Decision rigor is industry standard. Decision continuity is not.

Private markets industrialized how investment decisions are made. They have not industrialized how those decisions are tested over time.

Upfront, every firm acts rigorous. The IC memo is structured. The model is detailed. Assumptions are debated until they are explicit. Then the deal closes. Twelve months in, that same decision is no longer a system — it is institutional memory in a partner's head, plus a slide deck that gets updated when something needs explaining. The evidence is in different places than the original sources. The thesis got paraphrased. Nobody is forced to reconcile today's read against what was originally underwritten at the same level of precision.

That used to be tolerable. Hold periods were shorter, portfolios smaller, LPs less curious about provenance. None of those conditions hold now. Median buyout hold periods extended from roughly 4.1 years in 2005 to 6.6 years in 2023 (PitchBook via KPMG, Value Creation in Private Equity, October 2025). Global active buyout-backed unrealized value reached approximately US$3.6 trillion (Preqin via the same report). The PE distribution-to-contribution ratio fell roughly 37 percent from its 2013 peak through 2024, and fundraising contracted at a -12 percent CAGR from 2021 to 2025 (S&P Global via KPMG). Every firm is running more deals, more quarterly cycles, more covenant tests, and more LP letters than its operating model was designed for.

Then AI showed up. And here is where the thinking goes wrong almost everywhere we look: most of the industry is treating AI as the way to fix decision quality. AI is not the fix. AI is what you bolt on top of an investment process that was never auditable to begin with — and the bolt-on accelerates the same drift it was supposed to solve. Faster memos. More dashboards. More pages of plausible-sounding analysis. None of which can be reconciled, a year later, against what was actually underwritten.

The 2026 Stanford AI Index documents the receipts on this. We will come to that. But the diagnosis comes first, because the diagnosis is what makes the Stanford data matter.

Adoption is high. Impact is not.

88 percent of organizations have adopted AI. 70 percent use generative AI in at least one business function. Agent deployment across business functions is in the single digits. Productivity gains cluster in narrow, structured tasks. In judgment-heavy work, the measured effect is weaker or negative. A 2025 MIT study cited alongside the Index found that 95 percent of generative AI pilots produce no measurable financial impact.

The frontier is jagged.

Models that win gold at the International Mathematical Olympiad read analog clocks correctly half the time. Coding benchmarks jumped from 60 percent to near 100 percent of human baseline in twelve months. Agents on real computer tasks went from 12 percent to 66 percent success in the same window, and still fail one attempt in three.

Systems that look superhuman on one axis fail unpredictably on another. The wins do not predict the losses. That is why every “we built an agent that does what a VP does” demo is impressive in the room and ruinous in production.

Hallucination rates across 26 top models ranged from 22 percent to 94 percent on a new accuracy benchmark. When the same false premise was framed as user belief instead of third-party belief, GPT-4o's accuracy dropped from 98.2 to 64.4 percent. DeepSeek R1 dropped from above 90 to 14.4. A 30-point swing based on phrasing.

The most capable models now disclose the least.

This is the finding that should stop the room.

The Foundation Model Transparency Index fell from 58 to 40 in a single year. OpenAI, Anthropic, and Google have stopped disclosing parameter counts, dataset sizes, and training duration for their frontier systems. Data acquisition transparency averages 31 percent across major providers. Data properties disclosure sits at 15. Compute disclosure at 26.

You are building investment process on systems whose failure modes their own makers will not describe.

Your LPs are not going to accept the model said so as provenance. Your regulators are not either. The EU AI Act entered full enforcement in January 2026. Documented enforcement actions went from 43 in 2024 to 156 in 2025.

Five landmines.

The demo-to-deployment gap. Vendors show you an agent that drafts a memo in four minutes. Six months later your team is hand-correcting the output and decision quality has not improved. Demand post-deployment metrics, not demos.

Productivity is not decision quality. AI gains show up in structured tasks with fast feedback. Investment work is neither. If someone pitches “40 percent more productive diligence,” ask them to define productivity. Pages per hour is the wrong answer.

The entry-level collapse. U.S. developer employment among 22-to-25-year-olds fell nearly 20 percent in one year while older-developer headcount grew. Same pattern in customer support. The apprenticeship model your firm runs on — analyst to VP to partner — does not survive agent-first workflows untouched.

Model choice is a commodity. Workflow is the moat. The top four models on the Arena leaderboard are separated by fewer than 25 points. The U.S.–China gap is 2.7 percent. Firms whose AI strategy is “we use Claude” or “we use GPT-5” are one quarter from parity with everyone else.

Responsible AI benchmarks lag capability benchmarks. Labs report capability consistently. Safety, fairness, and factuality reporting is inconsistent or absent. And Stanford documents that improving one responsible-AI dimension consistently degrades another. There is no free lunch, and no standard way to track the tradeoff across vendors. You are on your own for that layer.

Two paths from here.

Most firms reading this already feel the problem. What they do next splits into two paths.

Path one. Add AI on top of existing workflows. Faster memos. More dashboards. More noise. Decision quality does not improve. The apprenticeship pipeline erodes, the IC deck looks slicker, and the first bad outcome is attributed to market conditions rather than to a process that was never auditable to begin with.

Path two. Rebuild the investment process as a governed decision loop. AI operates inside constraints. Decisions become traceable, comparable, and auditable. When something goes wrong, you can tell exactly where — in the data, in the thesis, in the timing, or in the conviction. When something goes right, you can tell why, and do it again.

Most firms will choose path one. They will not announce the choice. They will make it by default, through a series of small purchases and pilots that never resolve into a real system. They will not notice the impact immediately. It will surface in outcomes first, and only later in process.

The firms that outperform this cycle will choose path two. Not because they are braver, but because their LPs, their auditors, and their own risk committees will eventually force the question.

Stop trying to replicate your deal team. Start building a system that keeps your decisions testable.

That is the pivot.

The next category in private capital is not AI for deal teams.

It is decision continuity — a continuously maintained mapping between every decision made and the underlying evidence, through time. The IC memo is the artifact of decision rigor. Decision continuity is the discipline of keeping that artifact testable as the world moves.

Five non-negotiable pieces:

Provenance from source data through every transformation to the conclusion on the IC deck. If a number on page seven cannot be traced back to source instantly, you are asking your IC to trust something you cannot prove.

Decision aging, so theses do not quietly rot. A decision made in January on November data is not the same decision in April. The system should tell you before a covenant does.

Invalidation triggered by variance and drift. Not dashboards. Forced re-evaluation when facts move past tolerance.

Containment for escalated decisions. When a position goes sideways, you do not need more analysis tabs. You need a constrained environment where only the top decisions are visible, each has an explicit consequence, and nobody exits without resolving or explicitly deferring.

Outcome traceability. If you cannot tell, a year later, whether an AI-assisted decision became a good outcome, you cannot learn and your LPs cannot audit.

We see this pattern across live deals — where the numbers move, but the decision does not.

The spreadsheet is not the enemy.

The spreadsheet without a system around it is.

Most of private capital still runs IC out of Excel, a shared drive, and institutional memory in a partner's head. Those artifacts are not wrong. They are not auditable at the speed AI is forcing on the industry. A spreadsheet with a clean, traceable link from every cell to source, inside a system that knows when the decision is stale, is a reasonable part of a modern investment operating system. A spreadsheet sitting next to an AI-generated memo with no provenance is not.

The firms that win the next cycle will not be the ones that replaced their associates with agents. They will be the ones that treated their investment process as a governed decision loop and used AI inside it — not on top of it, not instead of it.

If you are pressure-testing your portfolio or evaluating how AI fits into your investment process, this is the work we are doing live.