We’re happy to share the latest in AltsTech’s series profiling how investment managers are using AI, tech, and analytics to generate alpha. We’re fortunate to interview Afsheen Afshar — Founder & Managing Partner, Pilot Wave Holdings.
Please give us an overview of your firm.
Pilot Wave is a lower middle market private equity firm in New York founded in 2019. We buy control positions in essential parts of the economy — including infrastructure services, manufacturing, and retail verticals. We create value by powering AI transformations of the underlying businesses.
We don’t invest in AI, we develop and deploy it, and we use the portfolio companies we control as both the investment and the data platform. This has produced an operating portfolio along with a venture-grade spinout, Detect, that now generates millions of ARR. While we’ve been completely boot-strapped and self-funded since our founding, we are now raising our first external capital. Our returns until now have been top decile (far over 10x MOIC), and we’re looking to scale rapidly from here.
Who are your peers/competitors, and how do you differ?
Lots of investors call themselves an “AI rollup” now. Almost none of them own the operating reality.
There are two camps. One is the venture and growth crowd — the General Catalyst / a16z / Thrive commentariat — applying AI to service businesses they mostly hold minority (non-control) or light-touch positions in. Smart capital, real conviction, but the AI is more often a thesis than a deployed system. The other camp is traditional lower middle market PE: excellent operators who treat AI as a vendor line item.
We bridge the gap. We have the control and operating muscle of a buyout shop and the actual capability to build — not merely buy — AI, because that’s the background of everyone at the firm. The moat is ownership: we own the companies, so we own the data exhaust and we’re our own first customer. Our competitors are renting theirs. Most AI companies are building copilots; we’re building the pilot.
What’s your background? How and why are you in your role today?
I’ve spent my whole career applying math to non-math fields, and I started Pilot Wave because I saw a unique opportunity to help Main St compete against Wall St and Silicon Valley.
I completed my academic training at Stanford where I earned an MD and PhD, with research in computational neuroscience and dynamical systems with Krishna Shenoy and Andrew Ng. From there: Goldman Sachs, where I cut my teeth in finance and got wide and deep exposure to a global investment bank; then the first Chief Data Science Officer in the history of Wall Street at JPMorgan, and most recently the first Chief AI Officer in the history of private equity at Cerberus Capital.
Same lesson at every stop. AI doesn’t fail because the technology is bad — it fails because it gets deployed into companies that were never prepared to use it, and inside a large institution you can build something extraordinary and still not control the business well enough to capture the value. Pilot Wave is that lesson made concrete: own the company, control the data, build the system, keep the upside. The timing is straightforward — AI got good enough to run real operations, the lower middle market is structurally inefficient, and the essential economy has durable tailwinds.
We’re very lucky to have recruited the best in the world: folks like Mark Spindel, who is a former CIO from the World Bank, DC Retirement Board, and a multi-billion dollar family office; founder partner Tzaras Christon, who has been a senior executive across multiple industries and companies; and many others who have been senior leaders at Deloitte, KPMG, Booz Allen, UBS, HSBC, JPMC, GS, and many other places.
What are the tools you’re using for your front office: sourcing, LP relations, investing analysis, etc.? What are the strengths and weaknesses of these providers?
We run deliberately lean and heavily AI-augmented, because the off-the-shelf PE stack is built and priced for funds ten times our current size.
For sourcing, we use a proprietary platform coupled with Linkedin and scrubbing of online data sources to identify potential targets. The majority of our deal flow is actually inbounds, given the market recognition we’ve earned over the past several years and our industry network.
For investing analysis, again we have a proprietary AI that we’ve built coupled with frontier models, namely Claude, which has proven itself to be the far and away market leader for Excel models. These platform analyzes CIMs, data rooms, legal documentation, and all other files that one might find in a dataroom and creates a knowledge graph that documents how the company represents information. Given this, the platform then automatically writes an investment memo that we then use for quick filtering and for due diligence questioning. On top of that it provides a natural chat interface for quick questions and hypothetical scenarios that we often use real-time while on-site with the potential target.
Finally, while do use open-source tools like OpenClaw to help manage various internal functions including LP relations, all of these agents do need to be regularly monitored with strong guardrails. Nothing can fully replace the human touch of outreach and phone calls quite yet.
These platforms analyze our incoming email and messaging traffic and keep updated CRMs for us automatically. In addition, they make sure we are managing all parties through our funnel, no one gets ignored, and we are actively following up. We find it at least as good if not superior to anything else on the market, most notably because it is constantly adapting to our team and workflow.
What are the tools you’re using for supporting your portfolio companies? What are the strengths and weaknesses of these providers?
Our primary portfolio support ecosystem is a set of products we built.
The center of it is PilotAI and AgentOS — an operational reasoning system we developed in-house that sits across the portfolio. It isn’t a dashboard; it’s a model that learns how an operating business behaves and reasons about the effect of interventions. We staff and deploy it through our own wholly-owned technology and operations subsidiaries, Pilot Wave Technologies and Charge Impact Group, which means our companies get senior AI and operating capability at well below market cost. Detect AI, which is a AI-powered powerline inspections start-up founded out of one of our portcos, is what happens when one of those internal builds becomes good enough to stand on its own.
Underneath that, our companies run the standard stack for their categories — commerce platforms, ERPs, marketplace tooling — and we don’t reinvent what already works. The honest tradeoff of building rather than buying is that it’s slower and we carry the cost on our own balance sheet. We accept that, because the alternative is handing the data and the value to someone else.
As a VC, I’ve been pitched various startups that also claim to aid CEOs (and/or their investors) in managing a company. If they were here in the room, they would no doubt argue that they have the advantage of more capital and more data across a much bigger portfolio. Why did you decide to build your own in-house tool to do this?
In short, because nothing we found could do the job. We have no religion about building technology ourselves; but as skilled technologists, we have no fear of doing so either. This was an instance in which a lot of startups claim to be offering something that can quickly and completely analyze data rooms, provide opportunities for transformation, etc., but none worked as well as it needed to. Confabulations abounded, inferences across structured and unstructured data didn’t gracefully integrate, and models weren’t complete. Given our AI expertise, we were confident that we could do better and do so quickly with minimal cost.
What technologies/databases have you found helpful in winning LPs?
The best database for raising a fund is still a good relationship graph plus honest signal on who actually writes checks. Most LP databases are table stakes, not edge.
We run fundraising as a sales funnel — pipeline, stages, conversion — in the CRM, which is exactly the discipline you’ve argued for. Where we get leverage is targeting. Rather than spray a generic LP list, we use network and event intelligence to prioritize: before a major conference, for example, we’ll analyze the attendee base to find the handful of allocators who actually back emerging managers in our strategy, map the warmest path to each, and tailor the materials to them specifically. LinkedIn Sales Navigator and relationship-path mapping handle the routing; AI does the research and the tailoring.
The candid weakness — and it’s the heart of your “15 Steps to Fundraising” essay — is that the big institutional databases (e.g. Preqin) tell you who exists, not who’s leaning in. Signal on actual appetite still lives in relationships, and no tool has solved that.
What tools do you find helpful for expediting due diligence?
LLM-assisted document review has compressed our first-pass diligence by something close to an order of magnitude.
We ingest the data room and use frontier models, driven by the PilotAI agentic harness I mentioned, to do first-pass review of contracts, financials, and tax: reconciling figures, surfacing inconsistencies, flagging what a human should look at hard. We’ve used this for genuinely forensic work — reconciling transaction-level data, pressure-testing working-capital and credit dynamics, reading complex agreements against each other. Document handling and data rooms sit in Box.
The strength is speed and recall — the model reads everything and never gets bored. Every analytical output that matters gets traced back to source and verified, and anything that touches an attestation goes to our accountants and is held to professional standards.
What are the tools you’re using for your middle office: tracking, risk management, etc.? What are the strengths and weaknesses of these providers?
Lean and mostly in-house, with a compliance layer on top.
Portfolio tracking and risk monitoring run largely through models maintained by the central team, with PilotAI increasingly feeding portfolio-level operational visibility from the bottom up.
The weakness is the same theme as the front office: the integrated portfolio-monitoring platforms — Chronograph, Cobalt, Allvue — are built and priced for funds much larger than ours, so the lower middle market is left stitching its own. We’d rather build than overpay.
What are the tools you’re using for your back office: settlements, records maintenance, accounting, human resources, etc.? What are the strengths and weaknesses of these providers?
Outsource what you can, keep it lean, and don’t pretend an emerging manager needs an enterprise back office.
Fund administration and fund-level accounting are handled with outside specialists and our accounting firms; the portfolio companies run on the standard small-business finance stack. HR and records are lightweight and cloud-based, with Google Drive for records, and for corporate and employment work — including standing up new entities and first hires — we use outside counsel.
The strengths and weaknesses here are boring, and that’s the point. The emerging-manager back office is a commodity, the providers are fine, and any time or money spent gold-plating it is time and money not spent on deals or on the AI.
A huge amount of valuable data flows through your pipes. What are you doing to capture that data and mine it? Can you share any patterns you have identified?
This isn’t a side project. It’s the reason the firm is structured the way it is.
Most PE firms let operational data flow past them. We built Pilot Wave so it flows through us — that’s precisely why we take control positions in operating companies rather than minority stakes. Every transaction, channel, and working-capital cycle in the portfolio gets captured and standardized, and Detect AI exists because that data and the capability around it became valuable enough to spin out. With real-time best-in-class data and data analytics, we are building robust, directed portfolios of companies, revenue streams, operating profits.
Patterns, at the level I can share: across our consumer and e-commerce holdings, the quiet driver of the business is channel mix — the same revenue produces very different cash and margin depending on where it’s sold, and a surprising amount of apparent performance variance turns out to be working-capital and channel composition rather than demand. In the asset-heavy businesses, the binding constraint is almost never the headline number; it’s the structure underneath it. But the real prize is the meta-pattern: in lower middle market companies, the operational intelligence lives in a few people’s heads and a pile of spreadsheets, and capturing it systematically is enormous, durable alpha.
PilotAI is an operational reasoning system, not a reporting tool. Technically it’s a hierarchical world-model: it learns the dynamics of an operating business and reasons about the effect of interventions — causal, not just correlated — which matters, because operators don’t need another chart, they need to know what happens if they pull a given lever. We’re deploying it across the portfolio, and Detect AI is commercial proof that the underlying capability stands on its own with paying revenue.
For the avoidance of doubt, Detect has real ARR; PilotAI is in active deployment, but not yet finished. What we’re really building is the loop — data to decision to measured outcome to a better model — because owning that loop across companies we control is a compounding advantage no one renting their data can match. We’re looking forward to open sourcing a good chunk of the central harness quite soon.
What are the most creative or unusual ways you’re using AI & analytics in your organization?
First, we treat each portfolio company as a running natural feedback loop and use modern methods to reliably measure whether our interventions worked, rather than asserting it in a board deck.
Second, we’ve put our own analysis on agents: a multi-step agentic harness, built on frontier models, that runs diligence and modeling work autonomously, and that Detect AI extends into genuinely autonomous, multi-agent operation inside our companies. Third, the architecture itself — PilotAI borrows from active inference and energy-based models, the math I worked on in my academic career, running on top of our agentic harness agentOS that reasons about running a business than a predictor that forecasts a metric.
The thread through all of it is the same line I use on stage: every AI company is building copilots; we’re building the pilot.
How do you address portco resistance to your building a ‘pilot’ to run their firms?
This is a critical question, and one that we have spent a lot of time thinking about. Our firm’s core philosophy is to lead with empathy, EQ, and bedside manner, so we put a tremendous amount of energy into developing a deep camaraderie with the operating team and, as such, demonstrating empathy for the value the people on the ground made every day.
At the same time, we require the empowerment that comes with being the owner to action technological transformation quickly and with conviction. The fundamental presumption of our thesis is that traditional business owners do not have the expertise or context to successfully execute an AI modernization on their own, and therefore we need to sometimes nudge the organization forward. Further, we always align economic successes with long-term growth so folks recognize the broader vision they are a part of.
As such, we have the most success when we first develop “street cred” with the team by initially using technology in perhaps more mundane ways: small wins that optimize workflows and make the lives of the operating team easier. These wins improve revenue and EBITDA, and indeed if they were the only thing we ever did, we would still have industry-leading returns. However, we use the trust developed from those early victories to then perform more ambitious changes, such as the creation of entirely new revenue streams that are data and technology based. That’s how we generate the truly outsized returns.
What are your unmet technology needs? Places in your firm where you’re seeking a solution and haven’t found an appropriate one?
Three, and the first is structural.
One: there is no integrated stack built for what we actually are — a control-PE buyer that is also a serious AI builder, at lower-middle-market scale. We fall between tools made for mega-funds and tools made for startups, so we build a lot ourselves. Two: connecting genuinely heterogeneous portfolio companies — every one on a different ERP and commerce stack — into a single data and reasoning layer is harder than any vendor admits. Data integration is the unglamorous bottleneck. Three, and the one I’d most like solved: rigorous, productized verification and evaluation tooling for AI outputs in high-stakes financial settings. I trust these models to read everything; I don’t yet have great tooling to prove a given output is right to the standard a number used in an attestation demand.
Honestly, the larger constraint than tools is talent (and capital) — people who can do the PE, the AI, and the operating work at once (and the finance needed to build and operate the tools) — but you asked about technology.
What processes are you focused on improving?
Closing the loop, and making fundraising as disciplined as our diligence.
Operationally, the priority is the full flywheel — capture the data, drive a real decision inside a portfolio company, measure the outcome cleanly, feed it back to improve the model — and making the onboarding of each new acquisition onto that infrastructure faster and more standardized.
On the firm side, I want our Fund II raise run with the same rigor we bring to diligence: a real pipeline, honest stage conversion, tight targeting, rather than the artisanal process most emerging managers tolerate. And underneath all of it, scaling our verification discipline, so that as we lean harder on AI the standard of proof goes up, not down.
