You have run a tender for an AI agent. Two vendors replied. The first proposes 18,000 euros in eight pages, the second 47,000 euros in thirty-two. Both describe the same request, both look credible, and you have to decide this week. The reflex is to compare the totals. That is the best way to get it wrong: on an AI project, the price gap rarely comes from the skill of the teams. It comes from scope, data, run costs, and from what each proposal does not say.
Key Takeaways
- Comparing two proposals for an AI project starts by bringing each offer back to the same unit: person-days, deliverables, responsibilities. The total comes last.
- The signed price covers only part of the cost. Licences, model consumption, maintenance and internal time are paid afterwards.
- A weighted scoring grid based on the full three-year cost decides better than a price table.
- According to Gartner (2025), over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls. Those three causes are already visible in the quotes.
- An offer that specifies neither acceptance criteria, nor code ownership, nor exit conditions is not cheaper. It is incomplete.
Two prices, two scopes
An AI project quote is not a catalogue: it is a working hypothesis written by someone who understood your need in their own way. Two serious vendors can answer the same tender with scopes that have almost nothing in common. One has priced an agent that answers questions over a document base. The other has priced the same agent, plus document clean-up, plus the interface inside your business tool, plus three months of adjustment.
Before comparing anything, take both documents and fill in the same grid for each.
| Axis | Question to ask both vendors | Effect on price |
|---|---|---|
| Scope | Which use cases are covered, and which are explicitly excluded? | An offer can look cheaper because it handles one use case where the other handles three. |
| Effort | How many days per role, and over what period? | It is the only unit comparable between a fixed price and time and materials. |
| Data | Who prepares, cleans and hosts the documents? Is that time billed? | Data preparation is often the first invisible line item. |
| Integrations | How many connectors, and who pays for subscriptions and model calls? | A forgotten connector is billed as an amendment, never at the initial price. |
| Acceptance | Which acceptance criteria are written down, and what happens if they are not met? | Without criteria, the project ends in a negotiation, not a delivery. |
| Run | What is the monthly cost after go-live? | It almost never appears in the initial quote. |
| Exit | What happens if you change vendor in two years? | It sets your future negotiating room. |
If a box is still empty after reading, it is not a wording detail: it is a risk you are buying without knowing it. A well-built tender pack reduces those grey areas, and a solid statement of work is worth more than an extra round of questions with each candidate.
The four lines nobody prices
Four items weigh heavily on the final bill and rarely appear in the summary table. Ignoring them distorts the comparison, always in the same direction: the most flattering offer on paper becomes the most expensive one in production.
Data. An agent only works well on clean, current and accessible documents. Sorting, reformatting, deduplication and defining access rights take time. If the vendor does not price them, you will pay for them, internally or through an amendment.
Integrations. Every connected tool has its API, its rate limits and its quotas. The number of model calls varies with real usage: the more your teams use the agent, the more consumption grows. It is a variable cost, to be modelled before signing, not after.
Internal time. An AI project ties up a business owner, a technical referent and pilot users for testing. That time does not leave your bank account, it leaves your production capacity. Two offers with the same scope can require very different levels of involvement.
Maintenance. A model changes, an API evolves, a document format shifts. Without a maintenance commitment, the agent degrades slowly and nobody notices before users stop using it.
Full three-year cost, the only honest basis
Comparing two commercial proposals for an AI project means comparing two trajectories, not two invoices. The formula fits on one line: build, plus run, plus licences, plus maintenance, plus internal time. Three years is a reasonable horizon for an SME, because it covers at least one model renewal cycle and one process review.
Take an arithmetic example, with fictitious amounts for the demonstration. Offer A costs 18,000 euros to build, then 900 euros per month in run costs and 4,000 euros of annual maintenance. Offer B costs 47,000 euros, then 150 euros per month, maintenance included. Over three years, A comes to 18,000 + 32,400 + 12,000, that is 62,400 euros. B comes to 47,000 + 5,400, that is 52,400 euros. The most expensive offer at the start is the cheapest at the end, provided the scope is identical. It rarely is, which is exactly why you should put both calculations side by side before discussing.
Two rules keep that calculation honest. First, ask for the run cost in writing, with the volume assumption behind it. Second, get confirmation of what happens if usage exceeds it: pay-per-use billing, a tier, or a cap.
Read the proposal as a contract
A commercial proposal binds both parties as soon as it is signed. Some clauses deserve careful reading, because they decide your freedom in two or three years.
- Ownership. Who owns the code, the prompts, the test sets and the documentation? A vague answer means you are renting your tool, with no way to have it taken over elsewhere.
- Confidentiality. Where do your documents go? Are they sent to a model provider outside the European Union? Are they used to train a model? Those answers must be written down, not given verbally in a meeting.
- Reversibility. Is data and configuration export planned, in which format, at what price? An agent's exit plan is prepared at signature, never at the moment of departure.
- Acceptance. Who signs off, against which criteria, and with what recourse if they are not met? A quote with no acceptance criteria leaves you as sole judge of a technical disagreement you do not master.
- Subcontractors. Can the vendor delegate part of the work without your agreement? On sensitive data, the question is not theoretical.
These are the same points as in a well-drafted AI agent quote. The difference, when you compare two offers, is that gaps show up by contrast: what one writes and the other avoids becomes a signal.
The scoring grid that replaces intuition
When two offers both hold up, the decision rests on criteria you cannot rank mentally. A weighted score, built before you read the prices, produces a decision you can defend in front of a partner or a banker.
| Criterion | Weight | What you score |
|---|---|---|
| Scope and deliverables | 25 | Number of use cases covered, precision of deliverables |
| Full three-year cost | 25 | Build, run, licences, maintenance |
| Data and security | 20 | Hosting, access rights, use of documents |
| Reversibility and ownership | 15 | Export, format, exit cost |
| Business understanding | 15 | Vocabulary, questions asked, restatement of the need |
Score each criterion from 1 to 5, multiply by the weight, add up. Two precautions. First, never score price on the figure at the top of the quote but on the full cost. Second, have two different people score the two offers separately: the gap between your two grids will show you where the real disagreement sits, often on understanding the need rather than on price.
Three signals that should rule out an offer
Some answers eliminate a candidate before any calculation.
A return on investment with no assumptions. A percentage of gains announced without volume, without upfront cost and without a measurement method is not a projection: it is a sales argument. Ask for the calculation. A serious vendor will show their assumptions, including the ones that might not materialise.
A deliberately vague scope. Formulas such as "tailored support" or "as needed" move the definition of the work to after signature, when you no longer have any negotiating room.
Agentic AI vocabulary without agentic capability. Gartner estimates that of the thousands of agentic AI vendors, only about 130 are real (2025), the rest being rebranded chatbots or existing automation. The question to ask is simple: what does the agent do alone, what does it do with human approval, and what happens when it is wrong?
The same Gartner document states that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024, and that in a January 2025 poll of 3,412 attendees, 19% of organisations reported significant investments in agentic AI and 42% conservative investments. In other words: the market moves fast, and declared caution does not prevent bad comparisons.
When the two offers are not at the same stage
Last trap, and the most common: comparing a pilot with a production rollout. One prices an agent running on fifty documents and three users, the other a platform with authentication, logging, monitoring and support. The second will look overpriced when it is simply answering a different question.
In that case, do not compare totals: compare trajectories. Ask each vendor for the phase breakdown, the cost of each phase and the criterion for moving to the next one. An offer that refuses to sequence the project is asking you to sign the whole thing with no checkpoint. Conversely, an offer that proposes a short first milestone with a clear stop criterion limits your exposure even if its final price is higher. The real cost of an AI pilot is worth reading before you approve that breakdown.
FAQ
Should you choose the cheapest proposal?
No, and not the most expensive on principle either. Choose the one whose full three-year cost is lowest at equal scope and quality. If the scope is not the same, comparing prices means nothing: first realign both offers on a common list of deliverables.
How do you compare a fixed price with time and materials?
Convert the fixed price into person-days from its effort breakdown, then compare both on the same volume. A fixed price with no effort breakdown is a price with no content: ask for the number of days per role before signing. Time and materials gives flexibility, a fixed price gives predictability, and the right choice depends on how clear your need is.
A vendor refuses to break down their person-days, what then?
Treat the refusal as information. You can ask for a breakdown by phase and deliverable without demanding the internal detail of each day. If the refusal persists, you will be unable to compare the offer, track its progress, or negotiate an amendment on a factual basis. That is reason enough to rule the candidate out.
How many proposals should you request?
Three is a good balance for an SME: enough to see different approaches, few enough to stay able to analyse them in depth. Beyond four, reading time exceeds the information gained, and the offers start to look alike because candidates end up copying the same framing. Prepare your scoring grid before the first one arrives.
Conclusion
Comparing two commercial proposals for an AI project is not about choosing a price, but choosing a scope, a level of dependency and a cost trajectory. Align the deliverables, price the run and maintenance, read the ownership and exit clauses, then score. The two offers that survive that exercise are usually closer than their totals suggest, and the one that collapses will have spared you a difficult negotiation a year later.
If you want an outside view on two proposals before signing, NexeAI can audit both offers with you, or frame the need upstream so the next quotes are genuinely comparable. For an agent or automation project, our AI agent services cover framing, delivery and follow-up. The simplest step is to talk: contact us.
