From traditional SME to AI-first: The Case of a 30-Employee Company

How does a normal SME—without a technology department or a multinational budget—actually become an AI-first company? This article tells the story of an industrial supply distributor with 30 employees: situational background, what they decided to automate, in what order, and what the numbers looked like before and after. This case is a composite and realistic portrait of such projects—the company is fictional and the figures are representative of what we see in distributors of this size—but the process is exactly what we follow in reality.

The starting point: growth through hiring

The protagonist: an electrical and industrial material distributor located on the outskirts of a provincial capital. Thirty employees, about €7 million in turnover, 1,900 references in their catalog, and around 620 orders per week. Out of the staff, eight people were dedicated purely to administration: orders, invoicing, and incidents.

Their operations mirrored thousands of Spanish SMEs. 70% of orders arrived via email—a PDF, a photo of a delivery note, a "give me the usual"—and three people spent most of their day interpreting and typing them into the ERP: about 14 minutes per order, with a 3% line error rate (wrong reference, misread quantity) that later resulted in returns and shipping costs. Quotes took 24–48 hours to be sent out because they passed through a single person who knew the rates and customer discounts. The monthly accounting close took 9 days.

The trigger wasn't AI: it was a spreadsheet. The administration manager asked for two more people to absorb the growth, and the CEO did the math: an additional annual cost of about €70,000 to keep doing the same thing. Before signing the job offers, he decided to look at the alternative.

Phase 1 (Weeks 1-4): Diagnosis, not tools

The first step wasn't buying anything. It was measuring: an inventory of 12 administrative and commercial processes, hours per month for each, cost, and error rate. The result: order entry consumed about 230 hours per month; quotes, 90; supplier invoice reconciliation, 60. Just three processes combined were equivalent to two and a half full-time jobs.

The diagnosis revealed two surprises. First: the process the manager wanted to automate—phone customer service—was one of the worst candidates: too many exceptions, high relationship value. Second: the quoting process wasn't slow due to a lack of hands, but because the knowledge of rates and discounts lived in one person's head. Before automating, they had to redesign: unifying discount criteria into a table approved by management. That redesign alone cut quoting times by 30%, without any AI involved.

Phase 2 (Weeks 5-12): Two agents in production

With the map in hand, two processes were chosen—high volume, clear patterns, corrigible errors—and moved to production:

• Order Agent: reads the order inbox (emails, PDFs, delivery note photos), interprets the lines, validates them against the catalog, rates, and stock, and creates the order in the ERP as a draft. A human reviews and confirms with a click; everything the agent doesn't understand with sufficient confidence is escalated and flagged in red. Connected to existing systems via native connectors, without changing the ERP.
• Quote Agent: based on the customer request, it assembles the quote with the approved rates and discounts and leaves it ready for commercial review. For quotes over €5,000, the offer requires approval from the sales manager: the limits and who approves what were defined on day one, not after the first mishap.

Both agents started in "propose" mode: employees continued to execute while the agent prepared. After four weeks of measuring their accuracy rate with real volume, they moved to execution with human confirmation. Team training was done using their own orders and customers, not generic examples—and with an explicit message from management: the freed-up hours were reinvested, no one was being fired.

Results after six months

Against the baseline measured in the diagnosis—not from memory—here is what changed:

• Order entry: from 14 minutes per order to about 2 minutes of review. From 230 hours per month to less than 50.
• Order line errors: from 3% to 0.4%, with corresponding savings in returns and rectification shipping costs.
• Quotes: from 24–48 hours to less than 4; 80% are sent the same day. The acceptance rate rose by about 6 points—in distribution, responding first wins orders.
• Administrative hours freed: about 210 per month across the two processes, the equivalent of 1.3 full-time jobs.
• The two planned hires weren't made: about €70,000 per year avoided. The entire project—diagnosis, two agents, training—paid for itself in less than six months.

And the people? No one left the company. Two administrative staff moved to commercial support—proactive follow-up on sent quotes, reactivating dormant customers, work that previously had no time allocated—and a third became the "owner" of the agents: reviewing their metrics, managing escalated cases, and deciding with the manager when an agent earns more autonomy. That role didn't exist before and is now one of the most valuable in the house.

Phase 3: From project to way of operating

With the credibility of the numbers, expansion funded itself: a supplier invoice reconciliation agent (monthly close went from 9 days to 4), a delivery incident follow-up agent, and a daily dashboard where the manager starts the morning seeing orders, margins, and exceptions without asking anyone for reports. The distributor stopped "having an AI project" and started operating as an AI-managed company: agents executing repetitive tasks, people deciding and selling.

A year later, the manager's question had changed. It was no longer "How many people do I need to grow by 20%?" but "Which process do we automate next?". This reversal of the question is, in practice, the definition of an AI-first company.

What this SME did right (and you can copy)

• Measured before building: the diagnosis with hours and costs was the compass and, later, the proof of ROI.
• Started with two processes, not twenty: high volume, clear rules, low cost to correct errors.
• Redesigned before automating: discount criteria were unified before putting the agent in place.
• Set limits and owners from day one: approval thresholds, escalations, and a named agent manager.
• Reinvested hours into selling more, and said so from the start: that's why the team supported the project.

The full sequence, with week-by-week milestones, is available in our 90-day roadmap for implementing AI in your company. And if you want to know what version of this story is possible with your processes and your numbers, start where this distributor started: with an automation diagnosis.

Frequently Asked Questions

Can an SME with 30 employees really be AI-first?

Yes, and they often move faster than large companies: fewer committees, more visible processes, and a manager who decides in days. The key isn't size but sequence—measured diagnosis, two high-volume processes first, governance from day one, and expansion funded by savings.

How long did the transformation take in this case?

Four weeks of diagnosis and eight more to put two agents into production: operational results in about 90 days. ROI was achieved in less than six months, and subsequent expansion (reconciliation, incidents, management dashboard) was paid for with the savings from the first phase.

Were there layoffs due to automation?

No. The two planned hires were no longer necessary, two administrative staff moved to sales tasks that previously had no time, and a third became the supervisor of the agents. Announcing from the beginning that freed-up hours were being reinvested was decisive for team collaboration.

Was it necessary to change the ERP or company systems?

No. The agents connected to existing systems via native connectors: they read the order inbox, validate against the catalog and price list, and create drafts in the same ERP as always. Changing systems before automating is, for most SMEs, an expensive and unnecessary delay.