AI agency or AI consultancy: how to choose for your custom projects

You have a custom AI project in mind: an agent that prepares sales files, a tool that sorts customer requests, a platform that scores prospects. Now you need to choose who will build it. AI agency, AI consultancy, independent consultant: on their websites, they all promise roughly the same thing.

At AI Sisters, an AI consultancy, we train teams and develop custom agents with the AI Sisters Lab. We read a lot of project briefs, and the projects that succeed look alike. Here are the six criteria we recommend you check, whatever the provider, including us.

In brief

• An AI agency usually starts from a need that has already been described and delivers it. An AI consultancy first helps you decide what to build, then follows the project through to actual use.
• If your need is written down and costed, a development partner may be enough. If you are torn between several ideas, start with a scoping phase.
• Ask for an estimated gain for each use case before any development, then a POC on your own data.
• Have the contract state that the code belongs to you.
• Check where your data goes: hosting, data processing agreement (DPA), no model training on your exchanges.
• Include user training in the project. A tool nobody opens brings in nothing.

Read also: integrating AI into a company, the 8-step method, how to create an AI agent and measuring the ROI of AI training.

AI agency, AI consultancy, AI consultant: who does what?

The labels overlap and many players do a bit of everything. Yet their starting point differs, and that is what shapes the project.

These portraits describe tendencies. Some agencies have a real consulting team, some consultancies subcontract all their development. So it is better to judge each provider on precise criteria than on its label.

To put it simply. For a house, the architect checks that the project stands up and fits the budget, then the building firm builds it. Some players do both. Trouble starts when the plans are drawn by someone who has not seen the site.

Six criteria for choosing your AI partner

These criteria apply to an agency as much as to a consultancy. For each one, we give you the question to ask in the meeting.

1. Scoping and ROI, before the first line of code

An AI project launched without figures is hard to stop when it does not pay off. Gartner predicts that more than 40% of agentic AI projects will be abandoned by the end of 2027, because of escalating costs or value that is never demonstrated. Scoping is there to sort projects before spending.

At JCDecaux’s IT department in France, several divisions each had their own AI project idea, and none of them was costed. In seven weeks, the AI Sisters Lab ran five use cases through the same grid: a co-design workshop with the teams, a technical and business analysis, and an impact and feasibility score out of ten. The IT department came away with €2.19M in estimated annual gains across all the cases and two MVPs selected. One case was dropped because its gains did not cover the investment. The full details are in the JCDecaux case study.

The workshops also produced very concrete measurements. Preparing a sales file took 2 h 35 min, and the MVP aims for 25 minutes. With this kind of figure, measured with the teams, management can make a decision.

The question to ask: “How do you put a figure on the gain from each use case, and in which case would you advise us not to develop it?” A partner who can say no saves you money.

2. Change management

A very well built agent is still useless if the teams keep working as before. Your partner must know how to bring needs to the surface with the business teams, then get the tool adopted.

ENGIE’s GET programme is a good example. The journey started with a Copilot workshop and an AI Show for the Adoption teams. On 24 March 2026, an Agent-a-thon morning brought out 16 use cases. Five months later, agents were running in production, including a Learning Coach that became the programme’s official training tool. The ideas came from the teams, and they use them. The full journey is described in the ENGIE case study.

The question to ask: “How do you involve the future users, and who measures usage after delivery?”

3. Code ownership

On a custom development, ask who owns the code at the end. Depending on the contract, you come away with a tool you can develop freely, or with a subscription that is hard to get out of.

At Energies France, an energy broker, the AI Sisters Lab developed a platform that scores each prospect out of 100 and prepares two batches of calls a day for 30 salespeople. The code belongs to the client. The platform reads the existing system without ever changing it, and it was delivered in increments over four months. Before, the salespeople spent 80% of their time looking for leads. Today, they spend 80% of their time on calls. See the Energies France AI sales platform.

The question to ask: “Who owns the code, the prompts and the documentation at the end of the project, and what happens if we end the contract?”

4. Data security

An AI agent reads your documents, your emails or your CRM. Before signing, you need to know where this data goes and what happens to it:

• where the tool is hosted: in your environment, on a European cloud, elsewhere;
• which AI models it uses, and under which contract. A DPA must guarantee that your data does not train the models;
• which access rights the agent gets. Read-only by default, write access only when it has been approved;
• how the project complies with the GDPR and the AI Act.

On Lab projects, these constraints are set from the scoping stage. The tools are built in your environment or on services covered by a DPA, and your data is not used to train the models.

To put it simply. Handing your data to an AI provider is like handing them the keys to your premises. You want to know who has a spare set, which doors they can open and whether anyone photocopies your files on the way out.

5. The ability to train your teams

An agent changes the way people work. Its users need to know how to talk to it, check what it produces and spot its mistakes. If your partner can only develop, you will need a second provider for training, and two contacts passing the buck when usage is slow to take off.

This is where a consultancy that trains adds the most. AI Sisters has worked with more than 900 companies and trained 60,000 employees. Our training is rated 9.5/10, and our Qualiopi certification means it can be funded by your OPCO.

The question to ask: “Who trains the users, on which cases, and who answers their questions after delivery?”

6. References you can check

Ask for case studies with a name, a date, a scope and a measured result. A logo on a home page says nothing about what was delivered. “Five use cases scored in seven weeks” or “agents in production in five months” can be checked, and you can ask to talk to the client about it.

The question to ask: “Can you put us in touch with a client whose project is similar to ours?”

Agency or consultancy: three situations, three answers

The right starting point depends mainly on where you are now.

Your need is written down and costed

You know which task to automate, how long it takes and what you want to gain. A development partner can start quickly, whether it is an agency or a consultancy’s Lab. Still insist on a POC on your data and ownership of the code. For a targeted need, our custom AI agent creation offer takes three days of work, with a testing phase by your teams in real conditions before the final iteration.

You have several ideas, none of them costed

That was the situation at JCDecaux’s IT department in France before scoping. Developing the first idea that comes along is a gamble. A scoping project runs each case through the same grid, and you only fund the ones that hold up. Our AI consulting support starts from your business divisions and delivers a prioritised roadmap.

Your AI tools are there, but little used

Copilot or ChatGPT licences sitting idle, teams hesitant to use them: a new custom tool will not change that. Train the teams first on their real use cases. The ideas for agents then come from the business teams themselves, as at ENGIE.

Launching your tender in five steps

Whether you are consulting an agency, a consultancy or both, these five steps make the answers comparable.

1. Describe the problem on one page

The task, who does it, how long it takes, what you want to get instead. Leave the solution aside: proposing it is the provider’s job.

2. List your constraints

The data involved, the tools in place (Microsoft 365, Google Workspace, your CRM) and your IT department’s security rules. A provider who asks no questions about this has surprises in store for you.

3. Ask for a short, costed scoping phase

A few weeks, a decision grid announced before the workshops, and a verdict for each use case. At JCDecaux, it took seven weeks from launch to final presentation.

4. Insist on a POC on your data before the go-ahead

A demo on dummy data tells you nothing about your files, with their gaps and exceptions. The POC is there to check this before committing the MVP budget.

5. Include adoption in the quote

User training, usage measurement at 30 and 90 days and a contact person after delivery should all appear in the proposal, alongside the development.

Not sure yet where your company stands? Our free AI assessment measures its maturity in 15 questions and about 4 minutes.

Key takeaways

• Judge a provider on precise criteria rather than on the “agency” or “consultancy” label.
• No development without an estimated gain per use case and a POC on your data.
• The code must belong to you, and the contract must say so.
• Hosting, DPA and access rights are settled at the scoping stage.
• User training is part of the project from the quote onwards.
• Ask to talk to a client whose project is similar to yours.

Sources

• AI Sisters, From idea to tool: JCDecaux’s IT department in France identifies €2.19M in annual gains with AI, case study, accessed 6 November 2026.
• AI Sisters, Sales prioritisation: Energies France equips its 30 salespeople with a custom AI platform, case study, accessed 6 November 2026.
• AI Sisters, AI agents in the workplace: the ENGIE case study, case study, accessed 6 November 2026.
• Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, press release of 25 June 2025.