Picking up the firm from the previous post. The AI licensing is in place. Usage is real and growing. The leadership team is satisfied with the program based on adoption metrics that don't tell them anything about operational change. Linda is still trying to surface workforce implications. Marcus is producing monthly dashboards that look good and mean little. The firm has settled into a comfortable equilibrium at stage two.
What disrupts the equilibrium isn't the AI program. It's a specific person getting specifically impatient with a specific problem.
Rita Chen is the head of the audit practice. She has been at the firm for eleven years, made partner six years in, and took the practice leader role two years ago. She is technically sharp, operationally demanding, and famously unimpressed with anything that doesn't reduce the friction of doing audit work. She has been using the firm's AI tool for about eight months. Her assessment of it is precisely calibrated: useful for some tasks, irrelevant for most of what the audit practice actually struggles with.
What the audit practice actually struggles with is engagement startup. The first three weeks of any new engagement involve enormous amounts of document collection, reconciliation, and population testing setup, work that staff find tedious, that consumes disproportionate hours in the engagement budget, and that can't be billed at premium rates because it's mechanical rather than analytical. Rita has been trying to reduce the cost of engagement startup for two years. She has not made meaningful progress on it.
In a Wednesday-morning practice leadership meeting, Rita asks her senior managers a specific question. If they had unlimited engineering time and unlimited budget, what part of engagement startup would they want a tool to handle. The answers are concrete: document categorization and indexing, prior-year-to-current-year reconciliation, sampling population generation, standard procedure documentation. The senior managers know the work intimately. They can describe it in operational detail.
Rita walks the list to Marcus's office that afternoon.
The conversation in Marcus's office is the firm's first operational-connection conversation, even though neither of them frames it that way at the moment.
Rita doesn't ask Marcus what the AI tool can do. She tells Marcus what the audit practice needs and asks whether the tool can do any of it. The framing is reversed from how Marcus has been thinking about the program. Marcus has been organizing the program around capabilities looking for applications. Rita is organizing the conversation around an application looking for capabilities. The difference is small in syntax and large in consequence.
Marcus listens. He asks clarifying questions: how much staff time does engagement startup consume, what percentage of engagement budget goes to it, what data sources are involved, what client systems would need to be connected. Rita has the answers. The practice has been measuring these numbers for years because they affect partner economics directly. Marcus realizes, as the conversation continues, that he has been running an AI program for a year without ever asking a practice leader these questions. He has been running it for IT, not for the firm.
They end the conversation with a defined problem statement: reduce the staff hours required for engagement startup by 30% within twelve months, measured against the prior year's baseline. The problem statement is specific enough to be evaluable. The reduction target is aggressive enough to require real change and modest enough to be plausible. The measurement methodology is something both Rita and Marcus can defend in front of the executive team.
Marcus walks back to his office with the first thing he can credibly call an AI project.
The project takes shape over the next several weeks. Marcus and Rita identify three specific workflows within engagement startup that are the highest-value candidates: document intake and categorization, prior-year reconciliation, and sampling population generation. For each one, they sketch out what an AI-assisted version of the workflow might look like, what data the AI would need access to, what the human reviewer's role would be, and what would constitute success.
The vendor evaluation is faster than it would have been a year earlier, because Marcus now has a defined problem to evaluate vendors against. The professional services SaaS pitch he had filed away during stage one is one of the vendors he revisits. He also evaluates two others, one specialized in document intake and one in workpaper automation. The evaluations include actual proof-of-concept work, not just demos. Each vendor is given a representative engagement workflow and asked to demonstrate the tool against it. Two of the three vendors produce results that are noticeably better than the firm's current manual process. One does not.
The firm selects the top-performing vendor. Dana approves the budget after asking Rita to commit to the engagement-budget savings the project is targeting. Rita commits. The implementation starts in earnest in Q2.
The implementation is messier than the proof-of-concept suggested. The AI tool handles the document intake workflow well on standard engagements but struggles with the firm's larger clients, whose document inventories include hundreds of non-standard file types and decades of inherited naming conventions. The reconciliation workflow requires more upfront configuration than the vendor's pitch implied. The sampling population generation is the area where the tool delivers the most clearly above expectations.
The first quarter of live use is mixed. Some engagements come in significantly under their startup budgets. Others come in at or above prior baseline because the engagement team is still learning the tool. Rita refuses to let the practice abandon the project during the rough period, partly because she has the partner credibility to absorb the criticism and partly because she suspects the rough period is temporary. By the end of Q2, the engagements that started after the tool was bedded in are reliably coming in 20% under the prior baseline. The 30% target looks achievable by end of fiscal year, possibly earlier.
Marcus has, for the first time, a real story to tell the executive team. Not "X% of staff are using AI" but "engagement startup costs in the audit practice are down 20% with a 30% target by year-end, achieved through a specific AI-supported workflow change." The Managing Partner asks why the firm hasn't been doing this kind of project all along. Marcus does not answer the question with the honest answer, which is that the firm hasn't been doing this kind of project because the firm hasn't been asking practice leaders what they needed.
Other vantages on the firm at this stage:
The senior associate doing audit startup work. Her experience of the rollout is the most complicated of anyone's at the firm. The work she used to do has changed. Document categorization that took her several days now takes a few hours of review, with the AI doing the first pass. Reconciliation that took her a week now takes a day and a half. The senior associates who adapted to the new workflow quickly are getting more interesting work, assignments earlier in the engagement, exposure to client conversations they wouldn't previously have been part of, more time on the analytical portions of the audit. The senior associates who haven't adapted are getting less. The differentiation is real and uncomfortable. Linda has started noticing it in retention conversations.
The advisory practice leader. He watches Rita's project closely. His practice has different bottlenecks and the audit playbook doesn't transfer directly, but he can see the shape of how Rita identified her problem and went after it. He starts running the same exercise with his own senior managers, asking what they would want an AI tool to handle if engineering and budget were unlimited. The answers in his practice point toward client-facing deliverable drafting, research synthesis, and proposal generation. He brings the list to Marcus three months after Rita did. The advisory project is initiated.
Linda's workforce conversation finally happens. The differentiation among senior associates in the audit practice produces specific questions about career paths that Linda has been waiting to be asked. In a one-on-one with Rita, she walks through what's happening to the entry-level work in audit, what skills are now more valuable, what skills are now less valuable, and what the firm's career ladder will need to look like in three years if the trend continues. Rita listens. Rita asks Linda to bring the same analysis to the next practice leaders' meeting. Linda has, for the first time, an operational invitation rather than a deferred one.
The tax practice leader's reluctance. He has been watching Rita's project too. His response is different. The tax practice has tighter regulatory constraints around AI use in client work, and the head of tax has interpreted those constraints as a reason to wait and see. The tax practice will be the last of the three to move into operational connection, and the gap between the tax practice and the others will become a recurring tension across the rest of the case study.
The firm has now moved into stage three for one of its three practices and is moving toward it for a second. The pattern that produced the movement is replicable: a practice leader gets specific about a specific bottleneck, IT translates the operational need into a tractable AI-supported project, the project ships with measurable outcomes, the success creates momentum.
What's notable about the pattern is that the AI program itself didn't drive the movement. Marcus's monthly dashboards, the firm-wide licensing, the training program, the policy documents, none of those were the cause of the shift to stage three. The cause was a practice leader walking down a hallway with a specific problem. The AI program was the substrate that made the conversation possible, but the substrate was not the project.
This is the most common pattern by which mid-market organizations actually move from stage two to stage three. It is not by enterprise strategy. It is by an operational leader who has the credibility to commit to a specific outcome, the operational specificity to define what success looks like, and the willingness to absorb the rough period that comes with any real implementation. Most stage-two organizations have all three of these ingredients somewhere in the building. Most don't connect them, because the AI program is being run by people who haven't learned to ask practice leaders what they actually need.
The next stage of the case study picks up the firm as the audit project becomes the template for a broader redesign, and as the workforce implications Linda has been tracking finally enter the firm's formal conversation about what jobs are going to look like in five years.
— Chris