What Parts of Financial Modelling Still Require Human Judgement?

August 21, 2026 By Yodaplus

Valuation assumptions, interpretation of management commentary, scenario weighting, and the final recommendation still require a human analyst, regardless of how much data extraction and model updating AI agents handle. A 2026 Connext Global survey found only 17% of U.S. professionals who use AI at work consider it reliable without human oversight, while 70% say reliable AI means either light review or dedicated human involvement at every step. Equity research reflects that pattern closely.

AI agents are strong at the mechanical layer of modelling. The judgement layer above it still sits with the analyst.

Setting and Defending Valuation Assumptions

An agent can update a discount rate or terminal growth figure once told what number to use. Deciding what that number should be is a different task entirely.

Analysts weigh:

• Company-specific risk factors that do not show up in historical data, such as management transition or pending litigation
• Industry cycle position, which changes how a “normal” margin or growth rate should be defined
• Competitive dynamics that might justify a premium or discount versus historical multiples

These calls depend on context an AI agent cannot fully access from filings alone, including conversations with management, channel checks, and comparisons across a coverage universe that only a human tracks holistically.

Interpreting What Management Actually Said

Earnings transcripts contain more than reported numbers. Tone, hedging language, and what management chooses not to address often carry as much signal as the figures themselves.

An AI agent can flag that a CFO mentioned “headwinds” three times on a call. It cannot reliably judge whether that phrasing signals genuine concern or standard caution ahead of a conservative guide. Reading between the lines on management intent remains a human skill, built from following a company across many quarters and comparing tone against past calls.

Choosing Which Scenarios Matter and How to Weight Them

Scenario and sensitivity automation can rebuild a bull, base, and bear case once assumptions shift. Deciding which scenario deserves more weight in a price target is a judgement call.

An analyst considers the following:

• Probability of a specific event playing out, such as a regulatory decision or a contract renewal
• How much weight recent management guidance deserves against the analyst’s own independent view
• Whether a scenario reflects a genuine risk or a low-probability tail case that should not drive the target

Multi-agent AI systems can present the range of outcomes clearly. The decision about where within that range to land still belongs to the analyst.

Making the Final Recommendation

This is the clearest line between automation and judgement. A model can show that a stock trades below its calculated fair value. Whether that gap justifies a buy rating depends on factors a model does not fully capture, including catalyst timing, market sentiment, and portfolio-level considerations the analyst is weighing across an entire coverage list.

Catching Context That Agents Miss

AI agents work from the data they are given. They do not independently know that a company just lost a major customer through a source outside its filings or that a regulatory change under discussion could affect an entire sector. Analysts bring that outside context into the model, something automation cannot originate on its own.

Where Automation Reliably Takes Over

To be clear about the other side of this line, automation handles a substantial share of the work reliably:

• Pulling and reconciling data from filings and transcripts
• Rolling forward historical actuals into updated model tabs
• Rebuilding scenario mechanics once assumptions are set
• Drafting a first-pass summary of what changed financially

This is where AI workflow automation delivers real time savings, freeing analysts to spend more time on the judgment calls above rather than data entry.

Common Challenges in Balancing Automation and Judgment

• Analysts sometimes over-trust an agent’s output simply because it arrived quickly, skipping the review step that catches errors
• Firms without clear guidelines on which decisions require sign-off risk inconsistent practices across a research team
• Newer analysts may lean on automated commentary as a substitute for developing their own interpretive skills over time
• Model updates that look complete can still carry a wrong assumption if the reconciliation step passed a figure that needed context, not just accuracy checking

Best Practices for Keeping Judgment in the Loop

• Require analyst sign-off on every valuation assumption change before a model update is finalized
• Treat AI-drafted commentary as a first draft, never as the final version sent to clients
• Build review checkpoints specifically around scenario weighting, not just data accuracy
• Document the reasoning behind a recommendation separately from the automated output that supported it
• Rotate junior analysts through manual modeling exercises so judgment skills develop alongside automation use
• Flag any model update tied to material outside information the agent could not have accessed
• Set clear internal rules for which decisions can proceed without senior review
• Compare automated variance commentary against the analyst’s own read of the transcript before publishing
• Track cases where automation missed context, and use them to refine what gets flagged for review
• Keep senior analysts involved in reviewing scenario probability weighting across the coverage list

Future Outlook

Expect the line between automation and judgement to hold steady even as agentic AI adoption grows. The same Connext Global survey found 64% of professionals expect the need for human review to increase, not decrease, as automation expands. Firms are not aiming to remove judgement from research. They are aiming to give analysts more time to apply it by removing the data work that used to consume the day.

Conclusion

Financial modelling automation changes how quickly a model gets updated. It does not change who decides what the numbers mean. Analysts remain responsible for the assumptions, the interpretation, and the recommendation that follows.

Yodaplus builds agentic AI systems for equity research and financial services teams that need this balance done correctly. Our enterprise AI solutions combine multi-agent AI, intelligent document processing, and secure enterprise integrations within a governance-first AI architecture, so automated steps are fast and human review points are built in exactly where judgement matters most. For research teams adopting AI workflow automation, that structure keeps the analyst firmly in control of the decisions that carry real risk.

FAQs