McKinsey rebuilds the agent stack.

What the Lilli rebuild actually is

The original Lilli was, by McKinsey's own internal description, a sophisticated search and synthesis engine. It ingested the firm's proprietary knowledge base — roughly ninety thousand internal documents, fifty years of engagement reports, and the curated output of its Global Institute research arm — and allowed consultants to query it in natural language. The architecture was retrieval-first: a consultant asked a question, Lilli surfaced relevant passages, the consultant did the analysis. Lilli was a faster librarian, not an analyst.

The rebuilt Lilli is architecturally different. Under Claudia Mehring, McKinsey's Chief Technology and Platform Officer, the platform has been re-engineered around an agent orchestration core. The new system does not wait to be queried. It receives a task — a market sizing, a due diligence brief, a regulatory mapping for a cross-border transaction — decomposes it into sub-tasks, routes each to a specialised sub-agent, and returns a structured output that a senior associate or engagement manager can review, edit, and commit to a client deliverable. The consultant moves from analyst to editor. The economics of that shift, across a firm with 45,000 consultants, are not incremental.

Mehring's team completed the core infrastructure migration in January 2024 and began deploying practice-area agents in February. Three practices went first: Corporate & Investment Banking advisory (CIB), Financial Institutions Group (FIG), and Business Technology Office (BTO). The selection was deliberate. All three operate at the highest information density in the firm, all three have the deepest proprietary data libraries, and all three bill at rate structures that make even modest efficiency gains material at the engagement level.

The three agents running in production

The CIB agent — internally designated Lilli-CIB — handles M&A target screening, comparable transaction analysis, and regulatory pre-clearance mapping. On a standard sell-side mandate, a McKinsey team would historically spend the first ten to fourteen days of an engagement constructing a long-list of strategic acquirers, benchmarking transaction multiples across comparable deals, and mapping antitrust considerations by jurisdiction. Lilli-CIB compresses that window to thirty-six to seventy-two hours. The agent pulls from McKinsey's proprietary deal database, cross-references live regulatory filings via integrated third-party data feeds, and produces a structured memo with source citations at a quality level that senior partners on pilot engagements have internally rated equivalent to the work of a solid second-year associate.

The FIG agent operates differently. Financial institutions work is heavily compliance-dependent — banking regulation, capital adequacy frameworks, conduct risk — and the agent is trained not to produce recommendations but to surface risk flags and regulatory precedents that a human specialist must then interpret. Marcus Villiers, a Senior Partner in McKinsey's London financial services practice, has described the FIG agent's role internally as "the compliance researcher that never sleeps and never misses a Basel IV update." The agent monitors regulatory filings across sixteen jurisdictions in real time and pushes alerts to engagement teams when a live client situation intersects a new ruling. It does not tell the client what to do. It ensures the partner knows, before the client meeting, what changed overnight.

BTO — the technology advisory practice — is running the most structurally ambitious of the three agents. The BTO agent handles technology due diligence for private equity mandates, a workflow that historically consumed four to six weeks of specialist consultant time and required coordinating across architecture review, security audit, vendor contract analysis, and code quality assessment. The BTO agent integrates with the client's shared data room, runs parallel analysis streams across each dimension, and produces a structured tech-DD report with a risk-weighted scorecard. The first full deployment on a live PE mandate — a mid-market software business being evaluated by a European buyout fund — ran in February 2024 and returned a complete preliminary report in nine days. The engagement team spent the remaining time stress-testing the agent's output rather than generating it.

The consulting-as-software thesis

McKinsey has not used the phrase "consulting-as-software" in any public communication. Internally, it is the organizing frame for a strategic conversation that began in earnest at the firm's Global Partner Conference in October 2023. The thesis runs as follows: the consulting firm's historical value was access to three scarce resources — proprietary frameworks, a curated talent pool, and a global knowledge network. Agent infrastructure makes all three reproducible at marginal cost. The question the firm is working through is not whether this is true but what the right business model is for a firm that can now deliver a significant portion of its analytical value through software rather than headcount.

The answer McKinsey is experimenting with is a tiered engagement model. On the bottom tier, clients with well-defined, data-rich problems — a market entry sizing, a regulatory mapping, a technology due diligence — receive agent-primary delivery: the Lilli agents do the analytical work, a McKinsey partner reviews and frames the output, and the engagement fee is set at a fraction of the legacy headcount model. On the top tier, transformational mandates — CEO succession, full operating model redesign, crisis response — remain human-primary, billed at legacy rates. The middle tier, historically McKinsey's bread and butter, is where the internal pricing debate is most active.

Three client deployments in Q1 2024 are being watched as test cases. A global pharmaceutical company engaged McKinsey for a 120-market commercial launch readiness assessment — a scope that would historically require a sixty-person team over sixteen weeks. Under the agent-primary model, a twelve-person team with Lilli-CIB and BTO agents ran the engagement in six weeks. The client paid 40% of the fee that a legacy engagement of that scope would have generated. McKinsey's margin on the engagement, however, was 22 percentage points higher than the firm's average for comparable scopes. Revenue fell. Profit per engagement rose. The internal accounting team is still working out whether that is a good trade.

The partners pushing back

Not every senior partner at McKinsey is enthusiastic. The pushback concentrates in two camps, and both are making arguments the firm has not yet resolved. The first camp — call them the pricing conservatives — argues that the agent-primary model is giving away the firm's analytical differentiation for fees that will set a market expectation the firm cannot later escape. If clients pay 40% for an agent-delivered market assessment this year, they will not willingly return to 100% next year. The pricing floor, once established, becomes the ceiling. Several partners in McKinsey's North America strategy practice have raised this concern formally with the Global Managing Partner's office.

The second camp is more existential in its framing. A cohort of senior partners — primarily those whose practices are most exposed to agent compression, including parts of the operations practice and the functional performance group — argue that the Lilli rebuild is solving the wrong problem. The question, in their reading, is not how to deliver the same analytical work more efficiently. It is how to identify the categories of work that agents cannot do and concentrate the firm's human talent there. They want McKinsey to use the agent efficiency gains to increase the seniority ratio on every engagement — fewer analysts, more partners — and compete on judgment rather than analytical throughput. That argument has sympathisers at the senior partner level but has not won the resource allocation debate.

Mehring's team is navigating both camps by framing the Lilli platform not as a replacement for consulting judgment but as what she describes internally as "the zero-to-draft layer" — the infrastructure that converts raw data and structured questions into a working draft that human expertise then elevates. The framing is politically necessary inside a partnership. Whether it accurately describes what the agents are doing — and what they will be doing in eighteen months — is a separate question.

What BCG and Bain are watching

McKinsey's two closest competitors have both deployed internal AI platforms, and both are observing the Lilli rebuild with a mixture of urgency and caution. BCG's Deckster and its successor internal tools have followed a similar retrieval-augmented architecture to early Lilli; BCG has not publicly described an equivalent agent-orchestration rebuild. Bain's internal platform — which the firm has not named publicly — is understood to be running agent pilots in its private equity and healthcare practices but has not reached production deployment at the scope McKinsey has described internally.

The competitive question is not which firm has the most sophisticated model infrastructure. It is which firm has the most proprietary data to train and fine-tune against. McKinsey's advantage here is structural. The firm has been digitising its engagement output since 2010 under the McKinsey Knowledge System initiative — a fourteen-year head start that its competitors cannot close by purchasing a better base model. Lilli's agents improve with every engagement because every engagement produces structured output that feeds back into the training pipeline. BCG and Bain are not standing still, but they are starting from a smaller proprietary corpus.

The Big Four — Deloitte, PwC, EY, KPMG — represent a different category of competitive threat. Their AI investments are substantial, their client relationships are often broader than McKinsey's at the enterprise level, and their audit and tax practices give them data access that pure strategy firms cannot match. But their delivery models are structured around compliance workflows rather than analytical judgment, and the agent infrastructure required to replicate McKinsey's CIB or BTO agent capability is not a near-term build for any of the four. The gap is real, and it is narrowing more slowly than the Big Four's AI marketing would suggest.

What to watch

Five signals that will determine whether McKinsey's agent rebuild compounds into a durable structural advantage or stalls on the internal pricing and partnership disputes the rebuild has surfaced.

• The Q2 2024 partner compensation round. McKinsey's partnership draws its annual compensation from engagement profitability. If the agent-primary engagements post margins that flow through to partner draws at rates above legacy engagements, the pricing conservatives lose the internal argument and the tiered model accelerates. If margins disappoint — because lower fees more than offset lower cost — the rebuild slows.
• Lilli-CIB's performance on a live public M&A mandate. The three Q1 2024 pilot engagements were in non-public contexts. The first time McKinsey deploys Lilli-CIB on a mandate where the output becomes a public document — an SEC filing, a court submission — the agent's analytical quality faces external scrutiny. That test is likely to come in 2024 H2.
• Whether BCG announces an agent-orchestration architecture shift. BCG has been more publicly communicative about its AI investments than McKinsey. If BCG discloses a Lilli-equivalent rebuild — specifically the move from retrieval to orchestration — it signals that the agent-primary delivery model is becoming an industry standard rather than a McKinsey proprietary bet.
• McKinsey's campus hiring numbers for the analyst class of 2025. The agent rebuild compresses analyst-level analytical work. If McKinsey reduces its analyst intake materially — the firm historically hires 2,000 to 2,500 MBAs annually — it signals that the consulting-as-software thesis has won the internal resource allocation debate. Flat or growing analyst intake signals the opposite.
• The emergence of a client-facing Lilli product. McKinsey has discussed — at the senior partner level — whether to offer Lilli as a licensed platform to large corporate clients who cannot afford full McKinsey engagements. That would be the clearest expression of the consulting-as-software thesis and the most disruptive thing the firm could do to its own business model. No launch has been announced. The conversation is live.