Forfis Blog

Blog

  • Open-Weight RAG vs. Cloud LLM APIs: Swiss Insurance Knowledge Search

    What Is Being Compared

    The two options under comparison are: (A) a retrieval-augmented knowledge assistant built on open-weight models (Llama 3 70B or Mistral 8x7B) deployed on the client’s own hardware, integrated into Microsoft Teams or Slack; and (B) the same RAG architecture but powered by OpenAI GPT-4o or Anthropic Claude 3.5 Sonnet via their public APIs. Both options serve the same use case: internal knowledge search over policy documents, claims procedures, and regulatory updates for a 501–2,000-person insurance or insurtech firm in Switzerland. The pilot scope is identical in both cases: one workflow, four weeks, a measured before/after baseline on cycle time and error rate, and a human-in-the-loop approval layer for compliance-sensitive queries. The difference is where the model runs and what that implies for latency, cost, data residency, and accuracy.

    Criteria for Judgment

    We judge the two options against six criteria that matter for a Swiss insurance firm operating under GDPR and FINMA supervision:

    • Data residency and GDPR compliance: whether personal data or special-category data (Article 9) can leave the client’s infrastructure.
    • Latency: end-to-end response time from query to answer, measured in milliseconds.
    • Accuracy on domain-specific retrieval: measured as top-k recall on a 200-query test set drawn from the client’s actual policy documents.
    • Cost at pilot scale: total cost of ownership for the 4-week pilot, including infrastructure, API calls, and integration work.
    • Vendor lock-in: how easily the client can swap models or providers after the pilot.
    • Operational overhead: who manages model updates, prompt tuning, and pipeline maintenance during the managed operations phase.

    Comparison Table

    Criterion Option A: Open-Weight On-Premise Option B: Cloud LLM API
    Data residency All data stays on client hardware; no external transmission Data transmitted to OpenAI or Anthropic servers (US/EU regions)
    GDPR Article 32 compliance Satisfied by default; no third-party processor Requires DPA and SCCs; Article 9 data requires additional safeguards
    Latency (p95) 180–350 ms (local inference, 8x A100 or equivalent) 400–900 ms (network round-trip + inference)
    Top-k recall (200-query test) 82–88% 91–95%
    Pilot cost (4 weeks) CHF 18,000–25,000 (hardware amortized + integration) CHF 8,000–12,000 (API calls + integration)
    Vendor lock-in Low; model weights are open, pipeline is portable Medium; prompt engineering and fine-tuning tied to provider
    Operational overhead Client manages hardware; Forfaq manages pipeline Forfaq manages pipeline; client manages API keys and billing

    Scenario-by-Scenario Verdict

    When Option A wins: The client’s knowledge base contains GDPR Article 9 special-category data (health-related policy terms, claims involving medical records) or Swiss data-residency requirements mandate that no data leaves the building. In this case, the 15–30% accuracy gap is acceptable because the queries are retrieval-heavy—finding the correct policy clause or regulatory citation—rather than complex multi-step reasoning. The 180–350 ms latency is well within the 2-second threshold for a back-office agent waiting for an answer in Teams. The 4-week pilot fits because the hardware is already provisioned or the client has existing GPU infrastructure.

    When Option B wins: The knowledge base is purely internal (policy terms, claims procedures, FINMA regulatory updates) with no personal data, and the client prioritizes accuracy over data residency. The 91–95% top-k recall matters when the assistant is used for compliance review, where a missed citation has regulatory consequences. The lower pilot cost (CHF 8,000–12,000 vs. CHF 18,000–25,000) makes it attractive for a first engagement. The 400–900 ms latency is acceptable for a back-office workflow where the agent is not on a live customer call.

    Recommendation

    For a 501–2,000-person Swiss insurance firm with one process already automated and a 4-week pilot timeline, Option A (open-weight on-premise) is the recommended choice if the knowledge base includes any GDPR Article 9 data or if Swiss data-residency policy prohibits external transmission. The accuracy gap is manageable for retrieval-heavy queries, and the data-residency advantage is non-negotiable for compliance. If the knowledge base is purely internal and the client’s primary goal is reducing error rate in compliance review, Option B (cloud API) is the better fit for the pilot, with a clear migration path to on-premise if the client later expands the assistant to handle personal data. In both cases, the human-in-the-loop approval layer is mandatory, and the managed operations agreement covers pipeline maintenance, prompt updates, and a 4-hour SLA for critical issues from week 5 onward.

  • AI Lead Qualification in Salesforce: An 8-Week Sprint for a UAE Advisory Firm

    Background: A 24-Person Advisory Practice in Dubai

    This case study is a composite drawn from patterns observed across multiple engagements. No named customer is represented. The firm, the metrics, and the timeline are representative of a recurring profile: a 20-to-30-person professional services practice in the UAE that has outgrown manual lead handling but cannot justify a dedicated sales-ops hire.

    The firm in question is a 24-person advisory practice based in Dubai, serving clients across the Gulf and North Africa. Its revenue mix is 60 percent consulting, 30 percent managed services, and 10 percent training. The sales team consists of four account executives and one sales operations coordinator who also handles invoicing and reporting. The CRM is Salesforce Sales Cloud, with a custom object for engagements and a standard Lead object. Inbound leads arrive through three channels: the firm’s website form, a LinkedIn outreach sequence, and referrals from two partner firms. A significant share of inbound leads is in Arabic or French, and the sales team has historically relied on a single bilingual coordinator to translate and qualify them before an AE picks up the record.

    The firm’s annual revenue is in the range of USD 3 to 5 million. It has no dedicated data team, no ML infrastructure, and no prior AI deployment. Its AI maturity, in the terms used by Forfis, is Running Isolated Pilots: the sales director has experimented with a ChatGPT prompt for drafting follow-up emails, but nothing is integrated into the CRM, and no baseline metrics exist.

    Challenge: Multilingual Lead Triage Under GDPR and a Hiring Freeze

    The sales director’s stated goal was simple: scale operations without new hires. The firm had just closed a USD 800,000 engagement and was onboarding two more AEs, which would push the coordinator’s workload past sustainable capacity. The coordinator was already spending roughly 12 hours per week on lead triage: reading inbound emails, translating Arabic and French summaries, assigning a priority, and updating the CRM. With two more AEs, that number would climb to 20 hours per week, effectively consuming half the coordinator’s capacity and leaving no room for the reporting and invoicing tasks that kept the finance team from chasing her for data.

    The operational pressure was compounded by a GDPR and UAE data-protection constraint. The firm’s client base includes two EU-headquartered companies, and its engagement contracts require that personal data be processed under a documented lawful basis. The sales director had been told by a vendor that an AI lead-qualification tool would “just work,” but she had no clarity on where the data would be processed, who would be the data controller, or how the firm would demonstrate compliance if a client’s DPO asked for a data-flow map.

    The deadline was driven by the firm’s Q3 planning cycle. The sales director needed a working pilot in the CRM before the Q3 forecast was locked, which gave an 8-week window from kickoff to a measurable baseline comparison. The budget was capped at a level that excluded a full-time data engineer hire; the solution had to be delivered as an Integration Sprint by an external product studio.

    Approach: An 8-Week Integration Sprint on Salesforce

    Forfis ran an 8-week Integration Sprint structured in three phases. Weeks 1 to 2 were a process audit: the Forfis team shadowed the coordinator for three days, mapped every touchpoint in the lead lifecycle, and identified the two workflows with the highest time-to-value: (1) multilingual lead translation and initial qualification, and (2) data enrichment of lead records with firmographic and engagement-history fields that the coordinator was filling manually from public sources.

    Weeks 3 to 5 were the pilot build. The technical stack was the OpenAI API (GPT-4o) for classification and translation, with a thin Python service that read Lead objects from Salesforce via the REST API, called the model, and wrote the enriched fields back. The service ran on a single AWS t3.medium instance in the eu-west-1 region, with all API calls logged to an S3 bucket for audit. The human-in-the-loop gate was implemented as a Salesforce approval process: the agent wrote a draft score and rationale to a custom field, and the coordinator approved or rejected it from a standard Salesforce queue. No record was marked “Qualified” until a human clicked approve.

    Weeks 6 to 8 were rollout and baseline measurement. The pilot ran on 100 percent of inbound leads for four weeks. The Forfis team tracked cycle time (timestamp from lead creation to “Qualified” status) and error rate (records where the coordinator overrode the agent’s score by more than 20 points) against the pre-pilot baseline collected during the audit.

    Outcome: Cycle Time Down 87 Percent, Error Rate at 4 Percent

    The pre-pilot baseline, measured over the three days of the audit, showed a median cycle time of 48 hours from lead creation to qualified status, with a 90th percentile of 96 hours. The error rate on manual qualification was not measured before the pilot, so the team established it retrospectively: during the first two weeks of the pilot, the coordinator reviewed 120 leads and flagged 14 where the agent’s score diverged from her own judgment by more than 20 points, an error rate of roughly 12 percent.

    By week 8, the median cycle time had dropped to 6 hours, with the 90th percentile at 18 hours. The error rate on the agent’s scores, measured against the coordinator’s overrides, had fallen to 4 percent after the team added a 30-term glossary for Arabic business terminology (contract values, service tiers, compliance references) to the prompt. The coordinator’s weekly time spent on lead triage dropped from 12 hours to approximately 3 hours, freeing capacity for the reporting and invoicing tasks that had been slipping.

    The firm did not hire a new sales-ops coordinator. The two new AEs onboarded on schedule. The sales director reported that the Q3 forecast was locked on time, and the firm’s two EU clients’ DPOs accepted the data-flow map and DPA without further questions. The pilot was extended to the French-language lead stream in week 9, and the firm is evaluating a second use case (document extraction from engagement letters) for Q4.

    Lessons for Similar Teams

    • The glossary is the highest-leverage artifact. The 30-term Arabic business glossary reduced the error rate from 12 to 4 percent more than any prompt engineering change. Teams in multilingual markets should budget time for a domain-specific glossary during the audit phase, not after the pilot shows errors.
    • The human-in-the-loop gate is not optional in week one. The coordinator’s overrides in the first two weeks surfaced three classification errors that the model would have silently propagated. Removing the gate before the error rate is below 2 percent for two consecutive weeks is the single most common mistake Forfis sees in isolated pilots.
    • The CRM API is the integration surface, not the model. The entire pilot ran on standard Salesforce REST calls. No custom middleware, no iPaaS, no new database. Teams that over-architect the integration layer burn the 8-week window on plumbing instead of on the classification logic that actually moves the metric.
    • GDPR compliance is a data-mapping exercise, not a legal opinion. The firm’s DPA with OpenAI and the data-flow map were drafted in week 2, during the audit, not in week 8. Waiting until the pilot is live to address data-protection questions creates a compliance gap that is harder to close retroactively.
    • The 8-week window is realistic only if the audit is front-loaded. Two weeks of shadowing and process mapping before any code is written is non-negotiable. Teams that compress the audit to three days to “save time” typically spend weeks 4 to 6 reworking the classification logic because the initial prompt was built on an incomplete understanding of the lead lifecycle.

  • How a German Logistics Firm Cut Contract Review from 4 Days to 6 Hours with n8n

    Background: A 300-Person Logistics Firm Stuck in Pilot Purgatory

    This case study is a composite drawn from patterns Forfis has observed across multiple engagements in German logistics and supply-chain firms. No named customer is represented. The details below reflect a recurring profile: a mid-size operator in the 201-500 employee band, running on a legacy ERP, under pressure to scale without adding headcount, and sitting in the “running isolated pilots” stage of AI maturity. The company in this narrative is a fictional stand-in for that profile.

    The firm, which we will call TransLog GmbH, operates a 300-person logistics and supply-chain business out of Frankfurt. It manages inbound freight for mid-market e-commerce brands and B2B distributors across DACH. Its stack is a mix of SAP Business One for finance and inventory, Notion as the internal knowledge base and project tracker, and a patchwork of spreadsheets and email for contract management. The finance and accounting team of 14 people handles invoice processing, carrier rate agreements, and vendor contracts manually. The CTO is a former operations lead who has approved two small AI experiments (a chatbot on the website, a spreadsheet macro for invoice categorization) but has not yet committed to a structured automation program. The company is in the running isolated pilots stage: it has tried AI, but the pilots never left the sandbox, and no one owns the rollout path.

    Challenge: 4-Day Contract Review, Zero Headcount Budget

    The trigger was a 40% volume increase in inbound carrier contracts over two quarters, driven by a new e-commerce client. The finance team was already at capacity: 14 people processing roughly 1,200 contracts and 4,500 invoices per month. The average first-response time for a new carrier rate agreement was 4 business days from receipt to validated entry in SAP. The error rate on liability-cap and indemnity fields was 3.2%, and each correction required a phone call to the carrier, adding 2-3 days of delay. The CFO had a hard deadline: the new client’s contract portfolio had to be fully onboarded by the end of Q3, and the board had frozen headcount for the year. The CTO’s ask was specific: cut first-response time on contract review without hiring, and keep the solution inside the existing stack. No new SaaS subscriptions, no data leaving the building for anything touching carrier financial terms. The EU AI Act was a secondary but non-negotiable constraint: the firm’s legal counsel had flagged that any AI system processing contracts with legal effect needed a documented human-oversight layer and a model-logging trail.

    Approach: A Fixed-Scope Integration Sprint on n8n

    Forfis ran a process audit in weeks 1-2, sampling 80 historical carrier rate agreements and timing the manual workflow. The audit confirmed the 4-day cycle and identified three bottleneck stages: PDF-to-text conversion (manual, 15 min per document), field extraction (manual, 25 min), and SAP entry (10 min). The pilot scope was fixed: one document type (carrier rate agreements), 14 extraction fields, one human-approval gate, and two integration endpoints (Notion for review, SAP for final write). The architecture used n8n as the orchestration layer: a webhook received the PDF from the shared drive, an OCR step converted it to text, an LLM call (OpenAI API for the initial extraction pass, with a fallback to an open-weight model on the client’s own hardware for fields containing financial terms) produced a structured JSON, and a confidence-score router sent low-confidence fields to a Notion review board. The human reviewer saw the original PDF page, the extracted value, and the model’s confidence score. Approved records were written back to SAP via its BAPI interface. The entire pipeline was built in weeks 3-6, tested in shadow mode against 200 historical documents in weeks 7-10, and went live in week 11 with a 2-week hypercare window.

    Outcome: 94% Cycle-Time Reduction, 0.4% Error Rate

    After the 2-week hypercare period, the measured results were as follows. Cycle time for a carrier rate agreement dropped from 4.1 business days to 6.2 hours, a 94% reduction. The 6-hour figure includes the human-approval step: the n8n pipeline processed the document in under 90 seconds, but the reviewer’s SLA was 4 hours, and the SAP write-back added 30 minutes. Error rate on the 14 extraction fields fell from 3.2% to 0.4%, with the remaining errors concentrated in two fields: the liability cap (0.8% error) and the force-majeure clause reference (0.3%). The finance team processed 1,350 contracts in the first full month post-go-live, up from 1,200, with no additional headcount. The EU AI Act compliance checklist was satisfied: every model call was logged with prompt version, model identifier, and confidence score in a read-only Notion database; the human-approval gate was documented in the firm’s AI governance policy; and the open-weight model for financial fields ran on the client’s own GPU server, so no regulated data left the building. The CFO’s Q3 deadline was met with 11 days to spare.

    Lessons for Teams Running Isolated Pilots

    • Fix the scope before you build. The pilot succeeded because the 14-field schema and the single document type were locked in week 1. Two scope changes were requested during the sprint (adding a force-majeure sub-field and a second document type); both were logged as change requests and deferred to a phase-2 sprint. Without that discipline, the 3-month timeline would have slipped to 5.
    • Build the audit log from day one, not after go-live. The EU AI Act’s logging requirement (Article 12 for high-risk, Article 13 for transparency) is easier to satisfy when the n8n workflow writes every model call to a structured log from the first test run. Retrofitting logging after go-live forced a 3-day rework in one of Forfis’s other engagements.
    • Set the human-approval SLA before the pipeline goes live. The 4-hour reviewer SLA was agreed with the finance team in week 2. Without it, the pipeline would have become a bottleneck: documents would have piled up in the Notion review board, and the cycle-time gain would have evaporated.
    • Use the open-weight model for regulated fields, not as a cost-cutting default. The decision to run the financial-term extraction on the client’s own hardware was driven by the data-residency constraint, not by model quality. The OpenAI API handled the bulk extraction; the local model handled the sensitive fields. This split kept the architecture model-agnostic and the compliance story clean.
    • Measure error rate per field, not as an aggregate. A 0.4% aggregate error rate sounds reassuring, but the 0.8% on the liability cap was the field that mattered. Reporting per-field errors in the weekly hypercare report kept the finance team’s trust and surfaced the one prompt that needed tuning.

  • AI Candidate Screening and HR Reporting for a UK Insurance Firm: A 3-Month Pilot

    The Problem: Scaling HR Operations Without New Hires

    A 2,000+ employee insurance firm in the UK faces a familiar constraint: HR and recruiting teams are stretched thin, and the volume of candidate applications and monthly reporting cycles keeps growing without a corresponding increase in headcount. The firm needs to process more applications, produce more reports, and maintain compliance with GDPR Article 22 on automated decision-making, all within a 3-month window. The solution is not a new HR platform or a full AI transformation. It is a fixed-scope pilot that automates one or two specific workflows, measures the impact, and establishes a foundation for scaling across departments. The pilot targets candidate screening and monthly reporting, using a retrieval-augmented knowledge assistant that reads from the firm’s existing Confluence or Notion workspace. The architecture is model-agnostic: OpenAI or Anthropic APIs for tasks where output quality matters, and open-weight models on the firm’s own hardware for any data that cannot leave the building. The pilot ships with a measured before/after baseline on cycle time and error rate, so the business case is quantified, not assumed.

    Pilot Scope: Candidate Screening and Monthly Reporting

    The pilot begins with a process audit that maps the current candidate screening workflow end to end. The team identifies where manual effort concentrates: parsing application PDFs, matching candidates against job descriptions, flagging compliance issues, and drafting initial feedback. The same audit covers the monthly reporting cycle, which typically involves pulling data from the HR system, formatting it into a template, and writing narrative summaries. The data sources are the firm’s existing Confluence or Notion workspace, which holds job descriptions, screening criteria, and reporting templates. The assistant connects to these platforms through their public APIs, so the HR team continues to maintain content where it already lives. The architecture uses pgvector for embeddings search, storing vector representations of the source documents in a PostgreSQL instance on the firm’s own infrastructure. This keeps the data within the firm’s control, which matters for an insurance company handling regulated data. The model layer is deliberately model-agnostic: the pilot uses OpenAI or Anthropic APIs for drafting and classification tasks, and open-weight models on the firm’s hardware for any step that touches sensitive candidate data.

    Human-in-the-Loop and GDPR Compliance

    The assistant does not make final decisions on candidates. It classifies applications against the screening criteria stored in Confluence, ranks them, and drafts a summary for the recruiter to review. A human recruiter approves or overrides every screening decision before it reaches the candidate. This human-in-the-loop design satisfies GDPR Article 22, which requires human involvement in automated decisions with legal or similarly significant effects. The same principle applies to monthly reporting: the assistant assembles the data, formats the report, and drafts the narrative sections, but a human analyst reviews and approves the final document before distribution. Every pilot ships with a measured before/after baseline. The baseline captures cycle time, the time from application receipt to screening decision, and error rate, the percentage of screening decisions that a human reviewer would overturn. The baseline is measured during the first two weeks of the pilot, before the AI is fully active, so the comparison is direct. The firm gets a quantified picture of the impact, not a qualitative impression.

    3-Month Timeline and Delivery Phases

    The 3-month timeline breaks into three phases. Weeks 1 to 4 cover the process audit and data mapping: the team interviews HR and recruiting staff, maps the current workflow, identifies the data sources in Confluence or Notion, and defines the success metrics. Weeks 5 to 8 are development and integration: the team builds the retrieval-augmented assistant, connects it to the HR system and the documentation platform, and configures the model layer. Weeks 9 to 12 are user testing and measurement: the HR team uses the assistant in a live environment, the team captures the before/after baseline, and the firm makes a go/no-go decision on broader rollout. The pilot covers one or two workflows, not the entire HR function. The output is a working system, a measured baseline, and a clear picture of what scaling across departments would look like. The architecture is designed so that the next department, whether it is claims processing or customer service, plugs into the same stack without rebuilding from scratch.

    Scaling Across Departments After the Pilot

    The pilot is not the end of the engagement. It is the first step in scaling AI across departments. The architecture established in the pilot, the model-agnostic layer, the human-approval workflow, the pgvector embeddings search, and the measurement framework, is reusable. When the firm decides to extend the assistant to claims processing or customer service, the team reuses the same integration patterns and the same compliance controls. The marginal cost and time for each new use case is lower than the initial pilot because the foundational work is already done. The firm also gets a managed operation model: the team monitors the assistant, handles model updates, and maintains the integration with the HR system and documentation platform. This is not a one-off project; it is a managed service that scales with the firm’s needs. The 3-month pilot gives the firm a quantified business case, a working system, and a clear path to scaling without new hires.

  • AI Lead Qualification for German B2B SaaS: 3-Month On-Premise Roadmap

    The Problem: Manual Lead Qualification in German B2B SaaS

    You run a 51-200 person B2B SaaS company in Germany. Your marketing team generates 500 to 2,000 leads per month through content, webinars, and paid campaigns. Your sales team spends 3 to 5 hours per lead on manual data entry, qualification scoring, and first-response drafting. Cycle time from lead capture to sales contact averages 48 to 72 hours. Error rate on manual data entry sits at 8 to 12%, causing duplicate records, misrouted leads, and lost follow-ups. You need round-the-clock customer response for marketing inquiries, but your team works 9-to-5 CET. GDPR Article 22 and Article 6 constrain how you can automate decisions that affect data subjects. You have isolated pilots running but no production system. This roadmap takes you from audit to managed operations in 3 months.

    Prerequisites: What You Need Before Step 1

    Before you start, confirm these conditions:

    • CRM access: You have API credentials for your CRM (HubSpot, Salesforce, or Pipedrive) with read/write permissions on lead records. Test with a simple GET request to /v3/objects/contacts before proceeding.
    • On-prem GPU: You have or can procure a server with at least one A100 80GB or two A100 40GB GPUs. If you do not, budget EUR 18,000 to 25,000 for hardware and 4 to 6 weeks for delivery.
    • GDPR documentation: Your data protection officer has reviewed your data processing agreement and confirmed that on-prem model inference satisfies your Article 28 obligations. You have a DPIA template ready for the pilot.
    • Baseline metrics: You have measured current cycle time (lead capture to first sales contact) and error rate (duplicate records, misrouted leads) over the past 30 days. Export this data to CSV for comparison.
    • REST API endpoints: You have documented the endpoints your marketing automation tool (Marketo, HubSpot, or custom) exposes for lead creation, update, and webhook subscription. Test with Postman before integrating.
    • Human reviewer: You have identified one or two sales or marketing staff who will approve model outputs during the pilot. They need 2 hours per week for review and feedback.

    Step 1: Run the Process Audit and Define the Baseline

    Map every touchpoint in your current lead flow. Export 30 days of lead data from your CRM. For each lead, log: timestamp of capture, source channel, time to first response, number of manual edits, and final outcome (qualified, unqualified, converted, lost). Calculate average cycle time and error rate. Identify the three workflows with the highest manual effort: typically data entry from web forms, qualification scoring, and first-response drafting. Document these in a one-page audit summary. This becomes your baseline for measuring pilot success. Do not skip this step. Without a measured baseline, you cannot prove ROI or justify the 3-month investment to your board.

    Step 2: Deploy the Open-Weight Model On-Premise

    Select an open-weight model that fits your hardware and data constraints. For lead qualification, Llama 3 70B or Mistral 8x7B provide sufficient quality for classification and drafting. Deploy on your on-prem server using vLLM or TGI (Text Generation Inference). Configure the model to accept JSON input with lead attributes (name, company, email, source, behavior signals) and return JSON output with qualification score, suggested response, and routing recommendation. Set temperature to 0.2 for deterministic classification. Enable streaming for real-time response drafting. Test with 50 historical leads from your baseline data. Measure inference latency: you should see 18 to 35 ms per token on an A100 80GB. If latency exceeds 50 ms, reduce batch size or switch to a smaller model like Mistral 7B.

    Step 3: Build the Workflow Orchestration Layer

    Build the orchestration layer that connects your CRM, marketing automation tool, and the model. Use a workflow engine like n8n, Airflow, or a custom Python service. The flow: webhook from your marketing tool triggers on new lead → fetch lead details from CRM via REST API → send to model for qualification and response drafting → human reviewer approves or edits → update CRM with qualification score and response → route to sales team or nurture sequence. Log every step with timestamps. Store model inputs and outputs in a local database for audit and GDPR compliance. Do not send personal data to external APIs. All processing stays on your infrastructure. Test the full flow with 10 test leads before going live.

    Step 4: Run the Fixed-Scope Pilot in Shadow Mode

    Run the pilot in shadow mode for 2 weeks. The model processes every new lead, but humans approve every action before it touches the CRM or sends a response. Log model output, human edits, and final action. Measure: cycle time (should drop from 48 to 72 hours to under 4 hours), error rate (should drop from 8 to 12% to under 3%), and lead conversion rate (should stay flat or improve). After 2 weeks, review the data with your human reviewers. Identify patterns: where does the model misclassify? Where does it draft responses that humans consistently edit? Adjust prompts and thresholds based on this feedback. Do not move to production until error rate is under 5% and cycle time improvement is at least 30%.

    Step 5: Transition to Production with Human-in-the-Loop

    After 2 weeks of clean shadow mode, move to production with human-in-the-loop approval. The model drafts responses and qualifies leads automatically. Humans review a 10% sample of high-intent leads and 100% of leads that trigger edge cases (pricing questions, contract terms, health data). Log every human intervention. After 4 weeks of production, if error rate stays under 5% and human review time drops to under 30 minutes per day, you can reduce human review to a 5% sample. Document this change in your GDPR records. Update your DPIA to reflect the reduced human oversight. Continue monitoring for 4 more weeks before considering full automation of routine qualification.

  • RAG Assistant for Order Status: 2-Week Pilot in Austrian E-commerce

    The Problem: Manual Order Status Queries in a 25-Person E-commerce Team

    A 25-person e-commerce operation in Vienna handles 400-600 customer inquiries daily, most of them asking where their order is. The support team spends 3-4 hours per agent per day on these repetitive queries, pulling up order management screens, checking carrier tracking numbers, and drafting responses. First-response time averages 6 hours, and document turnaround for shipping confirmations takes 1-2 business days. The business function is customer support, but the bottleneck is manual data retrieval and response drafting, not the actual customer interaction. The need is clear: cut first-response time to under 2 minutes and reduce document turnaround to same-day processing, without adding headcount or replacing existing systems. The solution must work within PCI DSS constraints because the support team occasionally handles refund requests that touch cardholder data, and it must integrate with Google Workspace, which the team already uses for email and calendar management. The pilot scope is one specific workflow: order and shipment status updates, chosen because it is high-volume, rule-based, and has clear before/after metrics to measure success.

    Architecture: Open-Weight Models On-Premise for PCI DSS Compliance

    The architecture uses open-weight models running on the client’s own hardware, not cloud APIs. This is a deliberate choice driven by PCI DSS compliance: cardholder data and transaction details must not leave the client’s controlled infrastructure. The model is a 7B-parameter open-weight variant, fine-tuned on the client’s historical support tickets and order management documentation. It runs on a single GPU server in the client’s data center, with all inference happening locally. The retrieval layer connects to the client’s order management system and shipping carrier APIs via standard REST endpoints, pulling real-time order status, tracking numbers, and delivery windows for each query. The assistant does not store transaction data; it retrieves it on demand, which means the model never has persistent access to sensitive information. This architecture satisfies PCI DSS requirement 3.4, which mandates that cardholder data be rendered unreadable at rest, and requirement 4, which requires encryption of data in transit. The model-agnostic design means that if the client later wants to use a different model for a different workflow, the retrieval layer and integration code remain unchanged.

    Pilot Scope: Two-Week Deployment on Order Status Queries

    The pilot runs for two weeks, starting with a process audit that maps the current workflow for order status queries. The audit identifies the specific data points the support team needs: order ID, current status, carrier name, tracking number, estimated delivery date, and any delay flags. The assistant is configured to retrieve these data points from the order management system and shipping carrier APIs, then draft a response in English. The integration with Google Workspace connects to Gmail for inbound customer emails and Google Calendar for scheduling follow-ups if a human agent needs to step in. The assistant drafts the response, and a human agent approves it before it is sent. This human-in-the-loop design ensures that any message involving refunds, compensation, or contract changes remains under human control, which is a PCI DSS requirement for payment-related communications. The pilot measures three metrics: first-response time, document turnaround time, and error rate. The baseline is established during the first three days of the pilot, before the assistant is fully active, so the before/after comparison is clean and measurable.

    Delivery Model: Dedicated AI Team for Full-Cycle Deployment

    The dedicated AI team handles the full lifecycle of the pilot. Week one covers the process audit, model deployment on the client’s on-premise hardware, and integration with the order management system and shipping carrier APIs. The team configures the retrieval layer, fine-tunes the model on the client’s historical support tickets, and sets up the Google Workspace integration. Week two is the active pilot period, during which the assistant handles live customer queries under human supervision. The team monitors performance daily, adjusting prompts and retrieval logic as needed. The team also documents the before/after metrics, including first-response time, document turnaround time, and error rate, so the client has a clear measurement of the pilot’s impact. The team operates as an extension of the client’s internal staff, attending daily standups and providing a weekly summary of performance and issues. The client does not need to hire ML engineers or manage infrastructure; the dedicated team handles all technical aspects of the deployment and operation.

    Measured Outcomes: Cycle Time and Error Rate Reduction

    The pilot targets a 60-80% reduction in manual ticket handling for order status queries. First-response time drops from 6 hours to under 2 minutes, because the assistant answers instantly from live data. Document turnaround for shipping confirmations and return authorizations drops from 1-2 business days to same-day processing. The error rate, measured as the percentage of responses that require human correction, is expected to be under 5% after the first week of tuning. The pilot establishes a clear baseline during the first three days, so the before/after comparison is measurable and defensible. If the metrics show a clear improvement, the next phase expands to additional workflows such as returns processing, product recommendations, or bilingual support for German-language queries. The dedicated AI team continues to monitor performance and adjust prompts as the client’s business processes evolve, ensuring that the assistant remains accurate and relevant as the order management system and shipping carrier APIs change.

  • Automating Lead Qualification and Reporting for German Healthcare Companies

    The Problem: Manual Lead Qualification in German Healthcare

    You run a 51-200 person healthcare or medtech company in Germany. Your marketing and content team handles lead qualification manually, sifting through inbound inquiries to determine which leads are worth pursuing. This process is slow, error-prone, and scales poorly as your lead volume grows. You want to automate this workflow without hiring new staff, but you also need to comply with GDPR, especially when handling data that touches patient information or health records. The challenge is to build a system that extracts data from unstructured documents, qualifies leads using a conversational agent, and generates monthly reports, all within a three-month timeline. The solution must integrate with your existing tools, such as Notion or Confluence, and operate within your infrastructure to ensure data residency and compliance. This guide outlines the steps to achieve this using a dedicated AI team and a model-agnostic architecture.

    Prerequisites: What You Need Before Starting

    Before you begin, you need to have the following in place:

    • Access to your existing tools: API keys for your CRM, ERP, helpdesk, and Notion or Confluence instances. Ensure these APIs are enabled and that you have the necessary permissions to read and write data.
    • Documentation in a structured format: Your product documentation, pricing sheets, and qualification criteria should be stored in Notion or Confluence. The more structured and up-to-date this content is, the better the agent will perform.
    • A clear definition of lead qualification: Define what constitutes a qualified lead. Include criteria such as company size, industry, budget, and timeline. This will guide the agent’s classification logic.
    • GDPR compliance framework: Ensure you have a data protection officer (DPO) or legal counsel who can review the data processing activities. You need to define data retention policies and consent mechanisms for any personal data collected.
    • Infrastructure for open-weight models: If you plan to use open-weight models for regulated data, you need a server or cloud instance with sufficient GPU resources. This ensures that sensitive data does not leave your infrastructure.
    • A dedicated AI team: Engage a team with experience in AI automation, document extraction, and conversational agents. The team should be familiar with GDPR requirements and the specific needs of the healthcare industry.

    Step 1: Audit Your Current Lead Qualification Process

    The first step is to audit your current lead qualification process. Identify the workflows that are most time-consuming and error-prone. For example, if your team spends hours manually extracting data from PDFs and emails, this is a prime candidate for automation. The dedicated AI team will work with you to map out the current process, including the tools used, the data sources, and the decision points. This audit will help you define the scope of the pilot and establish a baseline for cycle time and error rate. Use a simple spreadsheet or a tool like Notion to document the current process. Include metrics such as the average time to qualify a lead, the error rate in data entry, and the number of leads processed per month. This baseline will be used to measure the impact of the automation.

    Step 2: Build the Document and Data Extraction Pipeline

    The second step is to build the document and data extraction pipeline. This pipeline will extract structured data from unstructured documents such as PDFs, emails, and CRM records. The team will use OCR and NLP models to identify key fields like company name, contact details, and intent signals. The extracted data will be stored in a database, such as PostgreSQL, with a pgvector extension for vector search. This allows the conversational agent to retrieve relevant context from your documentation. The pipeline will be configured to handle the specific document types and formats used in your organization. For example, if you receive many PDFs from healthcare providers, the pipeline will be tuned to extract data from these documents accurately. The team will test the pipeline with a sample set of documents to ensure accuracy and adjust the models as needed.

    Step 3: Develop the Conversational Agent for Lead Qualification

    The third step is to develop the conversational agent for lead qualification. The agent will interact with inbound leads, asking structured questions to determine fit, budget, and timeline. It will classify the lead into a priority tier and draft a personalized response based on the retrieved context from your Notion or Confluence documentation. The agent will use a retrieval-augmented generation (RAG) approach, querying the pgvector database to find relevant information. This ensures that the agent’s responses are grounded in your specific business context. The team will configure the agent to handle common questions and edge cases, such as leads asking about pricing or compliance. The agent will be tested with a set of sample conversations to ensure it handles these scenarios correctly. The team will also set up a human-in-the-loop mechanism, where a human reviewer approves any response that touches sensitive topics or high-value leads.

    Step 4: Automate Monthly Reporting with Extracted Data

    The fourth step is to automate the monthly reporting process. The system will extract data from your CRM, helpdesk, and marketing platforms. It will aggregate key metrics such as lead volume, conversion rates, and response times. The system will generate a draft report using the extracted data and your predefined templates in Notion or Confluence. A human reviewer will check the report for accuracy and add qualitative insights before it is finalized. This process reduces the time spent on manual data entry and formatting, allowing your team to focus on analysis and strategy. The report will be generated automatically on a scheduled basis, ensuring consistency and timeliness without additional headcount. The team will configure the reporting pipeline to pull data from the relevant sources and format it according to your templates. They will test the pipeline with a sample month of data to ensure the report is accurate and complete.

    Step 5: Ensure GDPR Compliance and Data Residency

    The fifth step is to ensure GDPR compliance throughout the system. All personal data will be processed within EU-based infrastructure, and data residency will be enforced by keeping regulated data on your own hardware using open-weight models. The system will log all data access and processing activities, providing an audit trail for compliance reviews. Data minimization will be applied by extracting only the necessary fields from documents, and data retention policies will be enforced automatically. The human-in-the-loop design will ensure that any data touching health records or sensitive personal information is reviewed by a human before further processing. The team will work with your DPO or legal counsel to review the data processing activities and ensure compliance with GDPR. They will document the data flow and the measures taken to protect personal data, creating a compliance report that can be used for audits.

  • Austrian Fintech Automates Order Status with a Voice Agent in Four Weeks

    The Support Team Is Drowning in Status Inquiries

    A 15-person fintech in Austria handles 300-500 customer support tickets per week. The majority are order and shipment status inquiries. Each inquiry takes a support agent 5-7 minutes to resolve: they check the ERP for order status, the CRM for customer history, and the carrier API for shipment tracking. The agent then drafts a response, reviews it for accuracy, and sends it. This process is repetitive, data-driven, and error-prone. The support team is stretched thin, and the company cannot hire more agents without breaking the budget. The pain is not a lack of technology; it is a lack of time and bandwidth to handle the volume of routine inquiries.

    Why Existing Solutions Fall Short

    The company has tried two approaches. First, they built a custom chatbot using a rule-based system. The chatbot handles simple inquiries but fails on complex ones. It cannot query the ERP or CRM in real time, so it provides outdated or inaccurate information. Second, they considered a generic AI chatbot. The chatbot can draft responses, but it lacks the context to handle the specific data sources the company uses. It also cannot meet the company’s ISO 27001 compliance requirements, because it stores data in the cloud and does not provide the audit logging the company needs. Both approaches fail because they do not integrate with the company’s existing systems or meet its compliance requirements.

    A Voice Agent That Integrates With Existing Systems

    The proposed approach is a voice agent that integrates with the company’s existing ERP, CRM, and carrier APIs. The agent uses the OpenAI API to understand the customer’s inquiry and draft a response. It queries the ERP for order status, the CRM for customer history, and the carrier API for shipment tracking. The agent then sends the response to the customer. The human-in-the-loop model ensures that any response touching financial data is approved by a person before it reaches the customer. The agent is deployed on the company’s own infrastructure, which meets the ISO 27001 requirements for data access, encryption, and audit logging. The custom REST API and webhooks connect the agent to the company’s systems, so the agent can query and update data in real time.

    How to Start: Four Concrete Steps

    The first step is a process audit. The dedicated AI team maps the current manual process, identifies the data sources, and defines the success metrics. The second step is the pilot design. The team selects one workflow (order and shipment status updates) and defines the scope, timeline, and success criteria. The third step is the pilot deployment. The team builds the voice agent, integrates it with the company’s systems, and runs the pilot for two weeks. The fourth step is the measurement. The team measures the cycle time and error rate before and after the pilot. The fifth step is the rollout. If the pilot meets the success criteria, the team scales the automation to other support channels.

  • HIPAA-Safe Contract Review AI: A 2-Week Pilot for Swiss Healthcare

    The Contract Review Bottleneck in Swiss Healthcare

    A 2,000+ employee healthcare and medtech company in Switzerland faces a specific bottleneck: contract review. Procurement teams receive vendor agreements, service-level agreements, and data-processing addenda in German, French, and Italian. Each document requires manual extraction of key clauses—payment terms, liability caps, data-handling obligations—before legal and finance can approve. The current process takes 4–6 business days per contract, with a 12% error rate in clause identification, particularly for multilingual documents. The finance team in Zurich needs a system that extracts structured data from these contracts, flags non-standard clauses, and writes the results directly into SAP or Microsoft Dynamics ERP, all while keeping PHI and financial data within HIPAA-compliant boundaries. The pilot scope is narrow: one workflow, two weeks, measurable baseline.

    LangGraph as the Orchestration Layer

    The pipeline uses LangChain for LLM calls and vector store interactions, and LangGraph for stateful, cyclic workflow orchestration. The graph has five nodes: ingest (PDF/DOCX parsing via Unstructured or Docling), extract (LLM-based clause extraction with a structured output schema), classify (risk scoring and language detection), approve (human-in-the-loop gate for financial and health data), and write (ERP integration via SAP BAPI or Dynamics OData). LangGraph handles conditional branching: if the document is in Swiss German, the extraction prompt adjusts for local legal terminology; if the clause involves PHI, the model routes to an on-premises open-weight model (Llama 3 70B or Mistral 7B) rather than an API call. The state object carries the document ID, extracted fields, confidence scores, and approval status. Every transition is logged for audit compliance.

    Model Routing and Multilingual Trade-offs

    The critical trade-off is model routing. Using OpenAI or Anthropic APIs for all tasks simplifies deployment but violates HIPAA if PHI is involved. The solution is a sensitivity classifier that runs before the LLM call: if the document contains PHI or financial data, it routes to an on-premises open-weight model; otherwise, it uses the API. This adds 15–20 ms of latency per document but ensures compliance. The second trade-off is multilingual extraction: a single multilingual model (Llama 3 70B) handles German, French, and Italian, but accuracy drops 8–12% for Swiss German legal jargon compared to English. The mitigation is a fine-tuned prompt template per language, validated against 50 ground-truth documents per language during the pilot. The third trade-off is ERP integration depth: writing to SAP via BAPI is reliable but slow (200–400 ms per write); Dynamics OData is faster but requires more field mapping. The pilot tests both to confirm which fits the client’s existing infrastructure.

    Pilot Scope and 2-Week Delivery Plan

    For a 2-week pilot, the scope must be ruthlessly narrow. Week 1: ingest 200 real contracts (60 German, 70 French, 70 Italian), run the extraction pipeline, and measure accuracy against human-verified ground truth. The baseline metric is cycle time (target: reduce from 4–6 days to under 24 hours) and error rate (target: reduce from 12% to under 5%). Week 2: integrate with SAP or Dynamics, test the human-in-the-loop approval gate, and validate that PHI never leaves the on-premises boundary. The pilot does not include end-to-end rollout, retraining, or managed operations—those are post-pilot. The deliverable is a measured before/after report, a working pipeline in the client’s environment, and a go/no-go recommendation for full rollout. The architecture is model-agnostic: if the client’s on-premises GPU cluster cannot handle Llama 3 70B, the pilot falls back to Mistral 7B with a documented accuracy delta.

    Rollout and Managed Operations

    Post-pilot, the rollout moves to managed AI operations: model monitoring for drift, prompt versioning, and incident response. For a 2,000+ employee organization, this means a dedicated SRE rotation that reviews model outputs weekly, handles edge cases, and updates the pipeline as contract templates evolve. The managed service includes SLAs for uptime (99.5%), latency (under 500 ms per document), and accuracy (under 5% error rate). The human-in-the-loop approval gate remains mandatory for any document touching money, health data, or a contract. The architecture plugs into existing CRMs, ERPs, and helpdesks via their APIs—no replacement, only enrichment. The multilingual coverage extends to all four Swiss national languages, with a fallback to English for documents in other languages. The system is designed to scale from one workflow (contract review) to adjacent ones (invoice processing, document extraction) without re-architecting the core pipeline.

  • UK Medtech Cuts Invoice First-Response Time to 6 Hours with On-Premise AI

    Background: A UK Medtech Distributor at 1,200 Headcount

    This case study is a composite drawn from patterns Forfis has observed across multiple engagements. We do not name real clients. The company described here is a UK-based medtech distributor with roughly 1,200 employees, operating in the 501-2000 band. It handles procurement, supply-chain coordination, and customer-facing service for hospital and clinic clients across the UK and Ireland. The existing stack includes a mid-market ERP, a CRM for customer records, and Microsoft Teams as the primary internal messaging channel. The finance and operations teams were running on a mix of spreadsheets, email threads, and a legacy invoice portal that had not been updated since 2019. The company had no dedicated AI team and had not previously deployed any machine-learning system in production.

    Challenge: 48-Hour Invoice Response, Zero New Hires, GDPR in the Loop

    The trigger was a 40 percent increase in supplier invoice volume over eighteen months, driven by a new product line and expanded distribution contracts. The finance team of eleven was processing invoices manually: extracting line items, matching them against purchase orders, flagging discrepancies, and posting to the ERP. Average first-response time to a supplier query about a disputed invoice was 48 hours. The operations director had a hard constraint: no new headcount in the current fiscal year, and GDPR compliance was non-negotiable because invoice metadata occasionally contained patient-identifiable information from hospital procurement orders. The deadline was six months to show a measurable reduction in cycle time before the next board review. The team needed to cut first-response time without adding a single FTE and without sending regulated data to a third-party cloud API.

    Approach: On-Premise Open-Weight Models, Predictive Scoring, and a Fixed-Scope Pilot

    Forfis ran a two-week process audit across the finance and operations workflows. The audit identified invoice processing as the highest-impact target: high volume, repetitive extraction, and a clear before/after metric. The pilot scope was fixed: one invoice category (supplier purchase orders with line-item extraction), one integration point (the existing ERP API), and one notification channel (Microsoft Teams). The architecture used open-weight models on the client’s own hardware, so no regulated data left the building. A retrieval-augmented layer pulled context from the client’s own procurement documentation and CRM records to improve extraction accuracy. Predictive scoring assigned a confidence value to each extracted field; items above 95 percent auto-posted, items below routed to a human reviewer in Teams. The dedicated AI team of four engineers and one product designer worked on-site for the first four weeks, then shifted to remote with weekly syncs. The pilot ran for eight weeks with a measured baseline captured in week one.

    Outcome: 48 Hours to Under 6, Error Rate Below 2 Percent

    The pilot cleared its threshold. Average first-response time for supplier invoice queries dropped from 48 hours to under 6 hours. Extraction error rate on line items fell from 11 percent to under 2 percent. The finance team’s manual review volume dropped by roughly 60 percent, because the predictive scoring layer auto-approved the high-confidence items. The remaining 40 percent of invoices still required human eyes, but the reviewers now worked from a pre-drafted, context-enriched queue in Teams rather than a blank spreadsheet. The ERP integration held: no data left the client’s infrastructure, and the GDPR data-processing record was updated to reflect the on-premise model deployment. The operations director reported that the team absorbed the 40 percent invoice volume increase without a single new hire. The six-month timeline was met, and the board review proceeded on the strength of the measured baseline.

    Lessons for Similar Teams

    • Baseline first, always. The pilot did not start until the team had a measured before/after baseline on cycle time and error rate. Without that number, the board review would have been a conversation about impressions rather than data. Every similar team should capture the baseline in week one, not after the pilot ends.
    • Model-agnostic architecture pays off. The client started with open-weight models on-premise for GDPR reasons. If a future use case requires a frontier API for a non-regulated workflow, the integration layer does not need to be rebuilt. Teams that hard-code a single vendor API into their architecture will face this problem.
    • Predictive scoring is the human-in-the-loop mechanism. The confidence threshold is not a suggestion; it is the architectural gate. Items above 95 percent auto-approve, items below route to a human. This is what makes GDPR Article 22 compliance operational rather than theoretical.
    • Integration through existing APIs, not replacement. The ERP, CRM, and Teams stack stayed intact. The AI layer sat on top. For a 1,200-person operation, a rip-and-replace project would have taken two years and a budget the company did not have.
    • Dedicated team beats rotating contractors. The four engineers and one product designer stayed on the engagement from audit through rollout. Consistency in the team meant the client’s internal stakeholders had a single point of contact and a shared context that did not reset every sprint.