Leading a firm in the AI era comes down to a short list of decisions a managing partner cannot hand to anyone else. The research on which organizations get value from AI and which ones spend and stall points at leadership behavior rather than software choice: who owns governance, whether anybody rebuilt a workflow, and what the firm says the recovered hours are for.

This article sets out seven of those decisions. Each one carries a named study behind it, with the sample size stated so you can judge how far it travels. The evidence comes from McKinsey's State of AI surveys, BCG Institute's 2026 analysis of AI adopters, Thomson Reuters' 2026 Future of Professionals research, and a Stanford and BetterUp study on the quality of AI output at work.

None of the seven needs a budget approval or a technical background. Most fit inside one partner meeting. Firms that skip them buy tools and watch nothing change.

The Meeting That Ends Without a Decision

Think about the last two months at your firm. AI came up in a partner meeting, in the hallway, and in at least one client conversation. Accounting Today put a sharper version of that question to firm leaders in April: how many of those exchanges closed with a decision about what to do next. For most firms the honest answer is none of them.

That gap explains more wasted AI spending than any tooling mistake. BCG uses a rough split to describe where the work sits in an AI change program: around 10% on algorithms, 20% on technology and data, 70% on people and process. Treat it as a rule of thumb rather than a measured allocation. The shape still holds. Most firms run it backwards, because a partner group can approve a license in twenty minutes and needs six months to change how a client file moves through the office.

Seven decisions sit with you.

1. Decide That You Own This, and Say So Out Loud

McKinsey's State of AI survey tested a set of adoption practices against self-reported financial impact. A chief executive's oversight of AI governance came out as one of the elements most correlated with EBIT impact from generative AI, and the effect ran strongest at larger organizations. Twenty-eight percent of respondents said their CEO held that role. Seventeen percent said the board did.

The survey is self-reported and correlational, so read it as a pattern rather than proof. For a firm your size the translation is direct. Which tools touch client data, how you reprice an engagement when the labor content drops, what a reviewer puts a name to. Those sit with the partner group, and they stop being decisions the moment you route them to whoever manages your systems.

2. Rebuild One Workflow Before You Buy Anything Else

The same McKinsey work tested 25 organizational attributes. Redesigning workflows had the biggest effect on whether a company saw EBIT impact from generative AI. Twenty-one percent of organizations using generative AI said they had fundamentally redesigned any workflow at all.

Pointing a tool at your current process leaves the process where it was. Pick one workflow you can describe end to end. Month-end close prep, client document intake, engagement setup. Rebuild it around what the software can carry, then measure the same number you measured before. One rebuilt workflow will teach your firm more than a year of pilots.

3. Approve a Tool List This Week

Thomson Reuters found that 81% of tax and audit professionals now use AI tools on a regular basis, and that more than a third admitted to using tools their firm never authorized. Thomson Reuters sells software into this profession, so weigh it as research from an interested publisher. The direction is still hard to argue with.

Client financial data is already moving through software that nobody at your firm approved. You can close that exposure in an afternoon. Name three or four tools people may use, write down what may never be pasted into any of them, and circulate it. Cheapest decision on this list, and the one carrying real liability.

4. Say What the Recovered Hours Are For

BCG Institute scored more than 600 US public companies on AI adoption. Among the leading tier, 10% used AI mainly to cut cost. Fifty-nine percent used it to expand what each person could deliver, and 21% built new offerings on top. Those are large listed companies, so take the mechanism and leave the shareholder-return figures where they are.

Thomson Reuters frames the same fork for firms as three paths: elevate people into advisory work, scale capacity without hiring, or rebuild the service model. Pick one and announce it. Staff who hear nothing assume the answer is redundancies, and adoption stops there.

5. Spread Fluency Instead of Hiring One AI Person

That BCG analysis also compared the leading tier against the tier sitting just below it. Technology and deployment scores barely moved between the two. The talent score nearly tripled. At leaders, 13% of employees held AI-related skills against 1% at laggards, and specialist roles reached 3.5% of headcount against 0.1%.

Read the ratio rather than the absolute numbers. A ninety-person firm does not need a data science function. It needs the senior associate who runs the close to understand where the tool breaks, because she is the one who will spot the next thing worth automating. Every firm can buy the same software, so the difference sits with the people who understand the work.

6. Measure Rework, Not How Fast the First Draft Appears

BetterUp Labs and Stanford's Social Media Lab surveyed 1,150 US desk workers in September 2025. Forty percent said they had received AI output in the previous month that looked finished and left the thinking to them. Each instance took close to two hours to resolve. One snapshot, self-reported, so hold the figure loosely.

The finding that should concern a partnership is the social one. Around half of recipients rated the colleague who sent it less capable and reliable afterwards, and 42% rated them less trustworthy. In a firm where review time and compensation both run on partner judgment, that lands somewhere real. Track cycle time through review, not time to first draft.

7. Protect how your juniors learn judgment

Thomson Reuters lists this as its own priority for firm leaders. As AI absorbs routine work, firms have to preserve the structured development that turns a second-year into somebody who can review, and keep supervision in place before automation strips out the tasks that build that judgment.

Your pyramid depends on it. The work you are most tempted to automate first is the work your associates learn on. Decide what replaces it: earlier exposure to review, or protected time on judgment calls. Almost nobody writing about AI leadership raises the question, though your associates already have.

The Seven, on One Page

How We Approach This at Codebridge

Thomson Reuters interviewed audit practitioners this year about how firms build confidence in AI systems. The firms furthest along run tools in parallel on live engagements and build standardized tests so they can compare systems doing the same task. Partners trust what they watched happen on their own files. A vendor demonstration runs on the vendor's data.

We came out of KPMG and we work the same way. A fixed-fee three-week discovery starts with one of your workflows, prototypes on your data, and ends with something your people either used or rejected on evidence. Fixed price, fixed date. Your firm owns the code, so a supplier shutting down becomes a supplier problem rather than yours.

If a workflow came to mind while you read the seven, book a 30-minute call and we will tell you whether it fits.