Which process should a small business automate first with AI?

By Ranjit Bhinge, founder of Blur Studio · Updated · 7 min read

Short answer: Automate the process that eats the most weekly hours, has clear rules, costs little if a mistake slips through, and touches the fewest systems. Score each candidate from 1 to 5 on those four things, add them up, and start with the top score. AI here means software such as ChatGPT or Claude (chat assistants from OpenAI and Anthropic) that reads and writes text. Never automate a process that is broken or that nobody can describe, and keep a person checking the output.

Ranjit Bhinge

You run a small service business, you know AI could help, and you want a way to pick the first process that is not a guess.

I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.

If you would rather have someone score your processes with you, the AI Strategy Sprint is a paid 3-day first step that does exactly that.

Most small firms do not lack AI ideas. They lack a way to choose one. So they pick the exciting one, or the one a vendor demoed, and it stalls. The cheaper path is to score your processes on four plain questions and start with the winner. This page gives you the sheet, a worked example and the rules for when to stop.

If you want the whole sequence after the pick (setup, testing, rollout), see the AI implementation plan. This page only answers "which one first."

Why the first pick matters

RAND, a US research nonprofit, interviewed 65 experienced AI engineers about why AI projects fail. The top reason was that the people involved misunderstood or miscommunicated what problem the AI was meant to solve. Another frequent one was using AI on a problem simple rules could handle. Both are pick-the-wrong-process failures. RAND's study was mostly about building and training models, so it is not a perfect match for a small firm using ready-made tools. But the lesson carries: the choice of problem decides more than the choice of tool.

Anthropic, the company behind Claude, says in its guide to building AI agents that you should look for "the simplest solution possible" and add complexity only when it clearly helps. For a first project that means: one small process, not a company-wide change.

The four scores

Give each process a score from 1 to 5 on each line.

ScoreWhat it measures1 means5 means
Weekly hoursTotal person-hours across the teamUnder 1 hour10 hours or more
Rules clarityCould you write the steps so a new hire does it right?"It depends, I just know"Steps are written or easy to write
Cost of an errorWhat a mistake costs you in money, trust or lawA shrugA lost client or a legal problem
Systems touchedHow many tools or places the work passes throughOneFive or more

The total is: hours + rules clarity + (6 minus error cost) + (6 minus systems touched). The last two are flipped because a high error cost and many systems are bad for a first project. The maximum is 20.

Weekly hours are the easiest to inflate. Count the real time: ask the people who do it, then check against a week of actual work. Do not count the hours of a task that happens twice a year.

Worked example: a 20-person service firm

These figures are made up to show the method. Imagine a firm with 20 people that does consulting and design work for clients. It lists five processes.

ProcessHours/wkRulesError costSystemsTotal
Answering repeat client questions by email12 (5)32 (flipped 4)1 (flipped 5)17
Meeting notes and follow-up emails5 (3)41 (flipped 5)2 (flipped 4)16
Turning time sheets into invoices8 (4)44 (flipped 2)3 (flipped 3)13
New-client onboarding paperwork4 (3)53 (flipped 3)4 (flipped 2)13
Writing proposals from scratch6 (4)24 (flipped 2)2 (flipped 4)12

Ranking: repeat client questions (17), meeting notes (16), then a tie at 13 between invoices and onboarding. Break ties with the lower error cost, so onboarding is third. Proposals are last: the rules are fuzzy because every proposal depends on judgment.

What I would notice in this table: the winner is not glamorous. It is "answer the same eight questions again." The firm would pick it, build a drafting step that suggests replies from its own past answers and its own documents, and a person approves every reply before it goes out. Invoicing looks tempting because it has money attached, but a wrong invoice hurts, and that is exactly what the error score caught.

Rules before you start

Do not automate a broken process

If two people do the task two different ways, or nobody can say what "done" means, fix that first. Software repeats a mess faster. A short written version of the steps, even half a page, is enough to test whether the process is clear. Our guides to documenting business processes and a process documentation template show how. If writing it down feels hard, that is your answer: your rules clarity score is a 1 or 2.

Know what AI is bad at

AI output is probabilistic. That means the same input can give different results and it can be wrong with full confidence. One RAND interviewee said AI is poorly suited to automating internal processes where subjective human judgment decides how they should work. So things like "decide which client to drop" are not first projects. Things like "draft a first version a person edits" are.

Keep a person checking, and decide where

For each candidate, write down the exact step where a human looks at the output before it reaches a client, a bank or a filing. For low-risk work (internal meeting notes) that can be a quick skim. For anything customers see, check every item in the first month, then sample. Without that step, a rare mistake becomes your problem in public.

Watch what data goes in

Before you paste customer information into any AI tool, check which plan you are on. OpenAI says it does not train its models on data from ChatGPT Business, ChatGPT Enterprise or its API by default, and Anthropic says the same for its commercial products such as Claude for Work and the API. Personal plans can differ, so read the terms. See how to keep client work out of AI training.

Turn the score into a plan

  1. List 8 to 12 candidate processes by asking each team member for the most repetitive thing they do. Do not edit the list yet.
  2. Score the top five on the sheet. Do this with the people who do the work, not alone.
  3. Write the steps of the winner on one page, including the check step.
  4. Define "better" in one number before you start: hours per week, replies sent per day, days to invoice. Measure it for two weeks as your baseline.
  5. Build the smallest version. Often that is a prompt and a template, not a system. If you need wiring between tools, see AI workflow automation for small business and business process automation.
  6. Run it for 30 days next to the old way, then compare against the baseline number.
  7. Decide: keep, change or stop. Stopping is a valid result, and cheap.

If the process in question is the written procedures themselves, SOP automation and the AI SOP generator guide cover turning them into documents and checklists. For a broader look at where AI fits a small business, read AI for small business and the AI readiness checklist.

Common mistakes

  • Picking by excitement. The process the owner finds interesting is rarely the one with the most hours.
  • Counting AI hours saved before the review step. If a person must check every output, subtract that time.
  • Starting with the customer-facing step. Start behind the scenes, where a mistake stays inside.
  • Buying the tool first. Gartner, a research firm, said in June 2025 that many use cases sold as AI agents do not need an agent, and warned of vendors "rebranding" existing products. Choose the process, then the tool.
  • No baseline. Without a before number you cannot tell whether it worked.

When to get outside help

You can score processes yourself in an afternoon. Get help when the scores are close and the stakes are real, when the winner touches three or more systems, or when it involves regulated data. The NIST AI Risk Management Framework (NIST is a US standards agency) is a voluntary guide to building trustworthiness into how AI products and systems are designed, used and evaluated. You do not need to adopt it, but the habit fits a small firm: know what the AI is for, measure it, and have someone responsible.

How I'd approach this

I would not pick a process in a meeting. I would send the scoring sheet to the people who do the work, collect scores separately, and compare. Where people disagree on "rules clarity," I would take that as a warning sign that the process is not ready. I would start with the highest total, tie-broken by the lowest cost of a mistake, and I would run it in "draft only" mode for the first month so a person approves everything.

I also tell owners to expect the first win to be small and a little dull. That is a good sign. A dull first win that works earns the trust you need for the next one. If you want help scoring, the AI Strategy Sprint is a paid 3-day first step where we go through your processes together, find where AI could transform the business and generate new value, and write down the exact steps. It starts with a free 15-minute AI brainstorm call.

Questions people ask

How do I start using AI in a small business?

Pick one process, not a tool. Score your repetitive tasks on weekly hours, rules clarity, cost of an error and systems touched, then start with the top one in draft-only mode. The AI for small business guide gives wider context.

What makes a good first process to automate?

It eats real hours every week, has steps you can write down, costs little if a mistake slips through, and sits in one or two tools. Internal work such as meeting notes or first-draft replies usually beats customer-facing steps for a first project.

Which processes should I not automate with AI?

Ones where nobody can describe the rules, where one error is costly or illegal, or where judgment decides the outcome. Fix a broken process before automating it. Our process documentation guide helps you see which is which.

How long should a first AI project take?

Plan on a small version in days to a couple of weeks, then a 30-day trial beside the old way. If it needs months before anything works, the scope is too big for a first pick. The AI implementation plan shows the stages.

Do I need a consultant to choose?

Not always. You can score processes yourself in an afternoon. Outside help pays off when scores are close, the data is sensitive or several systems are involved. See questions to ask an AI consultant before hiring if you go that way.

How do I know the automation worked?

Measure one number before you start, such as weekly hours or days to invoice, then compare after 30 days. Subtract the time people spend checking the output. If the number did not move, stop or change it.

More in this topic: Problem solving: AI strategy and adoption (24 guides) →

Sources

Ready for the next step?

A free 15-minute AI brainstorm. We find 3 broad areas where AI could help your business.

Ranjit Bhinge You talk to me, Ranjit. I reply within a few hours.
See the AI Strategy Sprint

Not sure what you need? Tell me what is going on. I will point you the right way, even if it is not me.

Blur Studio · 30 minute call

Book a 30 minute call

You talk to me, Ranjit. If you do not need me, I will tell you.

  • We talk about your business and what is going on.
  • You leave with a clear next step, even if it is not me.
  • I reply to every message within a few hours.
Loading the calendar…