Automation & AI for Swiss SMEs.Book a workshop →Workshops →
Let’s talkContact
Services

Our services

Automation & AI

Automation & agents Enquiries to the CRM, invoices to the ERP, support drafts, with your approval.

Software & data

Tools & knowledge Connect CRM, ERP and files, make knowledge searchable with sources, build interfaces for AI agents.Privacy & secure operations Data flow, rights per person, Swiss or EU hosting, tests, costs in view.

Planning & training

Consulting & analysis Find bottlenecks on site, rank use cases by value and effort, define the first step.Workshops Use AI safely every day, rules for privacy and data entry, practice with your own tasks.

Everything from one partner: end-to-end support for planning, implementation and operation.

← All projects

Case study · Tools & knowledge · Automation & agents

One knowledge base: RFIs in minutes instead of days

Knowledge from Outlook mailboxes, OneDrive, SharePoint, HubSpot and past RFIs in one central knowledge base. An assistant drafts an entire RFI in a few minutes, with sources and in the style of earlier answers. Including review by sales, this saves one to two working days per RFI, depending on its size.

Area
Tools & knowledge · Automation & agents
Scenario
Special machine builder, Schaffhausen (demo data)
Technology
Microsoft Graph, HubSpot API, Claude API
Deployment
Microsoft 365 and hosting in Switzerland, approval by sales
RFI assistant for RFI-2026-118 from Demo Pharma AG: 27 questions, 23 sourced, 4 to check. On the left the central knowledge base with Outlook, OneDrive, SharePoint, RFI archive and HubSpot, the customer context and the style profile; in the middle the questions by topic with status; on the right the draft answer to question 2.1 with its source and a replaced older version.
RFI draft with knowledge base, customer context from HubSpot and style profile. Proof of concept with demo data.

Starting point

In the reference scenario, a special machine builder in Schaffhausen receives several RFIs a month from pharma and medtech customers, often with 40 to 80 questions on certificates, technology, service, IT security and sustainability. Most of the answers already exist: in emails of individual employees, on OneDrive and SharePoint, in the CRM and in previously submitted RFIs. Even so, the search starts from scratch every time, colleagues are asked, and sometimes an outdated certificate ends up in the document. A larger RFI can easily tie up two working days in sales.

The solution

andyou built a central knowledge base that brings together emails from employees’ mailboxes, documents from OneDrive and SharePoint and customer data from HubSpot. Previously answered RFIs are ingested and pseudonymised in the process: customer names, contact persons and contact details are replaced. From this archive, the assistant derives the answer style. For each new RFI, it finds the right sources per question, prefers the latest version, drafts an answer and marks its status: sourced, check or open. HubSpot provides the customer context and highlights the topics that matter most to the customer.

Flow: Outlook, OneDrive, SharePoint, HubSpot and the RFI archive are ingested with permissions, pseudonymisation and latest version into the central knowledge base (search index, style profile, customer context). For each RFI question, the RFI assistant (Claude) finds sources, sets the status and drafts in style; result sourced, check or open for an expert, and sales approves.
Architecture: sources are ingested continuously, with permissions and pseudonymisation; the assistant drafts each answer with sources.
A past RFI answer with customer name, contact person, email and phone number becomes [CUSTOMER], [CONTACT], [EMAIL] and [PHONE] in the knowledge base. Sentences with placeholders never flow into new answers.
Pseudonymisation of the RFI archive: names and contact details are replaced before ingestion.

Permissions & approval

The knowledge base adopts the access rights of the employees. Confidential folders such as costing or HR are blocked: if blocked information matches, sales only sees a note, never the content. Outdated versions are shown as replaced, such as the 2022 ISO 9001 certificate or a lead time from the chip shortage. Past RFI answers count as a suggestion for review when no current source exists. If a question names a figure that is missing from the source, such as ATEX Zone 1 instead of Zone 2, it is flagged. Sales approves every answer.

Question 4.2 on lead time, status check: the answer states 26 to 32 weeks from approval of the specification, noting that the question is only partly covered. The source is a 2026 OneDrive document; a 2023 Outlook figure of 36 to 40 weeks is marked as replaced.
Lead time: latest source used, 2023 version replaced, deviation flagged.
Question 3.6 on ATEX Zone 1, status check: the reused past answer states Zone 2. Three warnings point to the reuse, the different zone and a source older than 18 months; the question is passed to Lukas Frei (Engineering).
Edge case: the past answer only covers ATEX Zone 2, the question asks about Zone 1.

Result and cost

In the demo RFI from a pharma customer with 27 questions, 23 answers were backed by current sources and 4 were flagged for review; status and sources matched the expected result in all 27 questions. No confidential document made it into an answer. The model calculation assumes 15 minutes of research per question without the assistant and 2 to 6 minutes of review with it: for an RFI with 40 questions that saves around one working day, for 80 questions around two. The assumptions are measured against previously answered RFIs in the project. 10 automated tests cover pseudonymisation, blocking, versions and approval.

27 questions in the demo RFI: 23 sourced, 4 to check, 0 open. Flagged for review are an outdated reference list, the ATEX zone, the lead time “from approval” instead of “from order” and the hourly rates. Status and expected sources match in 27 of 27 questions; no confidential sources were used.
Status of the 27 questions in the demo RFI, proof of concept with demo data.
Model calculation: 27 questions (demo) around 6.8 hours manually, 1.2 hours with the assistant, around 0.7 days saved; 40 questions 10 instead of 1.7 hours, around 1 day; 80 questions 20 instead of 3.5 hours, around 2 days.
Model calculation of the time saved by RFI size, with our own assumptions.

Technology in brief

In the project, the sources are connected through Microsoft Graph and the HubSpot API and kept up to date. The search works at sentence level and recognises German compound words and technical synonyms; in production, a semantic search is added. Claude drafts the answers through the Anthropic API with structured output and may only tighten the status, never loosen it. The knowledge base and index run in Microsoft 365 or with a Swiss hosting provider.

  • Microsoft Graph
  • Outlook · OneDrive · SharePoint
  • HubSpot API
  • Claude API
  • Pseudonymisation
  • Source references
  • Swiss hosting
  • Human approval

Make your company knowledge usable for RFIs?

Using an RFI you have already answered, we measure how much an assistant can prepare with your sources.

Let’s talk about your project