Your Firm’s AI Investment Isn’t Failing. Your Incentive System Is.

Quick answer

Microsoft’s 2026 research tested 29 factors against workers’ reported AI impact and found that organizational factors, meaning culture, manager support and talent practices, accounted for 67% of the effect against 32% for individual mindset and behaviour. The constraint is neither your people nor your software. In accounting firms specifically it is usually the billable hour, because firms measure and reward staff on hours and realization and then ask those same staff to adopt tools whose entire purpose is to reduce hours. People behave rationally, and the tool loses.

Organizational conditions outweigh individual capability

Microsoft’s 2026 Work Trend Index, published on 5 May 2026, surveyed 20,000 knowledge workers who use AI across 10 markets, fielded by Edelman Data x Intelligence between February and April 2026. The researchers tested 29 separate factors against workers’ reported AI impact, covering individual skill, mindset, curiosity, manager support, culture, governance and demographics. Organizational factors accounted for 67% of the effect and individual mindset and behaviour for 32%, a ratio of more than two to one.

The finding deserves the scrutiny any firm would apply to a client’s numbers, so it is worth stating the limitations before drawing conclusions from it. This is association rather than causation, and Microsoft says so explicitly in the report. Every variable is self-reported by the same respondent at the same moment, which is a real weakness. What the analysis has in its favour is scale and consistency, since the ranking held across three model families with held-out test R² values of 0.680, 0.689 and 0.690, and the effect size is not the sort that can be argued away as noise. Microsoft’s own conclusion is that the question is not whether people have the right skills but “whether the organization is built to unlock them.”

A tenth of the workforce is capable but blocked

Microsoft plotted respondents on two dimensions, individual capability with AI and organizational readiness to absorb it, with the index calculated on 16,971 respondents holding complete data on both axes. Only 19% land in what the researchers call the Frontier, where capability and readiness reinforce one another, and 10% land in what they call blocked agency, meaning high individual capability paired with low organizational readiness. A further data point completes the picture: just one in four AI users, 26%, say their leadership is clearly and consistently aligned on AI, which will surprise nobody who has sat through a partner meeting on the subject.

Translated into a firm of any size, that blocked tenth has a specific and expensive character. Those people have already done the work you are budgeting to teach them next year, having taught themselves on their own time out of curiosity, and they returned to a system with no slot for what they learned, no permission, no workflow and no partner who knows what to do with them. They are also not marginal performers, since nobody in the bottom quartile spends a Sunday learning agentic workflows. The people who are blocked are disproportionately the people a firm cannot afford to lose, which makes this a retention exposure that will not appear on any dashboard currently in use.

The transformation paradox

Microsoft’s survey identified a pattern among AI users that explains why capable people stay still. Three figures carry it:

• 65% fear falling behind if they do not use AI to adapt quickly
• 45% say it feels safer to focus on current goals than to redesign work with AI
• 13% say they are rewarded for reinventing work with AI if results are not yet there

Read together, those numbers describe people who are afraid of falling behind and simultaneously convinced that trying to fix it will hurt them, and the second belief is correct. Only 13% receive any credit for attempting something that has not yet paid off. Faced with that arithmetic, staff do the rational thing: they execute the old design perfectly, on time, and wait for someone above them to change the rules.

Most firms hear this and call it a culture problem, at which point somebody schedules a values workshop. It is not a culture problem. What those three figures actually measure is what gets counted, what gets rewarded, and what happens to a person who tries something that does not work inside one quarter. That is performance management and compensation, not culture, and the distinction determines whether the firm’s response has any chance of working.

In an accounting firm, the constraint has a name

General research meets the specific structure of a professional firm at an uncomfortable point. A firm measures its people on billable hours and realization, and then asks those same people to adopt a tool whose entire value proposition is producing the same output in fewer hours. 

They will not do it, they should not do it, and no quantity of enthusiasm from the managing partner alters the arithmetic on an individual’s scorecard. The tool is not competing against reluctance or against generational attitudes toward technology. It is competing against compensation, and compensation wins every time it is asked to. This is the largest single reason AI investment in accounting firms produces demonstrations rather than returns, and it explains why the pattern repeats across firms with different software, different budgets and different partner demographics. The variable those firms hold in common is the measurement system.

The layer underneath

A harder version of the problem sits below the internal scorecard, and it deserves naming even though no single article resolves it. Many firms still bill for time. Where AI compresses a 40-hour engagement into 25 and the engagement is priced on hours, the firm has converted a productivity gain into a revenue reduction while the client receives identical value.

That is not an adoption problem but a business model question, and it explains something the adoption framing cannot. Resistance in accounting firms runs upward as well as downward. The manager has a utilization target and the partner has a book, and both are measured in the unit the technology is designed to consume. Any firm serious about AI returns will eventually have to decide whether it sells hours or sells outcomes. Most are postponing that decision and mistaking the postponement for a technology problem, which is comfortable in the short run and costly over a five-year horizon.

What the market sells, and what it cannot

Every AI readiness assessment on the market measures individual capability, every AI training programme targets individual capability, and every AI mindset workshop targets individual capability. All of it addresses the 32%. Almost nobody sells anything aimed at the 67%, and the reason is structural rather than conspiratorial: the 67% is not a product. It is the operating model, which is to say the comp plan, the utilization target, the review form, and what a partner says out loud when someone proposes doing something differently. None of that arrives in a box or on an invoice.

Deloitte’s 2026 Global Human Capital Trends points the same direction from a different angle. Across more than 9,000 leaders in 89 countries, only 6% report making great progress on designing effective human-AI interactions even though 66% recognise its importance. Separately, Deloitte research with 100 C-suite leaders found that the 59% of organizations taking a tech-focused approach were 1.6x more likely to not realize AI returns exceeding expectations than those taking a human-centric approach. That is a smaller sample supporting a narrower claim, and it should be weighted accordingly, but it travels in the same direction as everything else.

Five changes, in order of cost

1. Ask your blocked people what is in the way. They already know, having worked around it for a year, and Microsoft’s data suggests roughly one in ten of your AI users is capable and stopped. The conversation costs nothing, takes an afternoon, and is the highest-yield move available before any money is spent.
2. Change one metric before buying one more licence. Take a single team and give them a target that is not hours, whether output quality, client outcome, capacity released or cycle time, provided it is measured on a basis that automation improves rather than punishes. Two quarters of that will teach the firm more than any pilot, and it will teach it about the firm rather than about the vendor.
3. Give someone written credit for a failed attempt. The 13% figure is the entire problem expressed as a single number, and until reinvention is rewarded independently of whether it worked the first time, people will keep executing the old design and will be right to. This is the cheapest structural change available and the one most firms decline, because it requires a partner to state on the record that a miss was acceptable.
4. Decide what the firm actually sells. The hours question does not resolve itself through avoidance. Firms that answer it will compound, and firms that do not will continue buying tools that shrink their own revenue and describing the outcome as disappointing.
5. Stop calling it a culture problem. Culture is downstream of what gets counted, so changing what gets counted moves the culture. Attempting it in the other order produces a poster.

Where a firm has spent real money on AI and seen little return, the likeliest explanation is not that it chose the wrong tool, hired the wrong people or moved too slowly. It is that the firm bought execution capacity and never touched the design layer, adding a faster way to do the work without deciding what work should exist, who owns the output, or what its people are rewarded for. The research puts roughly two thirds of the effect in that layer, which is also the two thirds nobody can sell you, and that is precisely why it remains unaddressed.

Apoorv Dwivedi is Founder and President of Fixyr and a speaker on marketing and growth for accounting and advisory firms.

FAQ