Skip to main content
DoneThat

AI Adoption GuideConsultingDeliver

Reusable IP Extraction Agent

Agentic pipeline identifies and indexes reusable frameworks, templates, and models from project artifacts at delivery close.

Consulting processSellScopeStaffKickoffAnalyzeRecommendDeliverClose

By Don, DoneThat’s AI coach · updated

Method lives in client exhibits, not in a method folder

The job is to pull reusable method out of delivery artifacts and leave the client's operating system behind. At close, the useful template is almost never in a file named firm method. It is usually buried in a steering pack or a current-state exhibit that still carries the client's product names.

Index the exhibit as written and the next team retrieves another client's process as if it were yours. The next engagement should start from a stripped kit, not from a hunt through SharePoint.

Firm IP is transferable structure: how you run a decision-rights workshop, how you blank a RASCI, how you sequence a diagnostic. Client material is anything that would let a stranger name the account: their org chart, a process map with named sites, volumes, P&L lines, as-is swimlanes.

Engagement letters often assign work product to the client. If the IP clause is unclear, keep the facilitation template and discard the filled exhibit.

This is not a lessons-learned synthesizer. A postmortem explains what went wrong on this job. Extraction asks what, if anything, should survive as firm method.

The file named framework was their process map

A close-out folder will mix a reusable workshop kit with a client process map that someone labeled a framework. The following is an illustrative example, not a measured case.

A 250-person strategy and operations firm closes a 14-week operating-model redesign for a regional grocery chain. The project SharePoint site holds steering packs, workshop outputs, a current-state process map, an org chart of the commercial team, a financial model, and a deck named OM_Framework_v3.pptx. The agent queues reuse candidates at delivery close. A junior has already tagged three files as frameworks because the filenames said so.

Worth extracting, after a strip: a decision-rights workshop agenda, facilitator notes, and a blank RASCI grid. The method is the firm's. Strip the client's role titles and named people, and the next team has a session kit for Monday.

Not firm IP: OM_Framework_v3.pptx is the grocer's current-state process map. Boxes are named after distribution centers and SKU families. Calling it a framework does not make it yours. Another client's process map must never be indexed as firm IP.

Must never be indexed: the commercial org chart, the financial model with P&L lines and store counts, and any exhibit that only makes sense for that grocer. An org chart is not a span-of-control template until every identifying name and reporting line is gone.

The usual shortcut is to accept the filename tags and connect the whole project folder to enterprise search. The process map and the org chart would both have become retrievable firm-wide.

What changed: classify every flagged file as reusable method or client-confidential before anyone talks about search. Strip only the method pile. Dual-review, then ingest. The grocery process map stayed in the project archive under the engagement's access rules.

Classify method versus confidential at delivery close

Classification is the first gate, not redaction. Decide ownership before you touch names and numbers.

Run this on the delivery folder at close, not on a mid-project dump. Mid-flight files are drafts, client markups, and working numbers. Close is when the pack is complete and the engagement manager still remembers which slides were method.

You need the engagement letter's IP clause, the final pack plus working folder, a practice taxonomy for problem-type tags, and named reviewers (engagement manager and knowledge manager). Without the IP clause, stop. Without named reviewers, you do not have a process.

For each artifact, force one label:

  • Reusable method. Blank templates, facilitator guides, diagnostic trees, workshop agendas, slide skeletons, models with dummy inputs. Could another team use this on a different client in a different sector without recognizing the last account?
  • Client-confidential. Org charts, current-state process maps, as-is and to-be maps that name their functions, data rooms, financial models, interview notes, anything with employee names, site names, volumes, or prices. These stay in the project archive.
  • Partner call. A so-called framework that is half method and half their process. Most operating-model decks live here. A partner or practice lead splits the skeleton from the filled example. The agent does not split it unsupervised.

Filename is a weak signal. Consultants name client process maps "framework" out of habit. Score structure, not the title. Keep a reject log. Files the agent wanted and a human refused are how you tune the classifier. If the log is empty, nobody is reviewing.

Names and numbers are the easy strip

Named-entity redaction is necessary and not enough. Names, logos, legal entities, and obvious integers can be automated. What remains is structure that still identifies the account: swimlanes that match one grocer's DC network, a "generic" org that still has eight regions because they have eight, an anecdote in speaker notes.

Do the strip in this order:

  1. Automated pass. Names, emails, logos, obvious numeric facts, client codes in headers and file properties. Write a redacted copy. Never overwrite the project original.
  2. Structure pass. Would a practitioner who knows the sector recognize the client from the remaining boxes, geographies, or unique process steps? If yes, it is still confidential. Cut those shapes or do not index the file.
  3. Numbers that are method versus numbers that are theirs. A model with dummy drivers can be a template. A model whose drivers are that client's actuals is a data leak with formulas. Replace inputs with placeholders before the file is a candidate.

If stripping destroys the artifact, it was never method. A blank RASCI is method.

Case narratives are a different product. If you need an internal or public story, that is the reusable case study drafter, with its own consent and anonymization rules. Do not smuggle a case through the method index.

Nothing is searchable until two humans sign

Unreviewed files must not reach Glean, Hebbia, a SharePoint method library, a Notion wiki, or a firm knowledge assistant in the class of McKinsey Lilli. Those systems retrieve what you indexed. They will not know a client org chart from a span-of-control template. Connecting the project site for completeness is how confidential exhibits become answers for the next team.

Two reviewers, both required, before a file can be indexed:

  • Engagement manager: confirms the file is not that client's process, data, or org, and that the engagement letter allows method reuse.
  • Knowledge manager or practice librarian: confirms the stripped artifact is actually reusable and tagged by problem type, not by client name.

Until both signs exist, the file stays in the project archive. The agent can queue candidates and show the strip diff. It cannot publish.

Only after that gate should a project knowledge base ingestion agent classify, tag, and index. Extraction decides what is firm IP. Ingestion decides how it is found.

Do not index everything that survived the strip. Selection is the value. Trial this on one practice and one close. Measure whether the reject log has entries, whether the next kickoff used the kit, and whether anything identifiable reached search. Do not turn on a Glean connector to project sites as the trial.

Downstream retrieval inherits this gate

A comparable past scope retriever and a RAG proposal draft generator are only as clean as the method library they search. If a grocery process map is in the index, the next operating-model proposal will retrieve it as a framework, and leftover DC names become a confidentiality incident, not a copy-edit.

Judge the pipeline on incidents and reuse, not on files processed:

  • Incident: a retrieved artifact still identifies a prior client, or is their process labeled as method. Stop indexing until you know how it got through.
  • Reuse: engagement managers actually open the stripped templates at kickoff. If they keep hunting the last project's SharePoint, the library is the wrong stuff or too hard to trust.
  • Reject rate: a healthy queue has human refusals. A queue that always publishes is not being read.

Glean, Hebbia, Notion, SharePoint, and McKinsey Lilli sit in the same class: they accelerate a clean library and broadcast a dirty one. None of them is the classifier, the strip, or the signature.

The close-out checklist is short: classify, strip, dual review, then ingest. Skip a step and you have not extracted reusable IP. You have indexed someone else's company.

Is this worth automating for you?

Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.

DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.

Measure the baseline first