In 2026, the most boring problem in business is also the most expensive. It is the person who knows Excel better than the analyst but slower than the manager would like, opening the same file every Monday and writing the same VLOOKUP they have written since 2014. It is the operations lead who spends Tuesday reconciling two CSVs by hand. It is the finance team that does a monthly close that takes five business days because pivot tables and conditional formulas do not scale.

Every one of those workflows is now something an AI agent for Excel can take off your plate. Not "in two years" — today. With off-the-shelf tools. Without a data team. This guide walks through seven of them in detail, with the actual prompts and the actual formulas, so you can ship the first one before lunch.

This article is for the working professional who lives in spreadsheets — operations, finance, sales ops, marketing ops, founders, executive assistants, and the engineers who are tired of writing one-off scripts. If you have ever said "Excel can do this, I just don't have the time," this is the article that gives you the time back.

What an Excel agent actually is

Three words to keep straight, because vendors blur them.

• AI in Excel. A feature inside Excel. Microsoft Copilot is the most famous example. You click a button, Copilot suggests a formula, you accept or reject. The data never leaves the workbook context, the user never leaves the application, and the surface area is bounded by what Microsoft decided to expose. This is a feature, not an agent.
• AI for Excel. A separate tool that talks to Excel. You open ChatGPT or Claude in a tab next to your workbook, paste a sample of your data, ask a question, and copy the formula or the analysis back into Excel. Faster than a feature, but you are the integration layer. The agent does not see the workbook change; you do.
• AI agent for Excel. A software process that owns the workbook. It opens the file, reads the data, runs the analysis, writes the formulas or the new sheets, saves the file, and reports back. You give it a goal, not a script. It can loop, ask follow-up questions, and act on the file multiple times in a row. This is the new shape.

This article is about the third category. The first two are useful; the third is what you want in 2026, because the third is the one that can run the weekly reconciliation at 3am without you being there.

The 7 tasks that are worth the bet

We picked these seven by combining three filters: they are routine (every business does them), they are high-leverage (hours saved per week, not minutes), and they are tractable today (off-the-shelf tools can do them without a custom build). If you only adopt one AI agent in your company this quarter, pick the task that hurts the most on this list and start there.

The seven, in order of how often they come up in real businesses:

1. Formulas without knowing formulas. Lookup, match, regex, conditional logic, multi-criteria sumifs — described in English, written in spreadsheet.
2. Data cleaning and normalisation. Trim whitespace, fix dates, normalise phone numbers, dedupe rows, fill blanks, kill duplicates.
3. Reconciliation. Match rows in two spreadsheets, find what is in one but not the other, flag mismatches, produce a diff sheet.
4. Reports from raw tables. Turn a transaction log into a monthly P&L, a customer cohort table, a sales pipeline summary.
5. Pivot on demand. Group-by aggregations, multi-dimensional summaries, top-N reports — without dragging fields into the pivot table UI.
6. Mail-merge and personalisation. Pull rows from a sheet into a templated email, document, or message — at scale.
7. Anomaly and quality checks. Find outliers, detect duplicate customers, spot impossible values, audit the workbook before it reaches the CFO.

We will go through each one in depth.

Task 1: Formulas without knowing formulas

This is the lowest-risk entry point, and it is the one that sells the rest of the list to a skeptical colleague. The pattern: describe what you want in English, get a working formula back.

The classic: VLOOKUP / XLOOKUP / INDEX-MATCH

You have a list of customer IDs in column A, and you want the customer name from a second sheet. You do not remember the exact syntax. You do not need to.

Prompt you give the agent:

I have customer IDs in Orders!A2:A5000. I need the customer name from Customers!A:B (ID in A, name in B). Write a formula for column B of the Orders sheet that pulls the name. Use XLOOKUP. If the customer is not found, return "MISSING".

What the agent writes into Orders!B2:

=IFERROR(XLOOKUP(A2, Customers!A:A, Customers!B:B, "MISSING"), "MISSING")
          

The agent writes this formula into every cell from B2 to B5000 in one shot. You did not have to remember the argument order. You did not have to remember that IFERROR is a separate function. The agent did it.

The real win: multi-criteria and conditional logic

The boring-but-cumulative use case is conditional aggregation. You want the total sales for Region = West and Product = Pro Plan and Month = June 2026. The formula for that is genuinely hard to remember. The English is not.

Prompt:

Sum the Amount column (Orders!H) where Region = "West" AND Product = "Pro Plan" AND Date is in June 2026.

What the agent writes:

=SUMIFS(Orders!H:H, Orders!C:C, "West", Orders!D:D, "Pro Plan", Orders!E:E, ">="&DATE(2026,6,1), Orders!E:E, "<"&DATE(2026,7,1))
          

You just described the question in five seconds. The agent wrote a six-argument SUMIFS that you would have spent ten minutes on Stack Overflow for. The compounding effect, across an operations team that writes five of these a day, is the actual ROI.

What to be careful about

• Always test on a copy. The agent is good, not perfect. Have the agent run its own test cases against a sample of the data before you commit the formula to the live file.
• Use absolute references where they belong. If the formula is going into one cell and being copied down, ranges should be locked with $ so they do not drift. Tell the agent explicitly if you are copying down vs writing a single dynamic array.
• Demand the agent explain the formula in one sentence. A good agent will. If it cannot, the formula is too clever to ship without a second pair of eyes.

This task alone — a working professional who can describe a question in English and get a tested formula back — is worth the price of admission for everything else on this list. It is where you should start.

Task 2: Data cleaning and normalisation

The single biggest time-sink in any operations team is cleaning data that came in from somewhere else. The CRM exports phone numbers in five different formats. The accounting system exports dates as strings. The marketing platform double-counts rows. The agent eats this work for breakfast.

Trim, case, and dedupe

Prompt:

Sheet "Raw Leads" has 12,000 rows. Phone is in column F. Normalise it to E.164 format with a plus sign and country code 1 for US numbers. Trim leading/trailing whitespace. Convert "NULL", "n/a", "none", and empty strings to actual blanks. Remove exact-duplicate rows. Do not change any other column. Show me a 10-row before/after sample before you commit.

A good agent reads the sheet, makes a copy (Raw Leads cleaned), applies the rules, shows you the diff, and only then commits. The rules are:

• =TRIM(SUBSTITUTE(SUBSTITUTE(SUBSTITUTE(F2, CHAR(160), " "), " ", " "), " ", " ")) for the trim
• A small lookup that maps "NULL", "n/a", "none" to blank
• A phone normalisation that strips everything but digits, then re-adds +1 if the result is 10 digits

The point is not the formulas. The point is that you described the rules once, in English, and the agent applied them to all 12,000 rows consistently. A human would have given up at row 800.

Date normalisation

Dates are the worst offender. You have a column where some rows are 2026-06-15, some are 15/06/2026, some are June 15, 2026, and three are 6/15/26. The agent writes the right DATEVALUE plus a IFERROR fallback for each format, applies it, and reports the rows it could not parse (so you can fix them by hand, or so the agent can ask a follow-up question).

What an agent can do that a feature cannot

Microsoft Copilot will not run a 12-step normalisation on 12,000 rows in one pass. A standalone chat session will not write the cleaned file back to disk. An agent does both, and it does it in one task. The agent can also iterate: it can show you the diff, you can say "actually keep the original phone in a separate column for audit", and the agent adds the column and re-runs the clean. You are co-piloting the data, not running it by hand.

Task 3: Reconciliation

Reconciliation is the task that makes CFOs smile. The pattern: you have two spreadsheets (or two sheets, or a sheet and a database export) and you need to know what matches, what does not, and what is different.

Two sheets: customer list vs subscription list

Prompt:

Sheet "Customers" (5,000 rows) has customer_id in column A and name in column B. Sheet "Subscriptions" (4,800 rows) has customer_id in column A and plan in column B. Produce a sheet called "Recon" with three blocks: (1) IDs in Customers but not in Subscriptions, (2) IDs in Subscriptions but not in Customers, (3) IDs in both but with different plan values when joined with another sheet "Plan Catalog" that has plan in A and expected_plan in B. Show counts at the top.

This is a 30-minute job for a human. The agent does it in 90 seconds. The formulas it writes use XLOOKUP, COUNTIFS, and FILTER — but you do not need to know that, because the agent wrote them. You got back a sheet with three clean tables and a summary at the top. Your audit committee is happy.

Three sheets: the nightmare case

The more interesting reconciliation is the three-way match. You have an invoice, a purchase order, and a goods-received note. The agent reads all three, matches by PO number + line item, and reports:

• Invoices with no matching PO (potential fraud, or vendor billing wrong PO)
• POs with no matching invoice (vendor hasn't billed, you owe money or you need to chase)
• Quantity mismatches (invoice says 100, GRN says 98, PO says 100)
• Price mismatches (invoice at $4.50, PO at $4.20)
• Date mismatches (invoice before GRN — impossible, flag it)

This is the kind of work that an internal audit team would normally take a week to do on a quarter-end sample. The agent does the full population in minutes, and you spend your week on the exceptions, not the scanning.

Why this is the highest-ROI task on the list

Reconciliation is high-leverage because the cost of getting it wrong is asymmetric. If a reconciliation misses a $10,000 duplicate vendor payment, that is real money. If it flags a $40 mismatch that turns out to be a rounding issue, that is a five-minute follow-up. The agent is far cheaper to run on the full population than a human, and it never gets tired at row 4,000. The asymmetry favours the agent.

Task 4: Reports from raw tables

The finance team closes the month by turning a transactions spreadsheet into a P&L. The marketing team builds a weekly pipeline report by hand. The sales ops team joins three CSVs into a single dashboard. All of these are report-generation tasks, and all of them are agentic.

The pattern: rows in, summary out

Prompt:

Sheet "Transactions" has 30,000 rows. Columns: Date, Customer, Region, Product, Amount, Cost. Build a new sheet "Monthly P&L" with one row per (Month, Region) and columns for Revenue, COGS, Gross Profit, Gross Margin %. Use formulas, not hard-coded values, so it updates when new transactions arrive. Add a "Top 10 Customers by Revenue" block below.

The agent produces a sheet with SUMIFS formulas grouped by month and region, plus a LARGE + INDEX/MATCH lookup for the top 10. The whole thing is built on formulas that update live. You can hand this to a finance team member who does not know the agent is involved, and they will see "live P&L" and not realise it was an agent that built it.

The deeper move: cohort analysis and retention tables

A SaaS company wants to see how the March 2026 cohort is retaining. The agent takes the customer-events log, computes the cohort table (rows are cohort months, columns are months-since-signup, cells are % of original cohort still active), and writes a small heatmap of conditional formatting. The output is the kind of thing that would normally take an analyst a day to build in Looker. The agent builds it in five minutes, and the formulas are yours to inspect.

What the agent can do that Looker cannot

Looker, Tableau, and Power BI are great at the recurring report. They are bad at the one-off report that you need by Friday. An Excel agent is the opposite: it is best at the one-off, and it is not the right tool for the always-on dashboard. The two are complementary. The agent builds the one-off, and the recurring one gets promoted to a real BI tool once it has earned its keep.

Task 5: Pivot on demand

Pivot tables are powerful. Pivot tables are also the single most common cause of "I'll just do this in Python" in the operations world. The agent removes the UI friction.

Group-by with no pivot table

Prompt:

Sheet "Sales" has columns: Date, Rep, Product, Quantity, Unit Price, Region. Build a new sheet "By Rep by Product" that shows total revenue (Quantity * Unit Price) for each (Rep, Product) combination. Rows are Reps, columns are Products, cells are totals. Sort Reps by grand-total descending. Add a totals row at the bottom and a totals column on the right.

The agent writes a SUMPRODUCT (or a SUMIFS with two criteria) per cell, or it produces a single dynamic-array formula that spills. The output is a clean pivot-style summary, but it was built from English. No drag-and-drop. No "let me click Field Settings". No "why is Grand Total showing as 0".

The unsung hero: top-N reports

You want the top 20 customers by revenue, but only those in the East region, and you want them ranked by their Q2 numbers specifically. The agent does the filter, the sort, and the rank in one pass. It also has the good sense to ask whether you want the top-N within the region (i.e., the top 20 East-region customers) or the top-N overall that happen to be in East (i.e., the top 20 customers globally, filtered to East). The clarifying question is the difference between a useful agent and a frustrating one.

What to keep in the agent's hands vs your own

If you are doing a recurring pivot (the same dimensions every week), build a real PivotTable or promote it to a BI tool. The agent is for the ad-hoc pivot. The one where you need to slice by a new dimension, or compare two cohorts, or answer a "what if" question. The agent's superpower is the ad-hoc. Lean into that.

Task 6: Mail-merge and personalisation

Mail-merge is the oldest workflow on this list — it predates the web — and it is still one of the highest-volume uses of Excel. The agent makes it less painful and a lot more flexible.

The pattern: rows in, messages out

Prompt:

Sheet "Outreach" (200 rows) has columns: first_name, company, role, last_touch, talking_point. Sheet "Template" has an email body with placeholders {{first_name}}, {{company}}, {{talking_point}}. Produce a sheet "Outreach - Drafted" with one row per prospect and a column "email_body" containing the merged text. Use a formula or a script — your call, whichever is more robust. Do not actually send the email; just produce the drafts.

A simple version is a formula chain that does SUBSTITUTE three times. A more robust version is a small script the agent writes into the file (using Office Scripts or a small VBA) that loops through the rows. The agent picks the right approach for the size of the list and the complexity of the template. For 200 rows, formulas are fine. For 50,000, the agent will reach for a script.

Beyond email: documents, contracts, proposals

The same pattern works for personalised PDFs (contracts, proposals, certificates) and for templated Slack messages. The agent does the merge, but it also does the sanity check: it flags rows where talking_point is empty (the placeholder would render as talking about ), or where first_name is "TEST" (the QA row that should not go out). A good agent does not just do the merge; it does the QA pass on the merge.

The thing to remember about mail-merge

Mail-merge in 2026 is not the feature. The feature is the loop. The agent can draft 200 personalised emails, ask you to review 5 of them, edit the template based on your feedback, and re-generate the other 195. That loop — draft, review, refine, regenerate — is the part that turns a mail-merge tool into an agent. Use it.

Task 7: Anomaly and data-quality checks

The last task is the one that prevents disasters. Every workbook that reaches a decision-maker has at least one subtle error. The agent's job is to find them before the decision is made.

The classic checks

Prompt:

Audit the file "Q2 Forecast.xlsx". Look for: (1) rows where the same Customer ID appears more than once, (2) cells in the Revenue column that are > 3 standard deviations from the mean for the customer's segment, (3) rows where the Close Date is before the Create Date, (4) rows where Probability is empty or > 100% or < 0, (5) references to other sheets that no longer exist. Produce an "Audit" sheet with the offending rows and a one-line description of the issue for each.

The agent reads every sheet, runs the checks, and produces a clean audit list. The list goes to the analyst, who fixes 8 of the 11 issues, marks the other 3 as "intentional" with a note, and the file is clean. Total time: 10 minutes for the agent, 30 minutes for the analyst, 0 embarrassing moments in the board meeting.

The deeper checks: cross-file consistency

A subtler audit is cross-file. You have a forecast file, an actuals file, and a budget file. The agent checks that the customer list is the same in all three, that the date formats match, that the region codes are consistent (no EMEA in one file and Europe in another), and that the totals reconcile. This is the kind of audit that an analyst would do at quarter-end if they had time. The agent does it on demand, every time the file changes.

The ROI on audits

Audits have an asymmetric payoff. A good audit catches a $50,000 error and costs $50 of compute. A missed audit lets the $50,000 error reach the board. The agent makes "always run an audit before the file ships" cheap enough to be the default. That single behaviour change is worth more than the rest of the list combined.

How the agent actually sees your file

A short technical aside, because the right mental model is what separates a frustrating rollout from a smooth one.

In 2026, there are four common shapes for an Excel agent:

1. The in-Excel feature. Microsoft Copilot, Google Gemini in Sheets, WPS AI. The data does not leave the file. The agent is bounded by what the host product exposes. Best for ad-hoc, single-file, low-stakes work.
2. The cloud agent with file-upload. ChatGPT, Claude, Gemini, Perplexity. You upload the file (or paste a sample), the model reasons over it in a sandbox, you copy the output back. The file lives on the model provider's servers during the session. Fastest to start. Bounded by the upload size and the model's context window.
3. The dedicated agent with persistent file access. GolemWorkers, Manus, Lindy, Relay, n8n-with-an-LLM. The agent has a workspace, can save files, run scripts, and act on the file over time. Best for recurring work — the agent can run a weekly reconciliation because the file lives in its workspace.
4. The custom-built agent. Your team wires Claude or GPT to your own data warehouse via API, builds a small UI, and ships it as an internal tool. The most powerful and the most expensive. The right answer for regulated industries or for workflows that have to live inside your VPC.

For the seven tasks in this article, shapes 1 and 2 are enough to get started. Shape 3 is what you want once a task becomes a weekly ritual. Shape 4 is the right answer for the 2-3 tasks that become the backbone of your operations.

If you are using a cloud agent (shape 2 or 3), two non-obvious tips:

• Give the agent the schema, not the data. A short description of the columns and a 10-row sample is faster and more accurate than dumping 50,000 rows into the prompt. The agent can read the file directly.
• Always test on a copy first. A good agent will make a copy of the file before mutating it. If yours does not, ask it to. A 30-second insurance policy against an hour of "oh no".

Tools worth trying in 2026

You do not need to commit before you know which tool fits your team. A short list of the ones we have seen ship the seven tasks above in production.

• Microsoft Copilot for Excel. If your shop is on Microsoft 365 and your data is not regulated to the point of "no data leaves the tenant", Copilot is the lowest-friction path. It is shape 1, so it is bounded, but for tasks 1-3 it is genuinely fast.
• ChatGPT, Claude, Gemini. Shape 2. Upload a file, ask the question, get the answer. For one-off tasks and prototyping, the big three are interchangeable. Pick the one your team is already paying for.
• GolemWorkers (us). Shape 3. The agent has a persistent workspace, can read and write the same file across runs, and can be scheduled to run the reconciliation every Monday at 3am. The right answer when a task becomes a weekly ritual.
• Manus, Lindy, Relay, n8n. Shape 3 with different opinions about UI, scheduling, and integrations. Worth a trial if you are building a small ops automation layer.
• Custom agent on top of Claude or GPT. Shape 4. If you are at the scale where a vendor tool will not do (regulated data, custom workflow, internal-only), build it. Claude and GPT both have the file-handling primitives you need.

For each of the seven tasks, the same prompt works in any of the tools. Pick the tool based on where the file lives and how often the task runs, not on the model's brand.

How to start today: a 30-minute first run

A pragmatic first session, designed to ship a working result before lunch.

1. Pick the highest-pain task from the list. Do not start with the easiest. Start with the one that costs you the most hours. The agent's job is to remove pain, not to demo well.
2. Pick a tool you already have access to. If you have Microsoft 365, use Copilot. If you have ChatGPT Team, use ChatGPT. Do not start a procurement cycle.
3. Make a copy of the file. Q2 Forecast — TEST.xlsx. The agent will work on the copy. If anything goes wrong, the original is untouched.
4. Write a 4-sentence prompt. Describe the file, the task, the output format, and the safety constraints. "This is a test. Read sheet X. Produce sheet Y. Do not change any other sheet." A good prompt fits on one screen.
5. Review the output on a 10-row sample. Do not trust the agent on the first run. Verify the first 10 rows, the last 10 rows, and a random 10 rows in the middle. If they all look right, commit.
6. Commit the output, save the prompt. The prompt is your new SOP. Save it in a prompts/ folder next to the file. The next time you run the task, you are not starting from scratch.

That is 30 minutes. At the end, you have one task removed from your week. You can run the same pattern on task #2 next week. By the end of the quarter, you have six of the seven on the list automated. The seventh — anomaly detection — is the one you bolt on as a pre-flight check on the other six.

FAQ

What is the best AI agent for Excel in 2026? It depends on where the file lives and how often the task runs. For one-off work in a Microsoft 365 shop, Copilot is the lowest-friction option. For recurring work, a workspace-based agent like GolemWorkers pays for itself inside a month. For regulated or VPC-bound setups, a custom agent on top of Claude or GPT is the right answer. There is no single "best" — there is only "best for this task, this team, this compliance posture".

Can AI read Excel files? Yes. Modern agents read .xlsx natively. The model can also see formatting, comments, defined names, and sheet structure — not just the cell values. The bottleneck is rarely the read; it is the write. Make sure the tool you pick can save the file back, not just emit a CSV you have to copy-paste.

Is there an AI agent that opens Excel and edits it for me? Yes. The dedicated agent category (GolemWorkers, Manus, Relay, Lindy, n8n-with-an-LLM) does this. The agent opens the workbook in a workspace, applies the change, and saves the file. You review the result, not the steps. Microsoft Copilot edits the workbook too, but bounded by the features Microsoft chose to expose.

How does an AI agent handle very large Excel files? It depends on the architecture. A shape 2 cloud agent has a context-window limit — once you cross it, the agent has to read the file in chunks, which is slower and less accurate. A shape 3 workspace agent (GolemWorkers, Manus) usually has a code-execution sandbox that can stream the file, so size matters less. For files over ~100 MB, prefer a workspace-based agent with code execution over a pure LLM context.

Will AI replace the Excel power user on my team? No. The Excel power user is the person who knows which question to ask. The agent is the person who knows how to write the formula once the question is asked. The two are complementary. The power user becomes more valuable, not less, because their leverage per hour goes up by 5-10x.

Is my data safe when I upload it to an AI agent? Read the data-handling terms of whichever tool you pick. Microsoft Copilot inherits Microsoft 365's data terms (in-tenant by default). ChatGPT Team and Claude Team have enterprise data terms with explicit no-training clauses. For regulated workloads (healthcare, finance, government), use a self-hosted or VPC-bound option. Never upload a file containing PII to a free-tier consumer product without reading the privacy policy.

Can an AI agent write VBA macros? Yes. A capable agent writes, explains, and tests VBA macros. It can also refactor an existing macro, port a macro between Excel and Google Sheets, and convert a macro into a formula-based approach when that is the right answer. The agent is, in practice, a senior VBA developer on call.

Can I use an AI agent on Excel files in SharePoint or OneDrive? Yes. The shape 3 agents (GolemWorkers, Manus, Relay) integrate with SharePoint, OneDrive, Google Drive, and Box. The agent reads from the cloud location, writes back to it, and preserves the version history. This is the right pattern for team workflows.

Do I still need to know Excel to use an AI agent? Less than you think, but yes. The agent removes the formula-memorisation problem. It does not remove the question-formulation problem. If you do not know that the right question is "sum revenue by region for June", the agent will not think to ask. The good news: knowing what question to ask is the part that is hardest to outsource. The agent makes the rest of it cheap.

What is the difference between an Excel agent and Microsoft Copilot? Copilot is an in-Excel feature — a button you click. An Excel agent is a separate process that owns the workbook. Copilot will suggest a formula; the agent will write it into 5,000 cells, run the test, show you the diff, and save the file. Copilot is useful. The agent is more useful, and the two are not mutually exclusive.

Can I use an AI agent for Excel on Mac? Yes. All the agents in this article run on Mac, Windows, and Linux. The agent's workspace is a server, not your local machine. The file is uploaded to the agent's workspace, processed there, and the output is downloaded back to your Mac. Local Excel is not involved.

How long does it take to set up an Excel agent for a small team? An afternoon for the first task. A week for a full rollout to a 10-person operations team. The first task takes an afternoon because the prompt is the hard part, and you write it once. The next nine tasks take 30 minutes each, because the pattern is the same.

Conclusion

Excel was the first programming environment for business. The AI agent is the second. The seven tasks in this guide are the place to start because they are routine, high-leverage, and tractable today. Pick the one that hurts the most, run a 30-minute first session, and add the next one next week. By the end of the quarter, you have moved the most expensive hour of your week off your calendar and onto an agent that does not get tired at row 4,000.

If you want a concrete starting point, the GolemWorkers team has published a ready-made Excel agent and a starter spreadsheet you can drop into your workflow. The agent runs in your workspace, reads the file, performs any of the seven tasks, and saves the result. The link is in the description.