AI for consulting firms: proposals and knowledge.

02What does a consulting firm have to build on?

A knowledge firm’s raw material sits in a handful of systems, tied together by the engagement code:

• Documents. Proposals, statements of work, engagement letters, final reports and decks (Word, PowerPoint, Excel, PDF) in SharePoint, Google Drive or a document management system.
• Professional services automation (PSA) or ERP. Engagement codes, client, partner, team, dates and time entries, in platforms such as Deltek Vantagepoint, Kantata, NetSuite or Certinia.
• CRM. Opportunities, and whether each proposal won or lost, in Salesforce, HubSpot or Dynamics.
• HR and skills data. CVs, bios, certifications and self-declared skills, in Workday or a similar HR system.

Most of the work is joining these. A passage from an old report is only useful once you know which engagement, client, partner, date and outcome it came from, and whether it may be reused. That information lives in the PSA and CRM, not in the document.

03Where does AI pay off first?

Search over past work comes first. Proposals and expert finding are the same retrieval problem with different output.

04How should proposal drafting from past wins work?

The weak version pastes an RFP into a model and asks for a proposal. The version that survives partner review is built around structure:

• Start from the requirements, not the prose. Parse the RFP into numbered requirements and evaluation criteria, and map each to a section of the draft. A requirement with no matching section is flagged, not quietly dropped.
• Retrieve by section type. Pull approach, team and credentials passages separately, filtered to the same service line, rather than matching on whole documents.
• Use outcomes. The CRM knows which proposals won. Draft from wins, and keep lost proposals searchable so reviewers can see what was tried before.
• Cite every reused passage inside the draft. Each paragraph carries a link to its source engagement for the reviewer, and the links are removed from the final copy.
• Leave commitments blank. Fees, staffing and delivery dates are placeholders for the partner. The system can suggest a team through the expert finder, but it commits no one.

Test it on a set of past RFPs. Compare each generated draft with the proposal that was actually submitted, and score requirement coverage and the number of passages drawn from sources the author wasn’t allowed to see. The second number must be zero.

05How does an expert finder know who really knows a topic?

A useful expert finder answers “who has done this before, recently, and is free?” with evidence a manager can judge in seconds.

• Weight what people did over what they say. Staffing history from the PSA (who was on engagements tagged to the topic) and documents they wrote count for more than a skills field they filled in three years ago.
• Show the evidence. Each name comes with the engagements and documents behind it, so the asker can tell a lead from a bystander.
• Keep availability separate. Pull it from the PSA and show it beside the expertise, not inside the ranking, or the same few idle people always surface.
• Treat profile text as untrusted. OWASP’s guidance on vector and embedding weaknesses describes a resume with hidden white-on-white text that manipulates a retrieval system, and recommends validating data sources and using text extraction that detects hidden content. CVs and bios that feed an index can carry the same trick.
• Keep it a search tool, not a scoring tool. Annex III of the EU AI Act lists, among its high-risk employment uses, AI systems intended to allocate tasks in work-related relationships based on individual behaviour or personal traits or characteristics, or to monitor and evaluate the performance and behaviour of the people in them. A finder that suggests colleagues to ask is a different thing from one that assigns people to work or rates them. Design it as the first, and have a qualified professional review it before it drifts toward the second.

06What should “nothing on file” look like?

NIST’s Generative AI Profile defines confabulation as “the production of confidently stated but erroneous or false content”. In a firm that sells credibility, that is the worst failure. Our guide to retrieval that admits when it doesn’t know covers the general method. For a consulting firm, the specifics are:

• Every answer names its sources. Engagement, document, date and page, so a manager can open the original in one click.
• Answer only from retrieved passages. When nothing relevant is retrieved, or no passage supports the claim, the reply is “nothing on file for this”, with a suggestion of who to ask from the expert finder.
• Test the refusals. Write questions you know have no answer, such as a sector or country the firm has never worked in, and count how often the system answers anyway. This is the number most pilots never measure.
• Show age. Methods, market views and benchmarks go stale. Each result shows when the source was written and last reused.

07How do you keep one client’s work from reaching another client’s team?

Clients expect their material to stay confidential, and some deals also require that the team advising one party never sees another party’s material. An AI index built carelessly breaks that promise without anyone noticing, because it can summarize across everything it holds. OWASP names the risk: in a shared vector database there is “a risk of context leakage between users or queries”, and the fix it gives is “fine-grained access controls and permission-aware vector and embedding stores”. How we would build it:

1. Tag every chunk at ingestion with engagement ID, client, wall group and the source’s access list, copied from the document system and the PSA.
2. Filter before retrieval, not after generation. The query carries the user’s identity, and the search only considers chunks that user may open. Telling the model not to mention restricted material is not a control.
3. Make reuse opt-in. A document becomes reusable only when the engagement partner marks it at close-out, ideally in a sanitized version. The default is closed.
4. Give walled engagements their own index. Where a conflict wall is in force, a separate store is easier to prove than a filter.
5. Sync permissions continuously and check again at query time, and log what each query retrieved. People move between teams, and walls are added mid-engagement.
6. Test with decoys. Plant documents with unique phrases in a walled engagement, then try to pull them from an account on the wrong side with direct questions, indirect questions and requests for summaries. Ship at zero leaks.

If you run Microsoft 365, Purview Information Barriers may already be your wall for Teams and SharePoint. Microsoft’s own example is a lawyer’s data from one client that can’t be reached by a lawyer at the same firm who represents a different client. The barriers restrict two-way communication and collaboration between defined groups, and they don’t cover email. Don’t assume a separate AI index inherits them. Microsoft’s SharePoint documentation says apps running in app-only mode need an explicit opt-in to reach protected sites, and that people who had access before a policy was applied can still see items in search and Copilot results, with access denied when they open them. Use the segments as the source of truth for who sits on which side, and test the index against them.

A knowledge system for a consulting firm is judged less by what it finds than by what it refuses to show.

Law firms have run conflict walls for decades; see our guide to AI for law firms. Client files you ingest, such as RFPs and data-room documents, are also untrusted input: see our note on prompt injection.

08How do you check a deliverable before it reaches the client?

In October 2025 an Australian government department confirmed that a large consultancy would repay the final instalment under its contract for a report that contained several errors, after the firm admitted using generative AI to help produce it, as the Guardian reported. According to the Associated Press, a Sydney University researcher said the report misquoted a court case and contained a made-up quotation from a judge, along with references to work that doesn’t exist. The department said the substance of the report was maintained, with no changes to its recommendations, and the revised version disclosed that a generative AI language system was used in writing it. The checks have to be in the process, and some of them can be code.

What a deliverable-checking step can do well:

• Resolve every citation and quote in code. Does the cited document exist, and does the quoted text actually appear in it? A string match, not a model’s opinion.
• Reconcile numbers. The figure in the executive summary, the slide that repeats it and the spreadsheet it came from should agree. A model reads the three places, and code compares the values.
• Scan for other clients’ names. Compare the document against the full list of client and project names from the CRM. This is a deterministic search, and it catches the most embarrassing error a knowledge firm can make.

It doesn’t replace review, and the evidence says why. In a Harvard Business School field experiment with 758 knowledge workers, run with Boston Consulting Group on realistic consulting tasks, using GPT-4 on tasks inside the AI’s capability meant 12.2% more tasks completed and 25.1% faster, with higher quality. On a task chosen to sit outside it, people using AI were 19% less likely to produce a correct solution. That study used GPT-4, but the uneven frontier it describes is why the first month should measure errors, not just speed. NIST’s profile lists over-reliance and automation bias as a risk in how people and AI systems work together.

09Where is AI the wrong tool, and where does a person stay in charge?

• The recommendation. What the client should do, and why, is the product. AI can assemble evidence and draft the narrative. A partner owns the judgement.
• Fees, staffing and scope commitments. These bind the firm. The system proposes, a partner decides.
• Conflict clearance and reuse decisions. Whether a past engagement may inform a new one is a call for the engagement partner and risk team, recorded in the system, not inferred by it.
• Financial models and forecasts. Keep the model in a spreadsheet or code you can audit. A language model can explain it and check the inputs, not produce the numbers.

10Which rules apply?

As of October 2026, and not legal advice. A qualified professional should confirm how these apply to your firm.

• EU. The Digital Omnibus on AI, Regulation (EU) 2026/1744 of 8 July 2026, changes Article 4 so that providers and deployers must take measures to support AI literacy among their staff. The same Regulation (recital 40) sets 2 December 2027 as the date from which the rules for high-risk systems listed in Annex III, such as the employment uses above, apply. Our EU AI Act guide keeps the timeline.
• Personal data in the EU and UK. Staff profiles and named individuals in past reports are personal data. In the EU, the EDPB’s December 2024 opinion on AI models uses a three-step test for relying on legitimate interest and says the processing must be shown to be strictly necessary. In the UK, see the ICO’s guidance on AI and data protection, which the ICO says is under review following the Data (Use and Access) Act.
• US. NIST’s Generative AI Profile (July 2024) is voluntary guidance, useful as a checklist for governance, testing and disclosure. Client contracts usually matter more: engagement letters set confidentiality and can say what AI use must be disclosed.
• Gulf. Government and semi-government clients may set their own terms on where data is stored. Our guide to building AI in the UAE covers the data law there.

11How do you start, and what do you measure?

1. Pick one practice area with a few hundred past proposals and reports, and a partner who wants it.
2. Inventory and tag. Join documents to engagement code, client, outcome, wall group and a reusable flag from the PSA and CRM.
3. Build the index with permissions and run the decoy wall test before anyone uses it.
4. Write the test set. Questions with known answers, questions with no answer, and past RFPs with their submitted proposals.
5. Pilot cited search with proposal managers, then add drafting, then the expert finder. Set up the engineering around it, the evals, fallbacks and monitoring, from the start.

Track five numbers:

• Hours to a first draft of a proposal, before and after.
• Citation accuracy: the share of cited sources that exist and support the claim.
• Refusal accuracy: how often the system says “nothing on file” when it should.
• Wall leaks: must stay at zero in every test run.
• Errors caught before send versus errors found by the client.

12Questions consulting firms ask

Can’t we just use Copilot or ChatGPT Enterprise on our SharePoint?

They are a fair starting point for search over files people can already open. The gaps to check are the ones above: whether answers cite sources and refuse when nothing is on file, whether outcome and reuse flags from your CRM and PSA are respected, and whether the wall holds when you test it with decoys.

Will clients accept us using AI on their work?

That depends on the engagement letter and the client’s procurement terms, so read them before you start and say what you do in the proposal. Disclosure is a live question after cases like the Australian report above, and a log of what was retrieved helps in that conversation.

How long does a first version take?

Cited search over one practice area typically takes six to ten weeks to reach production. Connecting to the PSA and getting the permissions right is usually the biggest piece of work, not the model.