AI Agents for SMEs: What We Learned from Running Them Internally
Diesen Artikel auf Deutsch lesen
Since 2024 we have been running our own company as a test bed for AI agents. Not as a pilot project on the side of daily business, but as part of it. What started as an architecture for sales, administration and reporting is, in 2026, an organisation with 273 agents across five hierarchy levels running around the clock — with one human founder who sets priorities and grants approvals.
This article describes what worked over those two years, what did not, and what SMEs can take away for their own start.
Why we started internally
The obvious path would have been to introduce agents at a client first. We did the opposite. Internally there were three advantages: we knew the processes down to the detail, we had to live with our own mistakes, and we did not have to explain to anyone why an intermediate step took two weeks longer.
The first stage in 2024 was deliberately unspectacular: recurring tasks from sales, administration and reporting. Proposal drafts, status enquiries, weekly reports, documentation. All tasks where a person had previously followed the same pattern.
The setup: five levels, one human
The hierarchy is not an analogy, it is a technical structure. At the top sit strategy and targets, below that department owners, then specialists for individual tasks, a quality assurance layer, and an operational layer that monitors execution.
Operations run on heartbeats: an agent wakes up, checks its task, executes it and delivers. Anything missing gets escalated — not guessed. That rule is the most important part of the whole system. An agent that improvises when the situation is unclear produces work instead of completing it.
The human share is smaller than you might assume, but it has not disappeared: setting priorities, granting approvals, reviewing results. 24/7 operation does not mean nobody is watching. It means the company does not stand still while nobody is watching.
What worked
Tasks with clear inputs and outputs and a repeatable sequence proved reliable:
- Reporting: merge data, flag deviations, write a summary.
- Documentation: meeting notes, decision logs, handovers between participants.
- Monitoring: surface anomalies early, before they become a problem.
- Proposal and copy drafts: fill templates, generate variants, submit for approval.
The common denominator: the task had already been clear enough that a new employee could complete it after a briefing.
What did not work
The biggest disappointments did not come from missing models. They came from vague briefs.
- Goals like “make it better” produce output that sounds plausible and measures nothing.
- Poor data quality does not improve with agents — it becomes more visible.
- Without an escalation point you get loops that look like activity on the dashboard.
- Too much automation in judgement: an agent can prepare options, but it cannot take responsibility.
The learning effort therefore sat in scoping tasks and in deciding when a human has to take over — not in model selection.
What makes client projects different
With clients there is an additional constraint: the intervention has to pay off in months, not years, and it must not stop operations. That is why we almost always start with a slice rather than a full rebuild.
An HVAC group with five locations cut administrative effort by 52% and the error rate in quotes by 71% — over nine months, with 28 trained technicians and a workflow from first visit to invoice.
A logistics provider with 80 vehicles raised on-time delivery from 82% to 98% and reduced fuel costs by 12%, supported over twelve months in four rollout phases with 35 trained employees.
A components manufacturer with 240 employees achieved 38% more conversion and EUR 2.1 million in pipeline within six months — through positioning, a CRM workflow and sales enablement.
The pattern is the same in all three cases: one clear process, one measurable metric, one defined period.
What we do differently today
Four rules emerged from internal operations, and they apply unchanged in client projects:
- Measure first, then automate. Without a baseline, every improvement is a claim.
- Build every agent with an escalation point. The question “who takes over when it does not fit?” belongs in the design, not in the contingency plan.
- Treat quality assurance as its own layer, not as a final check before shipping.
- Take the human out of the routine, not out of the responsibility.
Conclusion
In 2026, AI agents for SMEs are no longer a future topic. They are a question of scope. Anyone who describes a process cleanly and defines a metric gets a result in weeks. Anyone who “introduces AI” gets a construction site.
Want to know which tasks in your business can be automated first? We will analyse it with you — and name the metric we will measure.
Book an initial call