Building effective enterprise agents
B · Useful, limited BCG AI Platforms Group engineering briefing drawn from its client agent builds; concrete practice on design, evaluation and platforms, but outcome metrics are labelled illustrative and come from unnamed clients.
Summary based on the full text·Original: about 40 minutes to read
In short
A November 2025 briefing from BCG's AI Platforms Group on building agents that work in large companies with old systems, messy data and heavy governance. Its central claim is that the limit is legacy systems and processes, not model capability. It recommends starting with the simplest agent that meets a business outcome, adding sub-agents only when failures show the need, and building evaluation harnesses and shared platform services before scaling.
Key numbers
- 300+agents BCG says it delivered across clients
- 75%technology leaders who fear 'silent failure', spending heavily on AI without real impact (BCG market research)
- ~50 to ~75F1 score on insurance entity extraction from sprint 2 to sprint 6 after adding a testing harness, RAG and tuned OCR
- 5blockers seen in two years of builds: brownfield integration, unreliable data, missing evaluations, governance overhead, scale frictions
As reported by the source.
English·PDF·opens media-publications.bcg.com
What it reports
- Five blockers from two years of builds: brownfield integrations, unreliable enterprise data, lack of evaluations, governance and audit overhead, and operating-model frictions when moving from proof of concept to production.
- A horizon model from constrained agents (at scale today) through single and "deep" agents (orchestrator plus specialists, recommended now) to role-based teams and self-organising meshes (early R&D).
- An agent suitability framework: use agents where goal and environment complexity are high and oversight needs are manageable; use rules or plain automation otherwise.
- Outcome-first design, with a loan-processing example targeting 30% faster approvals and 5% lower processing cost, decomposed into agent opportunities.
- Autonomy tiers for rollout: shadow (agent suggests, human acts), supervised (human approves), guided (human monitors), full autonomy.
- 14 platform building blocks, including memory, an AI gateway for model switching and cost control, context engineering, LLMOps observability and failure-mode guardrails.
- Insurance example: entity extraction F1 rose from about 50 to about 75 over six sprints once a testing harness was in place.
How they got it
Practitioner synthesis from BCG client work, with references to LangGraph, LangChain and METR. Client results (30-50% faster decisions, 30%+ shorter coding cycle times, 30-40% productivity uplift) are marked as illustrative, implementation-dependent and not guaranteed.
Limits
- Client outcomes are anonymised and unmeasured in the document.
- Slide-format briefing; some diagrams lose meaning without narration.
- Points to BCG's own assets, such as Agentic Outcome Maps and Agent Design Cards.
Why it matters for companies
The value is in the engineering discipline: define the business outcome, pick the least complex agent design that meets it, and do not deploy without an evaluation harness and pass/fail gates. The autonomy tiers give risk and business owners a shared language for how far an agent may act on its own at each rollout stage. Integration with legacy systems and data quality will take more of the budget than model choice.
Summary based on the full text. Read the original for the full argument and data.Open PDF on media-publications.bcg.com
- Tags
- AI agents, Evaluation