Five Analysts, One Metered Balance and a Monday Test

01 · Five Analysts, One Metered Balance and a Monday Test

A finance team carrying heavy, sustained workloads does not need another demo; it needs a workstation that survives the first hour of the week. We put five analysts on Claude Desktop for a full reporting cycle and ran it head-to-head against the incumbent workflow: exported workbooks in a spreadsheet stack, a shared drive of prior commentary, and a scatter of single-purpose scripts. The question was narrow. Can a desktop AI workstation carry real analysis volume without turning the monthly model bill into a guessing exercise? Every call was metered on real tokens, with cache writes and cache reads priced separately, so the work left an exact trail.

We scored both setups on the same six axes, and the head-to-head was decided by where each one bent first:

• › Throughput on repetitive close and reconciliation tasks
• › Context handling across long workbooks and multi-tab exports
• › Reviewability of the reasoning behind a number
• › Setup and maintenance burden for a five-person desk
• › Cost predictability once volume becomes sustained
• › Handover quality when one analyst picks up another person's work

Claude Desktop took the axes that reward judgment over messy inputs. The spreadsheet stack held its ground wherever the deliverable is a fixed template that an auditor expects to see unchanged. That split, more than any single feature, is what shapes the rollout.

02 · Reconciliation, Variance Commentary and the Long Workbook Problem

Where the desktop workstation pulls ahead is the explain-don't-format stretch of the job. Drop a trial balance export, a budget extract and last period's commentary into one window, ask for the variance drivers, and the analyst gets a ranked exception list plus a first draft of the narrative instead of a morning of manual joins. In our head-to-head, the spreadsheet workflow needed three separate steps to reach the same place, and the reasoning lived in the analyst's head rather than in the file.

The long workbook problem is the second win. Management packs arrive as dozens of tabs, inconsistent labels and formulas that reference tabs nobody remembers building. A desktop agent can walk the structure, call out broken links, and reconcile totals against a source extract in one pass. For the narrative layer we ran Claude Sonnet 5.5; for the fast triage of exception lists before the team's stand-up, Claude Haiku 4.5 was enough, and the cheaper tier mattered once every analyst was running it all day.

The tasks that favoured the workstation were consistent across the cycle:

• › Explaining month-over-month movement in a trial balance without manual pivoting
• › Matching invoices, accruals and reclass entries across mismatched export formats
• › Drafting board and lender commentary from a finished model output
• › Turning a research note into a clean assumptions list for a forecast

03 · Fixed Templates and Audit Trails: Where the Spreadsheet Stack Holds

Honest head-to-head reporting means naming the losses. Any deliverable whose value is the template itself stayed with the spreadsheet stack. Covenant compliance tables, statutory filing schedules and anything that gets loaded into a downstream system by position rather than by meaning are still built by hand, because the AI workstation optimises for understanding, not for cell-perfect repetition.

Two other places the incumbent kept ground:

• › Immutable formatting. A reviewer checking column widths and tick marks has nothing to gain from a rewritten sheet.
• › Line-level provenance. Where a regulator expects a formula chain, the source model stays authoritative and the desktop output becomes commentary, not evidence.
• › One-person scripts nobody owns. Migration cost sits here: a workflow that only one analyst understands is expensive to replace mid-quarter, so we sequenced those last.

The practical rule that emerged: let the desktop workstation do the interpretation and the drafting, and let the spreadsheet keep custody of the number itself. Analysts who respected that boundary moved faster and reviewed each other's work with fewer arguments about whether a figure had been restated.

04 · Cache Writes, Cache Reads and Why Sustained Volume Stays Predictable

A daily analysis workstation is a cache problem before it is a model problem. The same materials reappear every session: the chart of accounts, the accounting policy pack, the board template, last month's commentary, the entity list. Priced correctly, that repetition becomes an advantage rather than a cost. Because each call is metered on real tokens and cache writes are priced separately from cache reads, a workstation that reuses a stable context pays the write once and then reads it back cheaply, session after session. Teams that ignore this distinction see a bill that spikes every time a shared context is rebuilt; teams that structure their recurring inputs around it see a curve that flattens.

The other side of predictability is choice. With Quanta, one balance covers the major model families behind Claude Code, Codex, Gemini CLI and any OpenAI SDK-compatible client, so a team can route cheap triage to one tier and deep reasoning to another without opening a second vendor relationship or reconciling two invoices. Referring colleagues to the relay earns you 19 percent of their top-ups, which is a small but useful offset when you are the person who convinced five desks to change tools.

Cost head-to-head, then, came down to discipline rather than discounts: reuse what is stable, escalate only what needs escalation, and let the metering show you which of those habits you actually kept.

05 · A Five-Desk Rollout With Hard Limits and Clear Ownership

We did not migrate the department in a week. The pilot that worked had a shape any finance team can copy:

• › Scope by deliverable, not by person. Variance commentary, reconciliation narratives and research summarisation went first; statutory templates never moved.
• › Set a hard monthly balance per desk and let analysts see their own token usage against it, so the metering governs behaviour instead of surprising finance at month end.
• › Standardise the recurring context once. A shared chart of accounts and policy pack, loaded consistently, turns cache writes into a one-time investment for the whole team.
• › Review AI drafts like workpapers. One named reviewer per deliverable kept the audit trail intact and kept the desktop output in its lane.
• › Re-run the head-to-head each cycle on the same six axes, so tooling decisions rest on the team's own volume rather than on a vendor's story.

By the end of the cycle, the shape of the answer was stable: Claude Desktop as the daily analysis workstation for judgment-heavy work, the spreadsheet stack as the custodian of fixed outputs, and a single metered balance keeping both honest. If that mix matches the workload your team carries, open a Quanta account and run the same six-axis head-to-head on your own reporting cycle.