Consultancies create value through judgement, synthesis, challenge and relationships. They also spend substantial time finding material, reconstructing previous work and creating the first structure of a proposal or research pack.

Some of that surrounding work is necessary discipline. Some is, as the source article put it, administration dressed as consulting. AI may help distinguish the two, provided the firm measures quality and review effort rather than celebrating faster drafting.

A useful quarter can end with one controlled workflow evaluated and governed. It need not end with all three ideas in production. Proposals, research and knowledge retrieval each touch client information, intellectual property and professional reputation. Use approved tools and obtain legal, privacy, security, contractual and client-specific advice.

1. Test AI-assisted proposal scaffolding

Proposal work combines reusable firm material with new thinking about a buyer’s situation. AI may help with the reusable structure: organising requirements, locating approved credentials, checking coverage and drafting a starting outline.

It should not invent client insight, case experience, results, people or delivery commitments. A senior owner remains responsible for every claim and recommendation.

Prepare approved source material

Create a bounded library of current service descriptions, team biographies, methodologies and case evidence the firm is permitted to reuse. Remove superseded, confidential or unsupported material. Record the source and approval date.

Do not feed a client brief or previous proposal into a public or unapproved tool. Establish supplier terms, data handling, access and retention first. Client procurement terms may restrict AI use even when the input appears ordinary.

Separate scaffolding from proposition

Ask the tool to support tasks that can be checked:

  • map the request to required response sections;
  • identify missing evidence or unanswered questions;
  • retrieve approved material with source links;
  • suggest an outline within the page limit; and
  • check a draft for contradictions or unsupported placeholders.

The pitch lead should develop the commercial argument, buyer-specific judgement, delivery model and trade-offs. If generated language makes every proposal sound like the same consultancy, the process has damaged distinction while saving drafting time.

Run a paired evaluation

Compare several AI-assisted proposals with the current method. Measure total time from brief to approved submission, senior review and correction, unsupported claims caught, compliance with the request and quality judged before the outcome is known.

Win rate is commercially important and a noisy measure across small, different opportunities. Use it over time alongside independent review. Do not conclude that AI improved a proposal because the firm won, or harmed it because the firm lost.

Treat calibration as delivery work. Update prompts and sources only after examining what failed. Preserve strong writing from the firm’s best proposals instead of replacing it with generic generated copy.

2. Use AI to accelerate a verifiable research pass

AI-assisted search and synthesis can help a team explore terminology, identify possible sources, compare themes and organise a large reading list. It can also cite material that does not support the claim, flatten disagreement or produce a confident statement without evidence.

Define the boundary: the tool supports discovery and synthesis; consultants verify every material source and remain responsible for the analysis.

Build an evidence protocol

For each claim likely to reach a client deliverable, retain:

  • the primary or authoritative source where available;
  • title, publisher, date and direct location;
  • the relevant passage or data definition;
  • publication date versus period studied;
  • limitations and conflicting evidence; and
  • the consultant who verified it.

Open the source. A citation generated by a tool is a lead, not verification. Where the source cannot be accessed, omit or label the claim for further research rather than citing a search summary.

Respect licence and copyright conditions. Summarise appropriately and avoid copying substantial protected material into prompts or deliverables. Check client confidentiality before using engagement context.

Test the research task, not the tool’s fluency

Create a benchmark of realistic questions. Include niche topics, recent developments, ambiguous terms and questions where the correct response is uncertainty.

Measure time to a verified evidence set, unsupported-source rate, missed authoritative material, review effort and the quality of the resulting analysis. Faster search with slower verification may deliver little net benefit. A small number of material errors can outweigh many acceptable summaries.

Use researchers and subject experts to judge coverage. The aim is not to automate the first half and leave people with the same amount of work. It is to direct human effort towards evaluation and synthesis without weakening the chain of evidence.

3. Create permissions-aware retrieval for internal knowledge

Consultancies often locate expertise through memory and informal networks. “Who worked on that logistics project?” can be answered quickly when the right partner is available and poorly when they are not.

AI-supported retrieval may help find approved methods, prior deliverables and people with relevant experience. The human network still supplies context: what worked, what the client rejected, what has changed and whether reuse is appropriate.

Start with a governed collection

Choose one practice or problem area. Inventory documents, permissions, client restrictions, age, status and ownership. Decide which material is reusable, reference-only, restricted or due for deletion.

Retrieval over a disorganised drive can make the wrong document easier to find. Content hygiene, metadata and access control are part of the product.

Test that users cannot retrieve material outside their authority. Check source links, version and approval status in every result. Generated synthesis should distinguish current methodology from historical client work.

Include expertise location carefully

Project records and biographies may help identify colleagues with relevant experience. Avoid reducing expertise to a model-generated label. Let people correct their profile and account for recency, role and depth of involvement.

Use the system to support introductions, not to allocate people automatically. Staffing decisions involve availability, development, inclusion and client needs beyond document history.

Design around an existing behaviour

Observe how consultants currently find work and ask colleagues. Make the new route genuinely useful at the moment of need. Training should use live tasks and expose limitations.

Measure successful retrieval, time to useful source, reuse with appropriate approval, permission incidents, outdated results and continued dependence on a small number of knowledge brokers. Query count alone rewards activity rather than value.

Frame the value as quality and capacity

Time released from drafting, searching or document assembly is not automatically a cash saving. The consultancy needs to decide how it uses that capacity: deeper client analysis, more sustainable workloads, faster response, additional work or lower delivery cost.

Measure total workflow effort, including curation, verification, security, training, support and incident handling. Track quality and harm. A proposal produced faster with more unsupported claims is worse. Research completed faster with weaker sources is worse. A knowledge tool used widely with poor permissions is worse at scale.

Avoid presenting AI as a headcount answer. That framing can suppress reporting of errors and encourage employees to protect workarounds. Explain the task, intended benefit, controls and how role expectations may change. Employment and workforce decisions require their own consultation and specialist advice.

Decide client transparency in context

Should every client always be told about proposal and research assistance? Transparency is important, and a universal disclosure rule cannot be established in this article.

Review legal duties, contracts, procurement representations, client policies, confidentiality, data use and the materiality of AI’s role. Some uses will require or strongly support explicit disclosure and consent. Others may be covered by an agreed firm policy and ordinary quality controls. Obtain advice.

Have a position the team can explain:

  • which uses are approved and prohibited;
  • what client and firm information may be processed;
  • where human responsibility sits;
  • how sources and output are checked;
  • when client approval is required;
  • how incidents are handled; and
  • how a client can raise a concern.

Do not make “human reviewed” the whole assurance. Explain the competence, evidence and control that make the review meaningful.

Choose one quarter-sized evaluation

Rank the three options against:

  • materiality of the current problem;
  • quality and permission of the source material;
  • possible client and firm harm;
  • ease of measuring the complete workflow;
  • owner and protected reviewer time;
  • technical and security readiness; and
  • ability to stop or contain failure.

Set success and stop conditions before the test. Include an unassisted baseline. Report positive, negative and inconclusive results to the partnership without turning the exercise into a competition with other firms.

The consulting AI use case scorecard assesses knowledge quality, proposal organisation, research access and consultant capability. Treat it as an internal diagnostic rather than a market benchmark. Distinction may benefit commercially from AI advisory or delivery work; any assessment should be capable of recommending prerequisite content and governance work or no deployment.

The best use case is not automatically proposal generation. It is the bounded workflow where the firm has a real problem, permitted evidence, accountable judgement and a result leadership can evaluate within the quarter.