds four attestations, publishes two of the certificates as downloads, and links Security from the footer of every page.

Its security page states it is certified with ISO 27001 for its information security management system and ISO 27017 for security controls in the provision and use of cloud services, publishing the certificate PDF for each; SOC 2 Type II covering Security, Availability and Confidentiality, with the report available through an account manager; and a SOC 3 report published openly.

The AI product page repeats the set: "We hold ISO 27001, ISO 27017, and SOC 2 Type II certifications." ISO 27017 is the cloud specific standard and a narrower, more informative claim than a general ISMS badge.

The customer facing controls are named across both surfaces: single sign on, listed under advanced features on every plan and with its own help article; advanced access controls, custom user permissions and additional permission sets in the ATS feature list; audit trails in the HR feature list; and candidate data retention rules with automatic deletion when a retention period expires. A web application firewall sits above the cloud providers' own controls, security logs are analyzed for abnormalities, third party technical security assessments are commissioned and internal red teaming is run.

Governance at the agent level is separate and granular: the Agent is configurable per account, per job and per candidate, with an admin kill switch that disables it across the entire account at any time, leaving existing data, scorecards and pipelines intact.

The privacy program is a product rather than a policy: a Data Protection Officer is appointed, GDPR automation tools let a customer exercise a candidate's correction, deletion and access rights from the platform, consent is captured by default on every Workable generated application form, and EEO, OFCCP, CCPA and CPRA compliance features are listed.

Sourcehelp.workable.com security and data privacy article, with /workable-ai and /pricing feature listsread 2026-09-12

Workable publishes both an action log and an explanation layer, under a heading that names the standard. Its AI page has a section titled "Transparent, not black box" reading: "Every score has reasoning. Every candidate action is logged. You can audit exactly why the agent did what it did." The control section repeats the log and adds that every action, from sourcing to engagement, is logged and reversible. A log you can act on is a stronger artifact than one you can only read.

The scoring is decomposed rather than opaque. The agent evaluates against up to fourteen job specific criteria; "each criterion is scored independently and then normalized to a 0-100 scale," and "you see the score for each criterion, plus the overall score." A reader can see not only that a candidate scored 62 but which of the fourteen criteria produced it. That answers why this candidate; the action log answers what the agent did; Workable has both.

Audit trails are a named platform feature beyond the agent, listed in the HR information system feature set, alongside a full reporting suite and a custom report builder on every plan. Data export is listed in the ATS features, so the record can leave the system.

Workable positions explainability as part of its EU AI Act alignment and states it is committed to making its AI functionality as transparent and explainable as possible, so the audit surface underwrites its own compliance claim.

Sourceworkable.com/workable-ai transparency and control sections, with /pricing feature matrixread 2026-09-12

Three things could look like agent memory, and none is. The ATS is the system of record. Workable holds a talent CRM database, candidate profiles, scorecards, pipelines, interview kits, employee records and role history, and the agent reads all of it on every run. That is the customer's application data, persistent because it is a database, in the same way a CRM or an ERP is persistent, and it serves as the knowledge corpus the agent queries.

The ideal candidate profile is configuration, not accumulation. It is authored once through a guided intake conversation, with the customer marking must haves and disqualifiers, and it governs the run thereafter. Reevaluation as candidates respond is the agent rereading current facts against that fixed spec, which is state within a run.

Learning from outcomes is absorbed into the model, as the vendor describes plainly: Workable's AI tools are "trained using implicit or explicit feedback from daily user interactions and analyzing why candidates are hired or moved through the hiring stages," with customer data used to enhance the performance and accuracy of its in house systems. Nothing states that this learning is scoped per customer, inspectable or reapplied as recallable context.

The product page, a fourteen question FAQ and a pricing matrix listing roughly seventy features show no named memory component, no retained recruiter preference or correction the agent reapplies to the next requisition, no decision trace carried between roles, no stated scope or lifetime, and no surface to inspect, edit, export or purge agent state separately from the candidate records. The site is detailed enough to list audit trails and data retention rules and never mentions memory.

Sourceworkable.com/workable-ai FAQ and /pricing feature matrixread 2026-09-12

Cloud providers are named, regions are not. The security page refers to the security components provided by Workable's top level cloud providers, Google and AWS, so infrastructure is disclosed at provider level and stops there. No data center location, region list, residency election or customer hosted option is published, so this is plain multi tenant SaaS with no named region, customer environment or selection surface.

The absence stands out against a surface that itemizes everything else. The pricing page lists roughly seventy named features across recruiting, HR and time and attendance, down to auto fill resume applications, org chart generation and multi currency support. It includes multi entity and multi currency support, which are organizational and financial structure rather than data location, and carries no region, residency or hosting row at any tier, Enterprise included. Nine native language options are listed and no hosting region, which suggests region is not a choice on offer.

This matters for this vendor in particular. Workable is a UK registered company with a heavily European customer base named on its own AI page (Decathlon, Debenhams, Starling Bank, Boohoo, Grecotel, Korres, Ogilvy), holds ISO 27017 for cloud security controls, publishes GDPR and UK GDPR automation tooling, and states its agent is aligned with the EU AI Act. It processes candidate personal data at scale for those customers and publishes no residency position. A buyer with an EU localization obligation has nothing on the site to work from.

Sourcehelp.workable.com security and data privacy article, with /pricing feature matrixread 2026-09-12

One agent ships, not a set of agents to choose among.

The product page presents the agent as a single eight step pipeline: understand the role, fine tune the ideal candidate profile, source from 400M+ profiles, send personalized emails, chat to fill gaps, score against the criteria, move strong candidates to the shortlist or invite them to a screening call, and hand over to the team.

Sourcing is step three and scoring is step six of that sequence, not separate products. Workable's own framing throughout is singular, the Workable Agent, and the pricing page sells it as one add on line item on every plan.

Remove the screening step and sourcing does not stand as a whole product a buyer adopts instead; a customer does not choose the scoring agent over the outreach agent.

Still, the agent ships prebuilt and needs no build step, the job brief capability is available to every Workable Recruiting account rather than only to add on buyers, and reusable artifacts sit around it (interview kits and scorecards, offer letter templates, custom hiring pipelines, standardized evaluation tools and custom application forms), all listed as included features. Those are templates a customer selects and reuses, even though the customer authors most of them.

The add on list is a pricing structure, not a pack: Texting at $89, Video interviews at $109, Assessments at $59 and Performance and Engagement at $39 are priced modules of one platform, not working agents a buyer selects among.

Sourceworkable.com/workable-ai eight-step agent breakdown and /pricing feature matrixread 2026-09-12

Arrival is the trigger, and the vendor is emphatic about it: every applicant is read against the spec "the moment they enter," Sunday night included, set against the batching it replaces, where applications pile up and are reviewed in batches. The worked claim is "apply at 11pm, screening call booked by 9am." That is event driven processing with no person at the trigger point.

Three further event classes are named. As candidates reply to messages the agent re-evaluates them in real time, and assessment completion and pipeline progression both reopen scoring. Continuous sourcing runs on no external event at all: passive search, outreach and follow up "run continuously" for every role, not just the senior ones, a standing background process rather than a scheduled job. Follow up implies timed recontact.

Channels are the candidate facing ones and they are enumerated: personalized email as the outbound path, a chat the agent conducts directly with candidates to fill gaps and answer questions, screening call invitations landing in a calendar, plus texting, video interviews and assessments available as platform channels. Self scheduled interviews with two way Gmail and Microsoft 365 calendar sync close the booking loop. A careers page, jobs widget and 200 plus job boards are the inbound application channels.

No conversational channel for the recruiter, such as Slack or Teams, is offered, and the outbound email and calendar writes run through the integrations.

Sourceworkable.com/workable-ai operating-model sections and FAQ, with /pricing feature matrixread 2026-09-12

The models are Workable's own, by its own account.

The AI FAQ says Workable is committed to "developing our own in-house models," describes using "proprietary systems and metadata" to enhance generative AI output, states that its focus is to use customer data to enhance the performance and accuracy of "our in-house AI systems," and commits to self evaluating "our own in-house tools" for bias and model drift.

It contrasts itself explicitly with ChatGPT and Claude as black box applications. A vendor that says four times that the models are its own, and names no provider anywhere, has no model for a customer to choose.

One line reads like bring your own model and is not. Workable's homepage says: "You set the operating model and the guidelines, your data stays yours, and the platform works with whatever AI you bring." The same page resolves it two sentences later: "Access to the MCP lets your own AI plug in too." The AI a customer brings is their own assistant connecting inbound through Workable's MCP server, an extension point rather than a model choice. Bringing Claude to Workable does not change the model that scores a CV against fourteen criteria.

No routing across more than one model inside Workable is disclosed either, and no provider is named even once across the AI page, the pricing matrix or the security documentation.

Sourceworkable.com/workable-ai FAQ and workable.com homepageread 2026-09-12

Three inbound surfaces are published by Workable itself.

A developer API has its own footer entry on every page, at workable.com/developers, and API access is listed as an included feature twice in the pricing feature matrix, once under the applicant tracking system and again under time and attendance, so it is an entitlement on every plan rather than an enterprise upsell.

A partner and integrations estate runs at partners.workable.com. And an MCP server is stated on the homepage in the vendor's own words: "The autonomous Workable Agent runs the top of your hiring funnel, and access to the MCP lets your own AI plug in too."

The MCP server stands out: a customer's own AI can call it directly. Workable frames it as the customer's AI plugging into Workable, the inbound direction, and pairs it with the standing REST API rather than offering it instead of one. The agent acts on candidates, jobs and pipelines, so an outside caller reaches the working system rather than a reporting endpoint.

The endpoint list, the authentication model and the MCP tool surface are named rather than enumerated, and no SDK is claimed.

Sourceworkable.com/pricing feature matrix, workable.com homepage MCP statement and the site-wide Developer API footer linkread 2026-09-12

Scoring and bias research could pass for evaluation, but neither puts the agent under the customer's test. The scoring is the agent's output about candidates, not a measure of the agent. Workable scores each candidate against up to fourteen criteria, normalizes to 0 to 100 and shows the breakdown for each criterion. That is a readable result about the candidate, which answers a different question from whether the agent's own behavior is under test.

Bias and drift evaluation is real and is Workable's own. Workable states "we will consistently self-evaluate our own in-house tools for bias and model drift," and is active in the AIXPERT and AutoFair research programs on AI fairness and a member of the European AI Ethics Consortium. Model drift is the right object, but it is Workable evaluating Workable inside its own development process, with no surface exposed to the customer. The buyer receives the outcome, not the harness.

Nothing puts a customer's change under test. No test environment, sandbox, dry run, holdout, A/B comparison or versioning of the ideal candidate profile appears anywhere. A recruiter can change screening thresholds and scoring weights in seconds, and nothing exercises that change against a known set of candidates before it starts filtering live applicants, or reports whether a spec change improved the shortlist.

The one surface aimed at the agent itself, low confidence inferences routing to a person, is a runtime guardrail, a form of oversight rather than a measurement.

Sourceworkable.com/workable-ai FAQ and control sections, with /pricing feature matrixread 2026-09-12

The agent's action surface is the ATS, email and a chat, none of which is an interface it navigates. Workable's eight published steps are sourcing from its own profile index, sending personalized emails, chatting with candidates, scoring against criteria and moving records between pipeline stages. Every one of those is an operation inside Workable's own product or an outbound message. Nothing reads a rendered page belonging to someone else.

UI changes elsewhere do not break it. Workable's integrations are named API level connections (two way calendar sync with Gmail and Microsoft 365, Zoom, Teams and Meet, LinkedIn Recruiter System Connect, Indeed Easy Apply, background check and payroll integrations), so a third party changing its interface does not touch the agent.

The Chrome extension is not browser use. The pricing feature list includes a Chrome LinkedIn profile extension, but the recruiter opens a LinkedIn profile and the extension captures it. That is ingestion the recruiter operates. Workable publishes no LinkedIn send path; its LinkedIn connections are Recruiter System Connect and Apply with LinkedIn, both sanctioned integrations.

No hosted or local browser, desktop session, remote or local computer control, RPA or recorder appears across a product page, a fourteen question FAQ and a pricing matrix of roughly seventy line items that lists a mobile app and a jobs widget by name.

Sourceworkable.com/pricing feature matrix and /workable-ai agent stepsread 2026-09-12