Mapping consulting tasks to generative-AI strengths can boost consultant efficiency, but only with strong verification and governance; otherwise hallucinations and gradual loss of routine skills threaten quality and long‑term capabilities.

Main Finding

Generative AI (GenAI) acts as a double-edged sword in management consulting: it can materially raise efficiency and creativity for many routine and semi-routine tasks, but it also creates persistent accuracy, cognitive, organizational, and social tensions (e.g., hallucinations, overconfidence, reduced skill retention). The authors introduce Task-GenAI Fit (TGAIF)—an adaptation of Task-Technology Fit theory embedded in a dynamic equilibrium framework—to prescribe when GenAI should be automated, used for augmentation, or avoided for a given consulting task. Balancing proactive (integration/augmentation) and defensive (restriction/isolation) strategies at the task level improves performance while managing risks.

Key Points

• Double-edged nature: GenAI improves speed and generativity (e.g., ideation, draft creation) but can reduce factual accuracy and analytical quality (hallucinations, overconfidence).
• Tensions beyond hallucination: technical (misinformation, verifiability), organizational (role shifts, control), social/cognitive (trust, skill erosion), and confidentiality/transparency trade-offs.
• TGAIF concept: Task-GenAI Fit aligns GenAI strategy to task characteristics:
  • Automate → routine, well-defined tasks (e.g., formatting, note synthesis).
  • Augment → tasks requiring creativity plus judgment (e.g., strategy ideation) with human-in-the-loop.
  • Avoid/defend → high-stakes client-facing or highly ambiguous tasks where risk/ethical concerns dominate.
• Dynamic equilibrium framing: organizations must continuously balance proactive (embrace & integrate) and defensive (restrict & separate) responses rather than eliminate tensions.
• Practical mitigation strategies: domain-adapted models, prompt engineering, calibration of model parameters by task, validation chains/self-verification, and human oversight—each trade-offs between generativity and factual correctness.
• Empirical findings: based on 22 semi-structured interviews with 33 experts from leading German consulting firms; authors identified five main themes and 37 sub-tensions across consulting workflows.
• Research-practice integration: authors used ChatGPT (GPT-4o) as an assistive tool for thematic coding and saturation checks; they triangulated AI assistance with manual coding and preserved human interpretive control.
• Early-adopter behavior: many firms are developing proprietary, secure GenAI assistants to address confidentiality and control concerns.

Data & Methods

• Design: Qualitative, exploratory, interpretive study using semi-structured interviews and grounded-theory informed iteration.
• Sample: 22 interviews (individual or paired) with 33 consulting professionals from leading German consulting firms. Participants spanned junior to partner levels; mean consulting experience = 5.7 years.
• Sampling: Purposive selection of GenAI-experienced consultants followed by snowball sampling; measures taken to diversify firm size, consulting focus, and role level.
• Data collection: Interviews conducted Sept–Nov 2024, using an iteratively refined interview protocol; interviews continued until thematic saturation.
• Analysis: Two-phase thematic analysis combining manual coding and GenAI-assisted coding (ChatGPT GPT-4o). ChatGPT was used to detect recurring expressions, summarize transcripts, and support saturation checks; outputs were critically reviewed and reconciled by researchers.
• Outputs: Five high-level themes, 15 subthemes (3 per theme), and 37 identified task-level tensions. Developed the TGAIF framework integrating task characteristics with dynamic equilibrium responses.
• Limitations noted by authors: selection bias and network homogeneity risk from purposive/snowball sampling; sample limited to German consulting firms and early adopters; qualitative design limits generalizability; potential biases introduced by using GenAI in analysis (mitigated by triangulation and human oversight).

Implications for AI Economics

Policy, firm strategy, and research directions stemming from TGAIF and the paper’s findings:

• Productivity and TFP measurement

  • Implication: GenAI effects are task-specific; aggregate productivity estimates must be decomposed to task-level impacts (routine vs ambiguous tasks).
  • Research action: Use time-use microdata or firm-level task allocations to estimate differential productivity gains; construct task-weighted TFP measures.

• Labor reallocation and wage dynamics

  • Implication: GenAI automation pressures routine task labor while augmenting high-skill, judgment-intensive work—implying heterogeneous wage effects and potential skill-biased complementarities.
  • Research action: Estimate task-level substitution/complementarity elasticities (e.g., difference-in-differences exploiting staged GenAI rollouts across teams/firms).

• Skill dynamics and human capital investment

  • Implication: Risks of skill erosion (deskilling) and shifting skill demand (more emphasis on verification, synthesis, and client-facing judgment).
  • Research action: Longitudinal studies tracking skill sets, promotion patterns, and training investments pre/post GenAI adoption; evaluate returns to reskilling programs.

• Firm strategy, adoption, and market structure

  • Implication: Firms developing proprietary, secure GenAI assistants may gain product-market advantage (higher quality, confidentiality), creating potential lock-in and barriers to entry.
  • Research action: Study investment returns in firm-specific LLM infrastructure, impacts on entry, and price competition in consulting markets; model strategic complementarities between data assets and AI performance.

• Pricing, contracting, and quality control

  • Implication: Quality risk (hallucination) introduces liability and reputation externalities; contractual forms and insurance against AI-generated errors may change.
  • Research action: Analyze how firms price AI-assisted services, contract clauses for AI use, and how reputation/quality signals evolve.

• Externalities and regulation

  • Implication: Mis/ disinformation and confidentiality risks create negative externalities; dynamic equilibrium suggests partial restrictions may be optimal in some contexts.
  • Policy action: Consider targeted regulation (transparency, verification standards) and data governance rules for client-facing AI use; support standards for verification chains and auditability.

• Empirical strategies suggested by the paper that economists can use

  • Field experiments: Randomly assign teams or tasks to GenAI augmentation vs control to estimate task-level causal effects on speed, quality, and error rates.
  • Difference-in-differences / staggered adoption designs: Exploit phased deployments across firms or departments.
  • Structural task-allocation models: Calibrate models where firms choose automation vs augmentation given task characteristics, costs, and quality risks.
  • Cost-of-error accounting: Quantify expected costs from AI hallucinations (rework, reputational harm) to evaluate net benefits of automation vs augmentation.
  • Market-level analyses: Investigate how GenAI adoption alters market concentration, prices, and demand for standardized versus bespoke consulting offerings.

• Measurement and data needs

  • Task-level observables (time, output quality, rework)
  • Granular wage and promotion data to detect compositional effects
  • Firm-level AI investment, model fidelity (domain adaptation), and governance practices
  • Incident logs for AI errors/hallucinations and mitigation costs

• Broader theoretical implications

  • Refines task-based models of technological change (e.g., Brynjolfsson/Autor) by explicitly modeling a three-way strategy choice (automate, augment, avoid) conditional on task attributes and persistent tensions.
  • Suggests equilibria where firms optimally mix proactive and defensive strategies over time—implying dynamic path dependence in firm-level returns to AI.

Practical takeaway for economics-oriented stakeholders: treat GenAI as a task-contingent production technology whose net economic effects depend on the alignment of model capabilities, verification processes, organizational governance, and the costs of errors. Research and policy should prioritize task-level identification, measurement of error externalities, and incentives for firms to internalize quality and confidentiality risks.

If helpful, I can convert these implications into a short list of testable hypotheses and corresponding empirical designs (field experiment, diff-in-diff, structural estimation) tailored to a particular dataset or industry.