The agents draft. You still sign.
On May 5, 2026, Anthropic released ten named finance-services agents -- ready-to-run templates for investment banking, equity research, fund administration, and finance operations -- plus Microsoft 365 add-ins for Excel, PowerPoint, Word, and Outlook and eight new data connectors. The agents are not thought pieces. They are in the field today.
The most important sentence in the entire repository is the disclaimer. From anthropics/financial-services, verbatim:
Nothing in this repository constitutes investment, legal, tax, or accounting advice. These agents draft analyst work product -- models, memos, research notes, reconciliations -- for review by a qualified professional ... every output is staged for human sign-off.
Read it as a moat statement, not a disclaimer. The CPA's license, judgment, and relationship are exactly what the agent will not touch. Posting the JE, distributing the statement, concluding the audit, signing the return -- those four verbs don't move. The work that surrounds them does.
Three agents change CPA work in 2026.
Of the ten, three sit on top of work CPAs already do every day, every period, every quarter. The other seven (Pitch Agent, Meeting Prep Agent, Market Researcher, Earnings Reviewer, Model Builder, Valuation Reviewer, KYC Screener) belong more to bankers, analysts, and compliance teams.
GL Reconciler
Daily and month-end. Finds breaks. Traces them. Stages a sign-ready exception report.
- Inputs
- Trade date · asset classes
- Outputs
- Break list · root-cause trace · exception report
- Where you come in
- You approve every reclass and direct the resolutions; the agent never posts.
Month-End Closer
Accruals with calculation references. Roll-forwards tied to GL. Variance commentary you can actually use.
- Inputs
- Entity · period (YYYY-MM)
- Outputs
- Accrual schedule (with JE drafts) · roll-forwards · variance commentary · close package
- Where you come in
- You sign off on the JE drafts before they hit the ledger and own the variance narrative for stakeholders.
Statement Auditor
Last set of eyes on LP / investor statements before they leave. Tie-out table + sign-off sheet, pass-or-hold per LP.
- Inputs
- Statement batch · NAV pack
- Outputs
- Tie-out table · exception list · sign-off sheet (pass/hold per LP)
- Where you come in
- You hold every "hold" recommendation against your independent re-check before distribution.
Three surfaces. One agent.
Same system prompt, same skills, same guardrails -- picked up in three different surfaces. The value comes from picking the right one for the job: the analyst at her desk, the controller in a chat, the ops lead with a queue of fifty closes.
Inside Microsoft 365
The agent stays in the tools you already work in.
Add-ins for Excel, PowerPoint, Word and Outlook (rolling out). You build the model in Excel; the same context carries to the deck in PowerPoint without re-explaining. Numbers change in the model, the chart on slide six updates. Your tax memo in Word references the same workbook the partner just signed off on.
In Cowork
The chat-and-workspace surface for interactive work.
A live workspace for one-off questions and multi-step jobs. This is where you spend a Tuesday morning closing one tricky entity with the Month-End Closer at your elbow, asking it to walk through an unusual accrual, then pulling the result into the close package.
As Managed Agents
Headless, scheduled, queue-driven -- at firm scale.
The same agent deployed via the Managed Agents API, behind your own workflow engine. It runs on cron, on a webhook, or on a queue -- closing fifty entities the morning of business-day-three, processing every onboarding packet that arrives, or re-running statement tie-outs for a whole fund family before quarterly distribution.
Source-attributed outputs.
Every figure traces back to the filing, contract, or GL row it came from. Not a footnote habit -- a structural feature. Peer review survives. Regulator questions get answered without an evening of reverse-engineering.
Customizable to your firm.
Risk thresholds, materiality, valuation policy, approval routes, documentation standards -- your standards, configurable. Connectors for D&B, FactSet, S&P Global, Moody's, IBISWorld, Intralinks, Verisk arrive in the box.
One canonical instance.
The same prompts and skills run on all three surfaces. When your firm tunes the Month-End Closer to your standard, the tuning shows up in the analyst's Excel, the controller's chat, and the scheduled night-job -- all at once.
Where the hours go.
Illustrative, not promised -- every shop has its own bottlenecks. But the shape is consistent in early-adopter funds and small CFO-services firms. The work below the line is what actually changes.
The hours don't disappear. They migrate up the value chain -- to the conversation with the client, the judgment call on the unusual entry, the second opinion on a hard accrual, the tax-planning move that just became visible because you finally have the time to see it.
The next eight quarters.
Anthropic's repository seeds an obvious roadmap. These are the workflow areas most likely to ship vendor-grade or open-source agentic templates by the dates listed. None of this is promised -- but every line below is already in active development somewhere in the field.
- Q3 2026
Fund-admin month-end goes agentic by default.
Most early-adopter fund admins compress month-end close 60-80%. CPA-led shops follow within two quarters once the playbook is public.
- Q4 2026
Tax-provision draft agents land.
Year-end provision drafts (current + deferred) become standard agent workflows. CPA owns assumptions, sign-off, and disclosure judgment.
- Q1 2027
Audit lead schedules and confirmations.
Sampling, lead schedules, confirmations, and tie-outs become agent-prepared. The CPA's testing, evidence evaluation, and opinion remain human.
- Q2 2027
Tax research and structuring assistants.
Code-and-regs research with citations becomes agentic. Structuring (entity choice, S-corp / C-corp / LLC, succession) stays advisor-led -- and gets faster because the research lookup vanishes.
- Q3 2027
Vertical-specific accounting agents.
Real estate (cost-seg + 1031), e-commerce (sales-tax nexus + COGS modeling), healthcare (ASC 842 + payer mix), restaurants, dental, SaaS -- each gets a focused agent. Vertical depth becomes the most valuable thing a CPA owns.
- Q4 2027
Advisory copilots for CFO-services.
Dashboards, KPI commentary, and forecasting copilots scale fractional-CFO work. The CPA holds the strategic conversation; the agent does the prep.
- 2028+
The relationship layer is the moat.
Compliance work compresses. Tax compresses. Audit compresses. What does not compress: knowing this client, this industry, this family, this exit, this audit trail. Vertical depth × relationship × judgment compounds.
Prep time has been transformed into idea time, with faster workflows, richer client insights, and new use cases.
quoted in Anthropic's launch post
Translate to a CPA: prep becomes review. Review becomes judgment. Judgment becomes advisory. The hours your firm sells don't disappear -- they price up.
Where CPA value concentrates.
Compresses.
Bookkeeping. Reconciliation. Month-end close. Tax-return prep. Confirmations. Lead schedules. Documentation. Variance commentary. The preparation half of every workflow.
Concentrates.
Vertical specialization. Tax structuring. Audit conclusions. Controllership judgment. Family-office relationships. Exit planning. Exception handling. The judgment half of every workflow.
Compounds.
The CPA who serves dental practices for fifteen years and now uses agents to do the prep can serve 3× the clients with the same hours -- and command a premium because the relationship and vertical depth aren't reproducible.
Six concrete moves.
You don't need a transformation initiative. You need one workflow, one client, and one afternoon. The point is not to deploy AI. The point is to find out where your judgment is most valuable when the prep is no longer the bottleneck.
- 01Pick one client. Ideally one you know cold -- small entity, predictable close, where you can compare AI output to your usual workpapers.
- 02Pick one workflow. Month-end close, daily recon, or LP tie-out. The three agents above are your shortlist.
- 03Run the agent end-to-end on real (or anonymized) data. Don't tune. Don't pre-clean. See what it produces.
- 04Compare to your usual output. Where is it better, where is it worse, where is it genuinely surprising? Note the surprises -- that's where the moat lives.
- 05Iterate the prompt and the workflow. The agent improves; your judgment about when to trust it improves faster.
- 06Roll out to two more clients in the same vertical. Vertical specificity accelerates returns. After the third client, you're not running an experiment anymore.
Ready to put this to work in your firm?
QWave helps firms move from "interesting" to "in production" -- picking the right agent, wiring it to your tools, and rolling it out across clients without breaking what already works.
Get In TouchThese agents draft analyst work product -- models, memos, recons. They do not constitute investment, legal, tax, or accounting advice. Every output is staged for review by a qualified professional. Briefing prepared 2026-05-08, drawing on Anthropic's May 5, 2026 launch.