There's a clear push toward automation across valuation right now, and it isn't coming from one direction. Manual data inputs, one-at-a-time model updates, sequential review, all of it is being pulled toward speed and scale, and there's no shortage of solutions, applications and ideas competing to get you there. Different vendors, different approaches, different claims about how far along they are.
How Did Valuation Automation Get Here?
The first wave was valuation platforms, structured systems built to standardise data, calculations and workflow, built either internally or by third parties. These have existed for years. Historically, one of the criticisms of valuation platforms was that they imposed strict borders, and sometimes lacked the flexibility to fit the unique circumstances of each deal. The trade-off lies between a standardised, robust process and the ability to actually fit the reality of the transaction behind the valuation. But architecture does get better and platforms do provide more capabilities to their users. The challenge may still be true for a number of players operating in very specific strategies that require approaches and terms tailored to each deal, for example in distressed and special situations strategies, but for a large number of players, e.g. senior direct lending, straightforward leveraged buyouts, or even late-stage VC funds, the terms involved are fairly traditional, and platforms already handle things like PIK toggles and share class differences well. The nuances of those deals are well understood by developers, and the deal terms and specificities can now fit well into the fixed frameworks of standardised systems.
Today it's time for a new kid on the block, it’s the generative AI boom, and as generative AI has gotten better, a lot of people started to ask a question: can we automate the valuation process with AI instead of building sophisticated systems? And the AI models did get better. Originally designed for general analytical work, the models are now trained by teams of experts across various domains, including finance, to improve the models' capabilities in respective fields. As a result, the models themselves keep improving generation over generation, and that improvement may add a further layer of flexibility to the standardised systems described above. That's exciting news for the users.
Why Is AI Output Harder to Trust Than a Platform?
But there is an honest question underneath all of this. Can you trust the output? Or more precisely, what do you need to trust the output?
Let's think of different options in front of us. For a user, it’s easy to trust your own Excel model, because you know exactly how it's built, because you put it there yourself, with every formula and every assumption visible. It's already a step harder to trust a platform, because someone else built it and you're relying on their process instead of your own. However, a platform is still a closed system, at any given moment its architecture is fixed and documented.
Generative AI sits a step further again. It's genuinely harder to trust, not because its reasoning can't be inspected, you can ask it to lay out the steps it took, but because that reasoning path isn't fixed. Ask the same question twice and it might take a different route to the answer, and the next version of the model might reason differently again. That's an open system, by nature, and no amount of the model getting smarter changes that on its own. It's not a knock on capability, frontier models are already strong at this kind of work. It's a different kind of problem entirely.
Which is why in the world of financial reporting, you need to ask yourself two separate questions. Do you trust the model AI gave you? And does your auditor trust the model AI gave you? Because these will determine the amount of work and the number of questions on both ends. A team can convince itself an output is reasonable in twenty minutes. Convincing an external party who wasn't in the room, has no context, and is professionally obligated to be skeptical, takes something else entirely, evidence, not confidence. And that's really the point, automation without transparency doesn't reduce the amount of work, it just shifts who ends up doing it and when.
What Does It Take to Trust an Automated Valuation?
Regardless of whether we are thinking about platforms or AI, we need to think about the same things we'd think about in any standard valuation process. It comes down to transparency and control. We need to know where a number came from, that's a documented audit trail on the data pull. We need to know the calculations ran correctly, that's architecture design. We need to know the output coming out the other end is actually consistent with those steps. And once it's produced, it needs to be locked and time-stamped, so there's no ambiguity later about what was known and when. Put together, that's validated data, explainable architecture, and a reproducible, time-stamped trail. That's something that is achievable in a closed system. ISAE reports have a role here too, they give confidence not just in how a system was set up, but in how it actually operates day to day. Whether that same trust is achievable for generative AI specifically is a genuinely open question. We need to find a way to answer the same questions for an AI-generated output before we can rely on the numbers it produces.
The question of trust is not a problem we can postpone indefinitely. The traditional operating model, built around manual review at manual speed, will not scale to the demands of retailisation, semi-liquid structures, and more frequent valuation points. Try running a daily valuation across five hundred lines of credit by hand, and the ceiling shows up immediately. Therefore, technology is not optional, it is necessary. But we cannot move forward without being able to rely on the process. Speed without trust doesn't remove the risk, it just moves it downstream, to whoever has to sign off on the number without being able to fully explain where it came from.
A platform, or AI, is a tool. Neither one replaces governance, and neither one should be expected to. Governance is what actually ensures quality and an appropriate process, the technology is here to bring speed and efficiency. In practice, that means the same things any standard valuation process already asks for, just applied to something that now sits inside a system rather than a spreadsheet tab, a system whose code you can't always read or verify yourself, the way you could with your own formulas. You need confidence that the system was set up correctly, that it continues to operate as intended, and that the controls around the process actually work.
Which brings us back to the question we started with. Can we automate the valuation process with AI instead of building sophisticated systems? No, not instead, at least not yet. The governance a sophisticated system needs doesn't go away because the tool doing the work got smarter, it just moves to a different place. As for generative AI specifically, we're honest that part of that governance question is still unresolved. We don't yet know how to fully answer it for an AI-generated output the way we can for a closed system, though we believe it's a solvable problem, not an unsolvable one.
That's the honest version of trusting automation. Not a closed system versus an open one, not a platform versus generative AI, but a transparent process with a robust governance and control framework that produces explainable and replicable outcomes that will stand up to the scrutiny of auditors and regulators.
References
- IAASB, ISAE 3402 "Assurance Reports on Controls at a Service Organization", issued December 2009 — the international assurance standard behind the "ISAE reports" referred to above. A Type 1 report covers the design of a service organisation's controls at a point in time; a Type 2 report additionally covers whether those controls operated effectively over a period, which is the "how it actually operates day to day" assurance. iaasb.org — ISAE 3402 announcement
Related pages
- Product — smartZebra's valuation data platform, with the source and calculation log behind every data point.
- Facts & Figures — coverage, sources, update cadence and the standards the data is built for.
- Frequent Valuation Cycles: Does Technology Solve for Data? — the companion piece: what faster valuation technology fixes, and what it does not.
- Transfer Pricing in a World of AI: It All Comes Down to Accountable Data — the same trust question, asked of AI-drafted transfer pricing studies.
- PIK Loans and PIK Toggle: Private Credit Risk Monitoring — one of the deal terms platforms now handle as standard.
- Complex Equity Structures: Why Share Class Rights Matter — the share class differences a standardised system has to model correctly.
- AIFM & Fund Managers — the valuation data layer for fund-level NAV and reporting cycles.
