What 60 hours of CFO interviews taught us about variance bridges

Variance bridges look like reporting. They are actually organizational memory. A CFO asks why margin moved, and the answer may pass through sales mix, plant utilization, freight, rebates, raw-material indices, channel inventory, one large customer, and a spreadsheet maintained by somebody who has been in the company for seventeen years. The bridge is not a chart. It is a negotiation over causality.

We spent 60 hours interviewing finance leaders, business controllers, sales heads, planning teams, and analysts before writing the first production agent. The surprising part was not that the data was messy. It was that the mess had structure. The same five themes appeared in nearly every conversation.

The first answer is rarely the useful answer

Most systems can tell you that revenue is down or gross margin is lower. The useful question is why, and the useful answer is almost always a decomposition. Price, volume, mix, currency, freight, discounting, and one-offs need to be separated before a leader can act. A single generated paragraph is not enough. The product has to help the user walk the bridge.

That shaped the interface. We stopped treating the agent answer as a final summary and started treating it as an interactive path. The agent produces the first bridge, then lets the CFO open a driver, inspect the query, change the comparison period, or exclude a one-time item. The conversation is the control surface for analysis.

People trust familiar imperfections

Finance teams did not expect perfect data. They expected the system to understand the imperfections they already manage. They wanted to know whether the agent used provisional numbers, whether a plant had posted late, whether intercompany eliminations were included, and whether the sales hierarchy matched the latest management view. These caveats were not clutter. They were credibility.

So the agent learned to surface caveats early. If data was incomplete, the answer said so before the conclusion. If two sources disagreed, the trace showed both and identified the system of record chosen for the answer. This made the product feel less magical and more finance-native.

The real user is the meeting

A CFO rarely uses a variance bridge alone. The answer travels into a review, a board pack, a WhatsApp thread, or a call with a business head. That means the product has to produce durable artifacts, not just chat responses. Users wanted a chart they could pin, a paragraph they could paste, and a trace they could defend if challenged.

We added saved views, timestamped traces, and short executive summaries because the meeting demanded them. The best agent interfaces understand where the answer goes next. If the product ends at the chat bubble, it stops one step before the work is done.

Interviews are architecture

The interviews changed table selection, metric definitions, evaluation cases, and UI hierarchy. They also changed the contract. We narrowed the first release to a smaller set of variance questions because those were the ones with a clear owner and recurring pain. That narrowing made the system more valuable, not less.

This is the repeatable lesson. Before building an executive agent, sit in the decision long enough to hear the vocabulary, shortcuts, and anxieties. The software architecture will be better because the product team will know which imperfections matter and which can wait.

The most useful interview question was rarely technical. It was: "What would make you not trust this answer in a meeting?" That question surfaced stale snapshots, management overrides, unofficial product groupings, plants that posted late, and one-off adjustments that never appeared in clean documentation. Those details became product requirements. They told us where the agent needed to caveat, where the trace needed to be visible, and where a human review step was not negotiable. The interviews did not delay engineering. They prevented us from engineering the wrong certainty.

It also changed who needed to be in the room. The obvious users were finance leaders. The hidden users were the analysts who knew which spreadsheet was unofficially authoritative, the plant controller who understood late postings, and the sales operator who could explain why one customer should be excluded from a comparison. Without them, the agent would have reproduced the official process and missed the working process. Variance bridges are built from numbers, but they are trusted because the organization recognizes its own judgment inside them.

That recognition is the difference between adoption and polite applause.