AI agents in SMEs: what they actually solve, and what's just noise

AI agents in industrial SMEs solve, today and concretely, three classes of work: reading and matching non-standardised documents (invoices, delivery notes, contracts), filtering public tenders from BOE and regional platforms, and reconciling systems that don’t talk to each other (ERP, bank, accounting). They don’t solve, yet, operating without supervision, replacing entire functions, or governing themselves. Much of what gets sold as an agent in 2026 isn’t one: a single LLM call with a fancy prompt is a script, a chatbot over a FAQ is information retrieval, and an RPA with an LLM on top is deterministic automation with a touch of judgement. The rule is direct: if there’s no decision, no real action on tools, or no loop that adjusts the next decision based on the previous result, it’s not an agent. And the order matters: agents are step 5 of the Zero Friction method, not step 1.

There’s a word that has exploded this quarter in every pitch, every keynote and every deck landing on the desk of an SME general manager: AI agents. Tier-1 consultancies have shifted from talking about “digital transformation” to talking about “autonomous agents”. Webinars are full. The questions we get from the market have changed: it’s no longer “what is ChatGPT?”, it’s now “how do I govern five agents that interact with each other?”.

And most industrial SMEs that call us don’t quite know what an AI agent is. Which is reasonable, because the word is being used for three different things, two of which aren’t agents.

We’ve been deploying systems that actually are agents in Spanish SMEs for months. Here’s what we see.

What an AI agent really is

Without jargon: an AI agent is a system that makes a decision, takes an action and reviews the outcome, repeating that loop within defined boundaries. Three elements that all three have to be present:

Decision. The system chooses between options, not just executing what you tell it.
Action. The system interacts with real tools (an ERP, a database, an API, an email).
Loop. The result of the action shapes the next decision.

If one of the three is missing, it’s not an agent. It’s something else, perfectly useful, but something else.

Three things AI agents actually solve in industrial SMEs

Here are real cases we have in production or in advanced testing.

One. Reading and matching non-standardised documents. Supplier invoices in completely different formats. Delivery notes arriving as PDFs, as images, scanned from the carrier’s phone. Commercial contracts with clauses that change from supplier to supplier.

At an industrial distributor we moved from 3 to 5 manual minutes per invoice to a system that extracts the fields, matches them against the purchase order in the ERP, flags discrepancies in amount or quantity, and leaves the admin as the auditor of 100% of the flow instead of the typist of 60%. 4 hours per batch became 12 minutes. The agent isn’t the OCR: the OCR does the reading. The agent is what decides what to do with each anomaly and how to escalate it.

Two. Searching and filtering public tenders. The BOE, regional platforms and the OJEU publish thousands of tenders per week. Filtering the relevant ones for a specific SME, downloading the specifications, reading the hard clauses (solvency requirements, deadlines, place of execution), and discarding what doesn’t fit, is several hours of weekly work that often nobody has time for.

An agent can crawl the sources every night, apply the semantic filter (“equipment compatible with our industrial pump catalog”), download the specs, read the pages that matter, and leave in the sales rep’s inbox three real opportunities instead of four hundred links. The decision to bid still belongs to the sales rep. The screening work no longer does.

Three. Reconciliation between systems. The ERP says one thing, the bank statement says another, the scanned invoices say a third. Squaring the three at month-end is the silent work that pays for overtime in most SMEs.

An agent with read access to the three sources can crawl them, identify discrepancies, group them by probable cause (value-date difference, cents-level error, missing invoice, receipt without invoice) and hand the finance team a prioritised list instead of an 800-row spreadsheet. The decision on how to close each discrepancy is human. The cognitive load of detecting them no longer is.

These three cases share something: the agent does the repetitive work of detection and filtering, not decisions that commit the business.

Three things sold as agents that aren’t

This is where the noise confuses SMEs evaluating vendors.

One. A single LLM call with an elaborate prompt. If the system receives an input, sends a single request to the model, and returns the answer, it isn’t an agent. It’s a script on steroids. Useful, sometimes necessary, but not an agent.

Two. A conversational assistant that searches a FAQ database. A chatbot that answers questions by reading from a PDF uploaded to the system is information retrieval, not an agent. It decides nothing. It acts on nothing. It’s good for reducing tickets to the support team. Not for governing operations.

Three. RPA with an LLM on top. Classic click automation (fill the form, copy from A to B, send the email) with a model that decides which of three routes to follow is not an agent in the full sense. It’s deterministic automation with a small layer of judgement. The distinction matters because the promise of “autonomous agents” doesn’t hold when 95% of the flow is hard-wired by hand.

Three things the hype attributes to agents that aren’t true yet

One. “They learn the process on their own.” No, without explicit rules they don’t learn. An agent can adjust parameters within a defined margin, but real learning requires human supervision, labelled data and review cycles. Whoever sells “the agent learns alone and improves over time” is selling a promise that breaks in the first month.

Two. “They replace an entire function.” Also no. An agent replaces tasks, not functions. The admin function doesn’t disappear because you’ve put an agent on invoice reading. The admin moves from typing to auditing, and her judgement is still the critical filter. If anyone sells you “saving two people from the team”, the project is poorly framed before it starts.

Three. “They govern themselves.” No. Governance is set by the client team, with metrics, thresholds and exception criteria defined before deployment. An agent without human governance is a black box that, the day it makes a serious mistake, nobody can explain. And that day arrives.

Why the order matters: agents are step 5, not step 1

In the method we apply in every project there are five steps in this strict order: eliminate, standardise, simplify, automate, augment with AI. Agents live in step 5.

Skipping to step 5 without doing the previous ones is what produces the endless pilots in the market. If the process you’re going to “agent” has six steps that shouldn’t exist (step 1, eliminate), if the data lives in four different spreadsheets with the same information (step 2, standardise), if approvals take three days because of a mail that gets lost (step 3, simplify), or if the flow isn’t even deterministically automated (step 4, automate), putting an agent on top multiplies chaos instead of reducing it.

The order isn’t a methodological suggestion. It’s the difference between a system that works and a pilot that stalls for three quarters.

What to ask the vendor selling you agents

If you’re evaluating a proposal that puts the word “agent” in every slide, three questions that save weeks of project.

One. What specific decisions will this agent make, and within what boundaries? If the answer is generic (“optimises the process”, “manages the flow”), there’s no agent. There’s a script with marketing.

Two. What happens when the agent gets it wrong? Who detects it and how long does it take? If there’s no concrete error-detection mechanism and a human with authority to correct it, it’s not an agent deployable in production. It’s a time bomb.

Three. What have we eliminated, standardised, simplified and automated in the process before adding the agent? If the answer is “nothing, the agent solves everything”, the project is going to fail. Not because of the model. Because of the design.

The summary, without noise

AI agents are a real tool that solves concrete classes of work: detection, filtering and matching of information across heterogeneous sources. They work well when the underlying process is clean and the decision they take is bounded. They don’t replace teams. They don’t learn alone. They don’t govern themselves.

If an industrial SME is looking at where to put agents, the operational answer almost always starts with something more boring than the word agent: clean the process first, automate the deterministic part next, and reserve agents for the layer where bounded judgement adds value. Step 5, not step 1.

If your company is evaluating a project that includes AI agents and you want an honest second read before signing, the 2-week operations roadmap leaves it written down which steps are agents, which steps are classic automation, and which steps shouldn’t be touched yet.