What it will take to realize AI’s promise to simplify and speed federal acquisition?

Terry Gerton Let’s talk about AI. Because everyone is talking about using AI to simplify the contracting workload, both for government officials and also for industry. So as you think about that, and from your own experience, what is it about the current acquisition system, if you could pick one thing, that you would most want to change?

Steven Aberle I think that’s a great question and you framed it in the right way, because it’s two sides. There’s a bridge between both what we simply say is buyer and seller, right? Which is the commercial defense industrial base that supplies technology and services to the federal government. And then of course, the federal government on the other side. So at Rohirrim, what we believe is procurement speed is not an administrative convenience anymore. It needs to be a core element of both deterrence and mission assurance. So acquisition is not just one of many citizens in the federal infrastructure. It is the citizen. And that’s because time, or rather speed, has always been one of the most important battlefield currencies. And every month it takes to turn an idea into a contract is a month an adversary can field the next drone, that they can field the next missile, the next exploit. So if we can’t match that tempo, we risk losing technical overmatch before a fight even begins. So modernizing procurement with the application of artificial intelligence is how we turn time from what is now almost really seen as a liability into a strategic advantage.

Terry Gerton That’s a really interesting framing because I think the trade-off oftentimes is between speed and compliance. So how do you articulate those opposing values and how do you make the trade-off?

Steven Aberle For Rohirrim, at the simplest level, our mission is really to make work better for the people who keep our government running. But that’s really only the first half of the story. We’re really in an era, I think, that the FAR never contemplated, where adversaries can iterate in 18-month cycles, while we iterate in eight- to 11-year cycles. And that delta means adversaries will be fighting with third-generation hardware while we are still staffing acquisition paperwork. And you can look no further than Ukraine and before that conflict kicked off a non-existent drone program to what is now holding the line with a world power. There’s always a balance, Terry, between the application of new technologies and risk-mitigation approaches to ensure the sanctity of the Federal Acquisition Regulations and that process. There will always be that balance. And what we have in terms of the research that we perform and the technology that we have is an understanding that language models in particular, that specific area of artificial intelligence can be trained, can be controlled, by various training techniques, by various provenance and governance and guardrail techniques to ensure that the risk is mitigated as we adopt these new technologies.

Terry Gerton You made a statement early in that last response about the FAR never anticipating the current environment that we have and the deployment of AI. How do you see, and on what timeline, federal acquisition officers using AI to actually deliver speed?

Steven Aberle I would argue that we’re already there. So for the last 18 months, we’ve been quietly working with government agencies and program offices to build technology that they can utilize inside their acquisition workflows to remove some of that manual burden, some of that pain, and that’s just the first step. So if you look at a couple use cases inside the federal government. When an RFP is let, they’re going to have potentially dozens or hundreds of respondents write huge amounts of text, unstructured data, that the government then receives. And the FAR demands that we look through all of that text and check for compliance. Did they write this text in the confines of the rules that were set out in terms of compliance? And let’s say you’re at an intelligence agency and let’s say it’s a classified bid. The individuals, the people that have to read those tens of thousands of pages are not cheap. These are professionals that have security clearances and they get to page 398 of 400 and discover this bit it is not compliant, but they’ve run out of time to let the respondent know that they’re out of compliance. They don’t have enough time to send an EN for them to correct that. And that’s exactly what the government doesn’t want. They want the most amount of bids so that they have negotiating power and that presents better value to the government. So that’s just an example of one of the use cases. But there’s a second part of this, which we’re working on, which is part of Rohirrim’s mission. It’s certainly more ambitious, but at our core, Rohirrim and our system, RohanProcure, we exist to destroy that mountain of procedural pain that has for so long separated the buyer and seller. And I always like to say, in the place of that mountain, we’re going to build a cathedral, which is a bridge — getting back to the first comment you had, a bridge where government demand and commercial supply can meet in real time. And when that bridge is in place, then acquisition can move at the speed of need rather than the speed of paperwork.

Terry Gerton I’m speaking with Steve Aberle. He’s the founder and CEO of Rohirrim. All right, well, let’s go back to that. I love the cathedral metaphor, but it may be difficult to build that as the government is in the process of updating the FAR in little bits. So if someone is interested in deploying AI and the FAR is changing on a fairly rapid basis, how do they make sure that the AI that they’re using is staying current with the new rules that GSA and OMB are putting out?

Steven Aberle What we like to say is we accelerate through compliance. So the FAR, for all its faults, is simply risk mitigation, fairness and process, though it was built for kind of a different age. But what we build with Rohirrim and RohanProcure and our specific brand of language models, which we call organizational-specific language models, are pre-trained on the FAR, the DFARs, all the agency supplements. And also the thousands of historical, say, protest decisions. So the engine knows exactly where the cliffs are and places guardrails long before you reach them. And that includes not just all of this historical data but new data that it can pull in as these acquisition professionals are building their acquisition plans. These routine steps everyone goes through for clause selection, labor category mapping, market research — all of that can be automated. All of those new policy updates can be ingested in real time and presented to the human in a way where they can apply their specific brand of uniqueness and what a human does that no model can replicate, see around the corners to evaluate thoughts and ideas from multiple dimensions of approaches, to apply rhetoric and to make the best decision for the U.S. government. These are things as an applied engineer, Terry, I have not been able to replicate inside artificial intelligence models.

Terry Gerton So as you’re thinking about that transition, some folks argue that there’s a cultural opposition to the deployment of AI in the procurement workforce. One way to overcome that maybe is training. What do you see as the most important aspects of training for the current procurement workforce?

Steven Aberle This applies to both sides, the commercial and the government sector. On the commercial side, we have a system called RohanRFP, and that’s deployed to the army of proposal people at any given defense industrial base that brings efficiencies into the way that they respond to RFPs. And one thing that we’ve learned is it is all about trust. Can you prove exactly what the model is generating in terms of text? Do you have that provenance? Do you have that governance? And then can you build user interfaces and can you build the training apparatus and training systems that help a user understand when this is generated, it is all tracked to a lineage, to provenance and governance so they can simply trust it? Through training, really what we’re trying to do is tell the human that you maintain oversight, but what we can do is bring machine tempo and we will prove that governance, that provenance, that lineage of everything that’s being generated so you will trust it.