The problem with bottoms-up AI in staffing and how to fix it.
A few months ago I sat in on a call with a staffing firm's ops leader who was proud of what their desk had built. They were writing job descriptions with ChatGPT, had a resume screening prompt saved in a doc, and were even using AI to draft outreach messages. Usage was up everywhere.
She wanted to know if I thought her firm was ahead of the curve on AI.
I asked her a different question. Not how many people were using it, but what it had actually produced - something specific she could point to and say, “this is what AI got us this quarter”.
She could not answer that and neither could anyone else on the call. The tools were everywhere but nobody was tracking what any of them were actually worth.
Usage is easy to see. Value is not, and confusing the two is what turns a promising climb into a plateau that fails to deliver any real results.
User-led experimentation vs a strategic approach that’s built to scale
Give a staffing organization access to AI tools, encourage people to experiment, let people build their own workflows, and adoption climbs almost immediately. People are curious, the tools are easy to pick up, and everyone wants to be the one who found the shortcut. That's the bottoms-up curve, and it earns its reputation. It is the fastest way to get a room full of skeptical recruiters to actually try something new.
Then it plateaus. Not because people lose interest, but because the firm hits the ceiling of what individual experimentation can do on its own.
The other curve is slower to take shape. A top-down AI strategy means governance before rollout, a short list of prioritized use cases instead of a hundred small ones, and data that has been cleaned and unified before anyone builds anything on top of it. Although it looks less impactful at the start, it keeps climbing well past the point where the bottoms-up curve flattened out, because every gain compounds.
What the fast bottoms-up climb looks like up close
Here is a version of a conversation I have had more than once this year…
A firm's Chicago office builds a GPT that screens resumes against a job order. It works well enough that the branch manager is thrilled, and word spreads to two other offices.
Each of those offices builds its own version, because nobody owns this centrally and nobody knew the Chicago tool existed.
Three months later, the same candidate gets flagged as a strong match by one office's tool and a weak match by another, because the underlying prompts, the weighting, and the source data behind each one are all slightly different.
A recruiter in a fourth office notices the mismatch, stops trusting any of the tools, and goes back to screening resumes by hand.
Nobody did anything wrong here - every person involved was trying to move faster for their desk. That is exactly why bottoms-up AI is so hard to slow down once it starts, and exactly why it stalls out on its own. Fragmented use cases, duplicate agents solving the same problem three different ways, conflicting outputs on the same candidate, and experiments that never make it past the person who built them, that is not a hypothetical risk list.
The data backs up what we are seeing firsthand:
Microsoft's 2025 Work Trend Index found that the large majority of employees using AI at work are bringing their own tools rather than using anything sanctioned by their organization.
KPMG's 2025 research found that a large share of organizations still have no formal policy governing how employees use outside AI tools at all.
Adoption metrics do not tell you what any of it is worth
Staffing is not a special case here. It is a fairly typical example of what happens when a workforce this eager to move fast meets a technology this easy to pick up alone.
Kate Smaje, McKinsey's global leader of technology and AI, has pointed to the same failure showing up across industries: leadership teams build scorecards that compare adoption of tool A against tool B against tool C, market by market, and none of it says whether any of the work is creating value. Discussing that instinct in a recent McKinsey interview, Smaje called it plainly, "That's a highway to nowhere."
She has also started calling the tools nobody tracks anymore abandonware, prompts and agents that were useful once, left running, and never checked again to see whether they are still doing anything worth paying for. That is exactly what I saw on that call. Three offices, three tools, and not one person who could say which of them, if any, was actually finding better candidates faster.
What the slower top-down climb looks like once it gets going
Compare that to a firm we worked with that took the opposite approach. Instead of letting each desk build its own AI shortcuts, leadership picked one problem worth solving well, matching open job orders against the firm's own candidate pool, and built it on top of data that had been cleaned and standardized first.
It took longer to get it live, but once it launched, every recruiter in every office was working from the same trusted matching logic and the same underlying data, not five different versions of the same idea. The firm did not have to figure out why one office's tool disagreed with another's, because there was only one tool, built once, governed centrally, and improved centrally as real use came in.
That is the compounding part of the curve. Every improvement lifts the whole firm at once instead of staying wherever it was built.
Why staffing firms keep ending up on the bottoms-up trajectory anyway
None of this is a knock on the recruiters and ops managers building these tools on their own. Most of them are doing exactly what a good operator should do when the tools available to them clearly work.
The problem is that nobody in most staffing firms has the bandwidth to be the top-down layer. IT is already stretched managing the ATS itself. Ops leaders are already running the desk. Bullhorn Amplify has put real AI skills within reach, but someone still has to decide which skills matter most, map them into clean data, roll them out consistently, and keep them running well after launch week. Without that layer, momentum is lost and investment is wasted.
Where Dispatch fits, and why we built it this way
This is the exact gap Dispatch was built to close. Dispatch is our managed digital workforce service for Bullhorn Amplify, developed alongside Bullhorn's own Amplify team.
Amplify supplies the AI skills. Dispatch turns those skills into real, role-based digital workers that operate inside your Bullhorn instance the way a well-trained employee would, not five slightly different experiments running in parallel. Broad & Madison manages those digital workers, optimizes how they perform over time, and reports the results back in terms an operator can actually use, not a dashboard of usage metrics that never quite ties back to placements.
We built Dispatch this way because we watched the same pattern play out at client after client. The firms getting real, lasting value from AI were never the ones with the most individual tools in circulation. They were the ones who treated AI adoption the way they would treat any other operational change, with a clear owner, a prioritized rollout, and data worth building on.
If your firm already has a recruiter who built something clever on a Friday afternoon, that is not a bad sign. It means the fast climb has already started somewhere in your business. The real question is whether that climb is the whole story, or just the first thirty feet of a much longer one that somebody still needs to plan.
DISPATCH FOR BULLHORN AMPLIFY
Your Digital Workforce, dispatched.
Where skills become roles, and roles become results.
FAQ
-
Bottoms-up AI adoption starts with individual employees or offices building their own tools, which produces fast early results that plateau once fragmentation and inconsistent data catch up with it.
Top-down AI adoption starts with governance, prioritized use cases, and unified data, which takes longer to show results but compounds because every gain scales across the whole organization instead of staying with one person.
-
It depends on the use case, but it is rarely immediate. The tradeoff is that the results, once they arrive, apply to every desk at once instead of the one office that happened to build something on its own.
-
Start by asking what specific business result each tool is supposed to move, then check whether anyone is actually measuring it. If the honest answer is that you can only point to usage, not outcomes, that is a sign the work needs an owner and a reporting layer, not just more tools.
-
Dispatch takes the AI skills available through Bullhorn Amplify and turns them into managed, role-based digital workers inside your Bullhorn instance. Broad & Madison manages, optimizes, and reports on their performance, so the firm gets a consistent digital workforce rather than a collection of one-off experiments.
