ADL Brain: agentic AI for the world's first consulting firm

Our Hult MBA capstone, the Business Challenge, paired a team of six of us with Arthur D. Little (ADL), the world's first management consulting firm. The question was how a 130-year-old advisory firm should adopt AI, and how it should measure whether that is working. Over five deliverables we went from a diagnosis to an execution-ready Strategic Playbook, a pitch to ADL, and a clickable prototype of the product we proposed: ADL Brain.

My role: I was part of building the technology behind ADL Brain and the prototype, and worked on the pricing and the strategy: the subscription and value-share model, and the roadmap it sits in.

1. What is the problem statement?

AI isn't just making consulting faster. It's eating into how consulting makes money. Consulting firms are built as a pyramid: a few partners, many junior analysts, and fees driven by how many analyst days go into research, benchmarking, financial modelling and slide decks. That is exactly the work AI now compresses most. We estimated 60–80% of it can be automated.

• For ADL: the revenue base of the pyramid is shrinking, and competitors are already building their own AI platforms. ADL needed a response before its clients forced one.
• For the market: AI is also changing startups. There will be more of them, with smaller teams and less funding each, and many Series A/B companies can't afford a traditional consulting project.
• The brief: find where ADL should play, what AI should do and what people should still own, how to charge for it, how to reorganise the team, and which metrics prove it's working.

The work AI compresses first, research, benchmarking and first-draft modelling, is exactly the work consulting pyramids bill for.

2. What are the tradeoffs considered?

• Defend the current model vs build a new one. Protecting the existing project business felt safer, but it cedes the AI-native market to others. We chose to build a separate startup practice, with its own profit and loss ring-fenced for 18 months, so partners aren't penalised for cannibalising their own work.
• Which startups to serve first. We first assumed early-stage (pre-Series A) startups. A life-sciences corporate VC told us top-tier VCs already mentor that stage. So we targeted the underserved mid-segment (Series A/B) and pushed pre-Series A to a later phase.
• Project fees vs subscriptions. A VC partner told us investors won't let portfolio companies sign traditional consulting project fees. That ruled out the classic model. We moved to a monthly subscription plus a share of the value created, tied to growth in the startup's valuation.
• Affordable vs profitable. Our first price point looked attractive but lost money on every client once we loaded in real staff costs. We rebuilt pricing cost-first, with a target gross margin of at least 60%, and three tiers that grow with the startup.
• What AI does vs what people do. Clients trust individuals, not models. So AI surfaces the analysis, and a partner countersigns every client-facing recommendation, without exception. ADL keeps professional liability either way.
• Build, partner or buy. Building a foundation model is cost-prohibitive, so the core language model comes from an enterprise AI partner. The retrieval layer over ADL's archive and the model fine-tuning stay in-house, because that knowledge is ADL's moat. Compliance tooling and market-data licences are bought.
• One sales team vs sector pods. A founder we interviewed was blunt: a generalist can't credibly sell sector-specific advice. So we proposed a dedicated sales team per sector, starting with Healthcare and Energy, and adding Automotive later.

3. What is the solution proposed?

ADL Brain: an AI assistant built on ADL's own case archive, paired with a consultant who acts as a guide. It gives mid-segment startups a way to afford top-tier strategic advice. It also builds the AI delivery skills every ADL engagement will eventually need.

1. A five-stage engagement model. Discovery is led by people. Brain setup is AI plus people. Day-to-day delivery is AI-powered: market sizing, go-to-market, modelling, pitch scaffolding and KPI tracking. Monthly check-ins and quarterly milestone reviews are led by people. Each stage has a defined split between AI and human work.
2. A new business model. We moved from one-off project fees to recurring subscriptions in three tiers (Ignite, Accelerate, Scale). For later-stage startups, a "value-added incubator" option shares in the upside through introductions to investors and enterprise clients.
3. A new talent model. The pyramid flattens. Analysts reskill into a new AI Orchestrator role: part prompt engineer, part output curator, part quality checker. Partners focus on judgment and the relationship with the client.
4. Governance built in from day one. Each client gets their own isolated Brain instance. Partners countersign every recommendation. The Brain is trained only on ADL's archive, and data is handled to SOC 2 standards. Client data is deleted 90 days after the engagement ends.
5. A KPI system that decides whether it's working. Metrics span four areas: productivity, margin, quality and trust. The pilot has a 90-day gate that must be passed before a wider launch: client retention, Net Promoter Score and unit economics.
6. An 18-month roadmap. Four phases, in the order vision, structure, people, capital: invest and build, pilot with one hand-picked startup, scale in year one, then expand. Each phase has a gate that must be passed before the next begins.

We validated every major choice with five practitioner interviews: an ADL healthcare consultant, a VC partner, a life-sciences corporate VC, a startup founder and an IT-services industry advisor. Where the evidence disagreed with us, we changed the design. Two of our five starting hypotheses were rejected outright. We also built a working prototype of the client portal: adl-alchemy.netlify.app.

4. Any other improvements I could have done?

• Test it with real users sooner. Five interviews shaped the design, but a small paid pilot or a waitlist test would show real demand and price tolerance far better than conversations can.
• Keep one version of the numbers. Our pricing and financial assumptions changed between the pitch deck and the final playbook. A single versioned model feeding both would have kept the story consistent.
• Measure the AI's quality objectively. I'd define an evaluation set drawn from ADL's past cases, to score accuracy and the rate of made-up answers, instead of relying mainly on partner review.
• Plan the people side more carefully. Reducing some roles to zero is a big change for real people. A staged transition, with clear career paths into the AI Orchestrator role, would reduce the risk of losing talent.
• Go deeper on regulation. Healthcare and EMEA expansion need sector-by-sector legal work before launch, not just a rule keeping the Brain away from clinical advice.
• Model cannibalisation. I'd quantify how much existing enterprise revenue ADL Brain might pull away from traditional projects, not just the new revenue it brings in.