Why Forge Excel Intelligence — how structured context transforms AI data analysis

Excel & CSV Intelligence

Correct numbers.
Wrong
conclusions.

Every AI can read a spreadsheet. None of them know what the numbers mean — unless you tell them. Drop your file, tell Forge what's worrying you, and eight AI analysts turn your data into a decision. And every analysis you save, Forge remembers.

8
AI analysts working
your data simultaneously
7
Context questions that
change everything
5×
More accurate with
full context provided
forge.projectcoachai.com/excel
📊
municipal_budget_2025.xlsx
34 rows · 4 columns · Analysis complete
✓ Ready
F
I've read your file — 34 rows, 4 columns. Before I start analysing, tell me: what is this data about, and what are you trying to figure out?
Was good money management applied? Flag anything concerning.
F
Mixed practices — some good, some concerning.
✓ Personnel costs: CHF 268K under budget
✓ Operating expenses: CHF 387K under budget
⚠ Extraordinary expense: CHF 7M — zero budgeted
⚠ Net deficit: CHF 2.6M vs planned amount
What should I do?
Is this data reliable?
Tell me more
The problem with AI data analysis

The AI is right about
the numbers.
Wrong about what they mean.

Give any AI a spreadsheet without context and it will apply generic patterns to a specialist problem. The output is confident, readable, and wrong.

Flagged ≠ invalid

An AI without context sees a "flagged" status column and recommends deleting everything flagged. In most operational datasets, flagged means requires review — not remove. The AI turns a quality control signal into a deletion instruction.

The circular error trap

When both a calculated value and a stored reference value are derived from the same corrupted source, they match — and the AI concludes both are valid. Without knowing they share an origin, the AI confirms corrupt data as correct.

No domain baseline

A 42-million-hectare field is obviously impossible. But without knowing that a normal field in this dataset is 0.5–5 hectares, the AI has no basis to flag it as anomalous. Domain context is the calibration that makes anomaly detection meaningful.

The context behind the conversation

Seven questions.
One transformation.

When you chat with Forge about your data, these are the seven dimensions of context it explores. Each one gives eight AI analysts the domain knowledge they need to interpret your data correctly — not just compute it.

01
What is this data about?
Establishes the domain. Agricultural boundaries need different treatment than financial records or customer data. The AI calibrates its entire analysis to the domain you describe.
02
Where does this data come from?
Source determines reliability expectations. A KoboToolbox field collection export has different error profiles than a CRM export or a database backup. Context changes how anomalies are weighted.
03
What do status, category or flag columns mean?
The single question that prevented the deletion of 371 valid records in Run 1. Without this answer, every AI default-interprets "flagged" as a removal signal.
04
What are the normal or expected values?
Provides the baseline for anomaly detection. Without a known normal range, the AI cannot determine whether a value is anomalous — it can only report the distribution of what exists.
05
What decision will this analysis support?
Shapes the output structure. A boundary correction decision needs different emphasis than an audit report or a budget reforecast. The AI formats its conclusions for the actual use case.
06
Is there anything you already know or suspect?
"We suspect a coordinate order bug" is worth ten prompts of generic exploration. Priming the AI with domain hypotheses directs its attention immediately to the most important patterns.
07
The error-prevention question
Are stored or reference values independently verified?
This question catches the error class that no generic AI analysis ever catches on its own: circular error, where both a calculated value and a reference value are wrong for the same reason — because both were derived from the same corrupted source. When both values are large and closely matched, assume circular error. When stored size is blank, apply coordinate bounds check instead. This single question is the difference between confirming corrupt data and identifying it.
A real analysis — five runs

Same data. Same AI. Completely different conclusions.

The following five runs were performed on the same dataset: 376 flagged field boundary polygons from a carbon credit tree planting operation across verified project sites. The only variable was context.

Run 1
vs. Run 5
The difference between the first run and the fifth was not an incremental improvement. It was a different class of output entirely.
Run 1
All 371 polygons are invalid and should be deleted. The data cannot be used.
Run 5
4 geographic violators named by field ID. Circular error pattern identified. Three-phase action plan with 48-hour, 1-week, and 1-month timelines. Submittable to a carbon credit verifier as preliminary evidence of systematic data remediation.
1
No context
Delete everything
Conclusion: all 371 fields are invalid. Forge recommended deleting the entire dataset. The numbers were correct — the interpretation was completely wrong.
"The dataset contains 371 records with anomalously large polygon areas. These records are invalid and should be removed before further analysis."
2
Basic context added
Flagged ≠ invalid — first correct insight
First improvement. Adding a single line of context — "flagged means requires review, not invalid" — prevented the deletion recommendation. The AI now understood the status column correctly.
"Records flagged for large polygon area require review rather than immediate deletion. Further investigation is needed to determine which records are genuinely corrupted."
3
Six questions answered
Three-tier triage identified
The output flipped completely. With six questions answered, Forge correctly identified three severity categories, named the cascade effect, and recommended boundary correction over deletion. A document usable as a business report.
"Records fall into three categories: coordinate swap errors (immediate correction), boundary inflation (review and adjust), and genuinely large valid fields (retain with documentation)."
4
Six questions + analysis type
Named field IDs, geographic bounds applied
Significant precision jump. Adding the analysis type selector (Quality Review) produced named field IDs with hectare counts, applied geographic bounds reasoning for the project region independently, and cross-referenced the two uploaded files.
"FLD-M6S5C-N1D8R-001 reports a calculated area of 42,212,403 hectares — exceeding the entire landmass of multiple countries. Coordinates outside the project region bounds (lat −18° to −8°, lon 22° to 34°) confirm coordinate swap."
5
All seven questions answered
Report-ready: circular error detected, four geographic violators named
The most precise output of the entire progression. With the circular error question answered, Forge independently identified which fields had the error, distinguished it from coordinate swap errors, and produced a three-phase action plan with specific timelines.
Circular errors detected
Identified automatically
Geographic violators
4 fields named with exact coordinates
Action plan
48h · 1 week · 1 month phases
Output quality
Submittable to carbon credit verifier
What the output looks like

Not a summary.
A decision document.

Run 5 with full context produces findings you can act on immediately — with named field IDs, severity classification, and a phased action plan.

Critical findings
4 identified
Critical
Coordinate swap — 42.2M hectare polygon
Largest field exceeds the size of multiple countries. Latitude and longitude values are transposed, placing the centroid outside the project region entirely.
FLD-M6S5C-N1D8R-001 · Calculated: 42,212,403 ha · Expected: 0.5–5 ha
Critical
Circular error — stored matches corrupted calculated
Both stored size and calculated area are anomalously large and closely matched. Both values derived from the same corrupt polygon source. Neither can be used as reference.
FLD-R1W6D-R2Z7T-001 · Stored: 165,432 ha · Calculated: 165,447 ha
High
4 fields outside the project region geographic bounds
Coordinate bounds check (lat −18° to −8°, lon 22° to 34°) confirms centroid outside valid territory. Longitude violations in all four cases.
Valid the project region bounds violated · Immediate coordinate correction required
Medium
100% tree mismatch — cascade effect confirmed
8 fields show complete spatial allocation failures. Corrupt polygon boundaries are capturing trees assigned to adjacent valid fields.
FLD-R1W6D-R2Z7T-001 · 2,226 trees · 100% mismatch rate
Action plan
3 phases
48 hours — immediate
Isolate and quarantine the 4 records with centroids outside the project region bounds. Do not include in any audit submission.
Flag FLD-M6S5C-N1D8R-001 and all circular-error candidates for manual coordinate re-entry from original field collection records.
Halt any automated tree-matching process for affected fields until polygon boundaries are corrected.
One week — correction
Re-enter coordinates for coordinate-swap records from original KoboToolbox submissions. Verify against GPS tracks where available.
For circular-error fields: treat stored values as unreliable. Derive correct area from corrected polygon coordinates only.
Re-run spatial tree matching after boundary correction. Expect significant improvement in mismatch rates.
One month — verification
Submit corrected boundary dataset for independent spatial validation against the project region land registry coordinates.
Document correction methodology for carbon credit verifier. This analysis serves as preliminary evidence of systematic data quality remediation.

Export your analysis in any format

Every output is ready to share. White-label with your name and company on Work Like a Pro.

Six lenses on your data

One upload.
Six structured outputs.

Every analysis produces six tabbed outputs, each a different way of seeing the same data — from a one-page executive summary to ready-to-use Excel formulas.

📋
Executive Summary
Key findings in plain language, ready to present
ƒ
Excel Formulas
Ready-to-use formulas, click to expand and copy
📊
Chart Suggestions
Visualizations that reveal patterns in your data
Anomalies & Risks
Outliers, data quality issues, and red flags
Action Items
Phased plan with specific timelines and owners
Executive Summary
8 analyses synthesised
Key finding
The dataset of 376 field boundaries contains three distinct error categories requiring different responses. The most critical is a coordinate order bug confirmed in 4 records, with the largest polygon spanning 42.2 million hectares — larger than the entire country of Germany.
Severity breakdown
4 records — immediate quarantine (coordinate swap confirmed) · 8 records — cascade effect on tree matching · Remaining 364 — valid, retain with standard documentation
Consensus signal
All 8 AI analysts agree on the coordinate swap diagnosis. Divergence exists only on the correct remediation approach for circular-error records — reviewed in Divergence Spotlight.
Flag records outside the project region bounds
=IF(AND(B2>=-18,B2<=-8,C2>=22,C2<=34),"Valid","BOUNDARY VIOLATION")
Detect circular error candidates
=IF(AND(D2>10000,ABS(D2-E2)/D2<0.01),"CIRCULAR ERROR RISK","OK")
Classify severity tier
=IF(F2="BOUNDARY VIOLATION","CRITICAL",IF(G2="CIRCULAR ERROR RISK","HIGH",IF(D2>10,"REVIEW","VALID")))
Polygon area distribution by field
0–5ha
5–10ha
10–1k ha
>1M ha
Chart recommendation: Histogram of polygon areas on log scale reveals the bimodal distribution — valid cluster below 5ha and anomalous cluster above 100,000ha invisible on linear scale.
FLD-M6S5C-N1D8R-001 — 42.2M hectare polygon. Coordinate swap confirmed by centroid position outside the project region territory. Immediate quarantine required.
FLD-R1W6D-R2Z7T-001 — 165K hectare polygon. Stored size 165,432ha matches calculated 165,447ha — circular error pattern. Both values unreliable.
4 records — centroids outside the project region geographic bounds (lat −18° to −8°, lon 22° to 34°). Longitude violations in all 4 cases. Coordinate re-entry required.
8 records — 100% tree mismatch rate. Corrupt boundaries are spatially claiming trees belonging to adjacent valid fields. Cascade effect confirmed.
48h
Quarantine and stop all processing for the 4 records with confirmed coordinate violations. No audit submission until corrected.
48h
Halt automated tree-matching for all records with calculated area above 10 hectares until polygon boundaries are verified.
1w
Re-enter coordinates for coordinate-swap records from original KoboToolbox submissions. Do not use stored reference sizes for circular-error candidates.
1w
Re-run spatial tree matching after boundary correction. Expect resolution of the cascade mismatch effect across the 8 affected fields.
1mo
Submit corrected dataset for independent spatial validation. Use this analysis report as preliminary evidence of systematic remediation for the carbon credit verifier.
Who uses Forge Excel

Any data.
Any domain.
One conversation.

Forge adapts to your field because you tell it what your data means. The same intelligence that catches a coordinate swap in a carbon credit dataset reads a P&L or an HR export with the same precision.

Finance & Accounting

P&L variance explained,
not just reported.

Upload your monthly income statement. Ask Forge which variances matter, which are noise, and what to present to the board. It understands accruals, extraordinary items, and budget assumptions — because you told it.

"The CHF 7M extraordinary expense wasn't budgeted. Is that a red flag or routine?"
✓ Conservative revenue forecasting — tax revenue 13.5% above budget
✓ Personnel costs under budget by CHF 269K — well controlled
⚠ CHF 7M extraordinary expense — CHF 0 budgeted. Requires board explanation.
⚠ Net deficit diverges from plan — CHF 2.6M actual vs CHF 3.3M budgeted
Operations & Supply Chain

Inventory drift caught
before it becomes a crisis.

Drop your stock or production export. Tell Forge what normal lead time looks like, what your reorder threshold is, and what decisions are on the table. It surfaces the SKUs that matter — not a sorted list of everything.

"Which lines are at risk for next month's production run given current stock levels?"
⚠ 3 SKUs below reorder threshold with 18-day lead time — order window closes in 4 days
⚠ Component C-441 shows 12% variance between received and system quantities
✓ 87% of lines within normal buffer — no action required
✓ Supplier on-time rate: 94% — above category benchmark
Agriculture & Environment

The analysis that caught a
42-million-hectare field.

The Five Runs case study on this page is real. Without context, the AI recommended deleting 371 valid records. With context, it named four specific field IDs for coordinate correction and produced a report submittable to a carbon credit verifier.

"Which field boundaries need correction before the audit submission deadline?"
⚠ FLD-M6S5C-N1D8R-001 — 42.2M hectare polygon. Coordinate swap confirmed.
⚠ Circular error: stored size matches corrupted calculated area on 2 records
⚠ 8 fields with 100% tree mismatch — cascade effect from corrupt boundaries
✓ 364 records valid — retain with standard documentation
HR & People

Workforce patterns that
don't show up in dashboards.

Upload your headcount, turnover, or compensation export. Tell Forge which departments are under pressure, what tenure looks like in your industry, and what decision you're building toward. It reads between the numbers.

"Is our attrition in the 1–2 year band higher than it should be?"
⚠ 1–2 year attrition: 34% — 2.1× the industry median for this role category
⚠ Engineering exits concentrated in months 14–18 — onboarding cliff likely
✓ 5+ year retention: 91% — senior cohort highly stable
✓ Compensation bands within market range for 83% of roles
Healthcare & Research

Data quality issues found
before the submission.

Upload your trial data, patient records, or lab export. Tell Forge what a valid range looks like for each measurement, what the study protocol requires, and what you're preparing for. It flags the records that will fail peer review.

"Are there any values outside the protocol-defined range that would disqualify records from the analysis?"
⚠ 7 records with biomarker values outside protocol range — require adjudication
⚠ 3 visit date sequences out of order — data entry error likely
✓ 94.2% of records pass all protocol inclusion criteria
✓ Missing data rate: 1.8% — within acceptable threshold for imputation
Retail & E-commerce

Which SKUs are dragging
margin and why.

Drop your sales or margin export. Tell Forge what your target margin is, which product categories have different benchmarks, and what's on the table — a range review, a promotional plan, or a discontinuation decision.

"Which products are below margin target and what's driving the shortfall?"
⚠ 14 SKUs below 28% margin target — 9 attributable to supplier cost increases
⚠ Category C: average margin 19% — 9 points below category benchmark
✓ Top 20% of SKUs generating 67% of total contribution margin
✓ 3 underperforming SKUs show strong volume — margin fixable via pricing review
The memory companion

Forge Excel remembers
what matters.

Most AI tools forget everything the moment you close the tab. Forge Excel builds a relationship with your data world over time. Every analysis you save, every context you provide, every domain insight you share — it accumulates. The next time you upload, Forge already knows your field.

📁

Save every analysis to your profile

Every run is saved to your AI Diary automatically. Reload last month's P&L analysis, change one question, and run it on this month's data. The conversation doesn't end when you close the tab — it continues where you left off.

🔁

Reload, modify, re-run

Your saved analyses aren't static reports — they're living documents. Pull up a previous run, update the file, refine the question, and Forge produces a new analysis with full context intact. Month-on-month comparison at a fraction of the effort.

🧠

Gets smarter with your domain

Forge has a proprietary intelligence layer that accumulates domain knowledge across the platform. The more analysts in your field use Forge Excel, the better it gets at spotting what matters in data like yours. We don't share the technical details. You'll notice the difference.

📊
Your saved Excel analyses
Reload anytime — context preserved
CCS_Compliance_Trend_2026-05-18.csv
24 rows · 8 AI analyses · 18 May 2026
Reload →
Q1_PL_Variance_2026.xlsx
847 rows · 8 AI analyses · 2 Apr 2026
Reload →
HR_Attrition_Report_Mar2026.csv
312 rows · 8 AI analyses · 15 Mar 2026
Reload →
Inventory_Stock_Feb2026.xlsx
1,204 rows · 8 AI analyses · 3 Feb 2026
Reload →
"We don't reveal exactly how the intelligence layer works.
But every analyst who uses Forge Excel regularly notices the difference."

Your data.
Eight minds.
A companion that remembers.

Excel or CSV. Any dataset. Eight AI analysts working simultaneously, structured to your domain. Saved, reloadable, and getting smarter with every run.

Free to start. Upload your data and ask your first question — no setup required.