AI quotes are hard to compare because each supplier answers a different question. A one-page brief fixes that by stating: the process, with a two-week measured baseline in units and hours; the systems the work touches and how they can be reached; the inputs, with fifty real samples including the messy ones; the acceptance threshold the process needs, as a number; the constraints on data, people and law (what is personal or special-category, where it may be processed, what a person must decide); the running-cost question, per unit at current and ten-times volume; and what ownership and handover must look like. Send the same brief to every supplier. A good one replies with a fixed price, the threshold they will commit to, the per-unit running cost, the controls, a timeline in weeks and what is excluded. A vague brief gets a day rate and a discovery phase.
- Length
- One page, seven headings
- Attach
- Fifty real input samples
- Expect back
- Fixed price, threshold, per-unit cost, exclusions
Why quotes do not compare
One supplier quotes for an assistant with a template, another for a workflow, a third for an agent, because the brief said “automate our enquiries” and each heard a different thing. One assumes the CRM has an interface, another budgets to build one. One measures accuracy on its own demo examples, another on nothing. One states running cost, the rest do not. The buyer ends up comparing a day rate to a fixed price to a subscription, and picks on confidence rather than fact. The brief below removes the ambiguity that produces that.
RelatedFixed price or day rate How to choose an AI consultancy in the UK
The one-page brief
Copy the headings; fill in the brackets.
- 01
Process. [What happens today, step by step, who does it.] Baseline over two weeks: [units], [hours], [time from arrival to done], [error or rework rate]. Loaded hourly rate: [£].
- 02
Systems. [Each system touched: name, whether it has an API or import route, who administers it.] The system of record for the outcome is [name].
- 03
Inputs. [What arrives, in what form, from where.] Fifty real samples attached, anonymised where needed, including the messy ones.
- 04
Acceptance threshold. [The accuracy or outcome the process needs, as a number, measured how.] Exceptions go to [role].
- 05
Constraints. Personal data: [yes/no; special-category yes/no]. Processing region: [UK/EU/any with a transfer mechanism]. Decisions a person must make: [list]. Regulator or client rules: [list].
- 06
Running cost. Quote per [unit] at [current volume] and at ten times, with ceilings. State which model and provider you propose and why.
- 07
Ownership and handover. Code, prompts, data and accounts in our name from day one; runbook; credentials in our vault; an exit that another supplier could pick up. Fixed price, timeline in weeks, and a written list of what is excluded.
RelatedHow to measure AI return in 90 days Your data when a supplier builds your software
What a good supplier sends back
A fixed price with a scope and an exclusions list. The threshold they commit to, measured on your fifty samples before go-live, and what happens if it is not met. The running cost per unit at both volumes, with the ceilings they will set. The model and provider, and the processor position. The controls: least privilege, human confirmation where you specified, logging. A timeline in weeks with a demo cadence. Ownership as you asked, in writing. And an honest paragraph on what the brief gets wrong, because the brief always gets something wrong and a supplier who has read it will say so.
RelatedQuestions to ask a software partner AI vendor due diligence checklist
What a poor reply looks like
A day rate and a discovery phase before any number. A demo on the supplier’s own examples instead of your fifty. “Accuracy will be high.” No running cost, or a subscription that hides it. Full admin access requested up front. Ownership deferred to the contract. And no exclusions, which means everything is excluded once the work starts.
How we answer it
Every proposal we write follows the seven headings, because we ask clients for this brief. The price is fixed, the threshold is a number measured on the client’s samples, the running cost is stated per unit at two volumes, and ownership is in the client’s name from the first day. If the brief describes a process that does not need AI, the reply says so and costs nothing.
RelatedHow much AI automation costs in the UK AI implementation
Common questions
What should an AI project brief include?
The process with a measured baseline, the systems and how they can be reached, real input samples, the acceptance threshold as a number, data and human-decision constraints, the running-cost question at two volumes, and ownership and handover requirements.
Why do I get such different AI quotes?
Because a vague brief lets each supplier answer a different question: assistant, workflow or agent; with or without integration; with or without a measured threshold. The same one-page brief to every supplier produces comparable replies.
Should I send suppliers our real data?
Fifty real samples, anonymised where needed, under a non-disclosure agreement. A supplier who will not test on your samples is not committing to anything.
What should a good AI proposal contain?
A fixed price with exclusions, a committed threshold measured on your samples, running cost per unit at current and ten-times volume, the model and processor position, the controls, a timeline in weeks, and ownership in your name.
Is a discovery phase a bad sign?
A paid discovery that ends in a fixed-price proposal is fine and often right. An open-ended discovery on a day rate with no number in sight is the sign.
How many suppliers should get the brief?
Three is enough. With the same brief, the replies compare on price, threshold, running cost and ownership rather than on confidence.