AI Agents in Consulting: How McKinsey, BCG, Bain and Other Firms Are Using Them

In January 2026 Bob Sternfels, McKinsey's Global Managing Partner, told the HBR IdeaCast that his firm had 60,000 people: 40,000 humans and 20,000 agents. McKinsey later corrected the agent figure to 25,000. Two years earlier the same firm was running pilots. It now counts software in its headcount and updates the number monthly.

Behind figures like these are three questions: what the large firms have deployed, what the one firm that has published a payoff number reports, and why the agent count is a weak measure of any of it.

Chatbot or agent: the difference that matters

A chatbot answers the question it is asked. An agent is given an objective and works out the steps: search the knowledge base, pull the dataset, draft the document, check it against the source, return a result. For a consultant this is the difference between asking a question and delegating a task. It also changes the risk. When a system runs six steps on its own, a mistake in the first step has travelled a long way before a human reviews it.

McKinsey: Lilli, then 25,000 agents

McKinsey is the clearest case of a firm building its own stack, and it started early. Lilli, launched in August 2023, searches and synthesises the firm's internal knowledge across more than 100,000 documents. The AI work runs through QuantumBlack, which has more than 1,700 staff.

“Little over a year and a half ago, that [the number of agents deployed by McKinsey] was 3,000 agents and I originally thought it was going to take us to 2030 to get to one agent per human. I think we're going to be there in 18 months and we'll have every employee enabled by at least one or more agents.”^[1]

The current figures, as McKinsey reports them: a workforce of 60,000 including 25,000 AI agents (the firm calls this the accurate number, revised up from the 20,000 first cited), and a target of every employee enabled by at least one agent within roughly 18 months.^[2]

BCG: Deckster, GENE and 18,000 custom GPTs

BCG took a broader approach. Consultants use Deckster for drafting slides and GENE, a conversational assistant the firm uses for presentations, podcasts and outreach. BCG then rolled out ChatGPT Enterprise across the firm, and its consultants built more than 18,000 custom GPTs for research, slide production and routine queries. That figure is more informative than the platform names: it means thousands of consultants building their own small tools rather than a central team shipping a product for others to adopt.

BCG is also the only one of the three to have put a payoff number on the record. Its consultants spend about 15% less time on low-value work such as building slides, and reinvest roughly 70% of the saved hours in higher-value work such as deeper analysis. BCG X, the unit that builds custom AI systems for clients, has more than 3,000 employees.^[3]

Bain: buying rather than building

Bain chose to buy. Sage, built with OpenAI, lets consultants draw on Bain's own intellectual property. The firm also runs a customised version of ChatGPT and has more recently added Claude. Technology- and AI-enabled work was about 30% of Bain's business in 2025, with more than 2,500 AI projects delivered to date.^[4]

The build-or-buy choice reflects where each firm believes its advantage lies. McKinsey's position is that the connection between its knowledge and its people is the defensible asset. Bain's is that the models will become commodities within a few years, and that the advantage lies in the proprietary data fed to them and the people using them. Buying has one drawback: it leaves the firm dependent on a supplier that, as set out in are AI labs becoming consulting firms?, is turning into a competitor. Either way, the choice determines who the firm hires and how much of its position depends on the labs.

Beyond the big firms: agents for boutiques

Smaller firms use agents too. Tools such as Anthropic's Claude have put them within reach of small specialised boutiques. Queen's Tower Advisory runs what it calls an 80/20 model, 80% agents and 20% humans, and grows by adding agent capacity rather than people. Agents are also becoming a product in their own right: firms sell them to clients, and many start with one consultant solving one client's problem and adding the tool to the firm's repository for the next team that needs it.

Side by side: how the firms approach agents

From how many agents to how much value

The way firms assess agents is changing as well. In corporates as in consulting, the question has moved from how many agents exist, or how many people log in, to what the agents produce. PwC's Chief AI Officer Dan Priest states the position directly: people still run the workforce, and an agent is best measured by how well people use it rather than by what it could automate in theory.

“The human is still accountable. The humans are the ones who get certified. The humans are the ones who get licensed. The humans are the ones who get empowered.”^[5]

What this means for consulting firms

Agent counts say little on their own. Agents are cheap to create, so a firm can report thousands of them without any change in how it serves clients. BCG's 15% and 70% figures are more informative because they show where the freed time went. A firm that saves the time and simply bills fewer hours has automated away part of its own revenue.

Accountability stays with people. Every firm quoted here keeps a human responsible for the output, and after the hallucinated-citation episodes of 2025 and 2026 that is a minimum requirement rather than caution. More automation means more to check.

The skill gap also moves up the pyramid. Once agents do the production work, the scarce skill is directing them: framing the task, judging what comes back, and knowing what to delegate and what to keep. Firms tend to treat this as a technology roll-out. It is mainly a training problem.

This article expands chapter 4^[6] of The State of AI in Consulting 2026, which also covers the new competitors, the reshaping of the pyramid, revenue and pricing, hallucinated citations and the EU AI Act.