How to automate business processes with AI: where to start

Ask an operations manager where the week goes and you usually get a list of small jobs: copying enquiries from WhatsApp into a spreadsheet, chasing confirmations, re-typing invoice details into the accounting system, answering the same twenty questions about prices and opening hours. Each job takes a few minutes. Together they fill days, and they are the kind of work automation handles well.

Smaller firms have the most room to catch up. In the EU, 20.0% of enterprises with 10 or more employees used AI in 2025, including 55.0% of large enterprises but only 17.0% of small ones. The hard part is choosing well. Pick the wrong process and you spend months automating something that runs twice a week, or hand an AI a decision that needed a person. This guide covers how to choose, how to prepare, when rules beat AI, and the costs that rarely appear in the first quote.

Which business processes should you automate first?

Automate first the process that is high-volume, repetitive, rule-based and costly to delay. Check every process you are considering against these four traits before you look at a single tool.

• Volume. It happens often: dozens of times a day, or hundreds a month. A task that runs twice a week rarely repays the build and the upkeep.
• Repetition. Most cases follow the same path. If every case is a new puzzle, there is little to automate.
• Rules over judgement. The decision at each step can be written down, for example: invoices over 5,000 go to the finance manager. Steps that need judgement, such as a refund dispute or a clinical question, stay with a person or need a person's approval.
• Cost of delay. Waiting costs money or customers. An enquiry answered tomorrow is worth less than one answered now.

The last trait is the one most often underrated. In an audit of 2,241 US companies reported in Harvard Business Review in 2011, 37% answered a web enquiry within an hour, 23% never answered, and the average among firms that replied within 30 days was 42 hours. The same article reports that firms that tried to contact a lead within an hour were nearly seven times as likely to qualify it as firms that waited even an hour longer. When the work is answering people, speed is often where automation pays first.

To compare candidates, score each one from 1 to 5 on six criteria. Two of them work in reverse: a process that needs a lot of judgement, or does real damage when it goes wrong, scores low.

Add up the six scores. As a rule of thumb, 24 or more out of 30 is a strong first candidate, 18 to 23 usually works with a person approving the key steps, and below 18 you should fix the process or the data before automating it. Three common processes might score like this:

Map the process as it really runs today

The process in the operations manual and the process your team runs are rarely the same. The real one has workarounds: a side spreadsheet, a WhatsApp group where approvals happen, one colleague who knows which customers get special prices. Automate the documented version and the system will fail on the first case nobody told it about.

Map it with the people who do the work, using last month's real cases instead of memory:

1. Pick 20 to 30 recent cases and follow each one from the trigger (a message, a form, an email) to the finished result.
2. Write down every step, who does it, which tool they use and how long it takes.
3. Mark each decision point and the rule behind it. If nobody can state the rule, it is a judgement call.
4. List the exceptions and how often they happen. Exceptions decide how much of the process you can automate.
5. Note every place where data is typed twice or copied between systems. These are usually the cheapest wins.

The result fits on one page: steps, owners, tools, rules and exceptions. It also becomes the brief for whoever builds the system, and the core of scoping an AI project properly.

Measure a baseline before you build anything

Without a baseline you cannot tell whether the automation worked, and every later discussion about it turns into opinion. Measure for two to four weeks before building, from your systems where you can rather than from estimates.

Self-reported time savings are a weak substitute. In a November 2024 survey analysed by the Federal Reserve Bank of St. Louis, US workers who used generative AI said it saved them 5.4% of their work hours in the previous week, which the authors estimate implies a 1.1% productivity increase across the whole workforce. That is a real gain, and a modest one. Your own measured numbers will tell you far more than any national average.

Write the target down before you build: first response under five minutes at any hour, for example, or invoice entry under a minute with no more errors than today. A clear target also tells you when to stop adding features.

Rule-based automation or AI: which fits each step?

Rule-based automation follows fixed instructions written in advance; AI automation uses a language model to interpret inputs that come in no fixed format. Much of what gets sold as AI automation is plain rule-based automation: when this happens, do that. Rules are cheap to run, predictable and easy to audit. AI earns its place where the input is unstructured: a free-text message, a voice note, a scanned invoice, a question phrased in a way nobody anticipated.

That last risk has a formal name. NIST's Generative AI Profile calls it confabulation: “the production of confidently stated but erroneous or false content.” For that reason an AI step should answer from approved knowledge, act through a small set of well-defined actions, and hand off when it is unsure.

AI is also uneven in ways that are hard to predict. In a field experiment with 758 Boston Consulting Group consultants, Dell'Acqua and colleagues found that on 18 tasks inside the AI's capabilities, consultants using it completed 12.2% more tasks, 25.1% more quickly, with higher-quality results. On a task chosen to fall just outside those capabilities, consultants using AI were less likely to reach the correct answer than those working without it. The authors call this the jagged technological frontier. The practical lesson is to test AI on your own cases, including the awkward ones, before you trust it.

In practice the best systems combine both. AI reads the message and works out what the customer wants; rules decide what happens next and which systems are updated. An AI agent goes a step further and chooses among tools to complete a task, which makes clear limits even more important.

Keep a human in the loop, and design the hand-off first

Automation moves people to different work: the exceptions, the approvals and the judgement calls. Lisanne Bainbridge made the point in 1983, in a paper on industrial process control:

For every step in your map, decide which of three modes applies:

• A person approves before the system acts. For anything costly or hard to reverse: refunds, quotes above a threshold, a message to your whole customer list.
• The system acts and a person reviews a sample. For routine work where a mistake is cheap, such as tagging enquiries or updating a CRM field.
• The system acts and hands off the exceptions. For high-volume conversations, where most cases are simple and a few need a person quickly.

The evidence from customer service favours AI working alongside people. In a study of 5,172 customer-support agents published in the Quarterly Journal of Economics, AI assistance raised issues resolved per hour by 15% on average. Less experienced agents improved in both speed and quality, while the most experienced saw small gains in speed and small declines in quality.

A good hand-off carries the full context, so the customer never has to repeat themselves: who they are, what they asked, what the system already did and why it stopped. Some platforms require one. The WhatsApp Business Messaging Policy allows automated replies but requires prompt, clear and direct escalation paths, such as a transfer to a human agent.

Responsibility stays with the business. In Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline liable after its website chatbot gave a customer wrong information about bereavement fares. As the tribunal put it: “It makes no difference whether the information comes from a static page or a chatbot.”

Connect to the tools you already use

An automation that lives in a new app nobody opens becomes one more place to check. Build it into the systems your team already works in: the CRM, the calendar, the spreadsheet that really runs the business, the accounting package, email and WhatsApp.

• Name one system of record for each kind of data. Bookings live in the calendar and customers in the CRM. The automation reads from and writes to those, and keeps no private copy that drifts out of date.
• Check access early. Does each tool have an API, and does your plan include it? Can you create an account with only the permissions the automation needs? Many projects stall at this point.
• Read before you write. Let the automation read data and suggest actions first, and give it write access to live systems once its suggestions hold up.
• Learn each channel's rules. On WhatsApp, a business can reply freely once a customer writes in, but outside the 24-hour customer service window it may only send pre-approved message templates, under the WhatsApp Business Messaging Policy. Reminders and follow-ups have to be designed around that.
• Keep the accounts in your name. The WhatsApp number, the calendar, the hosting and the data should belong to your business, so the system stays yours if your supplier changes.

When connecting systems is most of the job, treat it as an integration project in its own right.

Your first AI automation project, step by step

Keep the first project to one process, end to end, and work in this order:

1. Score your candidates with the table above and pick the top one, even if another idea is more exciting.
2. Map it with the people who run it, using real cases and their exceptions.
3. Measure the baseline for two to four weeks and write down the target.
4. Decide rules or AI for each step, and which of the three human modes applies.
5. Write the hand-off before the happy path: what triggers it, who receives it and what context travels with it.
6. Build the smallest useful version on one channel or for one team, connected to the real systems.
7. Run it in shadow mode: the system drafts, a person approves, and you compare its output with what the team would have done.
8. Measure against the baseline, then widen the scope or give the system more autonomy one step at a time.

We do this work as projects, from scoring candidates to building and installing the system. See how we approach business automation, or describe the process that slows your team down.

What costs are easy to miss?

Five running costs are easy to miss: maintenance, monitoring, model changes, usage fees and people's time. The build is the visible cost; these arrive later, and they are the ones most often left out of a business case. Engineers at Google made the point about machine learning in 2015, when Sculley and colleagues wrote that “it is common to incur massive ongoing maintenance costs in real-world ML systems.”

Drift is measurable. Researchers at Stanford and UC Berkeley compared two versions of GPT-4 from March and June 2023 and found its accuracy at telling prime from composite numbers fell from 84% to 51%. Models are also retired outright: OpenAI's deprecations page lists gpt-3.5-turbo-0613 as shut down on 13 September 2024 and gpt-4-32k on 6 June 2025, each with a recommended replacement. A system built on one model needs a tested path to the next.

An example: an AI receptionist for dental clinics

Lina, the AI receptionist Flowramo built for dental clinics, shows how these choices fit together in one system. A clinic's WhatsApp scores high on the table: many messages, mostly the same questions and booking requests, and a real cost when a patient waits until morning for an answer.

AI handles the unstructured part: understanding messages and voice notes in whatever language patients write, including Gulf Arabic, English and Turkish, and answering only from knowledge the clinic has approved. Rules and integrations handle the rest: bookings and reschedules go straight into the clinic's calendar, and reminders go out on schedule. Urgent, clinical and sensitive cases go to the clinic team with the full context, and Lina does not diagnose. The system is installed on hosting the clinic chooses and owns, connected to the clinic's own accounts.

You can see how Lina works. The same method applies to any business whose work runs through repeated conversations, documents and hand-offs.