Automation & Workflows

AI Tools · Preconstruction & Construction Finance

Claude in Excel for Construction

What the Excel add-in is actually good for on a contractor's spreadsheets — and the verification rules that have to come with it.

Every contractor runs on two systems. There’s the ERP — Procore, Sage, Vista, CMiC — which is the system of record. And there’s Excel, which is where the actual thinking happens: the bid tab, the cost-to-complete, the WIP schedule someone rebuilds by hand every month because the built-in report doesn’t answer the question.

Claude in Excel is a sidebar that sits inside Excel and reads the workbook you have open. It doesn’t replace the ERP and it doesn’t connect to it. It works on what comes out of the ERP. That’s a narrower promise than the marketing suggests, and it’s also exactly where a construction team’s worst hours go.

This guide covers what it’s genuinely good at on construction spreadsheets, what it must never touch, and the verification discipline that separates the two. One thing up front, because it matters: there is no published case study of this add-in deployed at a commercial GC. The capabilities below are documented by Anthropic and the workflows are real, but the pairing is well-matched inference, not proven practice. Pilot it like you’d pilot anything else that touches a number going to a bank.

At a Glance

It’s generally available and it’s $20. Claude for Excel went GA on May 7, 2026, on Pro, Max, Team, and Enterprise. Pro is $20/month ($17 billed annually). The Free plan doesn’t include it.
You need a Microsoft 365 subscription. Excel on the web, Excel for Windows (build 16.0.13127.20296+), or Excel for Mac (16.46+). Excel 2016 and 2019 perpetual or volume licenses do not work. Neither does iPad or Android.
It reads your open workbook and cites cells. Every answer points at the cells it used, and you click through to check. That citation is the whole verification story.
It’s strongest on explaining and auditing, not authoring. Tracing a formula, finding the source of a #REF!, cleaning a messy takeoff export, normalizing a bid tab. Weakest on anything that ends in a signature block.
The governance gap is real. Claude for Excel activity does not appear in Enterprise audit logs or the Compliance API, and it doesn’t inherit your organization’s data-retention settings. Your IT director will ask. Have the answer ready.
Personal Pro seats are for evaluating, not deploying. Free, Pro, and Max run on consumer terms. Team and Enterprise run on commercial terms and are the only tiers inside Anthropic’s SOC 2 / ISO / HIPAA scope.

A bid leveling sheet is an externally-authored file. So is a sub’s quote, and an owner-supplied spec. That’s the one category Anthropic explicitly warns you not to hand it.

Three Ways to Put Claude on a Spreadsheet

Before the use cases, the choice that decides everything else. There are three ways to get Claude and a workbook in the same room, and most people only know the first one.

1. Upload the file to a chat. The obvious move, and the most expensive one. An .xlsx carries a lot of structural noise, and a big multi-tab workbook will burn through context fast — sometimes before it has read the tab you cared about. If you go this route, orient it first: a screenshot of the header row, or a short instruction document explaining what the tabs and column codes mean. Same as you’d do for a new hire who’s never seen the file.

2. Bring Claude into the workbook. This is the add-in, and it’s the mode this guide is about. Claude sits in a sidebar inside Excel, reads the open workbook directly, and cites the cells it used. No upload, no token cost for the file itself, no reformatting. If you’ve been told Excel files are a poor thing to hand an AI — true for uploads, and not what’s happening here. Different mode, different rules.

3. Have it build an artifact. For recurring analysis, don’t ask for a spreadsheet at all. Ask for a small interactive tool that reads your data and produces the view you need — one you can share with colleagues and re-run next month. Think of it as a pivot table that updates itself and doesn’t need you to rebuild it. Best for the report you regenerate every month for the same audience.

The right answer depends on what happens after. If the output has to leave your hands as a spreadsheet, work in the spreadsheet. If it’s analysis for you, upload or artifact may be faster. If it’s the same report every month for the same people, build the artifact once.

Before You Install: What IT Will Ask

Three questions, and you should have all three answers before you walk into that conversation.

“What version of Excel does it need?” A Microsoft 365 subscription — not a perpetual license with a big year on it. Excel 2016 and 2019 perpetual/volume are explicitly unsupported, as are iPad, Android, and older M365 builds below the SharedRuntime threshold. This is the single most common reason someone reports “it doesn’t work.”

“How do we deploy it?” Microsoft 365 Admin Center → Settings → Org Settings → User owned apps and services → turn on “Let users access the Office Store.” Then Settings → Integrated apps → Add-ins → search “Claude for Microsoft 365” → assign to the org or to specific groups. If the Office Store is locked down entirely, there’s a custom manifest XML path instead.

One gotcha worth naming, because it makes you sound like you’ve done this before: if your organization uses Microsoft Entra Privileged Identity Management for admin roles, the Integrated apps page doesn’t recognize PIM-activated roles and the deployment silently fails. It’s a known Microsoft issue, tracking ID 11126536. Deploy from an account with the role assigned as permanently active instead. Individual users can still self-install in the meantime.

“Where does our data go, and can we audit it?” Inputs and outputs are deleted on Anthropic’s backend within 30 days. Chat history is stored locally in your browser, last 50 conversations per app, not synced, not on Anthropic’s servers. Team and Enterprise plans are excluded from model training under commercial terms.

And then the hard part: activity is not included in Enterprise audit logs, the Compliance API, or data exports, and the add-in does not inherit custom retention settings. If you need an audit trail, someone has to stand up a custom OpenTelemetry collector — and that collector receives spans unfiltered, including prompt content and document references, so it becomes a sensitive endpoint in its own right. Don’t discover this in month three.

The Workbooks Worth Pointing It At

These map onto what the tool actually does well: multi-tab reasoning, formula tracing, template population, and cleaning messy data.

Bid leveling. The strongest fit in preconstruction. Normalizing a dozen sub quotes into comparable columns, flagging the scope lines nobody has priced yet, catching the bidder who excluded something everyone else included. Cell citations make the leveling defensible when someone asks why a number moved. Most leveling sheets are old, formula-heavy, and rebuilt every bid — exactly the kind of workbook that rewards having something read it for you.

Schedule of values. Building an SOV against a contract value, checking that line items sum to the contract, sanity-checking front-loaded early activities.

Monthly forecasting — ETC, cost-to-complete, EAC. Variance narration and outlier flagging across a portfolio of jobs. “Which cost codes are trending more than 10% over budget, and by how much” is a question Excel answers slowly and Claude answers immediately, with the cells to check. If your forecast cycle spans a job cost system, a payroll system, and a stack of Excel exports, the reconciliation between them is the part worth handing over.

The compliance chase. Every project controls group has a version of this: export the report, compare it against a master list, color-code who’s late, send the reminder. It’s monthly, it’s manual, and it’s deterministic — which is precisely the shape of work to automate first. Not glamorous. Highest hourly return in this guide.

Change order logs. Reconciling the log against the SOV and the current contract value, and keeping potential changes and executed owner changes straight. Finding the one that’s in a system and not on the log is a genuinely tedious hour that becomes a few minutes.

Labor productivity and unit costs. Comparing installed quantities against budgeted hours across crews or jobs. This is the “clean the messy data, then build the comparison” task the tool is most reliably good at.

Quantity takeoff cleanup. Unglamorous and the highest-confidence use case there is. Standardizing units, splitting merged columns, reconciling naming across sources.

Cash flow curves. Building an S-curve off a schedule of values and a duration, then flexing the assumptions. It updates values while keeping formula relationships intact, so the model still works after it edits.

Procurement and submittal logs. Status rollups, aging analysis, “what’s due in the next 30 days and who owns it.”

The Workbooks To Keep It Away From

Anthropic publishes a “not recommended for” list, and one line on it should stop a construction team cold: audit-critical calculations without verification.

For a contractor, “audit-critical” isn’t hypothetical. It’s most of the finance department.

AIA G702/G703 pay applications. Use it to check — does the G703 foot to the G702, is retainage consistent, do stored materials reconcile. Do not use it to author the document that goes to the owner.
WIP schedules going to a surety or a lender. Percentage-of-completion revenue, over/under billings, the work-in-place profit you agree to claim on a job. Checking the logic is fair game. Producing the schedule your bonding capacity rests on is not.
Certified payroll. Davis-Bacon and prevailing wage filings are legal documents with penalties attached.
Anything with a signature block. If a person signs it, a person builds it.
Final owner or client deliverables without human review. Anthropic’s words, not mine.

The rule that covers all of it: Claude can audit the number. A human owns the number.

Prompts That Work on Construction Spreadsheets

Copy these. They’re written to lead with explanation before edits, which is the order that keeps you out of trouble.

Understanding an inherited model

“Walk me through how the projected final cost in cell H42 is calculated. List every input cell it depends on.”
“What assumptions drive the cost-to-complete on this job? Cite the cells.”
“This workbook came from someone who left. Summarize what each tab does and how they connect.”

Bid leveling

“Compare these four subcontractor quotes. Build a leveling table with a row per scope item and a column per bidder, and flag any item where a bidder has no price.”
“Which of these bidders excluded something the others included? List the exclusions by bidder with the cell reference.”
“Normalize these quotes to the same unit basis and show me the adjusted low bidder.”

Job cost review

“List every cost code where actual-to-date exceeds the budget by more than 10%, sorted by dollar variance.”
“Add conditional formatting that turns any cost code over budget red.”
“Build a summary tab showing total budget, committed, actual, and projected final cost by division.”

Checking, not producing

“Does the total on the G703 continuation sheet match the G702 cover? Show me any line that doesn’t reconcile.”
“Check that retainage is calculated at 5% consistently across every line on this pay app, and flag exceptions.”
“In this WIP schedule, verify that earned revenue equals contract value times percent complete for every job, and list any that don’t.”

Cleaning

“Standardize the cost code format in column B to XX-XXX and flag anything that doesn’t fit the pattern.”
“This takeoff export has quantities in mixed units. Convert everything to square feet and show your conversion factors.”
“Find duplicate line items in this submittal log and flag them.”

The Subcontractor Bid File Problem

This is the section to read twice, because it’s the risk that’s specific to how construction actually works.

Anthropic ships an explicit warning: only use Claude for Excel with trusted spreadsheets. Files from outside your organization can carry hidden instructions — buried in cells, formulas, or comments — that manipulate the add-in into extracting data, modifying records, or taking destructive actions. Anthropic’s own testing confirmed those scenarios are possible.

Now look at a preconstruction inbox. Subcontractor bid spreadsheets. Vendor quote templates. Owner-supplied schedules of values. Supplier takeoffs. Manufacturer cut sheets with embedded tables. A general contractor’s Excel workflow is made of externally-authored files. That’s precisely the threat surface Anthropic is describing.

You don’t have to stop leveling bids. You have to change one habit:

Paste the data into a workbook you control. Don’t point Claude at the vendor’s file.

Open the sub’s file yourself, copy the numbers into your own bid tab, and let Claude work there. It’s the same move you’d make with a macro-enabled attachment from an address you don’t recognize, and it costs you thirty seconds.

When Claude does propose something risky, it asks you to confirm first. Read those confirmations rather than clicking through them — especially on anything that originated outside the building.

What It Can't Do

Set expectations before the room does it for you.

No macros or VBA. If your estimating workbook is held together by VBA, this isn’t the tool for that part.
No data tables. The Excel what-if feature specifically.
No ERP connection. It doesn’t talk to Procore, Sage, Vista, or CMiC. It works on the export.
Only open workbooks. It can’t create, open, close, or switch files on its own, and it can’t reach a linked workbook that isn’t open.
No scheduled or unattended runs. Everything happens with you sitting there.
Long sessions lose detail. Conversations get auto-compacted to avoid running out of context, and it doesn’t announce what it dropped. The reasoning behind a decision you made an hour ago may simply be gone. Capture decisions outside the chat.
No published size limits. Anthropic doesn’t publish a workbook size, row, or cell ceiling. Performance degrades on large models well before any hard wall. Test your own worst-case file rather than trusting a number.

Your Verification Checklist

Eight rules. Print them and put them next to the monitor.

Work on a copy. Anthropic’s own best practice: start from a trusted copy before asking it to edit widely.
Never point it at an unvetted external file. Paste the data into your workbook instead.
Ask before you tell. Explain and audit first, edit second.
Click every citation. If you don’t click through, you’ve given up the only verification affordance the tool offers.
Foot and cross-foot by hand. Every total, subtotal, and any number leaving the building.
Read the overwrite warning. It tells you before it overwrites. That warning is doing a job.
Never let it own the last mile. Pay apps, WIP to the surety, certified payroll, anything signed.
Capture decisions outside the chat. Compaction will eventually eat the reasoning. Write down why, not just what.

Where This Fits With Your ERP

Worth being clear about, because someone will ask whether this replaces something you already pay for. It doesn’t.

The ERP is the system of record. Job cost, commitments, AP/AR, payroll, contracts. That stays exactly where it is.

Excel is the system of analysis. Even with Procore in place, most finance teams still rebuild the WIP schedule in Excel every month, because built-in WIP reporting is thin. Sage 300 CRE in particular is known for pushing teams into Excel-heavy workflows.

Claude in Excel is a layer on the analysis, not the record. It makes the export faster to reason about. It competes with nothing you own, and it doesn’t integrate with anything you own either. That’s the honest shape of it.

Start Here This Week

Don’t roll it out. Pilot it on one workbook, with one person, on something that doesn’t matter yet.

Pick a finished job. A closed-out project’s cost report. Nothing live, nothing going anywhere.
Make a copy. Then open the copy.
Ask it to explain the workbook before you ask it to change anything. See whether the citations hold up when you click them.
Try one cleaning task. A takeoff export, a submittal log, something messy. This is where it will earn its keep first.
Then try one analysis task. Variance flagging on cost codes. Check its answer against your own.
Write down where it was wrong. It will be wrong somewhere. That list is worth more to the rest of the team than another success story.

Six steps, one afternoon, one seat. That’s the whole pilot.

And measure the right thing. The temptation is to track usage — how many people logged in, how many prompts ran. Usage is not value. Pick one number that means something, like hours returned per person per week, and hold the pilot to that instead.

Make It Yours

The teams that get value out of this aren’t the ones who trust it most. They’re the ones who built the checking habit first and then let it run.

Start with the spreadsheets nobody signs. Earn the trust one workbook at a time. And keep the rule that makes all of it safe: Claude can audit the number, but a human owns the number.

— Alec

Want the rest of the AI playbooks?

Every Friday I send one real workflow I'm running, broken down to the prompts and the receipts. No jargon, no hype.

Subscribe free