Your Best Junior Analyst Is Now a Machine

For years the junior analyst produced the first draft while the senior brought the judgement, and the machine now does the junior's work. That leaves one choice that decides whether the change improves the decisions or only speeds up the paperwork, whether the analyst's role and the way the firm measures a good analyst are redesigned to match.

When the first draft is free, the job changes

A capable first draft removes the need to produce the analysis and leaves the harder part, the judgement about what the output means and whether to trust it. The machine gathers the data, runs the first-pass modelling, and produces the first draft, while the judgement a senior brings stays with the human. Balyasny Asset Management, a multi-strategy hedge fund, reports that its central-bank-speech agent cut a macro scenario analysis from about two days to about thirty minutes. [1]

Boston Consulting Group's 2026 asset-management report finds that agentic AI moves analysts from data gathering and first-pass modelling toward management engagement and differentiated insight, redeploying 50% to 65% of traditional junior-heavy analyst capacity. [2] That redeployed capacity is junior work moving to the machine. The CFA Institute describes the same move, from data preparation to interrogating the model, checking data validity, correcting errors, and supplementing the output. [3]

The binding constraint was never producing the analysis. It was the judgement about what the analysis meant, and a faster first draft hands that judgement back to the human rather than away from them.

The human moves from producing to directing

Handing the judgement back rewrites the role rather than shrinking it. The analyst becomes the editor of the model's output, directing and reviewing the machine the way a senior has always directed and reviewed a junior, and owning the judgement a senior owns, what the output means, whether to trust it, and what to do about it. One Balyasny portfolio manager describes the engine as a teammate that never forgets, always cites its sources, and double-checks the details before sending anything back. [1] The person still owns the judgement, the machine owns the first pass.

This is the human-AI teaming pillar of Cognitive Transformation. As tasks move to the machine, roles stop being fixed bundles of work and become portfolios of judgement, oversight, and exception handling. [5] The unit of delegation drops to the task, and the human is redeployed onto the parts a model cannot own.

The design point is not that the machine is fast. It is that the analyst's day now starts where the analysis used to end.

The measure of a good analyst has to change

If the day starts in a new place, the measure cannot stay where it was. A firm that still measures analysts on how much they produce is optimising for a job the machine now does. At Wellington Management, Brian Barbetta reports that AI tools answer analysts' research queries within hours and give portfolio managers direct access to synthesised intelligence, shifting the analyst's job toward curating and supplying that intelligence rather than producing every research document by hand. [6] The yardstick moves from junior-style output volume to the senior judgement in the review, the quality of the calls the human makes on what the machine returns.

That is also where the authority line holds. Analysts edit the machine's output and the investment decision stays with the portfolio manager. Human-on-the-Loop Oversight means the human is not in the loop at every keystroke but on the loop, calibrated and ready to intervene where their judgement adds what the automation cannot. [7] This distinction is what keeps the machine's authorship of the research from ever becoming authorship of the decision.

The redesign only lands on an evaluation foundation

A redesigned role and a new measure still fail if nobody trusts what the machine returns. That trust is what Balyasny built before it scaled anything, and it is why adoption reached a level most firms never touch, with roughly 95% of its investment teams actively using the AI research engine. [1] The number is the payoff of the foundation, not the tool.

What sits underneath it is unglamorous. In late 2022 the firm stood up a dedicated Applied AI team of about 20 researchers, engineers, and domain experts, led by a Chief AI Officer recruited from Google. [1] Before scaling anything, it built a model-evaluation pipeline that tested outputs across more than 12 dimensions, including forecasting accuracy, numerical reasoning, and robustness to noisy inputs. [1]

Only then did specific agents absorb the routine synthesis, a central-bank-speech analyst and a merger-arbitrage forecaster that updates deal probabilities continuously in place of manual spreadsheets. [1] The critical step change was building the tools into each team's workflow rather than bolting them on as an optional add-on, and that is what produced near-universal adoption. Firms that want the number without the foundation get neither.

The 95% figure belongs to the research engine used by investment teams [1], and it should stay separate from a broader, later claim that 97% of all employees use the internal AI platform, which spans coding and back-office tools too [4].

How to run the same play in any analysis-heavy function

None of this is a hedge-fund-only move. The same shape works for credit analysis, actuarial review, due diligence, or any function where a human produces a document a decision rests on. Wherever the first draft becomes free, the role and the measure sit under the same pressure, and the same evaluation foundation decides whether the change holds.

It does not require a 20-person AI team either. The evaluation-then-redesign discipline is a way of working, not a headcount, and a smaller function can run the same sequence on one workflow before it touches the rest. The firm that skips straight to the tool inherits a faster version of a job the machine has already taken over.

Key Takeaways

• Do the evaluation discipline before you scale the model. Balyasny built a model-evaluation pipeline and a dedicated team before rolling the research engine out, and near-universal adoption followed. The boring foundation bought the result, not the tool.
• Redesign the role from producing analysis to directing and reviewing it, and re-measure the human on judgement rather than output volume. The machine takes the junior work and the first draft, the senior judgement stays with the analyst, so both the reviewing job and the measure that scores it have to be rebuilt.
• Keep the accountable decision with a named person. Balyasny's analysts edit the machine's output while the investment call stays with the portfolio manager, which is what stops "the model wrote the research" becoming "the model decided."

Sources

[1] OpenAI. "How Balyasny Asset Management built an AI research engine." Case study, 6 March 2026. https://openai.com/index/balyasny-asset-management/ (vendor-reported; all Balyasny figures are OpenAI's published account of its own customer.)

[2] Boston Consulting Group. "Global Asset Management Report 2026: Rebuilding Asset Management for an AI-First World." BCG, 19 June 2026. https://www.bcg.com/publications/2026/rebuilding-asset-management-for-an-ai-first-world

[3] CFA Institute (Rhodri Preece). "AI in Finance: Changing Workflows, Growing Demand for Human Judgment." Enterprising Investor, 5 January 2026. https://rpc.cfainstitute.org/blogs/enterprising-investor/2026/ai-in-finance-changing-workflows-growing-demand-for-human-judgment

[4] Pulse2. "Balyasny Asset Management: Chief AI Officer Highlights Firmwide Adoption Of OpenAI Tools." Pulse2, 10 June 2026. https://pulse2.com/balyasny-asset-management-chief-ai-officer-highlights-firmwide-adoption-of-openai-tools/

[5] Christopher Hughes. "From Digital Transformation to Cognitive Transformation." cgh.dev, 2026. https://cgh.dev/thinking/cognitive-transformation/

[6] Correlation One (Sham Mustafa). "How Asset Managers Are Using AI to Unlock the Analyst Knowledge Trapped in Company Calls." Correlation One, 18 May 2026. https://www.correlation-one.com/blog/ai-analyst-knowledge-asset-managers

[7] Christopher Hughes. "The Rubber-Stamp Problem: Why Human-in-the-Loop Is a False Promise, and What Should Replace It." cgh.dev, 2026. https://cgh.dev/thinking/rubber-stamp-problem/