How to Choose an AI Implementation Partner - 15 Questions, Warning Signs and Contract Clauses

In August 2025, MIT's NANDA team published a report that quickly spread through the industry: about 95% of corporate generative AI pilots had no measurable impact on the bottom line. The same report contains a less quoted number, though. Deployments built on tools bought from specialized vendors and on partnerships succeeded in about 67% of cases, while projects built entirely in-house succeeded roughly three times less often. Who implements AI, and how, largely determines the outcome - often more than the choice of model or tool.

According to MIT NANDA, 95% of AI pilots show no financial impact, yet deployments with partners succeed in 67% of cases. In-house, partner or hybrid, 15 questions to ask, warning signs, contract clauses and a pilot with success criteria.

The market doesn't make the choice easier. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs and unclear value, and that among the thousands of companies advertising "agents", only about 130 offer real agentic capabilities. In this post I cover when to build AI yourself and when to outsource it, which 15 questions to ask an implementation partner, how to spot warning signs, what has to be in the contract and how to plan a pilot with clear success criteria. You can ask me the same questions.

Three models: in-house team, partner or hybrid

A company that wants to implement AI has three options in practice: build the skills in-house, outsource the project to an agency, software house or freelancer, or combine both approaches.

The MIT finding doesn't mean in-house knowledge is unnecessary. The report also points out that companies did better when line managers from the business drove adoption, not just a central AI team. A hybrid most often works best: a partner builds the first deployments and transfers knowledge, while from day one there is a process owner in the company who will take the system over.

When to build yourself and when to outsource

An in-house team makes sense when AI is part of your product, when you plan many deployments a year and when you can hire at least two people, so knowledge doesn't depend on one. Outsourcing wins when you want to test one process quickly, have no experience with language models or need someone who has seen a similar project at another company. A freelancer works for a small, well-defined task, but for a system a whole department depends on, the risk of one person leaving is real.

Before deciding, count the total cost, not just the implementation price: maintenance, model costs, your people's time and changes after launch. I show how to calculate costs and returns in the posts on AI implementation costs and automation ROI. If you're still looking for your first process to automate, start with the post on where to start with AI implementation.

15 questions to ask a partner

Most conversations with vendors revolve around features and price. The questions below cover things few people ask about, but which decide whether the system still works a year later.

Problem and data

1. 1.What business problem are we solving and how will we measure success? A good answer includes a specific metric, such as ticket handling time, and its value before the project.
2. 2.What data do you need, and what if it's in poor shape? Gartner estimates that through 2026 companies will abandon 60% of AI projects that lack AI-ready data. I cover getting data in order in the post on data readiness.
3. 3.Can you show a similar deployment and put us in touch with the client? A conversation with a client tells you more than a presentation.

Quality and control

1. 4.How will you measure answer quality? Look for an answer about a test set prepared before launch, not impressions from a demo. I describe how such tests are built in the post on testing AI applications.
2. 5.What happens when the model gets it wrong? The partner should point to approval points, escalations and a safe default state - more in the post on human-in-the-loop.
3. 6.How do you monitor the system after launch? Logs, alerts and regular quality reviews are a requirement, not an extra - details in the post on monitoring AI in production.

Costs

1. 7.What will monthly maintenance cost at our volume? Model, hosting and tool costs should be calculated on your numbers. I show how to lower them in the post on API cost optimisation.
2. 8.What's included in the price, and what counts as a change? Prompt fixes after launch, new categories, a model switch - it's worth knowing upfront how they're billed.
3. 9.What does it cost to move to a different model provider? A well-designed system lets you change the model without rewriting everything.

Security and compliance

1. 10.Where does our data go and who processes it? If the partner or the model provider processes personal data, you need a data processing agreement under Article 28 of the GDPR - I cover this in the post on GDPR and AI.
2. 11.Can the model provider train on our data? The answer depends on the service and plan, so ask for the specific terms.
3. 12.How does the system fit the AI Act? The partner should be able to assess whether it's a high-risk system and what obligations your company has - more in the AI Act guide.

Ownership and exit

1. 13.Who owns the prompts, workflows, code and data? The answer must be in the contract, not just in the conversation.
2. 14.What does the handover look like? Documentation, access, training for your people and a support period after launch.
3. 15.What happens if we end the collaboration? This is the exit plan question: export of data and configurations, a transition period and help with the handover.

A partner who answers questions 7, 9, 13 and 15 with generalities is probably counting on you staying for a long time - regardless of results.

Warning signs

• A demo instead of a conversation about the process. A flashy presentation in the first week, before anyone has seen your data or understood the process.
• No questions about data. A partner who doesn't ask where the data will come from and what shape it's in doesn't yet know what they're building.
• An accuracy promise without tests. "99% accuracy" without a test set and a definition of an error means nothing.
• An agent for everything. Gartner calls this agent washing: an ordinary chatbot or assistant sold as an autonomous agent.
• A closed platform with no export. If prompts, workflows and data can only live with the partner, changing partners means starting over.
• An implementation price without maintenance costs. Silence about monthly model and support costs usually means a surprise after launch.
• No one on your side in the project plan. A deployment without a process owner in the company rarely survives its first year.

The contract: what it must include

An AI implementation contract differs from a regular software contract, because it also covers prompts, workflow configurations, test sets and data. A few clauses deserve particularly close attention.

The most important point concerns copyright. Under Polish copyright law, a transfer of economic rights requires written form to be valid (Article 53) and covers only the fields of use explicitly listed in it (Article 41(2)); many jurisdictions have similar formal requirements. A short prompt may not be a protected work at all, so its protection has to come from the contract: list prompts among the deliverables, transfer rights where possible and cover the rest with a licence and a confidentiality obligation.

If the partner delivers the system as a cloud service, some switching rules come straight from the EU Data Act. Since 12 September 2025, providers of data processing services must allow customers to switch providers with a notice period of no more than two months, and from 12 January 2027 they may not charge for the switch itself, including data transfer. It's still worth describing the exit plan in the contract, whether or not the rules cover your case.

A pilot with success criteria

A pilot has to answer one question: is it worth investing further. For the answer to be honest, the criteria must be written down before launch, not after seeing the results. A good pilot covers one process, usually lasts four to eight weeks, has a fixed price and ends with a decision: roll out, improve or close the project.

Process: routing incoming emails to departmentsBaseline: 4 hours of manual work a day, 6% misroutedPilot length: 6 weeks, fixed priceSuccess criteria (example):- accuracy on a set of 300 emails: at least 92%- handling time: down by at least 50%- model costs: up to 75 EUR a month at current volumeDecision after the pilot: roll out, improve or close

The numbers in the example are an illustration, not a norm. Your criteria follow from the baseline and from how much has to change for the investment to pay off. A pilot without criteria written down in advance almost always ends with "promising results" that settle nothing. Check the cost criterion at real volume, because model costs grow with the number of requests.

A step-by-step plan for choosing a partner

1. 1.Describe the process and the metric you want to improve, with its current value.
2. 2.Decide on the collaboration model: in-house team, partner or hybrid.
3. 3.Appoint a process owner on your side.
4. 4.Talk to two or three partners and ask each the same 15 questions.
5. 5.Ask for a client reference from a similar deployment.
6. 6.Review the draft contract: rights, deliverables, data, data processing, variable costs and exit plan.
7. 7.Start with a paid pilot with success criteria written down before launch.
8. 8.Decide based on the criteria, not on impressions from a presentation.

---

I help companies choose and implement AI so that a year later the system still works and belongs to them - from describing the process and success criteria, through the pilot, to a contract with clear rights and an exit plan. I do this as part of AI consulting and AI automation. Get in touch - and ask me the same 15 questions. You'll find a working demo with human approval and an audit trail in the portfolio: AI Document Automation.

Worth reading next:

• Where to Start with AI Implementation in Your Company
• How Much Does AI Implementation Cost? Debunking the Million-Dollar Invoice Myths
• AI Automation ROI - How to Calculate Returns Before You Sign Anything
• Data Readiness - How to Prepare Your Company's Data for AI Before You Spend a Dollar on Implementation