How to brief an AI agency so you get a fair quote

Knowing how to brief an AI agency is the difference between three quotes you can compare and three quotes that describe three different projects. Most owners send a paragraph and a hope. The agencies fill the gaps with their own guesses, and you get prices that are miles apart.

Fill it in once, send it unchanged to three agencies, and you will see quickly who is serious.

How to brief an AI agency: what the one page must contain

A brief has eight parts. Each one removes a guess the agency would otherwise make for you.

Section	What you write	Why it matters
Process	The job, its steps, who approves what	Stops the agency scoping a different job
Volume	Items, frequency, peaks	Drives run cost and tool choice
Systems	Tools, data, access	The biggest hidden cost is connecting things
Result	What done looks like, good and bad examples	Gives you something to test against
Budget	A range, split into build, run and review	Lets agencies size the answer honestly
Guardrails	What the agent must never do alone	Sets the approval points
Failure and logging	What happens when it gets it wrong	Shows how they handle problems
Timeline	When you need it live, who decides	Makes the quote a plan

The findainow guide to briefing an AI agency (https://www.findainow.com/blog/how-to-brief-an-ai-agency-quote-means-something) lists much the same core: the job, the tools you already run, the budget split, the data involved and what done looks like. Add volume, guardrails and failure handling and you have something an agency can price properly.

Keep it to one page. If you cannot say it briefly, you do not yet know what you want, and that is worth finding out before you pay anyone.

The process to fix: the job, its steps and where people approve

Pick one process. Not "use AI in our business". One job, such as "answer new enquiries and book a call" or "chase overdue invoices".

Write the steps as they happen today, in plain order. Who starts it, what they look at, what they decide, what they send, where it ends. Use the real messy version, not the tidy one in your head. If one person does it differently on Fridays, say so.

Mark every point where a person checks or approves something. Those marks become your approval gates later. Also add a short background paragraph: who your customers are, what tone you use, what the usual exceptions are. Sean Glennon's BRIEF framework on Medium (https://medium.com/glennonai/how-do-you-build-an-ai-agent-be-brief-6ac985d779bd) puts Background first and says most people skip it, then wonder why the agent produces mediocre output. He is right. An agent only knows what you tell it.

Write this in the brief:

• The one process, named in a sentence.
• Numbered steps as done today, with who does each.
• The steps that are slow, error-prone or boring.
• Where a person currently approves, and who.
• Three lines of background on your business and customers.

Volume and frequency: how many items, how often, and peak loads

Volume is the number that moves the price most, and it is the one owners leave out.

Give three numbers: a typical week, a busy week and the worst day you can remember. Say how long each item takes a person now. If you do not know, ask the person doing it to time five of them.

Also say what happens at peaks. Do items wait, or must they be answered within minutes? A quote for "most of the time" is useless if your month-end doubles the load.

Write this in the brief:

• Items per week, typical and busy.
• Peak day or month, and what causes it.
• Minutes a person spends per item today.
• Speed needed: instant, same day, or next day.
• Expected growth over the next 12 months, as a rough guess.

Systems and data: the tools, the connected software and the access each agent needs

List every tool the process touches: your CRM, inbox, accounting package, spreadsheet, booking tool, phone system, WhatsApp. Name the product and the plan you are on. "We use a CRM" tells an agency nothing.

Then say what the agent needs to read and what it may write. Read access to your calendar is a small ask. Write access to your accounts is a large one. Separate the two. Say where the data sits, whether it includes personal data, and which country it is held in. If it does, read our guide to AI data protection for UK small businesses before you send the brief, because the agency will need to explain where your data goes.

Also say who owns the logins. You should hold the admin accounts, not the agency.

Write this in the brief:

• Every tool in the process, with product name and plan.
• What the agent reads and what it writes, in two lists.
• Whether personal or financial data is involved.
• Who holds admin access to each tool.
• Anything that is messy: old data, duplicates, missing fields.

How the agent sits between your tools and your people

Inputs

InboxCRMBooking tool

Agent

Reads the itemDecides within your rules

Approval

A person signs off sensitive messages, payments and refunds

Output

Reply sentRecord updated

The result you want: what done looks like and examples of good and bad output

Define done in numbers or observable facts. "Every new enquiry gets a reply in under five minutes and a booking link" is testable. "Better customer experience" is not.

Then attach examples. Paste in three replies, invoices or reports you consider good, and two you consider bad, with a line on why. This does more than any description of tone. The agency can build against your examples, and you can judge the output against the same ones.

MATTER's guide to briefing an AI agency (https://www.matterstudio.ai/blog/how-to-brief-an-ai-agency) makes a point from the creative side that holds here too: if you do not specify tone and direction, those decisions do not disappear. They fall to whoever is doing the work, or get made by omission. In operations the same thing happens with wording, escalation and exceptions.

Write this in the brief:

• Done in one sentence, with a number if possible.
• Three good examples and two bad ones, with reasons.
• The tone in five words, and words to avoid.
• How you will measure success after 30 days.

Budget range and how to split it between build, run and review

Give a range. Owners fear this will anchor the price upward. In practice, a range lets an honest agency tell you what fits and a weak one reveal itself by quoting for everything you could possibly want.

Split the range three ways. Build is the one-off work to set it up. Run is the monthly cost of the tools, usage and upkeep. Review is the time someone spends checking output and adjusting it after launch. Most quotes cover build and gloss over the other two. Our piece on AI automation agency pricing explains retainer, project and per-workflow models, so you know which you are being offered.

Here is the worked example. The numbers are made up to show the method, not a benchmark for any real project.

Budget example: splitting a made-up first-year budget

Inputs, all invented:

• First-year total you are willing to spend: £12,000.
• Share for build: 50%.
• Share for run: 30%, spread over 12 months.
• Share for review: 20%, spread over 12 months.

The sum, line by line:

1. Build: £12,000 x 50% = £6,000.
2. Run: £12,000 x 30% = £3,600, which is £300 a month.
3. Review: £12,000 x 20% = £2,400, which is £200 a month.
4. Check: £6,000 + £3,600 + £2,400 = £12,000.

Your own split depends on volume, the number of systems and how much checking the work needs. The point is to put three lines in the brief so every agency answers all three. If one quote has no run or review line, ask why.

Guardrails and human approval points: what the agent must never do alone

This is the section most briefs miss, and the one that protects you. Write a plain list of what the agent must never do without a person approving it.

A sensible starting list, taken from how small firms use agents in practice: the Bitcoin Foundation's guide to AI agents for small businesses (https://bitcoinfoundation.org/news/ai/ai-agents-for-small-businesses) says a person still approves sensitive messages, payments, refunds and policy decisions. The same piece describes approval gates stopping an incorrect field before it reached a customer. That is the job of a gate: catching the error while it is still cheap.

MATTER also advises sign-off gates and treating "amends" with suspicion. Translate that into your brief: say who signs off the build at each stage, and how changes after sign-off are scoped and priced. Otherwise small changes become open-ended billing.

Then ask three questions the agency must answer in writing:

• What will you not do, or not recommend doing, for this process?
• What happens when the agent gets something wrong, and who is told?
• How are the agent's decisions logged, and can we read the log?

A good agency answers these without fuss. A weak one goes vague. Our list of AI automation mistakes to avoid shows what goes wrong when these are left out.

Write this in the brief:

• Actions needing human approval: money, refunds, legal or policy statements, first contact with a key client.
• Who approves, and how fast they must respond.
• Who signs off each stage of the build.
• How post-sign-off changes are scoped and priced.
• The three questions above, left open for the agency to answer.

How to send the brief to three agencies and compare the quotes like for like

Send the identical document to three agencies on the same day, with the same deadline and one named contact. Do not give one agency a call and the others an email. Different information gives you different quotes.

Allow questions, but share every answer with all three. The questions they ask are a test in themselves. An agency that asks about your volume, your systems and your exceptions has read the brief. One that replies with a price and a calendar link has not.

Then put the replies side by side:

Line	Agency A	Agency B	Agency C
Process they think they are fixing
Volume they assumed
Systems and access needed
Build price
Monthly run cost
Review and change cost
Approval points proposed
Logging and failure handling
What they will not do
Timeline and sign-off stages

A blank cell is a finding. If a quote assumes different volume from yours, it is pricing the wrong job. Before you shortlist, check each agency's track record using our guide on how to check AI agency reviews.

The quote variance you were worried about mostly vanishes once every agency prices the same page. What is left is real difference: approach, honesty and fit.

Where to start

1. Choose one process and write its steps today, in order.
2. Time five real items and count a typical and a busy week.
3. List every tool with product name and plan.
4. Collect three good and two bad examples of the output.
5. Set a build, run and review range.
6. Write your never-alone list and the three questions on failure, logging and limits.
7. Send the same page to three agencies, then fill in the comparison table.

If you would rather not write it alone, we can help. Our AI Systems & Automation service starts from $499 and takes one to three weeks for a defined process, and we are happy to tell you honestly if your job is smaller or bigger than that. You can get a free AI plan at our free AI plan page, which gives you a step-by-step plan you can keep and send to anyone, or book a call and we will go through your brief with you.

Sources

• Medium (Sean Glennon, Glennon AI)
• MATTER
• Digital Marketing Institute
• findainow
• Bitcoin Foundation