Anthropic just published a breakdown this week of where AI agents are actually deployed across nearly 1 million real production tool calls. Software engineering sits at 49.7%. Sales and CRM is at 4.3%. Finance at 4.0%. Legal at 0.9%.
At first it might seem to say AI works for coders, hasn’t cracked the rest of the enterprise yet.
That is exactly the wrong conclusion.
AI agents work in sales. The data is already there. 86% of sales teams using AI report positive ROI within their first year. 83% of sales teams using AI experienced revenue growth, compared to 66% of non-AI teams — a 17-point performance gap. Companies that implement AI in sales report revenue increases ranging between 3% and 15%, along with a 10-20% boost in sales ROI.
AI agents work in finance. Insurance — one of the most data-complex, compliance-heavy environments in the enterprise — went from 8% full AI adoption in 2024 to 34% in 2025, a 325% increase in a single year.
The low adoption numbers in Anthropic’s chart don’t mean AI doesn’t work in these domains. They mean the data infrastructure wasn’t ready. And that’s a very different problem — with a very different timeline for resolution.
The Reason Software Won First Isn’t What You Think
It’s tempting to say developers got there first because they’re tech-savvy early adopters. That’s part of it. But the real reason is structural.
Think about what a coding agent actually needs to operate. The code is right there — every file, every function, every dependency sitting in the repository, perfectly structured, version-controlled, instantly accessible. No data governance committee. No integration project. No six-month IT procurement cycle. You open the repo, the agent has everything it needs.
The feedback loop is equally clean. You run the code. The tests pass or they don’t. The build succeeds or it fails. You know within seconds whether the agent did something useful. That tightness — data access plus immediate verifiability — is what made software engineering the first domain to scale.
Customer support followed the same pattern. There’s a ticket. There’s a knowledge base. There’s a resolution event. The data universe is contained and the outcome is binary. Agentic AI is already expected to have the highest impact in customer support precisely because those conditions were already in place.
These two domains didn’t win because AI works better for them — per se. They won because they already had what every other domain is still working to build.
What Sales and Finance Actually Need (And Why It’s Harder — For The Moment)
Ask yourself what a genuinely useful AI sales agent requires.
It needs your CRM data — contacts, deal history, activity logs, pipeline stages. It needs email and calendar context to understand relationship history. It needs product usage data to know what customers are actually doing. It needs the recording from the last call, the champion’s title change on LinkedIn, the competitive intel from the last loss debrief. It needs to know which deals went dark last quarter and why.
None of that lives in one place. Most of it doesn’t have a clean API. Some of it exists only in people’s heads and inboxes. The integration surface is enormous — and getting it connected in a way that’s semantically coherent, not just technically linked, is genuinely difficult.
46% of organizations cite integration with existing systems as their primary challenge in deploying AI agents. That’s not a capability gap. That’s a plumbing gap.
Finance is more complex still. A useful AI agent in finance needs your ERP, banking data, payables, receivables, headcount planning, contract terms, and ideally some institutional memory about why historical decisions were made. It operates in a regulated environment where a mistake isn’t just rework — it can be a compliance event. 75% of enterprise leaders cite security, compliance, and auditability as the most critical requirements for agent deployment — and that bar is highest in finance and healthcare.
The AI isn’t the bottleneck. The data infrastructure and governance layer are.
The Containment Problem Is Just as Real
Beyond data access, there’s a second structural issue: feedback loops.
In coding, the agent’s output is verifiable in seconds. In sales, what does “it worked” mean? Did the email get a reply? Did the deal advance because of the agent’s outreach or because the champion finally got budget approval? Sales outcomes are noisy, lagged, and entangled with variables the agent didn’t control.
This is why early AI in sales looks like assistance — drafting emails, summarizing call recordings, filling CRM fields — rather than autonomous agents closing pipeline. Not because the models can’t do more, but because the feedback infrastructure to know whether autonomous decisions were good doesn’t exist yet at most companies.
Finance has the same problem compounded. When an AI agent flags a budget variance or generates a cash flow forecast, the ground truth arrives months later. The feedback loop is long and the consequences compound. You can’t iterate fast in that environment until you’ve built significant observability into what the agent did and why.
The Infrastructure Is Being Built Right Now
Here’s what makes the current moment interesting: these structural barriers are actively dissolving.
CRM vendors are opening their data aggressively. By mid-2026, most leading CRM vendors will offer native agent frameworks that integrate directly with enterprise workflows. HubSpot, Salesforce, and their competitors are racing to become the connective tissue AI agents need — and they have strong financial incentives to do it fast. Whoever owns the data layer for AI in sales captures enormous platform value.
87% of IT executives rate interoperability as either very important or crucial to successful AI agent adoption. Enterprise IT budgets in 2026 are flowing toward exactly the integration infrastructure that makes agents in sales and finance possible. The plumbing work is happening.
The numbers are already showing up in the forecasts. Gartner predicts 40% of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% in 2025. That’s not a marginal increase. That’s a step change. And it won’t all go to software engineering — the growth has to come from somewhere, and that somewhere is the domains currently sitting at 1-5% in Anthropic’s chart.
The Numbers in Sales and Finance Are Already Moving
The “low adoption” framing in Anthropic’s data obscures something important: within sales and finance, the companies that have actually deployed agents are reporting strong outcomes. The problem isn’t that deployments fail. The problem is that getting to deployment is still hard.
Salesforce drove a 15% increase in deals and shortened sales cycles by 25% with AI agents. AI cuts campaign launch times by 75% while boosting click-through rates by 47%. McKinsey projects that organizations integrating agentic AI into daily workflows can achieve productivity gains of up to 40% over the next decade, beginning with measurable improvements in service and response time as early as 2026.
These outcomes exist. They’re just not uniformly accessible yet because the data infrastructure required to replicate them isn’t in place everywhere.
Insurance is the case study worth watching. Historically one of the slowest-moving enterprise verticals, it went from 8% to 34% AI adoption in a single year. The reason: document-heavy workflows, structured data in core systems, and clear ROI signals from automated underwriting and claims triage. When the data conditions are right, adoption can accelerate dramatically regardless of how cautious the industry is by reputation.
What This Means for B2B Founders and Operators
If you’re building AI for sales, finance, or any data-complex vertical: the low Anthropic numbers are not a warning sign. They’re a market map. The companies that establish category leadership in these verticals during the infrastructure build-out — before the plumbing is fully mature — are the ones that will be impossible to displace once it is. This is the window.
If you’re a SaaS incumbent in CRM, ERP, or finance software: your urgency is higher than you probably feel. The integration work that’s happening right now — CRM vendors opening APIs, ERP vendors building agent frameworks — is what your replacement is being built on. If you’re not the one building the native agent layer in your category, someone else is. SaaS providers are already offering all-you-can-eat agentic enterprise license agreements — consumption models that turn agentic capability into a strategic retention play.
If you’re a buyer deploying AI in sales or finance today: over 40% of agentic AI projects are at risk of cancellation by 2027 if governance, observability, and ROI clarity are not established. The failure mode isn’t that the AI doesn’t work. It’s that the deployment isn’t scoped tightly enough, the data isn’t clean enough, and the success criteria aren’t defined clearly enough. Start narrow. Pick one workflow where the data is accessible and the outcome is measurable. Prove it. Then expand.
We’re Living This at SaaStr. The Agents Keep Getting Better.
I don’t want to just cite the data. I want to tell you what we’re actually seeing on the ground, because it tracks exactly with what all this research is pointing toward.
At SaaStr, we’ve been running AI GTM agents for months now. What’s been striking isn’t that they work — it’s how much better they keep getting. Month over month. Sometimes week over week. An agent that was adequate at outreach in Q3 is genuinely impressive at it now. The improvement curve is steeper than I expected, and I expected it to be steep.
But the moment that crystallized everything for me: Monaco — our AI sales agent — autonomously booked a $100k deal. Completely on its own. No human drafted the outreach. No human managed the follow-up. No human scheduled the meeting. Monaco identified the prospect, engaged them, nurtured the conversation, and got the deal to close.
A hundred thousand dollars. Autonomously.
That’s not a demo. That’s not a proof of concept. That happened.
And what made it possible wasn’t some breakthrough in model capability that happened overnight. It was the same thing the Anthropic research points to: the data was accessible, the workflow was defined, and the feedback loop was clear enough for the agent to navigate it. Monaco had access to the right information. It operated in a contained enough context. It knew what success looked like.
That’s the formula. And it’s increasingly replicable — not just for us, but for any B2B company willing to do the work of getting their data and their workflows in order.
The agents will keep getting better. That part is certain. The variable is whether your data infrastructure is ready to let them.
The Bottom Line
The Anthropic data showing software engineering at 50% of AI agent deployments is a snapshot of where the data infrastructure happened to be ready first — not a verdict on where AI works and doesn’t work.
- Coding had the data in the repo. Support had the data in the helpdesk. Both had tight feedback loops. Those conditions made early deployment straightforward.
- Sales, finance, and legal have scattered data, longer feedback loops, and harder containment. The adoption numbers reflect those structural realities, not the capability of the models.
- The infrastructure is being built right now. The integration APIs are opening. The CRM and ERP vendors are racing to become agent-ready. The companies deploying agents in these domains are already reporting strong outcomes.
67% of enterprise leaders say they will maintain AI investment even if a recession hits in the next 12 months. This isn’t hype anymore. It’s infrastructure spend.
The wave is coming to sales, finance, and every other vertical in Anthropic’s chart. The only question is whether you’ll be the one riding it or explaining why you missed it.
Sources: Anthropic “Measuring AI Agent Autonomy in Practice” (February 2026); Gartner (August 2025); KPMG Q4 AI Pulse Survey (January 2026); Deloitte State of AI in the Enterprise (2026); McKinsey; UiPath enterprise study; Sopro AI in Sales report.
