Two quotes land on your desk. Same stated need, same order of magnitude. The first runs to one page, the second to six. A year later, one project is still running without costing more than planned, the other has doubled its bill and nobody knows who is supposed to fix the errors. The difference almost never lies in the headline price. It lies in what the quote agreed to describe, and in what it left vague.
An AI agent is not a plumbing job. It consumes data, depends on a model provider, produces variable answers, and keeps costing money after delivery. A quote that says nothing about those four points is not merely incomplete: it is unusable as a contract basis.
Key Takeaways
- An AI agent quote describes use cases, not technology: triggers, input data, expected actions, failure cases, out-of-scope items.
- Recurring costs (model calls, hosting, maintenance) are priced separately, with an explicit volume assumption.
- The data section must be written down: purpose, legal basis, retention period, location, processor agreement under Article 28 of the GDPR.
- Since 2 August 2026, Article 50 of the EU AI Act requires informing people who interact with an AI system: the quote must state who carries that obligation.
- Without acceptance criteria, warranty and a reversibility clause, the quote leaves you with no remedy the day the agent gets something wrong.
What the law requires, and what it leaves to negotiation
The most common confusion comes from a simple misunderstanding. In France, a quote is not mandatory for every service. It is mandatory for a closed list of activities, mainly building work and repairs, and for certain regulated services such as personal assistance. An AI agent project is intellectual work: no general quote obligation applies between businesses.
That does not make the document a free-for-all. As soon as it is issued, it changes nature. Service Public Entreprendre (2022) states that a quote is a contract offer that binds the provider as soon as the client accepts it, and that the client is only bound once they sign it. The DGCCRF (2023) adds that the quote becomes a binding contract once signed by both parties, and that it must state the period during which the offer remains valid.
In other words, a signed quote is a contract. Anything missing from it will be negotiated in a hurry, to your disadvantage, when the dispute arises.
Where a quote is mandatory and absent, there is a penalty: the administrative fine can reach 3,000 EUR for an individual and 15,000 EUR for a company, according to Service Public Entreprendre (2022). The amount is not the point here. The point is that the particulars required for those activities, an itemised breakdown by quantity and price, the total excluding tax and including tax with the VAT rate, the validity period, show the level of precision a business owner is entitled to expect even outside a legal obligation.
The six blocks of a readable AI agent quote
A properly built AI agent quote has six blocks. The first three commit the provider's understanding of your need; the next three commit the economics of the project. If one is missing, ask before signing.
Scope, written as use cases
Forget the technology list. A good quote describes three to five scenarios, each with its trigger, its input data, the expected actions, the output format and the behaviour on failure. A workable phrasing: when a message arrives in the sales inbox and mentions a quote, the agent extracts the request, checks the customer record in the CRM, prepares a draft reply and places it in a human review queue. If the record is missing, it creates a task and writes nothing.
The block must also say what is out of scope. An agent that answers product questions does not handle complaints, does not negotiate discounts and does not speak on behalf of management. That list of refusals is often worth more than the list of features, because it prevents free additions along the way.
Add the expected volumes: requests per month, languages, channels, seasonal peaks. An agent tested on twenty examples and deployed on two thousand conversations a month does not behave the same way.
Architecture and technical dependencies
The quote must name the model used and its provider, the orchestration layer, the connected tools (CRM, email, document base), the hosting and the logging. Not to turn you into an architect, but so you know what the project depends on.
This section carries an immediate cost question: who pays for model calls, on which account, and what happens if the provider changes its prices or retires a model version? An agent built on a frozen version will need a migration one day. The quote must say whether that migration is included, billed, or left to you.
Data, the block providers forget
This is where an AI agent project differs from a classic software project. The agent reads messages, contracts, customer files. You therefore need a written statement of which data is processed, for what purpose, on what legal basis, how long it is kept, where it is hosted and who has access.
A provider processing that data on your behalf acts as a processor under Article 28 of the GDPR. That implies a processing agreement, a list of sub-processors (the model provider is often one), described security measures and a breach procedure. A quote that talks about AI without ever mentioning these elements leaves you carrying the consequences of an incident alone.
A second framework has recently been added. The European Commission (2026) states that Article 50 of the AI Act applies from 2 August 2026: providers must design systems so that individuals are explicitly informed whenever they interact with an AI, and deployers must inform people exposed to certain practices such as emotion recognition or deepfakes. In practice, if your agent answers customers, someone has to display that information, and it is not the model provider's job. The quote must say whether the notice, the wording and the audit trail are included or not.
Recurring costs, the line that never appears
An AI agent quote contains three cost families, and most documents show only one.
Build, billed once: scoping, design, development, integrations, testing. Go-live, often forgotten: access setup, user training, documentation. And operations, billed monthly: model calls, hosting, monitoring, fixes, improvements.
Operations is where budgets drift. Ask for a volume assumption, a unit price and a sample monthly calculation. A line such as "cost per conversation handled" or "cost per file analysed" beats an opaque lump sum, because it makes the quote verifiable from the first month. If the provider refuses to give that assumption, nobody has calculated it.
Acceptance testing, warranty and service level
A quote without acceptance criteria has no end. The project finishes when the provider decides it does.
Ask for measurable criteria on a test set you supply: share of compliant answers across one hundred real cases, escalation rate to a human, response time, integration failure rate. Specify the procedure for deviations: who fixes, within what delay, at what cost, and from when the project counts as delivered.
Add the warranty period after acceptance, support hours, the response time when the system stops, and the procedure if the model provider is unavailable. An agent that goes down on a Friday evening with no fallback plan costs more than its annual maintenance.
Reversibility, what actually belongs to you
The last page decides your freedom in two years. Who owns the instructions given to the model, the integration code, the test sets, the stored data? Can you retrieve all of it in a usable format? Does the provider commit to documenting the setup and supporting a handover for a defined period?
Without those answers, you are not buying an agent: you are renting a dependency. The clause looks harmless at signature time. It becomes decisive the day you want to change provider or bring maintenance in-house.
What a solid quote contains, and what should raise a flag
| Block | What to expect | Warning sign |
|---|---|---|
| Scope | Numbered use cases, volumes, explicit out-of-scope list | A sentence such as "intelligent agent to automate repetitive tasks" |
| Data | Purpose, legal basis, retention, hosting, sub-processors | No mention of GDPR or of the model provider |
| Recurring costs | Volume assumption and unit price | A single price with no split between build and operations |
| Acceptance | Measurable criteria on a test set supplied by the client | "Sign-off by the steering committee" |
| Warranty | Duration, hours, fix turnaround | A three-month warranty for a system running every day |
| Reversibility | Ownership of deliverables, export, documentation | Complete silence on exit |
| Liability | Clear split between provider, model vendor and client | A clause transferring all liability to the client |
Administrative particulars to check before signing
On the formal side, a consulting quote remains subject to the usual rules for commercial documents. Check the following systematically.
The provider's full identity: company name, address, SIREN or SIRET number, legal form, intra-community VAT number where applicable. A quote without those particulars will probably not bind the right entity in a dispute.
The validity period of the offer. It matters twice over on an AI project, because model pricing moves fast. An offer valid for three months is normal; an undated offer does not protect you.
The financial terms: price excluding tax, VAT rate, any deposit, payment schedule, late payment penalties, revision conditions if scope changes. Billing rules for out-of-scope requests must be written, with a day rate or a fixed price per batch.
The credibility checks: professional liability insurance certificate, confidentiality clause, commitment not to reassign the named team, and a list of sub-contractors.
Finally, the split of compliance roles. On an AI agent, three parties meet: your company, the provider who builds, and the model vendor who supplies the engine. A serious quote states who does what on GDPR, on the user information required by Article 50 and on output monitoring. Without that split, everyone will assume it is someone else's job.
Fixed price, time and materials, or subscription: read the price in the right place
Three pricing models coexist, and each shifts risk differently.
A fixed price fits when scope is stable and use cases are already identified. It protects the budget but pushes the provider to minimise changes. Check what is included in iterations and what an addition costs.
Time and materials, billed by the hour or day, fits exploratory projects where the need sharpens as you go. It protects quality but exposes the budget. Demand a cap, a weekly checkpoint and a right to stop without penalty.
Subscriptions, common for agents rented as SaaS, move the cost to operations. They reduce the initial investment but make exit harder. Look at the notice period, the data retrieval terms and the price increase policy.
A quote can combine all three: fixed price for the build, subscription for operations, time and materials for improvements. That is fine, provided every line is identifiable.
FAQ
Is a quote mandatory for an AI agent project?
No, not in general. The prior quote obligation covers listed activities, mainly building work and repairs, plus certain regulated services. For consulting or development work, a quote is optional. But once signed, it becomes a contract binding both parties.
How long should a quote remain valid?
No single duration is imposed. The DGCCRF states that the quote must mention the period during which the offer is valid. On an AI project, thirty to ninety days is common practice, the time needed to confirm the budget and the access rights. An offer with no validity date should be sent back to the provider.
Should maintenance and usage costs be quoted separately?
Yes, and it is the best transparency test. Build is a one-off cost, model calls and hosting are variable costs, maintenance is a fixed cost. Mixing them into a single lump sum prevents any comparison between two proposals and makes overruns invisible.
Who is liable if the agent gives a wrong answer?
The answer depends on what the contract says, which is precisely why the quote must address it. In practice, liability is split between the provider who built the system, the model vendor who supplies the engine and your company, which sets the rules and approves go-live. A quote that pushes all liability onto the client is not a quote, it is a waiver of warranty.
Can you ask for an audit before signing a quote?
It is often the best way to proceed. A short, paid audit verifies feasibility, identifies the data available and prices the scope against real cases. The build quote that follows is more precise, and you avoid paying for a full development on a badly framed need.
Conclusion
An AI agent quote is judged less on its price than on what it makes verifiable. Numbered use cases, stated volumes, described data processing, itemised recurring costs, measurable acceptance criteria, a reversibility clause: those six blocks are enough to tell a built proposal from a sales promise.
Before signing, take an hour to hold the document against that grid. If three blocks are missing, that is not a reason to reject the provider, it is a reason to write back. getting the need framed by a short audit before requesting a quote remains the cheapest way to avoid a project that drifts.
To place the different technical approaches before comparing proposals, comparing two solutions on the same grid helps you ask the right questions. The questions to ask before signing usefully complete that grid on the commercial side, and browsing concrete use cases gives you a sense of realistic scope.
If you want to see what a well-framed agent looks like before launching a tender, see how NexeAI designs an agent describes the approach, from scoping to go-live. And if you already have a quote in hand and doubts about its content, let's talk about your project: a second reading is often worth more than a fast signature.
