What do you do as the “excel” guy with AI now??

The New Reality: AI Can Draft Your Models in Minutes

If you've spent years mastering pivot tables and driver-based models, the ground just shifted: an AI assistant can now produce a near-analyst-quality first draft of your forecast in minutes. That's not a prediction — it's happening today, and it changes what your expertise is worth.

In a controlled head-to-head test on real finance tasks, one practitioner found that a manual forecast build that takes 3–4 hours could be compressed into a first draft in under 2 minutes at roughly "90% usability" (f9finance). The tools are fast. They are not, however, reliably right.

The same test exposed why speed alone is dangerous. Copilot produced clean-looking output built on the wrong methodology — a statistical projection with no assumptions to interrogate and no drivers to adjust — while ChatGPT omitted cost of goods sold entirely, leaving a structurally broken workbook. Polished formatting can hide wrong math, and a Copilot-style rate/volume error overstated margin compression by almost four percentage points.

The market context makes this urgent. Microsoft claims 70% of Fortune 500 companies have adopted M365 Copilot, and paid seats have grown from 15 million to 30+ million in six months. Yet adoption is broad but shallow: 74% of companies report no tangible value from AI yet, and a paid seat "can sit unopened" — one prompt in a 90-day window counts as an "active" user (Eyal Marcus).

So where does that leave the Excel expert? Precisely in the gap between fast output and correct output. The value-add is no longer typing formulas; it's the judgment AI lacks:

• Checking assumptions: is the model driver-based, and are the inputs defensible?
• Catching methodology errors, like a rate/volume mistake that quietly skews margins
• Validating structure — did the tool omit a cost line or bury everything in one tab?
• Deciding whether a draft is "hand-to-someone" quality or needs a rebuild

The expert's new job is to make AI actually deliver. That includes vetting the AI tools and agent skills you plug into your workflow, because quality varies enormously and subtle errors are common. SkillzBench is building exactly that layer: a free directory of agent skills with security scanning, version pinning, and published benchmark results showing how much each skill actually lifts performance — so you can choose assistants based on verified results, not marketing.

The tools will keep getting faster. The people who know how to check the math are the ones who turn that speed into value.

Polished Output Can Hide Broken Math: What the Head-to-Head Tests Reveal

The most dangerous AI output isn't the one with broken formatting — it's the one that looks perfect and is quietly wrong. A recent practitioner head-to-head test put ChatGPT, Claude, and Microsoft Copilot to work on the same coffee-shop dataset with the same prompts, and the results upended assumptions about which tool produces "good" spreadsheet work.

Claude won all three finance tasks — forecast modeling, scenario analysis, and CFO memo writing. Its edge wasn't speed; it was structure. Claude confirmed its plan before building (adding roughly 30 seconds of overhead), then produced driver-based models with separate assumptions tabs and notes explaining each assumption choice. The output came close to hand-to-someone quality with light review, and its first draft of a forecast that takes a human analyst 3–4 hours arrived in under two minutes at roughly "90% usability," according to the same-dataset comparison.

Copilot's failure mode is more instructive — and more worrying. It used FORECAST.LINEAR, Excel's native statistical projection, instead of driver-based assumptions. The formatting was clean; the methodology wasn't. As the test's author put it, "It's not wrong, exactly. It's just not how a finance team builds a forecast model" — no assumptions to interrogate, no drivers to adjust, no way to run a meaningful scenario. Worse, a Copilot-style rate/volume error overstated margin compression by almost four percentage points, and manually rebuilding the model took two hours.

ChatGPT was worse still: it omitted cost of goods sold entirely, despite the figure sitting in the source data. Its scenario output showed operating margin swinging from roughly zero to over 100% — a number no reviewer should have accepted at face value.

The lesson cuts three ways:

• Polish is not proof. Clean formatting can hide wrong math, and a wrong number in a polished model is harder to catch than an obviously messy one.
• Verifiable output beats beautiful output. Models that show their assumptions in separate tabs — ones you can interrogate, adjust, and stress-test — are worth more than one-click answers.
• Expert validation is the irreplaceable step. Checking drivers, catching rate/volume errors, and interrogating methodology are exactly the skills AI can't yet do for itself.

This quality variance is precisely why verification matters when layering AI onto spreadsheet workflows — and why SkillzBench is building benchmarking and security scanning for agent skills, with results published openly rather than hidden behind marketing claims. The tools are powerful; the people who check the math are what make them trustworthy.

Your New Role: From Model Builder to Model Validator

The scariest thing about AI-generated spreadsheets isn't that they're wrong — it's that they look right. In a practitioner head-to-head test of ChatGPT, Claude, and Microsoft Copilot on real finance tasks, Copilot produced a clean-looking forecast using the wrong methodology entirely — a statistical projection with "no assumptions to interrogate, no drivers to adjust, and no way to run a meaningful scenario." That's exactly the gap you now fill.

Your value has shifted from building models to validating them. The same test showed AI can compress a 3–4 hour manual forecast build into a 2-minute first draft at roughly 90% usability — but a Copilot-style rate/volume error overstated margin compression by almost four percentage points, and rebuilding it manually took 2 hours. The expert who catches that error before it reaches the CFO is worth more than the one who built the original model.

Adopt a simple workflow: first draft, never a finished model. Every AI output gets checked the way you'd check a junior analyst's work:

• Interrogate the assumptions — are the drivers real, or did the tool invent convenient numbers?
• Verify rate/volume math line by line, especially where margins look suspiciously smooth
• Confirm the methodology matches how your finance team actually models, not just what looks polished
• Check for silent omissions — in the same test, ChatGPT dropped cost of goods sold entirely

When choosing which AI assistant earns a place in your workflow, the f9finance test author offers a criterion worth taping to your monitor: "which one reduces the amount of work I have to do after it finishes." Not features, not polish — the tool you pick determines whether the analyst gets home before 2am. Claude won the head-to-head largely because it confirmed its plan before building and delivered driver-based models with separate assumptions tabs you could actually audit.

Here's the uncomfortable market context: 74% of companies report no tangible value from AI yet, despite 94% of leaders calling it critical. The gap isn't the technology — it's that most organizations bought licenses instead of building skills. As one adoption analysis puts it, "AI adoption is a people problem, and training is what makes a license worth it." A paid seat can sit unopened; a trained validator cannot.

The same discipline applies to the broader AI tooling ecosystem. Published research found roughly 26% of agent skills carry at least one vulnerability, which is why SkillzBench is building an independent directory that security-scans, version-pins, and benchmarks every skill before it touches your workflow. The principle is the one you already live by: trust output only after you've verified how it was built.

Governance First: Spreadsheets, Sensitive Data, and AI Risk

Before you let an AI assistant loose on company spreadsheets, there's a question worth pausing on: what exactly can it see? For the Excel expert handling sensitive financial data, governance isn't an afterthought — it's the prerequisite that determines whether AI adoption delivers value or creates liability.

The scale of the concern is measurable. Copilot reaches across SharePoint, Teams, OneDrive, and Exchange, which means it can surface anything anyone in the organization has ever shared too widely. According to Gartner data from June 2026, 51% of IT leaders name oversharing and data loss as the top deployment barrier — and roughly half of IT leaders lack confidence in managing Copilot's security and access risks at all.

Two recent developments sharpen the point. U.S. courts are beginning to treat Copilot prompts and responses as discoverable content, so a casual "what's our margin on this client?" typed into a prompt may resurface in litigation. And the Echoleak vulnerability, identified in early 2025 and since patched, demonstrated that sensitive email data could be silently exfiltrated from Copilot. As one analysis put it, "Copilot isn't just an AI project; it's a governance project."

For spreadsheet work specifically, the stakes compound. Financial models contain compensation data, pricing assumptions, and unreleased figures — exactly the material you don't want an AI tool summarizing for whoever asks. The same skepticism you apply to AI-generated math applies to AI access claims:

• What exactly was counted — seats, prompts, or actual data touched?
• Who was measured, and over what time window?
• When was the claim made, given how fast these tools change?
• Who is saying it, and could they sell something with it?

That four-question framework, developed by analyst Eyal Marcus, works for verifying any AI adoption claim — including claims about the skills and tools you connect to your agents. A skill package that promises to automate your reporting workflow is also a piece of code with access to your data, and published research has found roughly 26% of agent skills carry at least one vulnerability. This is why we're building SkillzBench: to security-scan, version-pin, and benchmark agent skills in a sandbox, so you can see a trust score and a measured performance Lift before anything touches your environment.

The Excel expert's role here is natural. You already interrogate assumptions, validate drivers, and catch rate/volume errors that polished output hides. Applying that same discipline to what your AI tools can access — and what you connect to them — is what turns AI adoption from a governance risk into a governed capability.

Getting Started: A Practical Playbook Plus Verified Agent Skills

Knowing where AI goes wrong is your advantage — the question is how to operationalize it. The rollout playbook is less about picking a "best" tool and more about building verification habits that scale.

Start small and start with believers. Research on Copilot adoption recommends licenses for leaders and AI champions first, then expanding based on actual usage — because "a paid seat can sit unopened," and one prompt in a 90-day window still counts a user as "active." Meanwhile, 74% of companies report no tangible value from AI yet, which usually means they scaled licenses before they built skills.

Your working rules look like this:

• Treat every AI output as a draft with a mandatory expert review step — in one practitioner test, a Copilot-style rate/volume error overstated margin compression by almost four percentage points.
• Prefer tools that expose assumptions. Claude won the same head-to-head because it built driver-based models with separate assumptions tabs and explained its reasoning, at ~90% usability from a 2-minute draft.
• Lock down governance before scaling — 51% of IT leaders name oversharing and data loss as the top deployment barrier, and Copilot prompts may be legally discoverable.

The same verification mindset now extends to agent skills. These SKILL.md instruction packages — for Claude Code, Codex, Cursor, and chat-app agents — promise to automate spreadsheet workflows, but published research found roughly 26% of skills carry at least one vulnerability. An Excel expert wouldn't trust an un-audited macro library; skills deserve the same skepticism.

That's the gap SkillzBench is being built to close. Its free directory will continuously crawl and dedupe skills from GitHub, pin every version to a commit SHA, and security-scan, license-check, and quality-judge each one. Every skill gets benchmarked in a sandbox with the same model, with and without the skill, so the Lift in percentage points is published — kept separate from ratings, with money never touching scores. A free MCP connector then wires your chosen skills into any agent, and ongoing tracking re-scans and re-benchmarks as upstream code changes.

The platform is pre-launch, but the waitlist is live — and for spreadsheet professionals, it's the natural home turf: your whole career has been about checking the math. Skills, proven.