AI Agents for Consulting Firms: The Jagged Frontier Is Now a Client Question

Consulting firms sell expertise, and for a long time the price of that expertise tracked the hours spent producing it. Generative AI unsettles that arrangement from two directions at once. It makes some of the work much faster, and it makes some of the work quietly worse.

Three primary sources, from a field experiment, a firm's own rollout, and a public refund, show both sides.

The experiment run inside a consulting firm

The best-known study of AI and consultants was run with Boston Consulting Group itself. The paper, "Navigating the Jagged Technological Frontier" by Fabrizio Dell'Acqua, Ethan Mollick, Karim Lakhani and colleagues, started as a Harvard Business School working paper in 2023 and is now published in Organization Science (March-April 2026).

The preregistered experiment involved 758 BCG consultants. After a baseline, they were randomly assigned to work with no AI, with GPT-4, or with GPT-4 plus a short prompt-engineering overview.

On 18 realistic consulting tasks that the researchers judged to be inside AI's capabilities, consultants using AI:

• completed 12.2% more tasks,
• finished them 25.1% faster on average, and
• produced work of significantly higher quality.

But on one complex business problem chosen to sit outside that frontier, consultants using AI were 19% less likely to reach the correct answer than those working without it.

The authors call this a "jagged frontier." Tasks that look equally hard to a person can fall on opposite sides of what the AI does well. The danger is not that AI is bad at consulting work. It is that the drop-off is invisible from the outside, and confident, polished output can hide it.

What a firm-wide rollout looks like

McKinsey has published its own case study of Lilli, the internal generative AI platform it rolled out firm-wide in July 2023. According to McKinsey:

• 72% of the firm is active on the platform,
• colleagues report up to 30% time savings in searching for and synthesizing knowledge, and
• Lilli answers more than 500,000 prompts a month.

The use case is telling. Lilli's main job is pulling together the firm's own knowledge from more than 40 sources. That is squarely inside the frontier: retrieval and synthesis from a known body of material. The same case study names the next step as "agents built to automate specific time-consuming tasks" and expanded slide building.

It also notes that "there's more work to do to boost regular engagement," which is a candid reminder that even a firm that sells AI transformation has to work at adoption internally.

What happens when the frontier is crossed in a client deliverable

In October 2025, Australia's Department of Employment and Workplace Relations confirmed that Deloitte would repay the final instalment of its contract for a A$440,000 assurance review, as reported by The Guardian. Researchers had found errors in the report, including references to nonexistent academic work and a misstated summary of a Federal Court case.

The revised version added a disclosure that part of the report had used a generative AI tool chain based on Azure OpenAI GPT-4o. The department said the substance and recommendations were unchanged, and Deloitte said the matter had been resolved with the client.

The point for every firm is not that one report had errors. It is the question that followed. Labor senator Deborah O'Neill said anyone contracting these firms "should be asking exactly who is doing the work they are paying for." Clients now have a reason to ask how AI was used, and a firm needs a credible answer.

Where AI agents fit in a consulting practice

Read together, the three sources sketch a practical map. AI agents are most useful where the work is retrieval, assembly, and checking, and riskiest where the work is judgment that only looks routine.

Good fits:

• Knowledge retrieval. Finding past proposals, case examples, benchmarks, and experts across internal sources, as Lilli does.
• First drafts of standard material. Proposal boilerplate, team bios, workplans, and meeting summaries.
• Citation and source checking. Verifying that every reference in a draft exists and says what the draft claims. The Deloitte case is an argument for running this check on every deliverable, AI-assisted or not.
• Engagement admin. Status reports, timesheet reminders, and pulling together client data requests.

Riskier ground:

• Final recommendations on messy, ambiguous problems, the kind the BCG experiment deliberately placed outside the frontier.
• Any claim that goes to a client without a named person who checked it.

What this means for consulting firms going forward

The jagged frontier is now a commercial issue, not just a productivity one. If AI makes some tasks 25% faster, clients will notice and ask how that shows up in fees. If AI makes some tasks less accurate, clients will ask how the firm caught it.

Firms that come out ahead will likely be the ones that can say precisely which parts of an engagement use AI, which parts are reviewed by whom, and how sources are verified. That is less a technology choice than an operating discipline, and it is one clients have started to expect.

FAQ

Does AI make consultants more productive? In the Harvard and BCG field experiment with 758 consultants, those using GPT-4 completed 12.2% more tasks, 25.1% faster, with higher quality, on tasks within AI's capabilities.

What is the "jagged frontier"? It is the researchers' term for AI's uneven abilities. On a task outside that frontier, consultants using AI were 19% less likely to get the correct answer.

How widely does McKinsey use AI internally? McKinsey says 72% of the firm is active on its Lilli platform, which answers more than 500,000 prompts a month.