Describe the job, not the technology: what starts it, each step a person does now, what the finished output looks like, and how often it happens. Add 20 to 30 real examples including awkward ones, a test that defines "good enough", the systems and accounts involved, a monthly running-cost ceiling, and what must always go to a human.
A good brief fits on two pages and takes about half a day to write, most of it spent collecting examples. That half-day pays for itself in the first week of development, because the developer quotes the real job instead of a guess and you both know what "finished" means. Below are the eight parts in order, a template to copy, and a filled-in example for a dog-grooming salon.
The eight parts of a workflow brief, in the order to write them
Write them in this order even if you present them differently. Each part depends on the one before, and the examples in part four will usually make you rewrite part two.
1. The job in one paragraph (15 minutes)
Say what the work is, who does it now and why you want it changed, in plain words. "We get booking requests through three channels. I answer each one by hand, often at night, and about a quarter need a follow-up because details are missing. I want replies drafted for me with the missing questions already asked." No mention of chatbots, agents or models. Leave the how to the developer.
The difference is easiest to see side by side. Here is a first draft from an illustrative kitchen showroom, and the rewrite:
FIRST DRAFT
We need an AI agent integrated with our CRM to automate
lead qualification using GPT.
REWRITE
Leads arrive from the website form and from phone calls
our showroom staff log in the CRM, about 45 a week. Our
sales lead reads each one and decides whether to book a
home survey, send a price range first, or politely
decline (budget under $8,000, or outside the areas we
fit). That takes her about six hours a week, and leads
from Friday afternoon wait until Monday. We want each
lead sorted into one of those three groups, with the
reason, ready for her at 9am.
The first draft names a tool and a buzzword. The rewrite gives the developer the volume, the three outcomes, the rule for declining and the delay that actually costs the showroom sales.
2. The trigger and the steps as they happen today (45 minutes)
Next time you do the task, write down every step as you do it: what arrives, what you open, what you look up, what you type, where it ends up. Include the checks you do without thinking, because those are the steps an automation will skip. If you've never written a process down, documenting your processes before adding AI gives you a format.
For an illustrative removals firm handling quote requests, the notes might look like this, with the unconscious checks marked once the owner reread them:
Starts when: quote request arrives (web form or email)
1. Open request, note move date, from/to, property size
2. [HIDDEN CHECK] Is the date a Friday or the last few days
of the month? If so, add the peak-day note to the quote
3. Look up the rate card by number of bedrooms
4. [HIDDEN CHECK] Any mention of a piano, a loft, or stairs
with no lift? Add the extra-crew line
5. [HIDDEN CHECK] Access looks tight (narrow street, no
parking)? Ask for photos before quoting
6. Draft the quote in the template, save as PDF
7. Email it, log it in the jobs spreadsheet
Ends when: quote sent and logged
Three of the seven steps are checks nobody would have mentioned in a meeting, and each one changes the price. Left out of the brief, the automation would have quoted every Friday move at the standard rate.
3. Volumes and timing (five minutes a day for a week)
Count how many items arrive each day, when, and how long each takes. Note the peaks. Volume decides the architecture and the running cost: 20 items a week and 2,000 a week are different builds, and the developer can't guess which you have.
A week's tally from an illustrative print shop that checks customers' uploaded artwork before printing shows why the daily pattern matters as much as the total:
ARTWORK CHECKS, one week Items Mins each Notes
Mon 31 8 weekend uploads, all before 11am
Tue 12 8
Wed 9 7
Thu 14 8 two re-uploads after rejection
Fri 18 9 rush jobs for Saturday events
Sat/Sun (uploads, unchecked) (26) - counted on Monday
Total 84 about 11 hours
A weekly figure of 84 suggests a steady trickle. The tally shows a Monday-morning pile of 31, most of which arrived over the weekend, and a Friday rush where speed matters more than anywhere else. That tells the developer the workflow can batch the weekend uploads overnight, but needs to handle Friday items within minutes, which is a different design from "check each file as it arrives".
4. Real examples, including the awkward ones (one to two hours)
Collect 20 to 30 real past cases with the output you actually produced for each. Make sure at least five are awkward: the rambling message, the one in poor English, the one that asks three things at once, the one you had to refuse. Remove or replace names, phone numbers and email addresses before sending, but keep the mess: the spelling, the missing details and the second question tucked in at the end are the point. One awkward case from an illustrative party-hire company, ready to send, reads like this:
EXAMPLE 17 (AWKWARD: two requests in one; venue not on our list)
Message: "hi its [name] [phone] need 60 chairs and 6 trestles for
the 14th at [farm venue], can u do linen?? also my sister wants a
price for her thing in june"
What we did: quoted chairs and trestles; said we don't stock
linen and suggested asking the venue; checked the venue had
vehicle access before confirming delivery; asked for the
sister's contact details and date as a separate enquiry.
The "what we did" line matters as much as the message. It shows the developer three rules nobody had written down: decline linen politely, check access for unfamiliar venues, and split a second enquiry off rather than answering it in the same reply. If you're unsure what counts as sensitive, classifying your business data before using AI sets out a simple scheme.
5. What "good" looks like (30 minutes)
Write acceptance tests: statements the finished workflow must pass before you sign it off. Make them countable. "Works well" is not a test; "in 30 test messages, the dog's size is extracted correctly in at least 27" is. Include the hard rules too: things that must never happen however good the rest is.
6. Systems, accounts and access (30 minutes)
List every piece of software the workflow touches, which plan you're on and who holds the admin login. The developer will check whether each has a connector or an API (a way for software to talk to other software). State that the build must live in accounts your business owns, with the developer invited as a user.
Fill in the "admin login held by" line by actually signing in, not from memory. An illustrative veterinary practice wrote "practice manager" against its website form, then discovered in the developer's first week that the form lived in an account the freelance web designer had set up years earlier under his own email. The build stalled for ten days while he was traced and ownership moved over. Ten minutes checking each login before you send the brief tells you which ones need sorting out first.
7. Constraints: data, money, speed (20 minutes)
Which data may go to an outside AI service and which may not. The most you're willing to pay each month to run it. How quickly a draft must appear. Whether you'll use it mostly on a phone. These change the design more than people expect.
A realistic way this goes wrong: an illustrative roofing contractor briefed a workflow that drafts replies to repair enquiries, and never mentioned devices. The developer built a tidy approval screen in a web dashboard designed for a desktop. The owner spends his days on roofs and checks messages on his phone at lunch, where the dashboard needed pinching and scrolling to read. Within a fortnight drafts were piling up unapproved and the old habit of replying by text had returned. One line in the brief ("approve from a phone, one tap, in the email app I already use") would have produced a different build at the same price.
8. Out of scope and handover (20 minutes)
Say what you're not asking for, so the quote doesn't quietly include it. Then list what you need at the end: written instructions, how to pause it, a list of monthly costs, and a walk-through for whoever will look after it.
A brief template you can copy
If writing isn't your strength, talk the process through into your phone's voice recorder while you do the task, then ask a business-plan AI assistant to sort the transcript into the template below. Tell it to mark every gap rather than filling it: "Put my words into this template. Where the transcript doesn't give an answer, write [ASK ME]. Don't invent steps, numbers or tests." A returned draft (illustrative) for a tile shop's sample-request process came back tidy, with "Step 4: check stock levels in the inventory system" in the middle. The shop has no inventory system; staff look at the shelf. The assistant had filled a gap with what shops usually do. Read every step against the transcript, and write part 5, the acceptance tests, yourself: they're your definition of "good", and an assistant can only guess at it.
AI WORKFLOW BRIEF Version 1, [date]
Business: [name, what you do, team size]
Contact for questions: [name]
1. THE JOB
What happens now, who does it, why it needs to change:
2. TRIGGER AND CURRENT STEPS
Starts when:
Step 1:
Step 2:
...
Ends when (the finished output is):
3. VOLUMES
Items per week (normal / peak):
Minutes per item now:
When they arrive (hours, days):
4. EXAMPLES
Attached: [number] real cases with our actual output,
including [number] awkward ones (marked). Personal details removed.
5. ACCEPTANCE TESTS
Must pass before sign-off:
- In [n] test cases, [field/result] is correct in at least [n].
- Never: [the things that must not happen].
- Always goes to a person: [the cases the AI must hand over].
6. SYSTEMS AND ACCESS
Software touched (and plan):
Admin login held by:
Built in our accounts: yes. Developer access removed at handover.
7. CONSTRAINTS
Data that must not leave our systems:
Monthly running-cost ceiling: $[amount]
Speed needed:
Devices it must work on:
8. OUT OF SCOPE
Not included:
9. HANDOVER
We need: written instructions, how to pause it, monthly cost list,
walk-through for [name], [n] days of fixes after go-live.
Filled in: booking requests at a dog-grooming salon
Here is how the template reads for an illustrative salon with an owner and two groomers.
The job. Booking requests arrive through the website form, Instagram messages and email: about 60 a week, rising to 90 before the summer and the December holidays. The owner answers each by hand, roughly five hours a week. About one in four is missing something essential (breed and size, coat condition, vaccination status, preferred days), which means a second message and a delay of a day or more.
Current steps. Read the request. Check the dog's record in the booking system if it's an existing client. Estimate the service and price band from the price list by size and coat. Check the diary for suitable slots. Reply with two or three options, a price range and any questions. When the client confirms, enter the booking. If the dog is new, ask for vaccination proof before the first appointment.
What the owner wants. For each request, a drafted reply waiting for approval, with the missing questions already asked, two or three real free slots, and a price range taken from the price list. The owner approves or edits on her phone, and the reply goes out from the salon's usual account. The booking itself still gets entered by a person.
Acceptance tests.
- In 30 test requests, breed, size and service are extracted correctly in at least 27.
- No draft ever quotes a price outside the price list's range for that size and service.
- No draft offers a slot that isn't free in the diary at the moment of drafting.
- Any request mentioning aggression, a medical condition, a recent operation, severe matting or a puppy's first groom goes to the owner with no draft, flagged.
- Drafts are ready within ten minutes of the request arriving during opening hours.
Systems. The salon's booking software, a shared email inbox, the website form and an Instagram business account. Whether Instagram messages can be read automatically depends on the account type and Meta's current rules for business messaging, so the developer is asked to confirm that before quoting. If they can't, the first version covers the form and email only.
Constraints. Clients' phone numbers and addresses stay in the booking system; only the message text and the dog's details go to the AI service, on a business plan or API that doesn't train on the data. Running-cost ceiling: $40 a month in total.
Is that ceiling realistic? The developer can check, but you can too. At 60 requests a week, call it 260 a month. If each draft reads about 3,000 tokens (the message, the price list, the instructions) and writes about 600, that's 0.78 million input and 0.16 million output tokens a month. On Claude Sonnet 5 at $2 input and $10 output per million tokens, that's about $3 a month. The AI is cheap; any connector platform and the developer's maintenance are the real running costs, so the ceiling leaves room for both.
Out of scope. Taking payments, sending appointment reminders (the booking system already does), and answering general questions about grooming. If collecting pet details before booking is your main problem rather than drafting replies, collecting pet details before booking with AI covers a simpler route.
What a good developer sends back
A brief starts a conversation; it doesn't end one. Within a few days, a developer worth hiring should reply with:
- Questions. Usually five to fifteen. "What happens when a client asks for a groomer by name?" is a good sign. No questions at all is not.
- An estimate with its assumptions written down, so you can see what would change the price.
- A monthly running-cost estimate broken down by service, checked against your ceiling.
- A test plan that uses your examples and your acceptance tests.
- What they'd drop to hit your budget, if the full brief doesn't fit it.
- Any concerns about data, reliability or channels that may not be automatable.
For the salon's brief, a useful set of questions back might read like this (illustrative):
1. Does your booking software offer an API, or only a
calendar export? That decides how "free slots" are read.
2. If a draft offers three slots, should they be held
until the client replies, and for how long?
3. How should a request for two dogs in one visit be
handled: one draft, or two?
4. Who approves drafts on your day off?
5. Are prices the same for new and existing clients?
6. If the client replies to the draft with a question,
does that come back to the workflow or to you?
Every one of those is a gap in the brief, and every answer changes the quote a little. A developer who asks them is pricing your salon; one who doesn't is pricing a generic booking bot.
When quotes arrive, evaluating an AI implementation proposal gives you a checklist and a scoring sheet for comparing them.
Where workflow briefs go wrong, and how it shows up later
Briefing a solution instead of a job. "Build us a chatbot" gets you a chatbot, whether or not your problem was one. The developer builds what you describe, so describe the work.
Sending only the tidy examples. The workflow passes every test and then fails in its first week on the messages that actually arrive. Awkward cases belong in the brief precisely because they're awkward.
"It should just know." The AI knows only what it's given. If the answer depends on your price list, cancellation policy or a rule in your head ("we never book two large doodles back to back"), write it down and attach it.
No owner for the exceptions. Every workflow has cases it hands back. If the brief doesn't say who picks them up and how fast, they sit in a folder nobody opens.
Rewriting the brief mid-build. You'll learn things once you see the first version. Keep version one, log changes with dates, and expect changes to cost something. A developer can't hit a moving target at a fixed price.
Automating a process that isn't settled. If the team does the job three different ways, the brief will be vague and the build will pick one of them. Checking whether a process is ready to automate takes twenty minutes and can save you the whole budget.
Further reads
- How to Write a Request for Proposal for an AI Project — Turn the brief into a request several builders can quote on.
- How to Automate Vaccination Checks for a Grooming Salon — A related grooming workflow worth briefing at the same time.
- How to Hire an AI Automation Freelancer on Upwork or Fiverr — Where to find a developer and how to shortlist them.
- Who Owns the AI Workflows a Consultant Builds for You? — Settle ownership of the finished workflow before work starts.
- How to Pilot AI in Shadow Mode Before Customers See It — Run the finished build quietly alongside the old way first.
- How Much Does It Cost to Automate One Workflow With AI? — What one automated workflow typically costs to build and run.
- Per-Seat SaaS Fees vs One Custom Tool: Five-Year Costs Compared — Five-year costs of per-seat software against a custom AI tool, year by year, with a break-even seat count and the upkeep costs quotes leave out.
- Is AI Too Complicated for Non-Technical Business Owners? — Four levels of AI use, from typing into a chat box to custom builds, with honest learning times and five tests for when to bring in help.
- What Is an API? Why It Matters When You Buy Software — A plain-English explanation of APIs for people buying software, with the vendor questions and pricing traps that decide whether tools can connect.
- How to Map a Business Process Before You Automate It — A practical process-mapping method for automation: six columns per step, swimlanes drawn with AI help, an exception count and a label for every step.
- How Much Does a Custom AI Assistant Cost a Small Business? — Three ways to get an AI assistant that knows your business, what each costs up front and every month, and the upkeep that most quotes leave out.
- No-Code AI Tools vs a Custom Build: When Templates Stop Working — Six signs a Zapier or Make workflow has outgrown no-code, the middle options to try first, and a worked cost comparison for a care agency's shift offers.
- Power Apps for Small Businesses: When It Replaces Custom Software — When Power Apps can stand in for custom software: what Microsoft 365 already covers, where Premium licences start, and a vet practice's ward app costed.
- Should a Small Business Build a Custom AI Tool or Buy One? — A seven-question decision table, one community interest company's two-year sums for buy, configure and build, and what a build costs after launch.
- How to Implement AI in a Small Business With No Tech Team — Running AI with no IT staff: who covers the technical jobs, which tools need no code, how to secure logins, and what to do when something breaks.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Anthropic API pricing page (checked September 2026).