Every firm I have worked in had one. The person who knew where the old policy document lived, which client had a complaint brewing, what the regulator said last month and why the platform migration was late. Ask them a question and you got an answer, usually with context nobody else had.

AI is often pitched as the thing that will replace that person, or at least spread their knowledge to everyone else. I think that is right, with one condition that gets far less attention than the models themselves. An AI can only outperform the smartest person in the room if it can see what they see. Give it a slice of the firm’s knowledge and data, and the human who holds the whole picture will beat it every time.

The people we rely on are rarely smarter in the abstract. They are better connected. They have read the email thread, sat in the meeting, remember the client’s divorce and know that the back office record is wrong. Their advantage is context, gathered across systems that don’t talk to each other.

That context is expensive to assemble. McKinsey estimated in 2012 that knowledge workers spent nearly 20% of their working week looking for internal information, and 28% managing email (McKinsey Global Institute, 2012). Microsoft’s 2023 Work Trend Index, a survey of 31,000 workers, found that 62% struggled with too much time spent searching for information, and the average employee spent 57% of their time communicating in meetings, email and chat (Microsoft, 2023). Its 2025 report found employees interrupted every two minutes during the core working day (Microsoft, 2025). In 1971 Herbert Simon summed up the problem: “a wealth of information creates a poverty of attention” (Simon, 1971).

Most of the working day is spent moving informationMicrosoft survey of 31,000 workers in 31 markets
Time spent communicatingMeetings, email and chat; share of the average employee’s time
57%
Struggle with time spent searching for informationShare of workers surveyed
62%

Source: Microsoft (2023), Will AI fix work? Work Trend Index.

In my own career, the most valuable meetings were often the ones where someone joined dots across teams that didn’t normally talk to each other. That joining of dots is exactly what an AI with partial access cannot do.

The best evidence for AI’s value comes from settings where the model could see the information that mattered. In a study of 5,179 customer-support agents, a generative AI assistant raised issues resolved per hour by 14% on average, and by around 34% for the least experienced staff. The authors suggest the tool spread the best practices of the most able workers to everyone else (Brynjolfsson, Li and Raymond, 2023). It could do that because it was trained on, and working inside, the firm’s own conversations.

The opposite is also visible. In a study with 758 BCG consultants, people using AI did more work, faster and at higher quality on tasks inside the model’s capability, but were 19 percentage points less likely to reach the right answer on a task outside it (Dell’Acqua et al., 2023). Part of what puts a task outside the frontier is information the model does not have. MIT’s NANDA initiative reported in 2025 that most organisations it studied saw no measurable profit-and-loss impact from generative AI, and it pointed to tools that don’t learn from the business and integrate poorly with existing workflows (MIT NANDA, 2025). That report’s method has been criticised, but its diagnosis matches what many firms describe.

In regulated firms, the gap between what AI could know and what it does know is often wide. The Bank of England and FCA’s 2024 survey found 75% of UK financial services firms using AI, but four of the top five perceived risks related to data, and 46% of firms said they had only a “partial understanding” of the AI they use (Bank of England and FCA, 2024). A model answering a client question from the platform alone, without the adviser’s file note, the complaint log or last week’s email, is answering a different question from the one the client asked.

Many firms don’t fully understand the AI they useUK financial services firms, 2024
Partial understandingof the AI they use
46%
Complete understandingof the AI they use
34%

Source: Bank of England and FCA (2024), Artificial intelligence in UK financial services.

This is where the human with context wins. An AI that sees one system gives an answer that is fluent, fast and incomplete. The person who knows the rest of the story spots what is missing. In that world the AI becomes another source of information for the smartest person to check, not a replacement for their work. Simon offered a test for this in 1971: an information system reduces the demand on an organisation’s attention “only if it absorbs more information previously received by others than it produces” (Simon, 1971). A tool that sees part of the picture usually fails it.

The High Court has already seen the consequences of AI answering without the right sources. In June 2025 it warned that freely available AI tools “are not capable of conducting reliable legal research”, after cases cited authorities that did not exist, and said leaders could expect the court to ask whether they had met their responsibilities (Ayinde v Haringey [2025] EWHC 1383 (Admin)).

The implication for firms is uncomfortable. Buying a better model matters less than connecting it, substantively, to the firm’s knowledge: the client record, correspondence, meeting notes, policies, research and the regulatory updates people are expected to know about. Microsoft describes the firms that get this right as “Frontier Firms”, organised around on-demand intelligence and hybrid teams of people and agents. In its 2025 survey, 36% of leaders expected managing agents to be part of their job within five years (Microsoft, 2025).

Giving AI access to everything raises obvious questions about control, and the regulator has been clear about who answers them. The FCA has said it will not introduce extra AI-specific rules and will rely on existing frameworks, including the Senior Managers and Certification Regime and the Consumer Duty (FCA, 2025). Its 2024 AI Update states that any use of AI in a business area “would fall within the scope of a SMF manager’s responsibilities” (FCA, 2024). Its October 2026 review of good practice recommends that firms register their AI agents in a central inventory so they are easier to oversee, and warns against human review becoming “a ’box-ticking’ exercise, rather than meaningful oversight” (FCA, 2026). The FCA’s Chief Data, Information and Intelligence Officer, Jessica Rusu, said the same month that “agentic systems therefore require continuous, conditional assurance”, and criticised firms that “retrofit governance after technology has already been deployed” (Rusu, 2026).

So the answer is not to open every system to every tool. It is to give AI governed access to the knowledge it needs, with permissions, logging and clear limits on what it may do on its own. That is a design decision, made up front.

There is a human risk too. Research from Microsoft and Carnegie Mellon found that knowledge workers who trusted AI more did less critical thinking (Lee et al., 2025), and a study in The Lancet Gastroenterology & Hepatology found doctors’ detection rates in procedures without AI fell after AI support was introduced, from 28.4% to 22.4% (Budzyń et al., 2025). If the smartest person stops exercising their judgment, the firm loses its best check on the machine.

My conclusion is simple. AI will not beat the smartest person in the room by being cleverer. It will beat them, at gathering and connecting information, only when it can see as much as they can, and ideally more. Until then, the person with the context will keep outperforming it, and firms will keep wondering why their AI pilots disappoint.

When it can see that much, the human role changes rather than disappears: less time as the firm’s switchboard, more time judging what matters, challenging plausible answers and owning decisions. The smartest person in the room becomes the person overseeing the smartest system in it.

References