This is a typical scenario. It shows how Remolda would approach research automation in a mid-size consulting firm, what we would build and what we would measure.
The situation
Picture a consulting firm whose partners sell analytical rigour on tight timelines. Before the analysis starts, analysts spend days pulling figures out of annual reports, regulatory filings, industry studies and folders of client documents. They copy numbers into spreadsheets, check them, and fix transcription errors late in the project.
The work is necessary. It is also the least interesting part of the job for the people who do it, and errors found late are expensive to fix.
Macros and data subscriptions help on one engagement type at a time. Most incoming material is unstructured, so the problem returns on the next project.
The approach
The scenario runs the Remolda Cycle over about four months on one pilot practice before any firm-wide rollout.
Audit and mapping (3 weeks). We follow the data from the moment a client sends a folder of documents to the final slide. Interviews with partners, analysts and the research team are combined with time logs on live projects. The baseline is the share of project time spent turning raw documents into structured datasets, and the number of data corrections found in review.
Three steps are the likely candidates: ingesting and classifying documents, extracting data against the team's questions, and assembling cited tables and summaries.
Build (8 weeks). The design is three agents, each with one job, running in a private cloud environment set up for the firm, with access limited by engagement.
- Agent A, the parser, receives PDFs, runs layout-aware OCR that keeps tables and headings, classifies each document and indexes it.
- Agent B, the extractor, answers structured questions from consultants, such as capital spending guidance by region over a date range, with the excerpt and page reference for every answer. It works only from the documents provided for that engagement.
- Agent C, the synthesizer, turns extractions into spreadsheet tables with citations, drafts a summary and lists what the documents do not contain.
Client data stays siloed by engagement, and every query and extraction is logged for review.
Empower (4 weeks, alongside the build). Working sessions with analysts on live projects: how to write extraction questions, how to spot-check cited output, and how to change the project kickoff so agents are configured before research begins.
What we would measure
Each item below is a target we would set with the firm and measure against the baseline from the audit. The size of each target is set only after the baseline is measured.
- Time to a usable dataset on comparable projects.
- Share of extracted facts with a working citation to a page in the source document.
- Corrections found in review per deliverable.
- Analyst time on analysis versus collection, from time logs.
- Adoption: share of pilot-practice projects that use the agents after the first month.
Rules that frame this scenario
Client files carry confidentiality duties under the engagement contract and, where they include personal information, PIPEDA or a provincial privacy law. Vendor terms matter: enterprise and API plans from OpenAI and Anthropic exclude training on business data by default, and the processing location is documented per vendor. The consulting firms page lists sources.
Key lessons
1. The bottleneck sits in the process. Partners often read the problem as training or tools. The fix starts with the order of work and who does each step.
2. Narrow agents are easier to check. An agent with one job and one output format is easier to test, and its errors are easier to find.
3. Adoption needs time next to the work. The build takes weeks. Changing how analysts start a project takes working sessions on real engagements.
For related work, see AI for consulting firms, AI workflow automation and executive AI training.