Audit Firms Face Talent Gap as AI Takes Over Entry-Level Work
His session sparked conversation about how firms are deploying AI to optimize for efficiency without designing what replaces the developmental value of the work being automated. The result, if left unaddressed, is a generation of auditors who learn how to confirm but not how to inquire.
Here’s the biggest takeaways.
The audit talent pipeline problem is bigger than the numbers show
Bukowski-Kruszyna opened with etymology. The root comes from Latin: "audio, it means I hear an auditor is one who listens listens with intent and should determine whether what they hear is consistent with the evidence."
That framing has held for two millennia. What has changed is what firms are asking junior auditors to actually do, and whether those tasks still build the listening skills the word implies.
The pipeline problem is bigger than the numbers show
But Bukowski-Kruszyna's concern is the part the data cannot capture: the people who are deterred before they even start leave no trace.
Many auditors do not enter the profession through a direct, planned route. They arrive through adjacent paths, discover they are good at the work, and stay. If the prevailing narrative around AI convinces potential entrants that audit has a limited shelf life, those routes narrow, and the profession loses people it never knew it was going to have.
Audit work has two layers. AI and auditing strategies target only one.
His core framework divides audit tasks into two categories.
- The mechanical layer covers repeatable, rule-based work: checking dates, matching invoices to the general ledger, reconciling confirmations, completeness checks. This is the work that fills files and is most amenable to automation.
- The analytical layer is where judgment lives: understanding assertions, interpreting what a discrepancy means for the audit opinion, reading what governance behavior implies, building professional skepticism through direct exposure to real situations.
The developmental value of junior staff has historically been embedded in the analytical layer, even when it appeared to live inside routine tasks. An auditor processing an invoice match is not just matching invoices. Done with attention, they are also building an instinct for what looks wrong.
The risk is not that firms automate the mechanical layer. The risk is automating it without preserving the analytical thinking that used to accompany it.
AI flags the risk before a junior has even formulated the thought
When AI flags a risk before a junior auditor has even formulated the thought, you haven't just automated the mechanical layer. You've automated the analytical layer too.
Bukowski-Kruszyna calls this the developmental blind spot. The behavioral pattern he observes in practice is that juniors are learning to confirm rather than to inquire.
Over time, that may produce technically competent practitioners. It does not necessarily produce auditors: people who can form an independent view about whether what they hear is consistent with the evidence.
Experience is accumulated scenarios, not a knowledge base
Bukowski-Kruszyna draws on a conversation with a partner who could resolve a complex technical question almost instantly. When asked how, the partner's answer was simple: competence came from accumulating scenarios.
That framing matters because it describes a formation model, not a training model. Competence in audit judgment is built through repeated, guided exposure to messy real situations, not through reviewing correct outputs.
This is where he challenges the human-in-the-loop assumption that many AI governance frameworks rely on. A human reviewer is assumed to add a meaningful check on AI output. But you cannot replicate that experience by handing someone an AI summary and asking them to validate it. A reviewer with no contextual awareness is not a meaningful check, regardless of where they sit in the process.
A reviewer who lacks the scenarios to interpret what they are looking at is not a meaningful check. They are a signature on a process.
Deterministic vs. probabilistic tools: matching the tool to the career stage
Rather than a firm-wide AI policy, Bukowski-Kruszyna recommends mapping tools to developmental readiness.
Deterministic tools are appropriate earlier in a career.
Centrally approved, repeatable, with guardrails built in. The quality of the output does not depend on the experience of the person using them. Sample selection tools and standard analytics applications fit this category. The output doesn't depend on the experience of the person cranking the handle.
Probabilistic tools, primarily generative AI, require domain experience to use well.
Output quality is tied directly to input quality. Without the contextual framing that experience provides, generative AI produces confident, plausible text that may be wrong in ways a junior auditor cannot detect.
His Copilot example is familiar enough to land. People complain about generic output, then you look at the prompt: "draft an email." Three words. "You gave it three words." The principle scales. The quality of the output is inherently tied to the quality of the input. That makes generative AI a tool that compounds experience rather than replaces it, most useful to practitioners who already have the scenarios to recognize a good result.
If AI creates space, leadership decides what happens to it
When automation returns time to managers and partners, a decision gets made, whether firms intend it or not.
Do we lump a whole load more work onto them? Oruse some of the space to have some genuine developmental conversations?
Bukowski-Kruszyna is direct about where accountability sits when firms get this wrong: That's not a technology failure. That's a very human failure and a leadership failure.
And the cost of leaving it unanswered: If you don't answer that question and set some direction, your firm's culture will answer it for you.
A structural shift already changing training economics
Bukowski-Kruszyna flags a concrete near-term disruption. From Jan. 1, Level 7 apprenticeship funding, the master's-level ACA route, was restricted to 16-to-21-year-olds. The problem: The average starting age of a Level 7 ICAEW apprentice is 22.
Most graduate entrants to the profession no longer qualify for this funding route. Training costs rise at exactly the moment firms are absorbing AI investment demands and competing for fewer candidates. It's not a distant risk. It's a cost you're already bearing whether you know it or not. His advice: model it quickly, then build strategy around the new economics rather than the old assumptions.
The pyramid-to-diamond model carries a formation risk
The restructuring many firms are moving toward looks efficient on a slide: fewer juniors, a wider mid-layer, the same narrow partner peak. Bukowski-Kruszyna's challenge is the premise it rests on.
Today's mid-layer was trained under the pyramid. They developed judgment by doing the work that AI agents are now being asked to take over. "All of those people honed their judgment by doing the work that we're proposing the AI agents take over," he says.
Firms hollowing out junior cohorts today may find in five years that the people responsible for overseeing AI output never built the judgment to do it well. They will have validated. They will have reviewed. They will not have done it.
His question for the room was clear: Who's going to supervise those agents in 2030?
What happened when an audit analytics tool looked like it failed
Bukowski-Kruszyna illustrated the point with a story about a revenue analytics tool that generated too many flagged outliers. The audit team's instinct was to abandon it and revert. Instead, a half-hour troubleshooting conversation produced something more valuable than a clean output.
His takeaway is the tool didn't fail. It created the conditions for a conversation that built understanding. And the harder truth is that culture can't be shipped as a software update.
The four questions every AI strategy should answer
In the session, four questions were brought forward that firms can use to test whether their AI strategy has a talent development chapter.
- Which tools, at which career stage, and why? "Most rollouts are driven by what's available, not what is necessarily developmentally appropriate." Without that mapping, firms are running an experiment on their people without a control group.
- What happens to the analytical question? Where analytical reasoning gets done is a design choice, not a default outcome of automation.
- Are managers using the space AI created? Or filling it with more work? The firm's culture will decide this if leadership does not.
- Are experienced practitioners using AI to develop? Generative AI scales with experience. The appropriate use is stress-testing thinking and preparing for difficult client conversations, not only producing text.
The profession has always trained more than its own staff. It has trained the financial leadership of the wider economy. If AI changes how auditors are trained, firms must deliberately design what replaces what was lost or discover too late that the pipeline did not just shrink. It changed shape in a way that produces fewer people who can listen with intent.
Connect '26 continues in New York.
If Bukowski-Kruszyna's session raised questions about how your firm is balancing AI adoption with auditor development, New York is where those conversations continue.
