Microsoft Copilot Excel Formulas: The Multi-Model Workflow Pakistani Analysts Use in 2026

Key takeaway: Microsoft Copilot Excel formulas are generated from the Copilot side pane and Agent Mode, not from an in-cell function, because Microsoft removed the experimental =COPILOT function on 14 September 2026. Copilot writes XLOOKUP, SUMIFS and nested IF logic dependably because it reads real column headers and data types. It falls short on unsupervised multi-sheet financial models. Analysts therefore route logic design to a long-context model such as Claude, execution to Copilot, quantitative work to Python in Excel, and handoffs to n8n.

Introduction: why single-model Excel work stopped being competitive

Microsoft Copilot Excel formulas no longer live inside cells — they are drafted in a side pane and executed through Agent Mode, because Microsoft pulled the experimental =COPILOT function out of Excel on 14 September 2026 after its Microsoft 365 Insider preview. That architectural change forced a working decision on every analyst who bills for spreadsheets: if the AI cannot sit inside the formula bar, it has to sit inside a pipeline. As a result, the stack that top-rated freelancers now run is four parts — Copilot in Excel, a long-context model such as Claude, @python-in-excel, and n8n for orchestration.

This guide is for analysts who deliver client-facing workbooks. It supplies the prompts that return usable formulas, the dated 2026 feature changes, the five-step multi-model pipeline, a model-selection table and a seven-step verification routine. Pakistan’s freelance market adopted the pattern early because Upwork client briefs increasingly specify AI-assisted deliverables while per-seat licence costs are dollar-denominated against PKR revenue — demand pressure plus licence economics, not novelty.

Definition: multi-model AI workflow
A multi-model AI workflow routes each stage of an analysis to the model best suited to it — retrieval and cleaning to a RAG pipeline, logic design to a long-context model, spreadsheet execution to Copilot in Excel, statistics to Python, and orchestration to an automation layer such as n8n.

What Microsoft Copilot in Excel actually does with formulas

Microsoft Copilot Excel formulas are produced when you describe an outcome in the Copilot side pane and Copilot returns an editable formula grounded in your workbook’s actual headers and data types. Because it reads the sheet rather than guessing at it, its column references land correctly far more often than a generic chatbot’s — but that same grounding means it inherits every inconsistency already present in your data.

The practical core of any Copilot in Excel tutorial is the reliability gradient, and training analysts at platforms such as Learnesy have made the same observation in their 2026 reviews of the tool. Copilot is dependable on single-sheet, single-objective formulas and degrades as dependent sheets multiply, because each additional reference adds one more assumption it has to infer on your behalf. The verification routine later in this guide exists precisely to catch what that inference gets wrong.

The formula families it writes reliably (XLOOKUP, SUMIFS, nested IF)

Copilot performs most consistently on XLOOKUP, SUMIFS and nested IF. XLOOKUP works because its argument order is positional and unambiguous, which leaves little room for misinterpretation. SUMIFS works because described conditions map one-to-one onto criteria pairs. Nested IF works because Copilot can restate the branching logic in prose before committing it to syntax, giving you a readable checkpoint. It also explains an inherited formula in plain language, which is why it functions as an Excel AI formula generator and a teaching tool at the same time. TEXTSPLIT, LET and dynamic-array outputs sit in the next tier, where results are usable but review time rises sharply.

XLOOKUP vs VLOOKUP — which one Copilot defaults to and why

Copilot defaults to XLOOKUP, and that default is correct. XLOOKUP does not break when a column is inserted, it searches left or right, and it takes an explicit if_not_found argument, which removes the IFERROR wrapper VLOOKUP requires. There is one genuine reason to accept VLOOKUP instead: compatibility with clients running older Excel builds where XLOOKUP is unavailable, which means the formula returns #NAME? on their machine and your deliverable looks broken. Control this by naming the target Excel version in the prompt itself — Copilot adjusts its output to the constraint you state.

Five prompts that produce usable formulas on the first attempt

The prompts that work share a shape: name the sheet, name the columns, state the output shape, state the error behaviour. Vague prompts force Copilot to infer all four, and inference is where the errors enter. The table below gives five patterns you can paste and test inside ten minutes.

Page not found – theverge.pk

Prompt pattern	Formula family returned	Review step before use
“On Sheet1, match Order ID against Sheet2 column A and return Amount; leave blank if missing”	XLOOKUP with if_not_found	Confirm the lookup column, not a similarly named one
“Total Sales by Region and Quarter from the Orders table”	SUMIFS	Check whether blanks and refunds are excluded
“Flag each row as Low, Medium or High using the thresholds in Limits”	Nested IF	Test boundary values at each threshold
“Split Full Name into first and last name columns”	TEXTSPLIT or TEXTBEFORE/TEXTAFTER	Inspect middle names and single-word entries
“Explain this formula in plain language, then simplify it”	LET or a rewritten nested formula	Compare old and new outputs on the full column

What changed in Copilot for Excel during 2026

During 2026, Copilot in Excel gained dashboard generation, reusable skills and a Python entry point, while losing in-cell AI entirely. The net effect is a shift in what the job actually requires: the valuable 2026 skill set is prompt design plus verification, not formula memorisation, because the tool now writes the syntax and the analyst owns the specification and the audit.

Date	Change	Effect on formula work
August 2026	@python-in-excel skill for Copilot	Statistical work stays in the workbook
14 September 2026	Experimental =COPILOT in-cell function removed	Existing workbooks using it return errors
October 2026	Interactive Canvas	Dashboards generated from natural language
October 2026	Skill Builder and Skill Recorder	Repetitive reporting becomes a reusable skill

Interactive Canvas (October 2026)

Microsoft rolled out Interactive Canvas in October 2026, and the Excel Interactive Canvas Copilot feature turns workbook data into interactive dashboards and executive-style reviews from natural-language prompts, with outputs that refresh when the underlying data changes, according to Microsoft’s own Copilot release communications. The consequence for formula work is uncomfortable but clarifying: presentation moves out of hand-built chart sheets, which means the analyst’s remaining differentiator is the correctness of the calculation layer beneath the dashboard.

Skill Builder and Skill Recorder (October 2026)

Also released in October 2026, Skill Builder lets you describe a workflow in words to create a reusable Copilot skill, while Skill Recorder captures your manual actions in Excel and converts them into a macro-like skill Copilot replays later. The economic cause of this shift is plain: repetitive monthly reporting was the bulk of junior analyst billing, therefore recording that work once collapses recurring hours into a one-time setup fee. Copilot Skill Builder Excel work consequently pushes freelancers upward into modelling and verification, where hours still defend themselves.

@python-in-excel (August 2026)

Microsoft introduced a Python skill for Copilot in August 2026, invoked by typing @python-in-excel, which generates editable, refreshable Python formulas directly inside the workbook for predictive and statistical modelling. Python in Excel Copilot matters because regression, clustering and distribution work that formula chains handle badly now lives in the same file as the audit trail. That removes the CSV round-trip to an external notebook, and the round-trip was the step that most reliably broke reproducibility.

The =COPILOT function removal (14 September 2026)

The Copilot function Excel deprecated event happened on 14 September 2026, when Microsoft removed the experimental in-cell =COPILOT function after its preview phase in the Microsoft 365 Insider program, as reported in coverage of the change. Workbooks built on it now return errors, and all AI interaction runs through the side pane and Agent Mode instead. The analytical reading is that an in-cell model call made every recalculation non-deterministic — the same sheet could return different values on different opens — therefore removing it protected auditability. For client-facing models, that is the correct trade.

The multi-model workflow, step by step

The pipeline exists because no single assistant covers retrieval, logic design, spreadsheet execution, statistics and scheduling at acceptable quality. Each stage is assigned to the tool whose failure mode does least damage there, and Microsoft Copilot Excel formulas occupy the execution link in that chain — the point at which an already-specified rule becomes working syntax inside a real workbook.

Be honest about the entry cost. This stack requires a Microsoft 365 Copilot licence, API access to at least one external model, and an n8n instance, self-hosted or cloud. That setup expense is real, which means it is recovered only on repeat client work rather than on a single engagement.

1. Ingest and clean source data through a RAG pipeline into one flat table
2. Draft the model logic in a long-context model and export a numbered logic spec
3. Execute the spec in the Copilot side pane and Agent Mode, rule by rule
4. Add regression, forecasting and simulation with @python-in-excel
5. Orchestrate triggers, retries and delivery in n8n

Step 1 — ingest and clean with RAG

A RAG pipeline feeding clean data into Excel retrieves from PDFs, invoices, client portals and databases, normalises units, date formats and entity names, then writes a single flat table. The causal logic is direct: Copilot grounds its formulas in column headers and data types, therefore inconsistent headers such as “Amt”, “Amount” and “Total Amount” produce formulas that are confidently wrong rather than visibly broken. Cleaning upstream is what makes the downstream prompt reliable.

Step 2 — structure logic in a long-context model

Analysts combine Copilot and Claude for Excel analysis because a long-context model holds the whole model specification at once — assumptions, dependencies, exclusion rules, edge cases — and returns a written logic map before any formula exists. Consequently the Copilot prompt stops being a request for analysis and becomes a narrow instruction with an unambiguous output shape, which is the condition under which Copilot performs best. The deliverable from this stage is concrete: a numbered logic spec naming each sheet, each column and each rule in order.

Step 3 — execute in Copilot and Agent Mode

Feed the logic spec into the side pane one rule at a time. Copilot Agent Mode Excel is the multi-step execution path that builds tables, applies the formulas and assembles structure across a workbook without a prompt for each action, and its constraint follows from its strength: longer chains mean more uninspected intermediate steps, therefore every agent run should end with a human check against a row you calculated by hand. Where the same workbook recurs monthly, capture the sequence with Skill Recorder rather than re-prompting it.

Step 4 — quantitative lift with Python in Excel

Call @python-in-excel for regression, seasonality decomposition, cohort retention and Monte Carlo ranges, because these are exactly the operations where formula chains become unreadable and therefore unverifiable. The audit benefit is the real argument: the Python block is visible, editable and refreshable inside the workbook, which means a reviewer reads the method as written instead of reverse-engineering six layers of nested functions.

Step 5 — automate the handoffs in n8n

Excel automation with n8n means triggers on a new file or a schedule, HTTP nodes calling the model APIs, a node writing the cleaned table to OneDrive or SharePoint, and a notification when the workbook is ready for review. The orchestration layer sits outside Excel because it must survive a failed model call and retry it, and a workbook cannot do that. The same reasoning drives scientific pipelines, and the structure is explained in detail in this guide to building an automated data analysis pipeline.

Claude vs Copilot vs GPT for spreadsheet work

For formulas written against a live workbook, Copilot in Excel wins, because it alone reads your actual headers, data types and change history. For multi-sheet logic design, a long-context model wins, because the specification exceeds what a side-pane prompt holds. For an ad-hoc ChatGPT Excel formulas request built from a pasted sample, GPT is fastest. The operating rule is to choose by task, not by whichever subscription you already pay for.

Task	Best model	Why	Failure mode
Writing a formula against a live workbook	Copilot in Excel	Reads real headers and data types	Inherits messy headers
Designing multi-sheet financial model logic	Long-context model (Claude)	Holds full spec and dependencies	No access to the live file
Explaining or debugging an inherited formula	Copilot in Excel	Sees the referenced ranges	Misses intent behind the design
Statistical analysis and forecasting	@python-in-excel	Editable, refreshable, auditable method	Requires Python literacy to review
Building an interactive dashboard from a table	Copilot Interactive Canvas	Refreshes with source data	Looks credible over wrong totals
Quick formula from a pasted data sample	GPT	Fastest turnaround, no file needed	Cannot verify column names

theverge.pk – AI Governance and Data Standards

The cost of mixing models is the part most guides skip: context does not transfer between them, therefore the analyst carries every assumption by hand from one tool to the next, and that manual carry is the single largest source of error in this stack. Interoperable formats and shared conventions reduce the friction, which is why open standards in AI are a practical concern and not an abstract one. On Claude vs Copilot for data analysis, and on the best AI model for financial modelling in 2026, the defensible position is the same: the model that writes the specification should not be the model that executes it.

Why Pakistani analysts adopted this stack first

Start with demand. Pakistan ranks among the fastest-growing freelance technology markets, and Upwork listings for top-rated analysts from hubs such as Lahore now advertise multi-model workflow orchestration rather than spreadsheet production alone. Briefs for data analyst jobs in Pakistan increasingly specify AI-assisted deliverables, which means a freelancer in Lahore or Karachi competes on pipeline delivery instead of formula typing speed. The direction of travel in freelance data analyst rates on Upwork follows scope rather than hourly skill — broader deliverables command broader engagements — though specific rate figures are not verified here.

Then the cost side. A Microsoft 365 Copilot price decision in Pakistan is a dollar-denominated per-seat commitment set against PKR revenue, therefore a solo analyst scrutinises it far harder than a salaried employee inside an already-licensed enterprise. That arithmetic drove the hybrid pattern directly: one Copilot seat for client-facing Excel execution, metered API calls for logic design and retrieval, and self-hosted n8n to avoid a third recurring subscription. The impact is a stack assembled around marginal cost rather than vendor convenience. Treat any pricing you find as indicative and exchange-rate sensitive.

Finally, knowledge. The Urdu and Hindi tutorial ecosystem on YouTube compressed the learning curve: full-length 2026 masterclasses from channels including AI Workspace teach VLOOKUP through Agent Mode and AI dashboard building in a single sequence. In this Excel course in Urdu with AI 2026 format, learners meet the AI layer and the formula layer together rather than years apart. As a result, adoption skipped the retraining resistance seen in organisations carrying legacy spreadsheet conventions. The decisive reading: the stack spread fastest where switching costs were lowest.

Where Copilot fails, and the verification routine that protects you

Can you trust Copilot Excel calculations? Trust the arithmetic, verify the assumptions. Copilot executes the formula it writes correctly, therefore errors enter earlier — when it infers the wrong column, the wrong date grain or the wrong exclusion rule. This matters more in finance than anywhere else, because an AI hallucination in a financial model does not look wrong. It looks plausible and it totals correctly, which means only a tie-out against a value you already know will catch it.

1. Tie the output to one row you have calculated by hand before trusting any aggregate
2. Re-read every cell reference Copilot chose and confirm it matches the intended column, not a similarly named one
3. Test the formula against edge cases: blanks, zeros, duplicates, negatives and text in numeric columns
4. Check date grain and timezone handling, because month-end and fiscal-year boundaries are where silent errors cluster
5. Confirm every assumption in the logic spec appears in the workbook, since omitted exclusions produce plausible wrong totals
6. Replace in-cell AI dependencies and stale references, including any workbook that still calls the removed =COPILOT function
7. Record prompt, model, version, date and reviewer in an audit tab before the file leaves your machine

5 Critical AI Governance Challenges in Multi-Institution Research Labs – theverge.pk

An audit trail for an AI-generated spreadsheet is a governance artefact, not paperwork. Record the prompt, the model, the date and the reviewer alongside the formula, because reproducibility requires knowing which system produced which cell. The risk-management framing that applies here is the same one US research teams use under the NIST AI Risk Management Framework, and retention expectations follow standard recordkeeping practice. For the implementation detail, see this AI data governance framework walkthrough, the analysis of model provenance challenges in collaborative environments, and the comparison of AI vs traditional data governance.

Frequently asked questions

Can Microsoft Copilot write Excel formulas for me?
Yes. Microsoft Copilot Excel formulas are generated from the side pane when you describe the outcome, and Copilot returns an editable formula grounded in your workbook’s real headers and data types. It is dependable for XLOOKUP, SUMIFS and nested IF, and requires human review on multi-sheet logic.

What is a multi-model workflow in Excel analysis?
It routes each stage to the best-suited model. A RAG pipeline cleans the data, a long-context model designs the logic, Copilot executes inside the workbook, @python-in-excel handles statistics, and n8n orchestrates the handoffs — because no single assistant performs all five stages at acceptable quality.

Why was the =COPILOT function removed from Excel?
Microsoft removed it on 14 September 2026, after its Insider preview phase. An in-cell model call made recalculation non-deterministic, which broke auditability; consequently AI interaction moved entirely to the side pane and Agent Mode. Workbooks still calling the function return errors and need rebuilding.

Is Copilot in Excel better than ChatGPT for formulas?
For live workbooks, yes. Copilot reads your actual columns, data types and change history, therefore its references are correct more often. ChatGPT is faster for a formula built from a pasted sample, but it cannot see the file, so column errors are more common.

How do Pakistani analysts use Copilot with Claude and GPT together?
By splitting the job. A long-context model writes the numbered logic spec, Copilot executes it rule by rule inside Excel, GPT handles retrieval and document parsing, and n8n connects the steps — which keeps each model inside the task where its failure mode does least damage.

Is Microsoft Copilot available in Pakistan and what does it cost?
Yes, through Microsoft 365 Copilot add-on licensing. Pricing is dollar-denominated per seat and billed monthly or annually, so the PKR figure moves with the exchange rate and reseller margin. Check Microsoft’s current Pakistan pricing page and your tenant eligibility before committing to seats.

Can Copilot build a dashboard from a spreadsheet automatically?
Yes, through Interactive Canvas, released in October 2026. Describe the data in a prompt and Copilot generates an interactive dashboard or executive review that refreshes when the underlying data changes. Verify the aggregation logic first, because a clean dashboard over wrong totals still reads as credible.

Where does Copilot in Excel still fail in 2026?
Unsupervised multi-sheet financial modelling. Accuracy falls as dependent references increase, because each one adds another inferred assumption. It also struggles with inconsistent headers, mixed data types and ambiguous date grains, which is exactly why experienced analysts design the logic outside Excel first.

How do I audit an AI-generated spreadsheet before sending it to a client?
Tie out one hand-calculated row first. Then re-read every cell reference, test blanks and duplicates, confirm the date grain, check that no removed function remains, and log prompt, model, date and reviewer in an audit tab. Follow the seven-step routine and the governance framework guide above.

Which Excel AI skills pay best for freelancers in 2026?
Pipeline orchestration and verification, not formula writing. Because Skill Recorder and Agent Mode collapsed repetitive reporting into one-time setup, paid work shifted toward RAG data preparation, Python in Excel modelling, n8n automation, and documented audit trails that a client’s compliance team will accept.

Limitations of this workflow and the alternatives worth considering

This stack has real limits. It requires three paid or hosted components, so it is uneconomic below roughly weekly recurring analysis. Feature availability in Copilot in Excel varies by Microsoft 365 channel, tenant and region, which means a documented feature may simply be absent from your build on the day you look for it. And context does not transfer between models, therefore the analyst remains the integration layer — and consequently the main point of failure in the whole chain.

Alternatives win under specific conditions. Power Query plus a single Copilot seat is sufficient for repeatable monthly reporting. A Python or R notebook under Git is stronger wherever reproducibility is formally audited, because version history is native rather than bolted on. A BI platform is the correct answer once more than five people consume the same numbers, since spreadsheet distribution itself becomes the governance risk.

The decisive takeaway

The 2026 changes form a single causal spine: @python-in-excel arrived in August, the =COPILOT function was removed on 14 September, and Interactive Canvas alongside Skill Builder and Skill Recorder landed in October. Together they moved the analyst’s value from writing formulas to specifying and verifying them. Consequently the competitive unit of work is the pipeline, not the cell.

Before adding a single model to your chain, run the seven-step verification routine on one existing workbook this week, because an unverified pipeline scales errors faster than it scales output. Treat Microsoft Copilot Excel formulas as the execution stage of a documented process rather than a shortcut around one. Read more AI coverage.

References

• Microsoft — Copilot in Excel feature announcements and release notes: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF5_LW6o0a3Rwe_mGkzGW2_nli-BuWsjJFU1JMYzeLCWr8dMcSQHdriXVTDqPeR5NSpGyw2HsbfQlD8HhEC7GkzceeHgf38EwHMGx87PrCkWhf7c5XwmFvJBneg4ACEN417AlatSO5zLELpQNA-losTVtDThMoEzYiO_NqY_EPdguU5oD1xpqO6pwz7eTh94g==
• Microsoft — Copilot skills and Excel integration documentation: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGrLIZ2PhecYhzeSY2GrJeprz-N3yyxiALEs3STnFKl1nO7UI0wdPGOTH7r3X7ssSxUUSyNXTmk4cOcBqiStKEQ-uoENSCK7qW4LfPe-tb5xmryB1z1bsShyd1hjZuj1jLFpEdyZQlefTTCGxf0iAomR7WJ12UQIQkFLg==
• Windows Report — coverage of the =COPILOT function removal from Excel: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEE82odGysfaxLKoUs0NUhLCO7sFqD6GAtapyMAmOzYNBp1hw16jwLgGrWwyXB1ZABWmEQoT2VDMtR7vDC5spSlt3KUDCiL3kRH1-xDe0s6g6WNhH8x_Eavjg_vZEyLtF03wYtsuseuzE1Wd7aJYEhWHxkmxbIyAHsKHnSARxWGiuuTxGZuoEVeM2M=
• Learnesy — 2026 analysis of Copilot in Excel capabilities and limits: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGWRBgqfjaZo0KFUwhCv_v5JPxOCqo-x0C2PQuJROtoKx7-9o-XfZdmkvQ2e9i6w8UDgx65acdC3kwhAMrJERZ55iU9_z9TszPnoJgahsbApqJ6tuZLVM0erhLk5WWdfUUZ4egQPS9rus024g==
• Upwork — top-rated Pakistani analyst listings advertising multi-model workflow orchestration: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHo_FupU9LwFKSzlsApwwTD7fNpoEBSDi64v8s3SVqGo__4lOxbjWHyiwqBLLVJPoMkSvbtG9LXm06Ch7lsyv5zx6XN-OgzHpc-avEvufL5DFmw9fOfMMs_4HNp9Zns13o1U7Ux8sZrq-91P-thwvNQaTA=
• Upwork — freelance data and AI service listings from Pakistani hubs: https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH4OUMpoivCqUtjHk-Gobjd7NKdHlzxKh91NmN1Cf5M5Z3EHbHJvT2BXGQH3bEZILbn-g5-7arW51gFypSfHImgxsazAQPd0cTglZ1SH3O_w-2BA4ZtdKBOf0UfHaipXVy-3g9I
• YouTube — AI Workspace Excel Full Course 2026 (Hindi/Urdu masterclass): https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGHgTUDIcRfMlKjPF6sdkxf8POeAXfUdsK9zkFEO2symInr5dOcxs4UoBS51ddwehpmhe4ATu8bG7ujpq0y4BUT3SiygVhxFl3oivTAEhA5MP6OhXlDx27MLmIwx3KVqcsT

Related Reading

• Home – theverge.pk
• What Are Self-Driving Labs? – theverge.pk