November 11, 2026, (Inside AI) — Accounting firms that deployed AI for tax preparation during the 2025-26 season cut practitioner time by roughly a third and lifted throughput by about 50%, according to data from four firms using the Current platform. One firm booked approximately $300,000 in additional tax work late in the season. Yet the same data shows review times barely moved, and in some cases did not fall at all. Faster preparation simply moved the bottleneck downstream.
The finding lands as managers across industries wrestle with a harder question than which tasks AI can perform. The real question is what it costs an organization to hand that work over. A new analysis of roughly 7,000 moderate-complexity individual and trust returns, including rental-property and K-1 income, identifies seven delegation costs that apply well beyond accounting. The research was conducted with access to Current through Nitin Nohria, chair of Thrive Capital, whose subsidiary Thrive Holdings owns the platform. The time and throughput figures are company-reported, and Current reviewed factual descriptions without influencing the analysis or conclusions.
The central insight is that task-level savings do not automatically become organizational gains. AI can produce a tax return faster than a reviewer can verify it. That single fact reframes the productivity debate. Leaders should weigh the total cost of delegation, meaning execution plus everything surrounding it, rather than comparing the hourly cost of a human against a machine.
Seven Hidden Costs of Handing Work to AI
The researchers group the costs into three clusters. The first concerns the work itself. Specification is the effort of describing a task clearly enough for a system to perform it. Verification is the cost of confirming accuracy quickly and reliably. Learning and lock-in are linked and shift over time: teaching an agent improves its output, but tailoring workflows to it makes switching harder.
Read: Gilbert + Tobin CEO Sam Nickless Deploys Agentic AI in Legal Workflows
The second cluster concerns organizational response. Incentive misalignment appears when AI changes the personal economics of the people handing work over. At one participating firm, billable hours had been a key performance measure, so a tool that compressed preparation time worked against the interests of the people expected to use it. The firm had to rethink those measures around quality and value rather than time spent.
Threats to identity formed another barrier. Accountants with decades in the role began asking whether their expertise would still be valued and whether their firm would need as many people. Much of the resistance was not to AI doing the work but to what it might mean for their future role. Goal misalignment covers systems that optimize a single metric while eroding client or employee experience.
The third cluster concerns ownership. Accountability is the cost of explaining the work and bearing the consequences when it is wrong. If a tax filing is incorrect and the IRS audits the return, a human still answers. AI speeds execution but does not materially reduce that burden.
Reviewers also noticed a pattern that differs from human error. Human mistakes tend to be idiosyncratic, while AI errors recur around particular document types or edge cases. Once reviewers recognized those patterns, they could target their attention and redesign the review process. The team supporting the deployment began making AI-generated work easier to trace and flagging where reviewers should look. Over time, each firm could lower verification costs by learning the characteristic mistakes of the system, much as a manager learns the strengths and weaknesses of a direct report.
The talent pipeline poses a longer-term problem. Junior employees traditionally learn pattern recognition and exception spotting through tax preparation and review. Managers and reviewers were already in short supply, and most came up through those junior ranks. If AI absorbs that work, where does the next generation of reviewers come from? One firm revamped its intern program to practice on dummy returns without AI. Another exposed junior staff to clients earlier to build people skills.
Partners gained a different benefit. Some used the recovered time for deeper client engagement. One partner phoned every client to walk through the return and explain each decision. The firms also learned that partners were not equally skilled at converting extra time into consultative advice or new business.
The accounting sector offers a useful test case because it combines structured document work with heavy regulatory accountability. Other professions with similar sign-off requirements, including law, medicine, and audit, face the same arithmetic. The execution layer compresses. The judgment and liability layers do not. That asymmetry explains why firm-wide productivity gains often lag the headline numbers from pilot programs, and why vendors and buyers can read the same deployment very differently.
Read: Northern Trust, Microsoft Deploy AI Agent to Cut Client Onboarding Time by 83%
The researchers propose four questions for leaders. What exactly are you delegating, given that jobs are bundles of activities rather than single tasks? Where will the delegation costs appear, since an hour saved in one workflow often reappears in another? Which costs can you reduce, through better traceability or revised performance measures? And what will you do with the capacity AI creates, since saved time does not by itself generate value?
Some costs resist reduction. A regulator may still require human sign-off, and a client may still want a professional standing behind a recommendation. Those responsibilities remain human because judgment and accountability cannot be delegated to software. The practical lesson from the tax season is that the salient decision is not where to use AI. It is what to delegate, what that delegation costs, and how to convert the resulting capacity into durable value.
Shamil has a low tolerance for AI that only works in controlled demos. He covers agentic systems and generative AI with the skepticism of someone who has seen too many “revolutionary” tools quietly disappear six months after launch.
