DEV Community: Barry Batsbak

Process automation for small companies: a practical checklist

Barry Batsbak — Sun, 11 Oct 2026 11:13:33 +0000

Every automation project that fails has the same cause: the process was never written down.

Here is the checklist we run before touching any tool.

1. Write the process as steps

Not as a goal ("handle incoming mail faster") but as steps. Who does what, in which order, with which input and which output. If two people describe the process differently, the automation will fail.

2. Count the time

Measure how many minutes per week this costs, per person. Automating something that costs ten minutes a month is a hobby project, not a business case.

3. Find the decisions

Mark every point where someone has to decide. Those points need rules, or a human in the loop. Almost all failed automations skipped this step.

4. Define done

What is the output? An answer, a completed file, a scheduled appointment, a prepared quote? Vague output means vague automation.

5. Start small, ship in two weeks

One process, fully working, beats five processes half-built. Build, measure, then expand.

6. Keep a human at the exceptions

Automation handles the normal path. People handle the exceptions. Every system needs a clear route back to a human.

If you want to see how we apply this with AI agents, the worked examples are here: Match-AI.

How we taught an AI colleague to handle customer questions (and where we drew the line)

Barry Batsbak — Sun, 11 Oct 2026 11:13:12 +0000

Most teams that try AI in customer contact start with the wrong question. They ask "can the AI answer this?" instead of "should the AI answer this on its own?"

We build managed AI colleagues for small and mid-sized companies. After running these setups in production, here is what actually matters.

Start with one repeating process, not with a chatbot

Pick a process that costs time every single week. Answering the same ten questions, completing incoming requests, preparing quotes, following up on leads. If a process does not repeat weekly, it is not a good starting point.

The first process should be live in two weeks. Not six months, not after a data project. Two weeks is enough to cover one workflow honestly.

The escalation matrix is the real product

The value is not in the answers the AI gives. It is in the rules for what it does not handle.

We define four levels for every process:

Level	Who handles it	Example
1	AI answers directly	opening hours, lead times, standard pricing
2	AI drafts, human approves	quotes, planning changes, complaints
3	AI routes to a person	contract questions, legal, exceptions
4	AI stays out	anything about money or health

Levels 3 and 4 are where trust is built. A colleague that knows when to stay silent is worth more than one that answers everything.

Measure before and after, per process

Pick three numbers before you start: time spent per week, response time, and error rate. If you cannot measure it, you cannot defend the project later.

What we would do differently

Do not start with the most complex process because it has the biggest payoff. Start with the process where the rules are already clear and written down. Complexity can come later, once the team trusts the setup.

If you want the full checklist we use for the intake and the first two weeks, it is here: Match-AI.

What we learned hiring an AI colleague for customer support

Barry Batsbak — Sun, 11 Oct 2026 09:48:55 +0000

Most small and mid-sized companies in the Netherlands do not need another dashboard. They need a colleague.

We build AI colleagues: one agent per role, working inside the tools a team already uses. This is what we learned after running these agents for more than 40 Dutch SMEs, and what we would tell anyone starting with one.

Start with one process, not with a platform

The failure mode is a company-wide AI programme. The working mode is one process with a clear owner:

pick a process that runs every day
make sure it has a measurable output
give one person ownership

Customer support triage is the most common first choice. It has volume, a clear output and an obvious owner.

The agent needs three things

Context. Product documentation, price list, internal procedures, tone of voice.
Tools. Mailbox, helpdesk, CRM. Read first, write later.
Escalation. A rule for when a human takes over, and a log of every handover.

Human control is not optional

An agent that answers 80 percent of questions is useful. An agent that answers 100 percent and is wrong 5 percent of the time is a liability. We keep a human owner who reviews the queue weekly, and we log every escalation.

What it returns

In our own deployments the pattern is consistent: the first response time drops from hours to minutes, the team stops answering the same ten questions, and the human owner spends the freed time on the cases that actually need judgement.

Doing it yourself

If you want to try this inside your own team, the practical route is:

write down the ten most common questions
document the answer for each one
decide the escalation rule
measure the first response time before you start

You can find the working templates and the process we use for this on Match-AI.

The short version: one process, three ingredients, one human owner.