Skip to content

Integrations and data17:07

Offboarding an AI vendor at an 86-person company and a 12-item exit checklist

An 86-person company migrated from a cloud AI assistant vendor. See what the exit package included, how long re-indexing a 1.5 GB corpus took, and why Data Act timelines did not apply.

AI-generated image.

Listen to the article summary

Synthetic voice.
0:00
0:00

Treat this as a free practical guide: solid for getting started and discussing with your team, but not sufficient as the sole basis for a major decision. B2B technical components distributor, 86 people, four warehouses in western Poland, about 1,400 service and technical tickets per month. In 2024, they purchased a cloud AI assistant with document search: 1.5 GB of datasheets, assembly manuals, price lists, and internal notes. Previously, a sales rep searched for a parameter in network folders, taking an average of 6–9 minutes per query, or called a product manager for unusual substitutes. The assistant reduced this to several dozen seconds, and within a year it became indispensable, integrated directly into the first-line support script. Two factors triggered the exit: a subscription increase from 2,900 to 4,600 PLN net per month and zero control over log retention. The board decided to migrate everything to a self-hosted instance. We describe one exit process; one metric, the cost of recomputing the corpus, is benchmarked against a second, similar project and noted accordingly in the text.

AI-generated image.

How they assembled it

Week 0 involved an audit of three documents: the main agreement, the data processing agreement, and the terms of service. An important finding emerged right away: the agreement was signed in 2024, meaning Chapter VI of Regulation (EU) 2023/2854 on switching data processing services, applicable from September 12, 2025, sets a separate date for earlier contracts: September 12, 2027. Assessing this clause in a specific contract requires legal counsel, but the practical conclusion was simple: the regulation calendar could not serve as leverage, leaving only contractual terms and Article 28(3)(g) of the GDPR, requiring the return or deletion of data after processing ends, an obligation also detailed in UODO guidelines. The team still adopted the regulation timelines as a negotiation baseline: 2 months notice to initiate the process, a 30-calendar-day transition period, 14 business days to notify technical infeasibility, and an upper extension limit of 7 months. Week 1: formal notice to the vendor detailing a 12-item exit package: source files in original formats, chunking structure with metadata, vectors along with the embedding model name and version, system prompts and agent configurations, transcripts and recordings with timestamps, query and response logs, fine-tuning dataset, adapters, permission matrix, dictionaries and taxonomies, integration definitions, and written confirmation of deletion. Weeks 2–3: vendor response and format negotiations. Weeks 3–5: export and completeness audit. Weeks 5–7: re-vectorization and regression tests across 120 benchmark queries drafted by support. Week 8: cutover. Weeks 9–12: resource download and deletion verification. Internal effort: IT administrator 90 hours, support lead 30 hours, external legal counsel 14 hours, COO 12 hours.

The invoice

The vendor charged 3,500 PLN net for the export, billed as out-of-scope services: 14 hours at the contract hourly rate. Add 14 hours of legal counsel (around 4,900 PLN), 132 hours of internal team time, and our technical support within the 18,000–26,000 PLN net range, depending on how many chunks required manual fixes. Vectorization itself proved the cheapest item: 1.5 GB of documents yielded about 47M tokens post-extraction, which cost under 300 PLN on an entry-level embedding model; a second similar project with a 2.1 GB corpus cost just over 500 PLN. Chunk overlap increased token count by roughly 20%, so every chunking strategy tweak shifts that figure. The true cost lies in labor hours, not API calls. Post-migration maintenance runs 1,100–1,800 PLN monthly (server, monitoring, model budget), roughly 3,000 PLN less per month than the renewed subscription price for the same volume of about 11,000 queries monthly. This yields ROI in 9–13 months, assuming the corpus needs no rebuild and no secondary project intervenes. Had the contract been signed after 12 September 2025, transition support fees would be legally capped at incurred costs, with switching fees banned entirely starting 12 January 2027. That 3,500 PLN line item would look quite different then.

What went wrong

First setback: the vector index arrived as a file with 68,000 vectors, missing the model name and version. Without this metadata, it is just a collection of numbers unable to process new queries; new inputs must be vectorized with the identical model. Resolving this took 9 business days and still ended in re-vectorizing the corpus from scratch. Second: ticket history export covered roughly 25,000 threads across 18 months, but lacked permissions matrices and consistent attachment references; 7% of files could not be mapped to tickets, burning 40 hours of manual reassembly. Call transcripts lacked timestamps, stripping their evidentiary value for dispute claims. Third: the vendor withheld system prompts and agent configurations as proprietary know-how. The team had to reverse-engineer them from query and response logs: 22 hours of work, and post-cutover 14 of 120 control test queries scored lower than before. Calibrating them took two additional weeks. Separate issue: nobody received the fine-tuned model weights, because technically there was nothing to export; downloading fine-tuned weights from an external provider is rarely supported, as detailed in OpenAI fine-tuning documentation. Written confirmation of data erasure arrived only after a follow-up notice, two weeks past the deadline stated in the demand letter.

AI-generated image.

What we would do differently

Three practical factors determine the success of such an exit. First, keep a local copy of source files and chunked data with metadata from day one, not from the termination date. That is the only part of the package nobody disputes later. Second, include the embedding model name, version, and chunking strategy in the technical documentation alongside passwords and endpoints. Missing details cost 9 days here. Third, run quarterly test exports on a random sample of 200 tickets to verify data ingestion elsewhere, not just button functionality. Review three topics with legal counsel before signing the next contract, treated as a baseline rather than a template: an itemized exit package inventory with exact file formats; delivery deadlines with at least 30 days to download post-termination, including a 14-business-day window to report technical feasibility issues; and fee structures limited to documented costs, aligned with Data Act timelines. Define vendor points of contact for handover to the new contractor in advance, along with formal confirmation of data deletion. Unilaterally imposed exit restrictions may be contestable in B2B contexts, subject to legal review. Designation of the national supervisory authority under the Data Act remains pending in draft legislation; track updates via the Ministerstwo Cyfryzacji. Personal data matters remain under UODO jurisdiction.

This scenario makes sense when monthly subscription fees exceed several thousand PLN, the document corpus remains stable, and internal staff can allocate 0.3 FTE over a year. Practical experience shows migrations below 1,500 PLN monthly rarely yield positive ROI: labor hours exceed cost savings, while risks of reduced response quality remain. Migrations also fail when documentation changes weekly without version control, where underlying file organization is the core bottleneck rather than the vendor. For contracts signed in 2024 evaluated for data portability, begin by reviewing three core documents and a 12-item checklist. Contact our team for technical audits; consult legal counsel for contract analysis.

Source materials

Sources consulted during research. The text above is our own; these third parties are not responsible for its content and have not authorised it.

No-obligation call

Schedule a workshop with Paul

Thirty minutes or a full workshop, choose what suits you. Book a slot directly in the calendar and get instant confirmation.

Paul Lazniak

HEXART Founder