A partner at a firm I spoke with recently built his own AI agent for deal analysis. He got it working reasonably well. Then he looked at his token bill and turned it off.
That story is more common than people admit, and it points to something worth examining properly. We are helping more and more clients work through the same question: is the right route for them to build or buy a solution?
Why building feels like the right call
The quality of general-purpose large language models in 2026 makes building your own AI workflow feel genuinely achievable. For many teams, it is. You can get a prototype running in a weekend, connect it to your data room, and start pulling useful output within days. For a small investment team that moves fast and has some technical resource, this is an attractive path.
And in most cases, teams do get impressively far. They build something that largely works and that seems to save a lot of time.
But we've often heard there's a ceiling.
The 60% problem
Most internal builds seem to reach around 60% of what the team originally set out to achieve. The remaining 40% is where the real complexity lives, and it tends to be the 40% that matters most for investment work.
Structured, reliable extraction from formatted Excels and large data rooms. Output that is reliable and consistent enough to put in front of an investment committee. Systems that are maintainable and explainable when something goes wrong. Costs that remain predictable as deal volume increases. These are not afterthoughts. They are the core requirements of any AI tool that needs to function in a professional investment context, and they are genuinely hard to build and harder to maintain.
The maintenance problem nobody talks about
Building an internal agent is one thing. Keeping it working is another.
General-purpose models update. Prompts that worked reliably six months ago start producing different outputs. Edge cases accumulate. Someone on the team who understood how it was built moves on. The token costs that seemed manageable at low volume become a real budget line as usage scales.
For most boutique PE and VC firms, maintenance becomes a recurring cost that is difficult to forecast and easy to underestimate. This is the part of the build versus buy calculation that rarely appears in the initial business case, and it is often the deciding factor.
What purpose-built actually means
Hebrides is built specifically for private markets workflows. Extraction, synthesis, and deal analysis designed around how investment teams actually work.
No token anxiety, no in-depth prompt engineering, and no maintenance overhead sitting with the investment team. The infrastructure is ours to manage. Teams get that time back for the things that require human judgment and expertise, like time with founders, with portfolio companies, with the unseen work that actually moves deals forward.
The question worth asking
Speed is rarely what firms tell us they're solely after. The question is whether the output is accurate enough, structured enough, and reliable enough to actually change how decisions get made. I am finding more and more that what investment firms want is AI that helps them make better decisions, not just faster ones.
That is a higher bar than most general-purpose tools are designed to meet.
If your team is evaluating build versus buy, or if you have already built something and are running into the ceiling, I am happy to share what we have built and what we have learned.