The Final Responsibility Remains With the Person

Who do you think keeps AI on a shorter leash: a global firm with a proprietary AI tool and a dedicated governance department, or a handful of students running a consultancy between lectures? Most people would bet on the students. We did too, until we interviewed Tommaso, an Associate at PwC’s Asset & Wealth Management Advisory practice, and Roland, Vice President of ConQuest Consulting, a Junior Enterprise. The answer turned out to be the exact opposite of what either of us expected. "The final responsibility remains with the person using the tool," Tommaso told us, describing PwC’s approach to AI, and by the end of both conversations we understood why the richest firm we spoke to was also the most cautious one. The reason has less to do with money than with bureaucracy. 

When More Resources Mean More Caution
PwC has almost everything a firm could want to move fast on AI: an internal proprietary tool, recurring training sessions, and a governance framework built specifically around responsible use. Yet, as Tommaso explained, none of this buys autonomy for the machine. "The AI is not supposed to make final professional judgments, approve decisions, or send client-facing outputs without human review," he told us, describing PwC’s own assistant as "not completely passive" but never left to close a loop alone. 
This caution is not PwC-specific, it is structural. A 2025 governance survey found that large enterprises are five times more likely than small firms to run multiple generative AI systems in production, yet they are also far more likely to employ dedicated governance officers and maintain incident-response playbooks (Gradient Flow, 2025). Scale, in other words, does not simply buy freedom for AI, it buys the machinery to control it more tightly. The bigger the firm, the more it has to lose if something goes wrong, and the more it spends making sure nothing does.

When Fewer Rules Mean Faster Decisions
ConQuest Consulting tells the opposite story, not because it is bolder, but because it has less to protect and less to untangle. Roland described an organisation still "in the process of preparing guidelines for the whole organisation," with no dedicated compliance function and no long-term contract tying it to a single AI provider. "We are not bound by any contracts regarding AI, so we can freely change tools whenever we want," he said, contrasting this with larger firms that are stuck with whatever a single partner provides, "even if it’s not necessarily the best option at a given moment." 
The absence of bureaucracy is not the absence of oversight. At every stage of a ConQuest project, Roland insisted, a human still checks the work: "There is always an element of human supervision throughout the project, really at every step." What changes is the speed of adaptation. The team can swap tools "over the course of a week," a change Roland assumes "probably takes months, quarters" inside a large corporation. Broader data backs this up: only 36% of small companies report having governance officers, compared with 62 to 64% of larger firms, and yet small companies are also far more likely to stay conservative in practice, adding at most one or two new AI use cases per year (Gradient Flow, 2025). Fewer rules, here, buy agility rather than recklessness. 

Two Different Meanings of Trust
The sharpest contrast was not about tools or budgets, but about what "trust" means. For Tommaso, trust is systemic: it lives in the fact that a tool has been tested, staff have been trained, and policies exist to catch mistakes before a client ever sees them. For Roland, trust is personal and concentrated: he put it at roughly "50-50" between the project manager and the client, with the board mainly judging results rather than process. PwC trusts a system. ConQuest trusts specific people to exercise judgement. 
Both sit on the cautious side of a wider industry hesitation: only 27% of executives say they would trust a fully autonomous agent in an enterprise setting (Elevate Consult, 2025). They arrive there by different roads, one paved with policy, the other with relationships, and neither is obviously the more advanced version of the other. PwC’s system was itself designed by people who had to decide, once, how much to trust their own colleagues. ConQuest’s reliance on individual judgement is not a phase to grow out of, it is a deliberate response to not yet having the budget to build a system. The two firms are not at different points on the same road. They are answering the same question, how much can we let AI decide, with whatever instruments happen to be in reach. 

Structure, Not Size, Sets the Ceiling
Our research question asked whether money or bureaucracy has a bigger impact on AI autonomy in consulting. After these two interviews, we are convinced it is bureaucracy. PwC’s budget goes almost entirely into containment: better tools, yes, but also stricter rules and narrower access. 
ConQuest’s lack of formal structure does not translate into recklessness either, it simply means decisions about AI get made faster, by fewer people, on the basis of common sense rather than procedure. 
Neither firm lets AI act without a human in the loop, and perhaps neither should, yet. What differs is who that human is, and how quickly they can change their mind. As AI keeps improving faster than most firms can write policy for it (EY, 2024), the real competitive question may not be how much AI a firm can afford, but how quickly its structure lets it decide when to use it.

• EY (2024) Business Leaders Plan to Double AI Investments by 2025, EY Survey. AI Business, 19 July. Available at: https://aibusiness.com/verticals/business-leaders-plan-to-double-aiinvestments-by-2025-ey-survey 
• Elevate Consult (2025) Selecting Your Technology Partner: Vetting AI Governance Consulting. Available at: https://elevateconsult.com/insights/selecting-your-technology-partner-vetting-aigovernance-consulting/ 
• Gradient Flow (2025) 2025 AI Governance Survey. Available at: https://gradientflow.com/2025ai-governance-survey/  

Article written by:

• BOFFETTI Stefano
• MOTTA Francesco
• RUDZINSKI Kacper