AI Workflow Setup for Small Businesses: Map, Build and Test One

Pick one task you repeat weekly with clear steps and cheap mistakes, write its steps on a page, build a draft-only version in Zapier or Make with one AI step, and run it on ten real past cases. Go live only when at least nine of the ten outputs would have been sent as they stood.

The mistake most first-timers make is starting with the tool. They open an automation platform, see a thousand app connections, and build something clever that nobody needed. Start on paper instead: what happens, who does it, how long it takes, and what a good result looks like. The tool is the last choice you make, and it matters least.

Follow me on Instagram@sagnikteaches

Step 1: choose the task with a scoring table

List five to ten jobs that make you groan. Then score each from 1 to 5 on the five questions below and add them up. A job needs at least 17 out of 25 to be a good first workflow.

Connect on LinkedInSagnik Bhattacharya

Here is a scored sample for an illustrative mortgage broker with a two-person team. Sorting new enquiry emails scores 5, 4, 4, 4, 5 = 22. Chasing missing documents scores 4, 4, 4, 3, 4 = 19. Preparing advice letters scores 3, 2, 2, 1, 3 = 11, so it is not a starter. The broker picks the enquiry sorting, because it is frequent, checkable and low-risk. For help scoring a longer list, our AI use-case template has a fuller one-page version.

Subscribe on YouTube@sagnikteaches

Jobs to leave until later

Skip anything involving payments, signed documents, medical or legal judgement, or decisions about individual people such as hiring. Skip jobs that fire once a quarter too, because you will never gather enough runs to test them. The first workflow exists to teach you how the pieces behave, so choose a forgiving one. The question of whether to start inside the business or facing customers is covered in customer-facing or back-office AI; for a first build, back-office is almost always safer.

Step 2: map the steps on one page

Take the chosen task and write down exactly what happens today, from trigger to finish. Use a plain list with four columns: step, who does it, what they look at, and what they produce. Do it for the real process, not the one you wish you had. Watch yourself or a colleague do the task once and note every small decision.

The broker's current process for a new enquiry email (illustrative) looks like this:

1. An enquiry arrives in the shared inbox (trigger).
2. Someone reads it and decides the type: purchase, remortgage, buy-to-let, or not a mortgage enquiry.
3. They copy name, contact details, property value and deposit into the client spreadsheet.
4. They send a short acknowledgement with the right list of documents to bring.
5. They add a follow-up task for two working days later.

Time per enquiry: about eight minutes, with 15 enquiries a week, so two hours a week. Mark which steps need judgement (step 2) and which are mechanical (3 to 5). Judgement steps are where an AI step might help; mechanical steps are where plain automation is cheaper and more reliable.

Define "good" before you build

Write one sentence for each output describing what counts as correct. For the broker: "The type is right, the five fields match the email exactly, no figure is invented, and the acknowledgement lists the documents for that type only." You will use this as your marking scheme in Step 5, and it keeps you honest when a run looks plausible but is wrong.

Step 3: decide where the AI step goes

Look at your map and ask, for each step: is this reading messy text? If yes, that is a candidate for an AI step. If the step is moving fixed fields between apps, use the platform's normal modules, which are cheaper and do not hallucinate.

In the broker example, one AI step does the reading: classify the enquiry type and extract five fields into a fixed structure. The acknowledgement can be a template per type, chosen by the classification, so the AI never writes the customer-facing text in the first version. That decision removes most of the risk.

Read the email below. Return only these fields, separated by line breaks, in this order:
Type: one of purchase, remortgage, buy-to-let, not-a-mortgage-enquiry
Name: the sender's name as written, or "unknown"
Property value: a number only, or "not stated"
Deposit: a number only, or "not stated"
Phone: as written, or "not stated"
If a field is not clearly in the email, write "not stated". Do not guess or calculate anything.

Email:
{{email body}}

Notice the three habits in this prompt: a fixed output format that later steps can read, an explicit "not stated" escape so the model is not forced to guess, and a ban on calculation. These keep the output checkable.

Step 4: build it in Zapier or Make

Both platforms follow the same pattern: a trigger starts the run, then a chain of steps does the work. In Zapier these chains are Zaps; in Make they are scenarios. Choose by trying the free plan with your two key apps.

Check the current plan limits before you commit. As of November 2026, Zapier's pricing page lists 100 tasks a month on Free, two-step Zaps only (one trigger and one action), a 15-minute check for new data and no Filters or Paths, while Make's Free plan gives 1,000 credits a month and Make says some AI features may use more than one credit per action. The broker build below has a trigger and five actions, so on Zapier it needs a paid plan with multi-step Zaps; on Make, check the Free plan's limits against the number of modules and runs you expect. Both vendors change plans often, so confirm on their pricing pages. For ten test cases and a modest live volume, a free or entry plan is usually enough to prove the idea.

Build it one step at a time and test each step on its own with sample data before adding the next. Name every step clearly ("Classify enquiry", "Add client row") so that you can read a failed run a month from now. Add a final step that logs the input, the AI output and the time, which gives you the evidence trail you need when something looks off.

Keep a human in the loop at first

For the first version, have the workflow create drafts and rows, not sent messages. A person scans the draft and clicks send. This keeps the speed benefit (no typing, no looking things up) and removes the risk of a wrong email going out unseen. The idea of reviewing without slowing everything down is covered in setting up human review for AI work.

Also add an error path. Both platforms let you send a message or create a task when a run fails, and without one a broken workflow can fail silently for weeks. Zapier turns off a Zap that errors on 95% of its runs once it has run more than 20 times in seven days (on Team and Company plans after a warning email and a grace period), so a quiet failure may also be a stopped one; check the run history weekly in the first month.

Step 5: test with ten real cases

Ten is a practical number: enough to catch the usual failures, few enough to check by hand in half an hour. Use real past enquiries, not clean examples you invent, and include awkward ones on purpose.

Pick them like this for the broker: six ordinary enquiries (two purchases, two remortgages, two buy-to-let), one with a missing deposit, one asking about a personal loan (not a mortgage), one in a forwarded chain with a long signature and one with a figure written in words ("two hundred and fifty thousand"). Remove names and personal details from the test copies if your tool settings are not yet confirmed, as covered in this guide on customer data in ChatGPT.

Mark each output against your definition of good from Step 2 in a simple grid: type right, fields right, nothing invented, template right. An illustrative result for the first ten:

Seven of ten passed, below the nine we set. That is a useful result, not a failure of the project. The two fails point at specific fixes: tell the model to take the sender from the first message in a forward, and to convert numbers written in words. After editing the prompt, rerun the same ten, and add five new cases so you are not tuning to the test.

Step 6: go live in stages

Passing the ten-case test earns a live trial, not full trust. Stage it.

1. Week 1: run live but draft-only. A person reviews every output and notes each correction.
2. Week 2: review a sample of every third output, and every one where the type is "not stated" or the figure looks unusual.
3. Week 3 onwards: if corrections have stayed under one in ten, allow the acknowledgement to send automatically for ordinary types only, and keep humans on exceptions.

Keep the draft-only route for anything unusual permanently. A workflow does not need to do everything to be worth having; handling the ordinary eight in ten and flagging the rest already saves most of the time. For the wider idea of running a trial quietly before customers see output, see piloting AI in shadow mode.

A second example: an HR consultancy's CV intake

To show the method is not specific to mortgages, consider an illustrative three-person HR consultancy that receives candidate CVs by email for client roles. The scoring table puts "log new CVs and summarise them" at 21 out of 25 and "shortlist candidates" at 9, because shortlisting is a decision about individuals and a mistake is costly.

So the workflow does only the first job: a new CV email triggers a step that saves the attachment to a folder, an AI step extracts name, current role, years of experience stated and location of work as written, and a row is added to the tracking sheet. It never ranks or rejects anyone. The consultant reads the summaries and makes every decision. Choosing the right boundary for what AI does is the design work, not the wiring.

Three more small workflows to try next

Meeting notes to action list

A recording transcript lands in a folder; an AI step extracts action items with owner and date; a task is created per item in your task tool. Check: owners and dates must be stated in the transcript, so tell the model to write "no owner stated" rather than invent one.

Invoice query triage for an interior designer

A client email about an invoice triggers a classification (query, dispute, payment promise, other); the row goes in a sheet and a draft reply is created for queries only. Disputes and anything about money owed go straight to the owner. A mistake in this area is expensive, so start with drafts and strict categories.

Weekly summary for the owner

A Friday schedule collects the week's rows from your sheet and asks the model for a five-line summary with counts. The check is simple: counts must match the sheet, which you can verify by formula in a minute.

Three failures a first workflow usually shows

The duplicate row

In week one the broker's sheet shows the same enquiry twice. The cause is that the trigger watches the whole inbox, and the acknowledgement draft, saved in the same mailbox, fires the workflow a second time. The fix is a filter at the start: continue only if the email is not from your own address and does not carry the label the workflow adds. On Zapier's Free plan filters are not available, so this is one of the situations that pushes people to a paid plan or to Make; check which features your plan includes before you design around them.

The quietly wrong field

A client writes "deposit of 25k" and the model returns 25. The row looks tidy and nobody notices until a review. The fix is an instruction to write full numbers ("25000") and a check step that flags any deposit under 1,000. Add one sanity rule like this for every numeric field, and the worst slips are caught without a person reading each row.

The silent stop

A connected account's login expires and the workflow stops. Nothing breaks visibly; enquiries simply stop appearing in the sheet. Guard against it with a weekly "heartbeat" check: if the sheet has received no new rows for three working days, send yourself a message. A simple scheduled step that counts rows does the job.

A filled-in process map

Here is the broker's map as a completed table, so you can copy the layout (illustrative):

The last column is the decision record. Five rows, three of them plain automation, one AI step, and one template: a simple design, and the reason the workflow is easy to test and to explain to a colleague.

What to write down so it keeps working

A workflow that only you understand is a liability. Keep a one-page record: purpose, trigger, steps, prompt text, who owns it, which accounts it connects through, what it costs a month, and how to switch it off. Use a shared login or a company account for the connections, not a personal one, so the workflow does not die when someone leaves; Make's Gmail documentation, for instance, says connections for personal @gmail.com accounts must be re-authorised every six months, so put a reminder in the calendar. Review the prompt every quarter and whenever an app changes its fields.

Also note a stop rule in advance: "If more than two in ten drafts need correcting in a week, we switch back to manual and review the prompt." Having that rule written down makes it easy to stop without drama. The next step, once the first workflow is steady, is described in scaling AI from one workflow to the whole business.

What it costs

For this example (illustrative), the plan costs are low next to the labour saved. Fifteen enquiries a week at eight minutes each is two hours; with the workflow, a person spends about two minutes checking each draft, so 30 minutes. That saves 90 minutes a week, about 78 hours a year. At $25 an hour of staff time, that is roughly $1,950 a year of time, against a platform subscription that might run from free to a few hundred dollars a year plus the AI usage. Your own numbers will vary, and the first month includes the setup time of perhaps six to eight hours, so the payback is a matter of months, not days. If you outgrow no-code tools or need integrations they do not offer, take the process map, the ten test cases and the stop rule to the developer; together they make a far better brief than a list of features. Or, if you would rather choose and map the first workflow with someone, that is what my AI implementation consultation covers; you can get in touch here.

Questions about a first AI workflow

Should I use Zapier or Make for my first workflow?

Either works. Zapier is usually quicker to learn for simple chains and has many ready-made app connections; Make shows branching and data more visually and can be cheaper for many-step scenarios. Pick the one whose free plan connects to your two key apps, build one workflow, and only compare plans after it runs reliably.

What makes a task a good first choice?

It repeats at least weekly, follows steps you can write down, has a clear right answer you can check, and mistakes are cheap to fix. Sorting enquiries, drafting replies for approval and filing documents fit. Anything involving payments, legal decisions or sensitive personal data does not fit as a first project.

Do I need an AI step at all?

Not always. If the task is moving data between apps in a fixed way, a plain automation without AI is cheaper and more predictable. Add an AI step only where the input is messy text that needs reading, sorting, summarising or drafting. Many good workflows are 80 per cent plain automation with one AI step.

How do I stop the workflow sending something wrong to a customer?

Make the first version draft, not send. Have it save a drafted reply or create a task for a person to approve, and only allow automatic sending after ten or more cases have passed your check. Keep a log of every run so you can trace a mistake back to its input.

Further reads

• Can AI Work With the Tools Your Business Already Uses? — Check whether your current apps already connect.
• How to Calculate AI ROI for Your Business (Worked Example) — Work out whether the time saved justifies the plan.
• How to Check AI Is Doing Good Work, Not Just Fast Work — A routine for marking outputs beyond speed.
• From Enquiry to Fact Find: An AI Workflow for Mortgage Brokers — A sector example of an enquiry workflow from start to finish.
• A Wedding Planner's AI Workflow From Enquiry to Final Timeline — Another worked example of a complete AI workflow.
• How to Brief a Developer on a Custom AI Workflow You Need — What to put in a developer brief when no-code tools run out.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Zapier's and Make's own plan, pricing and help pages, checked in November 2026; limits and prices on both change often, so confirm them on the vendors' pages.