AI tools are spreading. Capability decides who leads

What's new:

BILL partnered with NewtonX to survey 207 accounting firm leaders on how they're approaching AI in 2026. The headline: AI awareness is nearly universal. Almost all (92%) say they're familiar with AI applications in accounting, yet only 14% consider themselves extremely familiar. Nearly half (48%) plan to dedicate 10% or more of staff time to AI adoption this year, signaling that firms are moving beyond curiosity into active experimentation.

How it works:

Firms cluster into four adoption stages: awareness (10%), early adoption (34%), active adoption (38%) and mature adoption (18%). According to the survey, most firms are no longer asking "what is AI?"— they're deciding how far and how fast to push it. The ROI story is already tangible with 92% of firm leaders reporting at least one hour saved per week with AI, and a median of five hours saved weekly. That equates to roughly 260 hours per professional per year, more than a month of reclaimed capacity.

Top priorities remain pragmatic:

• 65% are focused on automating routine tasks
• 61% on increasing operational efficiency
• 56% on improving client service

In other words, most firms are using AI to do existing work faster and better. Many report reinvesting that time into advisory conversations, strategic planning and deeper client engagement.

Behind the news:

The data reveals subtle tension with 49% describing their firm's AI ambition as high, and 53% define AI primarily as an efficiency tool for automating specific tasks, while 40% see it as transformative. Knowledge gaps help explain the difference. Nearly half of high-ambition firms cite insufficient end-user expertise as a constraint. Trust is also conditional, with "trust somewhat" being the middle ground which suggests firms are willing to experiment, but not to fully delegate. Barriers also vary by size. Smaller firms struggle with system integration. Midsize firms face change management and talent gaps. Larger firms wrestle with data governance and infrastructure complexity. The constraints differ, but the pattern is consistent: capability is the gating factor.

Why it matters:

The profession is largely treating AI as an efficiency lever, not yet as a structural redesign of service delivery. At five hours saved per week, the real question isn't whether AI works, but rather what firms do with reclaimed capacity. One path is linear: process more compliance work with fewer hours. Another path is compounding: reinvest that time into advisory, analytics, and proactive client engagement that wasn't previously scalable.

The 18% in mature adoption are not just automating tasks; they're building organizational muscle around training, governance, and workflow redesign. Those capabilities compound, and they are harder to replicate than a single tool deployment. The widening gap won't be defined by access to AI, it will be defined by depth of capability.

We're thinking:

This survey reflects cautious optimism—and that caution is rational in a profession built on accuracy, compliance and trust. The firms that treat AI purely as a productivity upgrade may miss the larger shift underway. Efficiency is the entry point, while strategy is the unlock. The next volumes will show how ambitious firms are allocating capital, reshaping services, and rethinking pricing. The early signal is clear: most firms are still in the efficiency phase, while a smaller cohort is beginning to redesign how value is created and delivered. Over the next 24–36 months, that difference will matter.

What's new:

Paul Ford, former CEO of a software services firm, has done what no consultant report has managed: put a precise dollar figure on AI's displacement of knowledge work. A data conversion project he would have billed at $350,000 in 2021 – requiring a product manager, designer, two engineers and up to six months of work – now takes him a weekend on a $200/month Claude subscription. Markets are already pricing in the reckoning: software stocks including Salesforce, Monday.com and Adobe shed half a trillion dollars from the Nasdaq 100 in two days, and legal software stocks slumped the moment Anthropic announced tools to automate legal work.

How it works:

Ford describes “vibe coding” – issuing plain-English prompts to an agent like Claude Code, walking away, and returning to finished, deployable software. He issues a prompt on his phone entering the subway, loses signal underground and resurfaces on the Manhattan Bridge to find the work largely done. Claude Code crossed a capability threshold in November 2025, going from useful-but-halting to autonomously producing “whole, designed websites and apps that may be flawed, but credible.” The key shift isn't quality. It's that software now wants to ship, collapsing the elaborate bureaucracy the industry built specifically to manage the risk that it wouldn't.

Behind the news:

Claude Code generating $1 billion in revenue in its first six months is the number the industry should be staring at — not as an Anthropic success metric, but as a demand signal. That revenue came largely from professionals paying to replace what they used to outsource. Ford's framing cuts to the bone: “I would have no idea how to bill for their time.” This is categorically different from previous automation waves, which increased work volume while reducing cost. What's collapsing isn't the hourly rate — it's the premise that the hours were ever the point.

Why it matters:

Finance and accounting professionals who read this as a software developer's problem are making a category error. The "ship risk" principle Ford identifies — the coordinated expertise required to produce a reliable, high-stakes deliverable exists across audits, tax work, financial reporting and compliance. If AI compresses parts of that production cycle from weeks to hours, the economics will change. But in regulated environments, the value has never been just the document — it's the professional judgment, regulatory interpretation, and accountability behind it. The real question isn't whether CPAs remain essential; it's how firms define and price the expertise that AI can't automate.

We're thinking:

Ford doesn't pretend the transition is painless, distant, or manageable with a change management plan. He watched half a trillion dollars evaporate from software stocks and called it “not subtle.” The uncomfortable parallel for accounting and finance: markets are already discounting the future value of human-at-desk work in financial services — the same way they just did to Salesforce and Adobe. The firms that navigate this aren't scrambling to bolt “AI features” onto existing workflows. They're asking the harder question: if AI collapses the cost of the deliverable toward zero, what is the client actually paying for? The answer — accountability, judgment under genuine ambiguity, the professional who signs their name — has always been the real product. The disruption isn't erasing that. It's making it explicit.

What's new:

Anthropic's Feb. 24 Cowork rollout didn't just expand a product — it detonated a $285 billion software stock selloff before the announcement even landed. Thomson Reuters dropped nearly 16% in a single day, LegalZoom fell 20% and FactSet shed more than 10%. The market wasn't reacting to a feature update; it was repricing the future value of human-at-desk work across professional services. Cowork's new finance-specific plugins — covering financial analysis, investment banking, equity research, private equity and wealth management — make explicit what the market had already intuited: this product targets the knowledge worker's core workflow, not just their productivity at the margins.

How it works:

Cowork uses the same agentic architecture as Claude Code, now applied to office work without a terminal. It runs locally on your desktop, accesses your files directly, breaks complex tasks into parallel sub-agent workstreams, and delivers finished outputs – formatted Excel models, PowerPoint decks, synthesized research – while you step away. The Feb. 24 update lets Claude carry context natively between Excel and PowerPoint, run on a schedule without manual triggers, and connect to Google Drive, Gmail, DocuSign and FactSet through deep integrations that eliminate copy-pasting. Enterprise IT can govern access and audit what Claude touches via the open Model Context Protocol standard.

Behind the news:

The Cowork rollout is the third act of a month-long blitz: Claude Opus 4.6 on Feb. 5, Claude Code Security on Feb. 20, then Cowork's enterprise expansion on Feb. 24 – each detonating a different sector of the stock market. That sequencing isn't accidental. Anthropic is methodically demonstrating that agentic AI can execute the full stack of professional knowledge work, not just assist with it. Anthropic's head of economics noted during the launch that there's no evidence yet of widespread labor displacement. That caveat is doing a lot of work – and is almost certainly temporary.

Why it matters:

The finance plugins deserve a close read. The equity research plugin parses earnings transcripts, updates financial models and drafts research notes. The private equity plugin reviews large document sets, models scenarios and scores opportunities against investment criteria. The wealth management plugin generates rebalancing recommendations in the firm's own voice. These aren't chatbot features dressed up as workflows – they're governed, audit-trailed automations of the exact tasks junior analysts spend most of their billable hours performing. FactSet, MSCI, S&P Global and LSEG Connectors mean Cowork now has access to the licensed data those workflows depend on.

We're thinking:

Anthropic says it's “a platform, not a product, trying to own every workflow.” That framing deserves scrutiny – because owning the workflow is exactly how you eventually replace the product. The firms most at risk aren't the Salesforces already pricing in disruption; it's an accounting or advisory practice whose economics depend heavily on the labor-intensity of the deliverable. Cowork's scheduled tasks capability is the tell: when an AI runs your Friday variance report, updates the model and drafts the client memo before you arrive at the office, the engagement model inevitably evolves. The strategic question isn't whether that changes delivery. It's how firms reposition around oversight, interpretation and client counsel as production becomes automated.

What's new:

A growing body of real-world evidence highlights a significant gap in enterprise AI strategy. Martin Alderson, who works with both enterprise and smaller-company AI deployments, observes that some of the most capable AI users are not technical specialists. Finance directors may run Python scripts in terminal windows, while marketers deploy agentic workflows to automate tasks. Meanwhile, many users in large enterprises remain limited to tools such as Microsoft Copilot. Alderson describes the experience as similar to a simplified version of the ChatGPT interface. One indicator of the gap: Microsoft itself is reportedly rolling out Claude Code to internal teams despite broad internal access to Copilot.

How it works:

The divide isn't AI versus no AI — it's chat interfaces versus agents. A non-technical executive who's got their head around Claude Code can convert a 30-sheet Excel financial model to Python in a session or two. Once in Python, they effectively have a data science team in their pocket: Monte Carlo simulations, live external data via APIs, web dashboards. That's not “better Excel.” That's a different species of work. Meanwhile, enterprise IT enforces locked-down environments with no local script execution, no internal APIs for agents to connect to, and siloed engineering departments that couldn't build the infrastructure even if leadership wanted them to.

Behind the news:

The security concerns are real – you don't want people running coding agents over production databases with no control – but the result is that enterprises don't have the engineering capacity to build safely sandboxed agents either. The governance apparatus designed to manage risk has become the primary source of competitive risk. And the irony cuts deep: senior decision makers using Copilot with poor results are writing off AI entirely, while simultaneously spending a fortune with large consultancies to get not very far.

Why it matters:

For accounting and finance professionals, Alderson's finance director example is the case study to internalize. The "Consumer" path – Copilot, sanctioned chat tools, AI-assisted formulas in Excel – yields 10-20% productivity gains. The "Builder" path yields 10-100x. Smaller companies that don't carry the enterprise baggage are absolutely flying, and the gap is obvious when you can see both sides of it. Firms that have normalized Copilot as their AI answer have not solved the AI problem. They've installed a faster typewriter and called it transformation.

We're thinking:

The BlackBerry analogy holds, but the timeline is more compressed than 2008. BlackBerry had years before the iPhone made the gap undeniable. This bifurcation is accelerating in months. For firm leaders, the structural question is not simply whether their organizations are “AI-ready.” It is whether their governance models enable responsible adoption while protecting client outcomes. Over the next five years, success will not be defined solely by large-scale deployments such as thousands of Copilot licenses approved through IT. Instead, leading firms will focus on enabling their most capable professionals to use advanced builder tools with appropriate guardrails. Otherwise, clients may increasingly turn to smaller firms that can deliver deeper analysis more efficiently.