Web agencies: selling credible AI offers and delivering them white label

Your clients are asking you for “AI”. A credible web agency AI offer is not sold like a demo: it is qualified, priced in three lines, stopped if the criteria are not met, and properly contracted. Here are six offers to package, the grid for sorting requests, the pricing method and how white label works in practice.

Why your clients are asking you for AI

The use of AI is growing fast in businesses. According to Eurostat, 20.0% of EU enterprises with at least 10 employees used at least one AI technology in 2025, up from 13.5% in 2024. The most common uses: analysing text (11.8%), generating images, videos or sounds (9.5%) and generating text or speech (8.8%) [1].

Your clients ask you because you already know their website and their data. But behind “we would like some AI”, there is almost always a more concrete problem: an overloaded support desk, product sheets to produce, an internal search that finds nothing, re-keying between two tools. A credible offer starts from that problem, promises a measurable result and says in advance what will happen if the result is not achieved.

Six AI offers to package

1. Document assistant (RAG)

This is the most requested offer and the one most often botched. The iones lab measured on 180 questions that a hybrid search followed by reranking raises precision@5 from 0.61 to 0.84, and that mandatory citation and the right to refuse bring hallucinations down from 11% to 2% (sheet EXP-05). The technical detail is in our article Enterprise RAG: why reranking matters. What you need to sell: a rate of correct answers measured on the client’s real questions, not “a ChatGPT on your documents”.

2. Triage agents

Qualifying and prioritising tickets, emails or leads. In iones lab sheet EXP-06, on 412 real anonymised tickets, the agent correctly qualified 91% of the tickets and found the technical lead’s priority in 87% of cases, for €0.04 per ticket triaged. Sell it with human validation before any message to the end client and actions limited to an allowlist.

3. Assisted content

The riskiest request for your clients’ SEO. In sheet EXP-03, fully generated articles were indexed only at 58%, against 92% for an outline produced by AI and then enriched by an expert, with 65% less human time than fully human writing (2 hours 10 minutes against 6 hours 15 minutes). The offer to sell is therefore a production chain with human expertise, not a mass generator. For the visibility of content in Google’s AI answers, see our guide AI Overviews in France: staying visible when position 1 is no longer enough.

4. Intelligent product search

For an e-commerce site, semantic search understands “waterproof hiking shoes for children”, but lexical search remains indispensable for references, brands and sizes. Hybrid is the right architecture; the deliverable must include a set of real queries taken from the search logs to prove the gain.

5. n8n automations

n8n is very effective for connecting CRM, forms, email and LLMs. Mind the licence: the Sustainable Use License allows use for internal needs and explicitly allows consulting services, such as building workflows for a client. On the other hand, it lists as not permitted offering n8n white label for a fee and hosting n8n while charging for access [2]. In practice: deploy the instance in the client’s name, invoice the building and maintenance of the workflows, and for any other arrangement (a shared instance that you operate for several clients, n8n embedded in your product), ask n8n (license@n8n.io) or plan for a commercial licence.

6. Audit and awareness building

Article 4 of the AI Act, on AI literacy, has applied since 2 February 2025; Omnibus Regulation 2026/1744 softened it into an obligation to encourage the upskilling of teams [3][4]. It is a legitimate entry point: mapping existing uses, identifying three priority use cases, training the teams. We offer it in our AI strategy and awareness building offer. Condition for credibility: the audit leads to a costed POC, not to a presentation.

Qualifying a request: the value, feasibility, risk grid

Score each criterion from 1 (low) to 3 (high) with the client, during the scoping workshop.

Decision rule: total ≥ 22 out of 27, go ahead; between 16 and 21, start with an audit or a very short POC; below that, decline or reformulate. Any score of 1 on the “impact on a person” question requires a prior regulatory analysis: some uses (recruitment, credit, education, etc.) fall under the high-risk systems of Annex III, whose obligations have been postponed to 2 December 2027 by the Omnibus [4], but not removed.

Pricing: build, run and inference costs

An AI project has three price lines, to be presented separately.

The inference line is calculated, not estimated by guesswork:

coût mensuel ≈ nb de requêtes/mois
             × (tokens d'entrée × prix d'entrée + tokens de sortie × prix de sortie)
             + reclassement + embeddings des nouveaux documents

Two reference points from the iones lab: triaging a ticket cost €0.04 in EXP-06, or about €200 per month for a hypothetical volume of 5,000 tickets; a document assistant that sends 5 passages of 400 tokens consumes about 2,000 tokens of context per question (EXP-05). Measure these values during the POC rather than promising them beforehand.

Run is the line that agencies forget. Yet an AI system drifts (corpus, models, questions): without run, quality falls silently and it is your agency that the client will hold responsible.

Setting the stop criteria for a POC before you start

A POC without stop criteria becomes a project without end. Write the thresholds into the commercial proposal and have the client approve them.

Stop or redirect if the data is not accessible after two weeks, if the thresholds are not approached after two iterations, if the unit cost exceeds the value created, or if the regulatory analysis reveals a high-risk use that the client is not ready to manage. A POC stopped cleanly strengthens your credibility; a POC that drags on destroys it.

Contracting: NDA, ownership, data, AI Act

• Confidentiality down the chain: NDA between the client and the agency, and between the agency and each subcontractor, with the same scope.
• Ownership: specify what is assigned (code, prompts, workflows, evaluation set, documentation) and what is not (third-party models, subject to their providers’ terms; open-source components, subject to their licence).
• Personal data: if the agency processes data on behalf of the client, a contract compliant with article 28 of the GDPR is mandatory; using a sub-processor requires the client’s prior written authorisation [5]. If people or suppliers located outside the EU have access to the data, check the transfer framework (adequacy decision or standard contractual clauses) [6]. The simplest approach: develop on anonymised or synthetic data and keep production data in the EU.
• AI Act, article 50: applicable since 2 August 2026. A conversational assistant must be designed so that the user knows they are interacting with an AI; synthetic content must be marked in a machine-readable format; deepfakes and texts published to inform the public must be disclosed, except, for the latter, where there is human review with editorial responsibility [3]. Determine in the contract who is the provider and who is the deployer: depending on who places the system on the market under their own name, the agency may carry provider obligations. Details and timetable in our article AI Act: what has applied since August 2026.
• Liability: an AI makes mistakes; write it down, describe the guardrails (human validation, scope of action) and limit the commitment to what you measure.

Margin and resale model

Our recommendation: take your margin on build and run, where your value lies, and re-invoice inference transparently. A hidden margin on tokens always ends up showing.

How white label works with iones

White label answers a simple observation: you have the client relationship and the knowledge of the business; recruiting a senior AI team for a few projects a year is not profitable. The principle, which we also detail for white-label peak load cover:

• You keep the client: you are the single point of contact, under your brand, with your project management tools.
• We qualify the request together with the grid above; response within 48 to 72 hours, discussion without commitment.
• NDA signed before any exchange of data.
• A senior French-speaking team, in the same time zone as France, builds the POC and then the build, with the methods measured at the iones lab.
• Run stays in your hands or we provide it for you, always under your brand.

For one-off or long-term reinforcement, see our white-label dedicated developers offer.

Frequently asked questions

Launching your first AI offer without risking the client relationship

Start small and measurable: choose one of the six offers, qualify two or three client requests with the grid, write a proposal with three price lines and quantified stop criteria, and lock down NDA, data and AI Act in the contract. A first measured success is worth more than three demos.

If you want to sell AI without recruiting a dedicated team, discover our white-label dedicated developers or let’s talk about your first AI project: response within 48 to 72 hours, discussion without commitment, under NDA.