AI development contracts are rarely one-size-fits-all, yet many teams sign deals without negotiating critical terms around data ownership, model performance guarantees, and maintenance responsibilities. The difference between a favorable agreement and a costly mistake often comes down to five to ten specific negotiation points that most buyers overlook. This guide walks you through exactly what to push back on.
Ownership and Intellectual Property Rights
Before signing, determine who owns the trained models, training data, and underlying code. Some vendors retain ownership and license the model to you; others transfer full ownership after payment. This matters enormously if you want to switch providers later or modify the model in-house.
Push for explicit clauses stating whether you receive:
- The final trained model weights and architecture
- Access to training scripts and preprocessing pipelines
- Rights to retrain or fine-tune without additional licensing fees
- Ownership of synthetic or augmented data created during development
If the vendor retains ownership, negotiate a perpetual, irrevocable license at minimum—not a subscription you must renew annually.
Data Handling and Security
Your training data is sensitive. Specify in writing how the vendor handles it during and after the project. Ask whether data is encrypted in transit and at rest, where servers are physically located, and whether the vendor uses your data for any other purpose (including model improvement across clients).
Include a clause requiring data deletion within a defined timeframe (typically 30–90 days) after project completion, unless you've agreed otherwise. If regulations like GDPR or HIPAA apply, require explicit compliance statements and audit rights.
Model Performance SLAs
Vague promises like "highly accurate" don't hold up when your model underperforms in production. Define measurable success metrics upfront: accuracy, precision, recall, F1-score, latency, or whatever aligns with your use case.
Typical structures include:
- Baseline metrics: Model must achieve ≥95% accuracy on your test set
- Production SLAs: Model inference must complete within 200ms for 99% of requests
- Retraining triggers: If accuracy drops below 90% in production, vendor retrains at no additional cost within 30 days
Tie remedies to missed targets—credits, retraining, or scope extension. Avoid contracts that offer "best effort" without consequences.
Timeline and Milestone Payments
Link payments to concrete deliverables, not just calendar dates. Instead of paying 50% upfront and 50% at completion, structure it around milestones:
- 25% upon contract signature
- 25% upon delivery of exploratory data analysis and feature engineering plan
- 25% upon delivery of trained model meeting baseline SLAs
- 25% upon 30-day production stability (zero critical bugs)
This aligns incentives and gives you leverage if the project stalls. Also negotiate a realistic timeline—most custom ML projects take 3–6 months for scope discovery through production deployment. Any vendor promising a full-stack solution in 4 weeks is cutting corners.
Maintenance and Model Decay
Production models degrade over time as data distributions shift. Clarify who monitors performance and who pays for retraining. A common approach:
- Vendor provides 90 days of free monitoring and one retraining cycle included
- After 90 days, retraining costs are à la carte or covered under a support contract (typically 15–25% of original project cost annually)
- Vendor alerts you if performance drops >5% month-over-month
Without this clause, you inherit all responsibility and cost the moment deployment occurs.
Change Requests and Scope Creep
Define the original scope in granular detail: number of target variables, expected training data volume, number of model iterations, and specific algorithms or frameworks. Any request outside this scope triggers a change order with revised timelines and fees.
Example: "Three classification models using tabular data up to 10GB" is clear. "Build an AI solution" is not. Use detailed project briefs or statement-of-work documents as contract appendices.
Limitations of Liability and Indemnification
Negotiate reasonable liability caps. Most vendors won't accept unlimited liability for ML projects—the risk is simply too high. A typical middle ground: each party's liability is capped at the total contract value or 12 months of support fees, whichever is larger. Carve out exceptions for gross negligence or data breaches.
Require the vendor to indemnify you if the model infringes third-party IP or uses open-source code in violation of its license.
Frequently Asked Questions
Q: Should I require the vendor to guarantee minimum accuracy on unseen data? Yes, within reason. Tie guarantees to your hold-out test set, and include retraining obligations if the model underperforms, but avoid unrealistic targets like 99% accuracy on highly imbalanced datasets.
Q: What happens if the vendor's model performs well in staging but fails in production? Include a post-deployment support period (30–90 days) where the vendor debugs performance issues at no extra cost, and define what "fails" means using production monitoring metrics you've agreed on upfront.
Q: Can I negotiate data residency requirements if I'm in a regulated industry? Absolutely. Specify which data centers or geographic regions are acceptable, and require contractual warranties that data never leaves those regions without written consent.
Use Mercoly to compare AI and Machine Learning Development providers side-by-side, review their standard contracts, and identify which terms they're willing to negotiate before you enter discussions.