Read time: 9 Minutes

A collaboration with Geoff Momin from Interlay.

Somebody above you wants an AI plan 🥴

Maybe it was a keynote, maybe it was a board deck, maybe it was a competitor’s press release. However it played out, the request landed on your desk, and it’s up to you to ground it in reality.

So let’s ground one together today 🤓

But first, if you’re headed to Workday Rising and feel as wary about the onslaught of AI hype as this reddit thread, but also have an inkling there’s something real to this wave, you’ll want to register for this event…

Now back to it!

You're hiring for a securities role at a financial institution with branches across the U.S.

Candidates for this role have to pass the Series 7 exam before they can go customer-facing, so you bring them in as trainees while they work toward it. To satisfy the onboarding policy, your org just needs confirmation the candidate has booked the exam before you can move them onto the next stage.

Simple enough.

Except, there’s a looot of friction! The available tools in Workday (document upload, questionnaires, etc.) are clunky, disconnected, and still require regular human review.

And you want something well built 🦄

So, you consider…

…is this a good use case for an AI agent? How would you test it? And once you’re clear, how do you decide whether to buy a product or build it yourself?

We’ve got 5 steps and a framework to help you get there.

The entire purpose of AI is to automate or outsource the work you don’t want to do. Which usually means the best place to direct your attention is towards the processes that cause high friction.

Think: People chasing responses. Correcting the same records over and over. Digging facts out of documents by hand. Lots of emails, lots of repeated clicks, and steps chronically forgotten.

Then, go ask the people actually doing the work where it stalls, and what they'd love to never do again. They'll have a lot to tell you 😆

In our example, we’re willing to bet your recruiters don’t love the process of verifying submitted exam documentation, scrolling through the policy diligently, and requesting resubmission when it isn’t up to snuff. They also would rather not be on the hook if any candidate malfeasance sneaks through and surfaces in your audits (hellooo, fines 🤑).

Interlay’s developed a framework called PSDE: Process, Specifics, Data, and Evaluation.

It helps you turn a broad request like “use AI for candidate compliance” into a description you can test.

This is where you define what actually happens from start to finish.

In our example, a candidate submits evidence of an exam booking. So, you’d define: who reviews it, what gets recorded in Workday, and how do incomplete submissions or exceptions get handled (or resubmitted)?

Take special notice of places where someone has to interpret documents and work out how a policy applies (this is called “reasoning”), and whether those patterns are numerical (thresholds, distributions, outlier detection) or linguistic (interpreting unstructured text, classifying intent, extracting meaning from documents)!

This is where your AI can be most useful in removing the burden from your people.

Now you put numbers to it! 🔢

How many candidates go through this each month? How much time does each review take? How often does someone need to follow up?

And don’t forget the ugly bits!

Include time spent fixing errors, and the number of hires delayed because a candidate slipped through the cracks.

These specifics help you prioritize your problems and measure the before/after (your CFO lives for a high ROI 🤓).

This is where you map what data goes in and what comes out.

In our example, this is the candidate’s job requisition, the submitted documents, the dates, and the applicable policy for the role.

With the data established, you then decide where that information should live and who gets access to it.

This builds your technical requirements list. What gets exchanged, between which systems, under whose security.

Your evaluation criteria is the last piece of the framework (you success meter ⏲).

If success criteria can't be articulated explicitly, it’ll be difficult to qualify the solution or justify pursuing it at all. Some evaluation metrics you can measure include review time, and the count of corrections or requests for resubmission.

Basically, how is the solution (with AI playing its part) performing end to end?

With your PSDE in place, it’s time to question everything!

Don’t be in the business of just replicating an old solution with a new host. Take the opportunity to ruthlessly delete what no longer serves your org or can’t be defended.

Question every step, condition rule, approval, and notification. Focus on favoring governance, user experience, adoption, and approvals (where you still need a “human in the loop” 🤪). Look for the baggage that has accumulated around those justified steps — removing it might end up bringing the process closer to what your policy intended.

We can’t overstate how much this stage matters.

AI can quickly become the most expensive way to solve a problem if you end up paying to automate steps you could have easily deleted. Plus, simplifying and testing your requirements gives your vendors or in-house developers a clear understanding of what qualifies as a deal-breaker vs. a nice-to-have.

AI prototyping tools are a dime a dozen these days, and you might already be paying for a few of them. Which one you decide to lean on depends on the mock-up you want to build and your technical background.

Here are just a few, roughly ordered from lowest-touch to highest-touch. Start as far up this list as your problem allows.

(Noting a pattern in these bios? So did we 👀)

Whichever direction you move in, the purpose of a prototype is the same: elicit real-time reactions from your stakeholders and possible vendors.

Watching someone use the thing (or analyze it) gives you information that might stay hidden beneath the surface across multiple meetings. It’s an amazing “get on the same page” tool and can even spark fresh ideas.

This is where all your work pays off — or doesn’t. You can build the most elegant solution in the world, but if people don’t naturally reach for it, it just sits there! 🦗

During this step, you want to think about how people will encounter the solution.

If a recruiter has to remember to ask an agent to collect each candidate’s exam confirmation, the process still depends on their memory. For back-office work tied to a known event or deadline, the agents should start automatically.

There are two main kinds of agents:

The same process can use both.

In our example, a candidate should receive an email automatically when it’s time to submit their confirmation (ambient). When they respond, the agent extracts the facts, identifies the gaps and notifies the reviewer, with the evidence, policy requirements, and any discrepancies ready to inspect. Nothing falls through the cracks.

A manager might ask an assistant for the status of a compliance check already in progress for their candidate (delegated). The agent would query the candidate’s compliance event decisions and return their status with options to send follow-ups and request resubmissions.

As you work through this exercise, you’ll quickly notice that most of the time, a delegated agent complements an ambient agent. The ambient agent gets the work done, the delegated agent reports on it.

This step is also a great time to ask, how much of an interface does this solution really need?

Could it rely entirely on email? A Teams message? Those tools already exist, they’re already secured within your org, cost nothing to stand up, and, paired with agents, could service the entire shebang.

Build what the process needs for adoption and not one interface more.

Did some of these steps feel familiar? They sure should!

Great solution design didn’t get reinvented when AI showed up. AI just became another tool you can intentionally incorporate along the way to reduce friction, remove unwanted tasks, and operate at a higher level of fidelity ✨

So when those vague requests to “incorporate more AI” land on your desk, you can pair your groan with these 5 steps to craft a legitimate case study. You’ll have a problem statement, a set of requirements, and a clearer view of what it takes and what it’ll win ya 🏆

And if you'd rather hear all this from people who're at various stages of this journey, we've still got some spots left for our panel in Vegas 🎟️ Grab one here.