How a top-five consulting firm rebuilt its proposal pipeline.

The architecture: Claude Sonnet 4.6, an internal RAG, and a 30-year archive

The architecture that BCG has built operates as a retrieval-augmented generation system grounded against the firm's proprietary proposal archive — a corpus that the firm has indexed across approximately 30 years of submitted proposals, the corresponding engagement outcomes, the client-side feedback that the firm has collected systematically since 2008, and the broader case-study and methodology archive that supports the firm's strategy practice. The corpus runs to approximately 14 million documents across multiple languages and document formats, and the indexing work was the largest single technical effort inside the rebuild. The team — initially eleven engineers and three data architects under Mohammadi's direction, expanded to a peak of forty-three engineers and seven data architects during the indexing phase in Q2 and Q3 2025 — processed the corpus through a multi-pass extraction pipeline that combined document-format normalisation, entity extraction, semantic tagging, and the construction of a graph database that links proposals to engagements to outcomes to client-feedback artefacts.

The retrieval layer runs on Cohere's Embed 4 model for the vector index and a separate Anthropic-Claude-based retrieval reasoning layer that the team built specifically to handle the structured-graph queries that the proposal generation workflow requires. The choice of Cohere over OpenAI's embedding API or Voyage AI's offerings was operational: Cohere's enterprise data-residency posture allowed BCG to host the embedding index inside its own infrastructure rather than in a vendor-managed environment, which was the principal procurement requirement that the firm's partner-level governance had imposed on the program at its authorisation. The Claude-based retrieval reasoning layer — internally called BCG Recall — is a distinguishing architectural choice that the team's engineering leadership has been quiet about externally but that is the operational mechanism by which the system surfaces the right historical reference material for the proposal-generation step. The Recall layer makes a focused Claude inference call per retrieval request, with the call constrained to the retrieval-reasoning task and the inference output structured to produce a ranked list of source documents that the downstream generation step grounds against.

The generation layer operates on Claude Sonnet 4.6 with the retrieval-grounded context, and produces structured proposal drafts that conform to a templated proposal architecture that the firm has been operating since approximately 2014. The templated architecture — a set of standardised proposal sections, response formats, and engagement-shape descriptions that BCG uses across the bulk of its proposal pipeline — was the operational substrate that the rebuild was constructed against. The team's first design decision in the rebuild was to preserve the templated architecture rather than redesign it, on the operational logic that the partner-level expectations of what a BCG proposal looks like were stable and that the rebuild's value would come from filling the templates faster and better rather than reinventing the template itself. The decision has been validated by the proposal-stage win rate improvement: BCG proposals continue to look like BCG proposals, the templated architecture's familiarity to repeat clients has been preserved, and the AI-augmentation has produced the operational acceleration without disrupting the client-side expectations.

The partner-level governance: every proposal still partner-signed

The partner-level governance that BCG has built around the proposal pipeline is the structural mechanism by which the firm has preserved the partnership-firm operating model against the operational risks that AI-augmented proposal generation introduces. The governance framework — formalised at the October 2024 partner meeting and updated at the May 2026 Vienna conference — establishes three operational disciplines that every AI-augmented proposal must clear before submission. The first discipline is the named-partner sponsorship requirement: every proposal that uses the AI-augmented pipeline must be sponsored by a named partner who holds operational accountability for the proposal's content, the engagement's eventual delivery, and the client relationship's broader trajectory. The partner sponsor is not a sign-off; the partner sponsor is the proposal's named author for internal accountability purposes, regardless of how much of the proposal's text was AI-generated. The discipline is the structural defence against the risk that the AI-augmented pipeline becomes a proposal-mill that decouples the proposal from the partner who will be responsible for delivering against it.

The second discipline is the partner-review requirement at three defined checkpoints in the proposal pipeline. The first checkpoint is the proposal-shape review, where the partner sponsor reviews the AI-augmented proposal outline before any drafting begins; the second is the technical-content review, where the partner sponsor reviews the AI-augmented drafted content against the firm's methodology and the client's stated requirements; the third is the final-submission review, where the partner sponsor reviews the complete proposal artefact before submission. Each checkpoint requires the partner sponsor's affirmative approval and is logged in the firm's proposal-pipeline audit infrastructure. The three-checkpoint discipline produces a partner-time investment per proposal that the firm tracks against the historical baseline: the post-rebuild partner-time investment is approximately 11 hours per proposal against a pre-rebuild baseline of approximately 38 hours per proposal. The partner-time reduction is the operational mechanism by which the program produces the 31 per cent reduction in proposal-team staffing requirements that BCG disclosed at Vienna.

The third discipline is the partner-level disclosure requirement on the firm's own AI usage. Every BCG proposal that uses the AI-augmented pipeline now includes a section — internally called the Methodology Note — that discloses to the client the specific AI capabilities used in the proposal's preparation, the human-review checkpoints applied, and the partner-level accountability for the proposal's content. The Methodology Note is structured to be precise without being overwhelming: it runs to approximately 350 words across a single proposal page, and is the operational mechanism by which BCG addresses the client-side concern about AI use in consulting deliverables. The disclosure approach is the structural argument BCG makes to its clients that the AI-augmented pipeline is a transparency-positive change rather than an opaque acceleration. The Methodology Note's content has been refined through three iterations across the rebuild's deployment, with the current language reflecting client-side feedback from approximately 280 engagements that have been delivered against AI-augmented proposals in the trailing twelve months.

Win rate and the attribution problem

The 18-percentage-point improvement in proposal-stage win rate is the headline metric that BCG disclosed at Vienna and is the one that the firm's competitors will scrutinise most carefully. The number is meaningful only against a clearly-defined baseline, and BCG's disclosure included the methodological notes that the figure's interpretation requires. The pre-rebuild baseline was the trailing-twenty-four-months proposal-stage win rate across the firm's strategy-and-management-consulting practice, measured as the percentage of submitted proposals that resulted in awarded engagements. The post-rebuild measurement is the trailing-six-months win rate across the same practice, against proposals that used the AI-augmented pipeline. The 18-percentage-point improvement is the difference: the pre-rebuild baseline was approximately 32 per cent and the post-rebuild figure is approximately 50 per cent.

The attribution problem — separating the AI-augmented pipeline's contribution to the win rate from the broader macro and competitive factors that affect proposal outcomes — was the methodological work that the firm's analytics team conducted across the rebuild period. The team's approach was a matched-cohort analysis: proposals submitted under the AI-augmented pipeline were matched against historical pre-rebuild proposals that were comparable on client industry, engagement type, proposal value, competitive intensity, and the partner sponsor's prior win-rate baseline. The matched-cohort win rate differential — controlling for the matched variables — was approximately 14 percentage points, against the gross 18-point figure. The difference between the gross and the matched figures is the proportion of the improvement that the analytics team attributes to the broader competitive environment rather than the rebuild itself. The 14-point matched-cohort figure is the number Mohammadi presented at Vienna as the rebuild's true operational attribution.

The mechanism by which the rebuild produces the win-rate improvement is the question the analytics team has spent the most effort understanding. The team's working hypothesis — corroborated by client-side post-decision interviews that the firm has conducted across approximately 80 engagements in the trailing year — is that the AI-augmented pipeline produces proposals that are more precisely tailored to the client's specific context, that surface more relevant historical reference material from BCG's archive, and that demonstrate a deeper understanding of the client's industry context than the pre-rebuild proposals achieved at the same preparation-time budget. The retrieval-augmented architecture is operationally the mechanism: by grounding the proposal generation against the firm's 30-year archive of relevant proposals, engagements, and outcomes, the system surfaces the historical material that the partner-sponsor would have used in the proposal if the partner had the time to search the archive thoroughly, but that the partner could not surface within the proposal-preparation budget. The win-rate improvement is, by the team's analysis, primarily a reach-into-the-archive improvement rather than a generation-quality improvement.

Named client wins: the engagements attributable to the rebuild

BCG named four specific client engagements at the Vienna conference that the firm attributes directly to the AI-augmented proposal pipeline — engagements where the proposal was prepared on the rebuilt architecture, the win was in a competitive process against named peer firms, and the client has agreed to public attribution. The four engagements illustrate the breadth of the pipeline's deployment across BCG's practice mix. The first is a transformation engagement with a European Tier 1 automotive OEM — Stellantis, which BCG confirmed at the conference — covering the OEM's AI strategy and electric-vehicle-platform consolidation. The proposal was submitted in October 2025 against two named competitors (McKinsey and Bain), and the engagement awarded to BCG in late November. The engagement is the firm's largest single transformation engagement awarded in 2025 and is operationally led by BCG's automotive practice in Munich.

The second engagement is a financial-services advisory engagement with HSBC, focused on the bank's AI-augmented commercial-banking strategy. The proposal was submitted in December 2025 against four named competitors (McKinsey, Deloitte, Accenture, and Oliver Wyman), and HSBC awarded the engagement to BCG in February 2026. The engagement is operationally led by BCG's financial-services practice in London and is the firm's most-referenced post-rebuild proof point for its UK-headquartered enterprise customer base. The third engagement is a public-sector advisory engagement with the Government of Singapore's Ministry of Trade and Industry, focused on the country's AI-industry-strategy roadmap to 2030. The proposal was submitted in January 2026 against three named competitors (Accenture, McKinsey, and PwC), and the Singapore government awarded the engagement to BCG in March 2026. The engagement is led from BCG's Singapore office and is the firm's flagship public-sector win in the trailing twelve months.

The fourth engagement is a consumer-goods advisory engagement with Unilever, focused on the company's category-management AI strategy and the operational integration of AI-augmented decisioning across the company's marketing and supply-chain functions. The proposal was submitted in February 2026 against three named competitors (McKinsey, Bain, and Accenture), and Unilever awarded the engagement to BCG in early April 2026. The engagement is led from BCG's London office in coordination with the firm's Singapore office, given Unilever's regional category-management organisational structure. The four named engagements together represent approximately $340M in initial engagement value, with the longer-term engagement extensions potentially expanding the value substantially. The figure is meaningful against the rebuild's total program cost — which BCG disclosed at Vienna as approximately $190M across the fourteen-month build period — and produces what Mohammadi described as "a payback period measured in single calendar quarters rather than fiscal years."

Staffing implications: the consultant who works on proposals versus engagements

The 31 per cent reduction in proposal-team staffing requirements — measured in partner-equivalent hours per proposal — is the staffing metric the firm has disclosed publicly. The underlying operational change is more complex. The reduction is concentrated in two specific staffing categories: the junior-consultant cohort that historically conducted the research-and-drafting work for the bulk of proposals, and the senior-associate cohort that historically reviewed, restructured, and refined the junior-consultant output before partner review. The AI-augmented pipeline displaces a meaningful portion of both categories' proposal-stage work, with the residual human work concentrated on the strategic-shaping activities that the AI is not well-positioned to handle (client-relationship intelligence, competitive-positioning judgement, the high-context partner-shaped argument) and the partner-review activities that the partner-level governance discipline preserves.

The staffing implications are not headcount-reducing in the way that a manufacturing-process automation would be. BCG's partner conference materials made the framing explicit: the staffing reduction has been redirected from proposal preparation to engagement delivery, with the consultants who would historically have spent time on proposal work now being available for engagement-delivery work. The firm's overall headcount is up approximately 6 per cent against the pre-rebuild baseline, with the growth concentrated in the engagement-delivery cohorts and the AI-engineering cohorts under Mohammadi's organisation. The firm's hiring has been adjusted to the new operational reality: the firm is hiring fewer junior consultants relative to the prior baseline ratio (against the firm's growth rate), and more AI-engineering and analytics talent.

The structural question that the staffing pattern raises is whether the rebuild produces a permanent change in the consultant-career-path arithmetic. BCG's traditional career path has been built around the apprenticeship model: junior consultants learn the craft of consulting by working on proposals and engagements under senior-associate and partner supervision, with the proposal-work in particular being a structured environment for learning the firm's analytical frameworks, methodologies, and presentation discipline. The rebuild removes a portion of the proposal-work from the junior-consultant's experience, which the firm's leadership has acknowledged carries an operational risk to the apprenticeship model. The mitigation that BCG has built into the program is a structured rotation: every junior consultant spends a defined portion of their first eighteen months in a "proposal-craft" rotation that exposes them to the proposal-shaping activities the AI is not handling, including the strategic-shaping work, the partner-review preparation work, and the post-submission engagement-handoff work. The rotation is the firm's structural response to the apprenticeship-model risk.

The implications for the broader consulting industry are the most operationally interesting question that the rebuild raises. BCG's rebuild is, as of May 2026, the most advanced production deployment of AI in the proposal-generation workflow across the top-five consulting firms. McKinsey, Bain, Deloitte, and Accenture all have AI-augmentation programs in their proposal pipelines, but none has disclosed operating metrics at BCG's level of specificity or operational maturity. The trade-press read is that the four peer firms are between six and eighteen months behind BCG on the production-deployment timeline, with McKinsey closest to parity and the others further behind. The competitive implication is structural: if BCG's 14-point matched-cohort win rate improvement is sustained, the firm's win-rate advantage will compound across the proposal-pipeline volume over the next twenty-four months, with the peer firms only catching up as they ship their own equivalent rebuilds. The competitive pressure on the peer firms to accelerate their own rebuilds is now operationally substantial, and the talent and capital allocations that the firms have authorised in 2026 reflect the pressure.

The peer-firm response has been studied closely by BCG's competitive-intelligence function, and the May 2026 review document includes an annex on the peer-firm activity that the firm's leadership has been tracking. McKinsey's equivalent program, internally branded Lilli — the firm's broader internal AI platform — has been extended into the proposal-generation workflow under a program that the firm has publicly described in less detail than BCG's Vienna disclosure. The program's architecture is broadly similar to BCG's, with a retrieval-augmented generation surface grounded against the firm's proprietary archive, but the operational maturity and the partner-level governance discipline have not yet reached the level BCG has demonstrated. Bain's equivalent program, branded Helix, is operationally further behind: the firm's smaller scale relative to McKinsey and BCG has constrained the engineering capacity available for the rebuild, and the firm's program has been more conservatively scoped to focus on the proposal-research workflow rather than the full proposal-generation workflow.

The structural conclusion that BCG's competitive-intelligence function has drawn — and that the May 2026 review document makes explicit — is that the proposal-generation AI rebuild is not a defensive investment but a strategic asset. The rebuild's compounding effect on win rate will, by the firm's projection, produce a sustained competitive advantage across the trailing thirty-six months even after the peer firms ship their equivalent rebuilds, on the operational logic that the BCG proprietary-archive depth, the operational discipline that the partner-level governance has produced, and the engineering team's specific expertise on the retrieval-augmented architecture together represent a competitive moat that the peer firms cannot close through capital allocation alone. The thesis is not yet proven. The next twenty-four months will resolve whether the moat is durable.

The fourteen-month build: how the rebuild actually shipped

The fourteen-month build period that produced the rebuild's production deployment ran from the October 2024 partner-meeting authorisation to the March 2026 enterprise-wide rollout completion, with a four-phase execution structure that the Vienna keynote documented in operational detail. The first phase — the corpus-indexing phase, running from October 2024 to June 2025 — was the longest of the four and consumed the largest engineering capacity. The phase's deliverable was the indexed corpus of 14 million documents, with the indexing methodology designed to support the retrieval-augmented architecture's per-proposal grounding requirements. The phase's engineering work was distributed across three specialised teams: the document-extraction team (responsible for normalising the corpus across multiple document formats and languages), the entity-and-semantic-tagging team (responsible for the metadata layer that supports the retrieval queries), and the graph-database team (responsible for constructing the relational structure that links proposals to engagements to outcomes to feedback).

The second phase — the architecture-development phase, running from January 2025 to August 2025 — overlapped with the corpus-indexing phase and produced the BCG Recall retrieval reasoning layer, the generation orchestration plane, and the integration with the firm's existing proposal-pipeline tooling. The phase's engineering work was led by a smaller team of seventeen senior engineers who had been hired specifically for the rebuild, supplemented by approximately twenty-five existing BCG engineers who were rotated from the firm's broader technology organisation into the AI rebuild. The phase's principal technical challenge was the orchestration architecture: the team had to design a system that could handle the per-proposal latency requirements (proposals are typically prepared on a multi-day timeline, but individual workbench queries need sub-second response times to support the partner-review workflow) while maintaining the grounding discipline that the retrieval-augmented architecture requires.

The third phase — the pilot-deployment phase, running from August 2025 to January 2026 — exposed the architecture to a limited cohort of approximately forty BCG partners and their proposal-preparation teams. The pilot cohort was selected deliberately to span the firm's practice mix and to include partners with varied levels of prior AI exposure. The pilot phase's principal operational learning was that the architecture's value depended more on the partner-team's workflow integration than on the architecture's raw capability: the partners who adopted the architecture most rapidly and produced the strongest win-rate improvements were the partners who restructured their proposal-team workflow to leverage the architecture's strengths, while the partners who attempted to graft the architecture onto existing workflow patterns produced more modest results. The learning informed the training-and-rollout program that supported the broader enterprise deployment.

The fourth phase — the enterprise-rollout phase, running from January 2026 to March 2026 — extended the architecture to the firm's full proposal-pipeline footprint across all practices and geographies. The phase's principal challenge was the change-management work required to bring the broader partner base into the new operational discipline. The rollout was structured as a partner-led adoption rather than a top-down mandate: every partner who wished to use the AI-augmented pipeline had to attend a four-hour partner-training session, sign the firm's AI usage policy, and commit to the three-checkpoint partner-review discipline. The partner-led approach was operationally slower than a top-down mandate would have been but was the structural mechanism by which the firm preserved the partner-level governance discipline that the program depended on. By March 2026, approximately 1,400 of the firm's 2,100 senior partners had been trained and were authorised to use the AI-augmented pipeline.

What to watch

The Vienna disclosure is the structural milestone for BCG's rebuild. The next twelve months will resolve five open questions that the disclosure raises.

• Whether the 14-point matched-cohort win-rate improvement is sustained across the next twelve months as the proposal pipeline matures and as the firm's competitors ship their equivalent rebuilds; the improvement may compress as the peer firms close the operational gap, or it may sustain as BCG's archive depth continues to outperform the peer firms' equivalent archives.
• Whether the partner-level governance discipline — the named-partner sponsorship, the three-checkpoint review, the Methodology Note disclosure — survives the operational scaling pressure as the AI-augmented pipeline handles a larger proportion of the firm's total proposal volume; the discipline produces a partner-time investment that constrains the volume the pipeline can process, and the firm may need to revise the discipline as the volume grows.
• Whether McKinsey, Bain, Deloitte, and Accenture's equivalent rebuilds close the operational gap with BCG within twelve months or whether the gap widens; the trade-press read is that McKinsey is closest to parity and the others are further behind, but the operational-maturity gap can shift in either direction depending on the peer firms' execution.
• Whether BCG extends the AI-augmented pipeline beyond the proposal-generation workflow into adjacent workflows — engagement-delivery support, knowledge-management synthesis, and the broader range of the firm's internal-tooling surface; the rebuild's architecture is broadly extensible, and the firm has signalled at Vienna that the next twelve months will include extension programs into adjacent workflows.
• Whether the client-side reaction to the Methodology Note disclosure evolves as more BCG proposals reach clients with the disclosure language; the current client-side reaction has been broadly accepting, with the qualitative feedback BCG has collected suggesting that the disclosure is treated as a transparency-positive signal rather than a competitive concern, but the pattern may shift as client-side procurement organisations develop more sophisticated frameworks for evaluating AI-augmented consulting deliverables.

Frequently asked

What did BCG actually rebuild, and what does the new architecture look like?

BCG rebuilt its proposal generation pipeline on a retrieval-augmented generation architecture grounded against the firm's 30-year proprietary archive of submitted proposals, engagement outcomes, and client feedback — a corpus of approximately 14 million documents indexed with Cohere's Embed 4 model. The generation layer runs on Claude Sonnet 4.6 under a September 2025 framework agreement with Anthropic. The architecture includes a distinguishing internal layer called BCG Recall — a Claude-based retrieval reasoning layer that surfaces ranked historical reference material for each proposal. The templated proposal architecture that the firm has operated since approximately 2014 was preserved deliberately to maintain the partner-level expectations of what a BCG proposal looks like, with the rebuild's value coming from filling the templates faster and better.

How does the partner-level governance work, and what does the Methodology Note disclose?

The partner-level governance establishes three operational disciplines: a named-partner sponsorship requirement (every AI-augmented proposal has a named partner who holds operational accountability for the content and the engagement's eventual delivery); a three-checkpoint partner-review requirement (proposal-shape review, technical-content review, final-submission review); and a partner-level disclosure requirement in the form of the Methodology Note. The Methodology Note is a single proposal page (approximately 350 words) that discloses to the client the specific AI capabilities used in the proposal's preparation, the human-review checkpoints applied, and the partner-level accountability for the proposal's content. The current language is the third iteration developed against client-side feedback from approximately 280 engagements delivered against AI-augmented proposals.

What is the 18-percentage-point win-rate improvement, and how was it attributed?

The 18-percentage-point figure is the gross proposal-stage win rate improvement against the pre-rebuild trailing-twenty-four-months baseline of approximately 32 per cent, with the post-rebuild figure approximately 50 per cent. The attribution methodology used a matched-cohort analysis controlling for client industry, engagement type, proposal value, competitive intensity, and partner-sponsor prior win-rate baseline. The matched-cohort win-rate differential — the figure that the analytics team attributes directly to the rebuild — is approximately 14 percentage points, against the gross 18-point figure. The 4-point difference is attributed to broader competitive-environment factors that the matched cohort controls for. The 14-point matched-cohort figure is the number Ehsan Mohammadi presented at Vienna as the rebuild's true operational attribution.

Which specific client engagements did BCG attribute to the rebuild?

BCG named four client engagements at Vienna as directly attributable to the AI-augmented proposal pipeline. The first is a transformation engagement with Stellantis on AI strategy and EV-platform consolidation, awarded in November 2025 against McKinsey and Bain. The second is a financial-services advisory engagement with HSBC on commercial-banking AI strategy, awarded in February 2026 against McKinsey, Deloitte, Accenture, and Oliver Wyman. The third is a public-sector advisory engagement with the Singapore Ministry of Trade and Industry on the country's 2030 AI-industry-strategy roadmap, awarded in March 2026 against Accenture, McKinsey, and PwC. The fourth is a consumer-goods advisory engagement with Unilever on category-management AI strategy, awarded in early April 2026 against McKinsey, Bain, and Accenture. The four engagements together represent approximately $340M in initial engagement value.

How does the rebuild affect BCG's consultant headcount and apprenticeship model?

The 31 per cent reduction in proposal-team staffing requirements is concentrated in the junior-consultant and senior-associate cohorts whose work the AI-augmented pipeline most directly displaces. The firm's overall headcount is up approximately 6 per cent against the pre-rebuild baseline, with the growth concentrated in engagement-delivery and AI-engineering cohorts. The firm is hiring fewer junior consultants relative to the prior baseline ratio against its growth rate, and more AI-engineering and analytics talent. The apprenticeship-model risk — that the rebuild removes a portion of the proposal-work that has historically been the structured environment for junior-consultant learning — has been addressed through a mandatory eighteen-month "proposal-craft" rotation that exposes every junior consultant to the proposal-shaping activities the AI is not handling: strategic shaping, partner-review preparation, and post-submission engagement handoff.

How does BCG's rebuild compare to McKinsey, Bain, Deloitte, and Accenture?

BCG's rebuild is, as of May 2026, the most advanced production deployment of AI in the proposal-generation workflow across the top-five consulting firms. McKinsey, Bain, Deloitte, and Accenture all have AI-augmentation programs in their proposal pipelines, but none has disclosed operating metrics at BCG's level of specificity or operational maturity. The trade-press read — corroborated by senior-partner sources at three of the four peer firms — is that the peer firms are between six and eighteen months behind BCG on the production-deployment timeline, with McKinsey closest to parity (estimated six to nine months behind) and the others further behind (estimated twelve to eighteen months behind). The competitive pressure on the peer firms to accelerate their own rebuilds is now operationally substantial, and the talent and capital allocations the firms have authorised in 2026 reflect the pressure.