Automation8 min read

Stop Writing an AI Strategy. Automate One Boring Workflow First.

Most AI strategies are long decks with zero automations. Pick one boring workflow, ship it in 30 days, then write the strategy from what you learned.

A thick strategy binder pushed aside next to a single simple workflow diagram on a desk

The fastest way to start with AI automation in a small business is to skip the strategy deck and automate one boring workflow. Pick a task with high volume, simple rules, text in and text out, quick human checks, and someone who hates doing it. Ship it in 30 days, then plan from real results.

Most AI strategies we are asked to look at are long decks. They have a vision slide, a maturity model, twenty use cases ranked on a two-by-two grid, and a governance framework. What they rarely have is a single automation running in production.

That is not a criticism of the people who wrote them. Strategy feels like the responsible first step. But in a small or mid-sized business, a strategy written before you have shipped anything is mostly guesswork. You do not yet know what your data looks like, where the edge cases hide, or how your team reacts when software starts doing part of their job.

So here is the contrarian advice: stop writing the AI strategy. Pick one boring workflow, automate it properly in 30 days, and then write the strategy from what you learned. This is how AI automation for small business actually gets started, and it costs less than the deck.

Why do most AI strategies produce zero automations?

The problem is structural. A strategy document is judged by how complete it looks, so it grows. Every department adds its wish list. Legal adds a risk section. Someone adds a slide about agents. By the time it is approved, the list of priorities is so long that nobody owns the first one.

Meanwhile, the questions that decide whether AI works for you are practical ones that no deck can answer:

  • How messy is the data in your helpdesk, inbox or ERP once you actually export it?

  • What share of cases follow the rules, and what share are exceptions?

  • How long does it take a person to check an AI draft, and will they trust it?

  • What does it cost to run per case once real volume hits it?

You only learn these by building something small and real. One shipped workflow teaches you more about your readiness than three months of workshops.

How do you pick the right first workflow?

The best first workflow is boring. Not the flashy demo, not the one the board mentioned, but the repetitive task that quietly eats hours every week. Run each candidate through five tests. If it passes all five, it is a strong first project.

  1. High volume. It happens dozens or hundreds of times a week. Saving two minutes on something that happens twice a month is not worth a build.

  2. Rules you can explain in a page. If a new hire could learn to handle most cases from one page of instructions, an AI system can usually follow them too. If the honest answer is that it depends on years of experience, wait.

  3. Text in, decision or text out. Emails, forms, PDFs, chat messages and tickets go in. A reply, a category, a field value or a yes or no comes out. This is where language models are strongest.

  4. A human can check the output in seconds. You want to glance at a draft reply or an extracted invoice line and know whether it is right. That keeps the review step cheap and safe.

  5. Someone already hates doing it. This matters more than people expect. If the person doing the task is relieved to hand it over, adoption takes care of itself. If they feel threatened, it stalls.

Notice what is not on the list: strategic importance. The first workflow is a learning vehicle. It needs to be safe, measurable and quick to ship, not transformative.

What are good first workflows, team by team?

Here are ten workflows that tend to pass the five tests. You will not need all of them. You need one.

  • Support: first replies. Draft the first response to common tickets from your help center and order data, for an agent to approve.

  • Sales: lead qualification. Read inbound form fills and emails, work out the qualifying answers, and route hot leads to a rep within minutes.

  • Finance: invoice entry. Extract supplier, amounts, tax and line items from PDF invoices and create draft bills in Odoo, QuickBooks or Xero for someone to post.

  • Customer service: order status on WhatsApp. Answer where-is-my-order questions by looking up the order and replying through the WhatsApp Business Platform.

  • HR: CV screening. Score applications against the must-have criteria in the job brief and explain why, so a recruiter reviews a ranked shortlist instead of a pile.

  • E-commerce: product descriptions. Turn supplier specs and a few attributes into consistent product copy in your brand voice, ready for review.

  • Real estate: viewing requests. Read enquiries, match them to listings and open slots, and propose viewing times.

  • Logistics: proof-of-delivery checks. Check that delivery photos and signatures match the order and flag the ones that do not.

  • Legal and procurement: contract first pass. Compare incoming contracts with your standard positions and highlight the clauses that differ.

  • Management: weekly reporting. Pull numbers from your systems and draft the weekly summary with the changes that matter called out.

A 40-person logistics firm in Manchester might start with proof-of-delivery checks. A Shopify brand in Austin might start with order status replies. A real estate brokerage in Dubai might start with viewing requests in English and Arabic. Same method, different workflow.

What does a 30-day plan for one workflow look like?

Thirty days is enough to go from talking about AI to a workflow running on real cases. The trick is to stay narrow and to keep a person in the loop before you let the system act alone.

Week 1: measure the workflow as it is

Before anyone writes a prompt, measure. How many cases arrive each week? How long does each one take? What are the categories, and which ones repeat? Export a few hundred real examples, including the messy ones, and write down what correct looks like for each type.

This week produces two things: a baseline you will compare against later, and a test set of past cases you can run the system on before it touches anything live.

Week 2: build narrow, for the common cases

Build for the common 60 to 70 percent of cases and nothing else. If most of your tickets are order status questions and return policy questions, those are your scope. Everything else goes to a person, exactly as it does today.

Narrow is not a compromise. It is what makes the system reliable enough to trust. Trying to cover every edge case in week two is how pilots die.

Week 3: run it beside a person

Now the system drafts and a person approves. Every case still goes through a human, but they are editing or approving instead of writing from scratch. Track how often they accept the draft unchanged, edit it a little, or throw it away.

This is your real evaluation. It also builds trust, because the team sees exactly what the system does before it does anything on its own.

Week 4: let it act on the cases it gets right

By now you know which categories the system handles well. Let it act on those directly and keep the approval step for the rest. Keep logging, keep sampling outputs, and keep a clear path for a person to take over.

At the end of the month you have a live workflow, real numbers on time saved and accuracy, and a list of what went wrong. That is worth more than any maturity model.

What does a first-workflow brief look like in practice?

Here is an illustrative brief for a 25-person distributor in Riyadh that wants to automate supplier invoice entry. It is an example of the format, not a client story, and the numbers are made up to show the level of detail you want.

Workflow: supplier invoices arrive by email as PDFs and are typed into the ERP by hand. Volume: say 400 invoices a month from about 60 suppliers. Time today: 4 to 6 minutes each. Month one scope: the 15 suppliers who send most of the invoices. Output: a draft vendor bill with supplier, invoice number, date, VAT, totals and line items. Rule: nothing posts automatically in month one; an accountant approves every draft. Success: 9 out of 10 drafts approved without edits for in-scope suppliers by week four.

Notice how specific it is: a named input, a named output, a narrow scope, a human check, and a number that decides whether it worked. If you can write a paragraph like that for one of your workflows, you are ready to build.

When should you write the AI strategy, then?

After the first workflow, not before. Once one automation is live, you can write a strategy grounded in your own evidence rather than vendor slides. It tends to be shorter and far more useful, because it answers questions like these:

  • Which data sources were clean enough, and which need work before the next project?

  • How much review time did the team actually need, and where can it drop?

  • What did it cost to run per case, and what does that mean for the next three workflows?

  • Which model and hosting setup suits your data rules, including whether anything must stay inside your own cloud?

  • Who owns the automation now that it is live, and who fixes it when something changes?

Then pick the next workflow using the same five tests. Second and third projects usually move faster, because the integrations, logging and review habits already exist. Over time, single-task automations can grow into agents that handle several steps and choose between them. If you are not sure where that line sits, our post on AI agents vs chatbots explains the difference.

What should you avoid in your first AI project?

  • Starting with the hardest problem. Complex judgment calls make poor first projects. Save them until you know your data.

  • Skipping the measurement week. Without a baseline, you cannot show the automation saved anything.

  • Letting the system act on day one. Draft first, act later. Trust is earned case by case.

  • Buying a platform before you know the workflow. Tools are much easier to choose once you know exactly what they must do.

  • Ignoring ownership. Make sure you own the prompts, code, test set and deployment, so the work is never stuck with one vendor.

If you want help scoping or building that first workflow, this is the work we do in AI workflow automation, and in AI agent development for workflows that need to reason across several steps and tools. We usually put a working prototype in front of you in two to four weeks, with a demo every week, and you own the code at the end. We work remotely with teams across the United States, the UK, Europe and the Gulf.

Already have a workflow in mind? Book a free discovery call and we will check it against the five tests with you before you spend anything on a build.

Frequently asked questions

What is the best first AI automation for a small business?

The best first automation is a boring, high-volume task with simple rules, such as drafting first replies to support tickets, entering supplier invoices, or answering order status questions. It should take text in, produce a decision or text out, and be easy for a person to check in seconds.

Do I need an AI strategy before starting with AI automation?

Not a long one. A short note on goals and data rules is enough to start. A detailed strategy is far more useful after your first workflow is live, because it can be based on real costs, accuracy and team feedback instead of assumptions.

How long does it take to automate one workflow with AI?

A narrow workflow can usually go from measurement to running on real cases in about 30 days: one week to measure, one to build for the common cases, one running beside a person, and one letting it act on the cases it handles well.

Will AI automation replace my team?

For most small businesses, the first automations remove repetitive work rather than roles. People move from writing and typing to reviewing and handling exceptions. Choosing a task that someone already dislikes makes this transition much smoother.

What if the AI gets something wrong?

That is why the first weeks run in draft mode, with a person approving every output. You only let the system act on categories it has proven it handles well, and you keep logging, sampling and a clear way for a person to take over.