The 2026 AI Agent Stack for Deal Teams: What to Use and When

In 2025, the finance world used AI to chat. We pasted transcripts into ChatGPT, asked for summaries, and treated the technology like a very smart, very hallucination-prone librarian. In 2026, we use AI to work. The problem is no longer "can AI help me?" It is "I have nine agents on my desk and no idea which one to hand the task to."

This post maps the current landscape. Not every tool, but the ones that matter for deal teams in IB, PE, CRE, and consulting — what each one does, where it fits, and when to reach for it instead of the others.

The Three Layers of the Stack

The 2026 agent landscape for deal teams breaks down into three functional layers. Understanding this hierarchy is more useful than memorizing individual product features, because the layers tell you what kind of work you are delegating.

Layer 1: General-Purpose Agents. These are the workhorses — Claude Cowork, Microsoft Copilot Cowork, ChatGPT, and Google Gemini. They handle broad, multi-step tasks across file types and applications. Think of them as the generalist associates on your team: capable of almost anything, but not specialized in any one deliverable.

Layer 2: Specialized Execution Tools. These are narrower and deeper — Crunched for Excel modeling, Rogo for sell-side research and comps, Hebbia for large-scale document analysis. They do one category of work exceptionally well and integrate into the general agents' output.

Layer 3: Infrastructure. NemoClaw and OpenClaw sit here. They are not tools you interact with directly — they are the plumbing that IT departments use to deploy and secure agents across the enterprise. Most deal professionals will never touch this layer, but your firm's CISO is thinking about it constantly.

Layer 1: The General-Purpose Agents

Claude Cowork — The Desktop Analyst

Claude Cowork is a local-first desktop agent powered by Opus 4.6. It runs inside a secure virtual machine on your machine, mounts your local folders, and operates directly on the files sitting on your hard drive. When you tell Claude to "read the 20 PDFs in the Data Room folder and build a lease abstract in Excel," it does the work locally. Your files never leave your machine.

Best for: Deep, multi-file work sessions. Cross-referencing a folder of lease PDFs against a rent roll. Building a 3-statement model from raw financials. Cleaning and restructuring messy data room exports. Any task where the deliverable is an .xlsx, .docx, or .pptx file that needs to be production-quality.

The edge: File-system depth. Claude reads, writes, and edits actual Office files with working formulas and linked tabs. The 1M token context window (in beta) means it can hold an entire data room in memory at once. For deal teams handling confidential materials, the local-first architecture is often the deciding factor.

The gap: It lives on your desktop, not in your browser. If the work requires pulling live market data, browsing the web for comps, or coordinating across cloud-based apps, Claude has to work harder for it. And the learning curve for prompt engineering complex deliverables — getting the sensitivity table increments right, matching your firm's IC memo format — is real.

Microsoft Copilot Cowork — The Enterprise Backbone

Copilot Cowork is the Wave 3 evolution of Microsoft 365 Copilot. It is not a chatbot living in a sidebar anymore. It is an autonomous agent that plans and executes multi-step workflows across Word, Excel, PowerPoint, Outlook, and Teams — and now, through MCP support, across Salesforce, Adobe, and other third-party tools.

Best for: Cross-application orchestration in Microsoft-native firms. Preparing the quarterly vendor review by pulling data from Excel, drafting the summary in Word, building the deck in PowerPoint, and scheduling the meeting in Outlook — all from a single instruction. Managing email triage, meeting prep, and recurring reporting workflows.

The edge: If your firm runs on Microsoft 365, Copilot Cowork has the deepest integration by default. The "Work IQ" layer grounds its actions in your organization's actual hierarchy, project relationships, and collaboration patterns. The multi-model architecture (Claude + GPT under the hood) means it can select the right reasoning engine per task automatically.

The gap: Copilot Cowork is a cloud service. Your data flows through Microsoft's infrastructure. For firms with strict data residency requirements or air-gapped deal rooms, this is a non-starter. And while it is excellent at orchestrating across apps, its depth within any single app — particularly Excel — still lags behind tools that are purpose-built for modeling.

ChatGPT — The Research Librarian

ChatGPT, powered by GPT-5.4, remains the strongest general-purpose tool for web-connected research, visual content generation, and broad knowledge work. It lives in the browser and has the widest plugin and custom GPT ecosystem of any platform.

Best for: Market research, industry primers, and anything that requires synthesizing public information at speed. Drafting client-facing memos where tone and narrative matter more than formula precision. Generating VBA macros for legacy Excel workflows. Visual content — charts, infographics, presentation graphics — where aesthetics matter.

The edge: The web browsing and data analysis capabilities are unmatched. If you need to pull together a competitive landscape, summarize 30 earnings call transcripts, or draft a teaser for a sell-side process, ChatGPT gets to a usable first draft faster than anything else. The custom GPT ecosystem also means your firm can build and share specialized prompts at scale.

The gap: ChatGPT does not touch your file system. Everything goes through a browser upload-download cycle. The Excel files it produces are functional but often need formatting cleanup. For work that requires sustained, multi-file context — like cross-referencing a 50-document data room — it runs into context limitations faster than Claude.

Google Gemini — The Workspace Brain

Gemini is Google's cloud-native workspace agent, now running Gemini 3 Pro for complex reasoning and Gemini 3 Flash for real-time interaction. The standout feature is Workspace Studio, a no-code environment where business users build custom automation agents across Gmail, Drive, Sheets, Slides, and Meet.

Best for: Firms that run on Google Workspace. Research synthesis across large document sets (the context window scales to 10M tokens with NotebookLM). Cross-app automation — summarizing a week of Google Chat threads, pulling status updates from Drive, and generating a Slides deck, all on a Friday schedule. Video content creation through the Veo integration.

The edge: Context window size. Gemini can hold more raw material in memory than any other general agent, which makes it formidable for due diligence and research-heavy workflows. The A2A and MCP protocol support means it can coordinate with agents from other vendors — a Gemini agent can talk to a Salesforce agent to close a deal workflow.

The gap: Google Sheets is not Excel. For institutional finance, where the deliverable is a .xlsx with XLOOKUP chains, conditional formatting, and VBA modules, Sheets falls short. The Workspace ecosystem is also less common in finance than Microsoft — most bulge brackets and PE firms are Microsoft shops.

Layer 2: The Specialists

Crunched — The Speed Analyst

Crunched is an Excel-native AI add-in tuned specifically for financial modeling. It reads entire workbooks — not just the active sheet — and understands the interlinked logic of multi-tab models. It can build a 5-year three-statement model from a 10-K, audit an existing workbook for hidden errors, and map extracted PDF data directly into your firm's standardized templates.

When to reach for it: When the task is specifically about building or auditing a financial model in Excel. Crunched is not trying to be your general assistant. It does one thing — Excel modeling in finance — and it does it with a level of formula precision and auditability that general agents cannot match. If you need a fast first-draft LBO or DCF, or you need to QC a model before IC, this is the tool.

The honest caveat: Crunched produces fast first drafts, not finished products. The models need senior review, particularly around judgment-heavy assumptions like terminal multiples, WACC inputs, and working capital normalization. It also does not handle narrative work — it builds the numbers, not the memo around them.

Rogo — The Sell-Side Research Engine

Rogo is the dominant research platform for investment banks. It has the widest data partnership network in the space — S&P Capital IQ, LSEG, FactSet, PitchBook, Preqin, and as of March 2026, Fitch Solutions. Its Sheets Agent can build exportable comps tables and precedent transaction analyses with full source citations. The March 2026 acquisition of Offset added "learning agents" that live inside your spreadsheets and presentations.

When to reach for it: Pitchbook assembly, comp set generation, industry research, and any task that requires pulling structured data from multiple financial databases and presenting it in a client-ready format. If you are a sell-side analyst building an industry overview or a banker assembling CIM exhibits, Rogo's data integrations save hours of manual Capital IQ pulls.

The honest caveat: Rogo is a browser-based platform. The data and analysis live in Rogo's environment, not on your local machine. For firms with strict information barriers or data residency requirements, this matters. And while the data breadth is impressive, the modeling depth is shallower than Crunched — Rogo is better at populating exhibits than building linked financial models.

Hebbia — The Buy-Side Document Brain

Hebbia's Matrix platform is built for processing massive unstructured document sets — hundreds or thousands of pages of CIMs, 10-Ks, credit agreements, and VDR files. It serves 30% of the top 50 global asset managers. The core capability is multi-document reasoning: ask a question across a thousand pages and get a sourced, structured answer.

When to reach for it: Due diligence on the buy side. Screening a universe of 1,000+ public companies for specific covenant triggers. Analyzing a VDR with 500 documents to find every instance of a change-of-control provision. Monthly portfolio reporting where the input is a stack of disparate financial statements. Any workflow where the bottleneck is "reading a lot of documents and finding specific things."

The honest caveat: Hebbia is a research and extraction tool, not a production tool. It finds the data and structures it, but it does not produce the final deliverable — the IC memo, the model, the lender package. You still need a general agent or a modeling tool to turn Hebbia's output into the finished work product.

Manus AI — The Right Brain

Manus, acquired by Meta for roughly $2 billion in December 2025, is a cloud-native autonomous agent that excels at visual and creative work. Its "Design View" produces high-fidelity presentations that look like they were made by a graphic designer, not an AI. The March 2026 "My Computer" desktop app extends its reach to local file management.

When to reach for it: Consultant-grade slide decks where visual quality matters as much as content. Client-facing materials for pitches and presentations where the default PowerPoint templates look too generic. Web research and competitive analysis where you want the output delivered as a polished, visual report rather than raw text.

The honest caveat: Manus runs in the cloud. Your files are processed on Meta's infrastructure. For deal teams handling confidential transaction data, this is a hard stop. The financial modeling capabilities are also shallow compared to Claude or Crunched — Manus is a presentation and research tool, not an analyst.

Layer 3: The Infrastructure

NemoClaw and OpenClaw — The Plumbing

NemoClaw is NVIDIA's enterprise agent stack, announced at GTC 2026. OpenClaw is the open-source framework it wraps. Together, they allow IT departments to deploy secure, locally-hosted AI agents on NVIDIA hardware — DGX Stations, Mac Mini clusters, or on-premise servers.

Who cares about this layer: Your CISO, your CTO, and your IT infrastructure team. Deal professionals will not interact with NemoClaw directly. But if your firm wants to run AI agents on-premise with air-gapped security — no data leaving the building, full audit trails, and IT-controlled permissions — NemoClaw is the reference architecture for making that happen.

The honest caveat: NemoClaw is in early access as of March 2026. It is not production-ready for highly regulated environments. The Nemotron models that power it are competent but do not match Opus 4.6 or GPT-5.4 on reasoning benchmarks. This is a long-term infrastructure bet, not a tool you deploy Monday morning.

How to Think About the Stack

The mistake most firms make is treating this as an either/or decision. It is not. The stack is complementary.

A realistic 2026 workflow looks like this: Hebbia screens the VDR and extracts the key data points. Crunched builds the first-draft model in Excel. Claude Cowork cross-references the model against the source documents, flags discrepancies, and builds the sensitivity tables. Copilot Cowork assembles the IC memo across Word and PowerPoint, pulls in the relevant email threads, and schedules the committee meeting. ChatGPT drafts the market overview section and generates the industry charts.

No single tool does all of that well. The firms that are moving fastest in 2026 are the ones that stopped looking for one agent to rule them all and started assembling a stack where each tool handles the work it was built for.

The Gap That Remains

Here is the problem with every tool described above: they are all general-purpose within their category. Claude Cowork is the best general desktop agent. Copilot Cowork is the best general enterprise agent. Rogo is the best general research platform for finance. But none of them inherently knows how your firm formats an IC memo, what your acquisitions team expects in a stacking analysis, or that your lender package requires DSCR schedules with 25bps rate increments.

Every time you use these tools, you teach them. You write the prompt. You specify the columns. You describe the format. And next week, you do it again.

Purpose-built AI coworkers — like those from Lumetric — are designed to close that gap. They take the raw power of frontier models and make it opinionated about specific industries and deliverables. Not a general agent you configure for CRE work or PE workflows. A coworker that was built for the job, deployed as a specialized worker your team reaches by email. The stack gives you capability. Purpose-built gives you expertise.