The challenge
The firm places physicians and advanced practice providers into temporary assignments, and their ATS does what an ATS does well: it houses candidates, tracks jobs, runs placements. The three parts of the work that actually require judgment were all happening in recruiters' heads.
Matching is the clearest example. An open hematology-oncology assignment at a rural critical-access hospital is not best served by the candidate with the strongest academic resume. It is best served by the one who has done three critical-access placements and knows what that environment is, holds an active license in the right state, and has a compensation history in the range the assignment can support. Keyword search surfaces the impressive resume and buries the right person. As the buyer put it, most staffing is checkers and this is five-dimensional chess.
Sourcing was named as the single biggest day-to-day obstacle. Pulling a list of licensed physicians in a specialty and a state is easy and public. Finding a way to actually reach them is not, and it is getting harder. Recruiters were spending their hours hunting for contact information rather than making contact.
Outreach was generic templates going out and no intelligence coming back. A recruiter could not tell, without reading every reply themselves, which responses were genuine interest and which were politeness. Providers went cold while warm ones waited in an inbox.
And then the constraint that shaped the entire design: the firm had previously rolled out a general-purpose AI product company-wide and it had failed. Not because the technology did not work, but because adoption required every employee to learn a new tool and work out for themselves where it helped. The buyer described change management as his biggest challenge in his own words. Any system that asked recruiters to open something new was going to fail the same way.
The solution
Three modules that run underneath the ATS the recruiters already use. No new interface, no new login, no training on a second system. That constraint drove every other decision.
The matching engine scores every provider in the database against an open assignment on semantic fit rather than keyword overlap: subspecialty alignment, license states, prior placement patterns, compensation history, geographic preference, availability signal. The top ten to twenty appear as candidate suggestions on the job record itself, each with one plain sentence explaining why. Something a recruiter can agree or disagree with in three seconds, rather than a similarity score they have to take on faith.
Outreach drafts the first touch across email and SMS in language that reads like somebody who understands the specialty, and the recruiter approves or edits before it sends. For pipelines the recruiter has already marked warm, it can send inside guardrails they set. The part that changes the working day is reply handling: every response gets scored for intent, and only genuine interest interrupts the recruiter. Polite declines and wrong contacts get logged without a notification.
The sourcing engine feeds normalized provider records back into the pipeline so the matching has current data to work against, with compliance scaffolding for TCPA and CCPA built in rather than added later.
The design is explicit about what the data can and cannot do, because that is where these projects usually break their promises. NPI, specialty and license state resolve at 95 to 100%. Practice address and phone at 90% or better. Current employer at 70 to 85%. Personal email at 30 to 50% with licensed data sources behind it and 10 to 20% without. Personal cell at 15 to 30% with, under 10% without. Anyone promising better than that is either reselling something or guessing.
The results
The day-to-day change is narrower than the module list suggests, and it is better for being narrow. A recruiter opens a job record and the ten to twenty best-fit providers are already sitting there, each with one sentence explaining the match that they can agree or disagree with in three seconds. First-touch drafts are waiting for approval rather than for composition. The only replies that interrupt them are the warm ones; polite declines and wrong numbers get logged without a notification. Nothing about the recruiter's Monday looks different from their Friday, which is the entire point.
The architecture decision carrying the most weight is modularity. Rather than a platform claiming to handle every specialty on day one, the engine launches with two and takes each additional specialty as a module at $1,500 to $2,500 with a small retainer adjustment. Known cost, known timeline, known outcome per specialty. A firm that has already watched one AI rollout die does not need another all-or-nothing bet. It needs to see one desk work before committing the next.
The volumes are real ceilings rather than marketing numbers: up to 500 enriched provider records and 2,000 outreach activities a month, with prompt tuning and a weekly digest included. Licensed data subscriptions sit with the client rather than being resold at a markup, because a vendor who profits from the data layer has every incentive to recommend more of it than the desk needs.
Eight weeks to a live first desk. The strongest argument in the whole design is not any of the AI in it: the previous failure was a change-management failure, so the answer had to be architectural rather than a better training plan.