The bridge applied to SMEs

The owner of a small distributor comes to the meeting with an idea he has heard many times in recent months: he wants artificial intelligence to anticipate which customers are about to stop buying. One vendor showed him a convincing demo; another warned him that whoever does not adopt soon will be left behind. He does not want to miss the opportunity, but he cannot fund an open-ended exploration either. Every investment competes with stock, vehicles, wages and credit.

In the conversation, the data shows up. Sales live in the invoicing system; complaints, in messages; sales visits, in each rep’s memory. “Active customer” does not mean the same thing to accounting as it does to sales. Some branches record a customer as lost after thirty days without a purchase; others wait ninety. The model he is asking for could be built. The question is what it would learn.

Many SMEs stand in that same place with respect to AI. They face intense pressure to adopt and have few buffers if adoption goes wrong. A large company can run several pilots at once, hire specialized teams and wait for an investment to pay off. In an SME, a failed project can tie up months of cash and wear down the team’s trust for years.

That is why the bridge does not start from what technology has to offer, but from protecting the capacity to decide. Before asking which model to use, it asks which problem deserves solving, what data describes it and what improvement would be valuable enough to justify the risk.

A technology that amplifies differences

AI does not benefit every person or every organization equally. An experiment with 640 entrepreneurs in Kenya offers a useful warning: generative assistance improved the performance of those who were already doing well by around 15% and lowered that of those starting from weaker results by about 8%.¹ The latter tended to follow generic advice on tasks where it took judgment to see that the advice did not apply.

The study cannot predict what will happen in an Argentine company, but it does illustrate a mechanism. When the tool offers a plausible answer, the prior capability of whoever receives it determines how much they can make of it and when they need to discard it. AI brings knowledge closer, and it can also lend confidence to a mediocre recommendation. Without context or judgment, it widens the gap that already existed. → Authors and currents.

That asymmetry weighs most on the SME that most needs help. The organization with orderly data, known processes and people able to review gets more value at lower risk. The one working under urgency, relying on someone’s personal spreadsheet and with no way of checking a recommendation receives the same advice to use the same technology, as if the tool were landing on neutral ground.

Thinking of AI as a multiplier with a sign helps avoid easy promises. The model’s power multiplies whatever it finds: order or confusion, judgment or dependence, a good definition or a repeated error. The sign does not come with the license; it is decided in the diagnosis, in the data and in the design of how the tool is used. → Concepts of our own.

The regional gap, up close

Adoption also takes place amid regional inequality. CEPAL estimated AI penetration in Latin America in 2024 at below 4%, against more than 20% in Europe. In Brazil, 41% of large companies reported using AI; among SMEs, only 11%.² The figures do not describe every sector, but they show that the distance between companies within the same country can matter as much as the distance between regions.

The nodo nadIA survey of 402 Argentine SMEs in 2025 adds a valuable nuance. 41.6% already used at least one AI tool, mostly for basic functions such as generating text or code, but governance and internal-capability indicators were low.³ The picture is no longer that of an SME unaware of the technology: many people have already tried it and taken it up informally. Visible adoption arrived before any agreements on data, risk, learning and responsibility.

The Latin American Artificial Intelligence Index (ILIA), by CEPAL and CENIA, makes it possible to place these experiences within broader national ecosystems.⁴ It classifies nineteen countries as pioneers, adopters and explorers based on their infrastructure, talent, research and governance. For an SME, those conditions are not an abstract backdrop. They shape connectivity, the supply of professionals, financing and the quality of the support within reach.

Talking about a gap, then, is not the same as blaming a company for arriving late. Adoption depends on resources no SME produces on its own. Even so, a difficult context does not take away its room to decide; it makes that room more important. With scarce capital, choosing a bounded problem and building capability around it is more sensible than imitating a large firm’s strategy.

The triple constraint of emerging economies

In emerging economies, three constraints tend to coincide. The first is limited infrastructure: unstable connectivity, old equipment, little redundancy and vendors that do not always offer the same service away from the big centers. The second is scarce talent, with one or two people concentrating systems, support, data and security. The third is cultural: leadership says it wants to innovate, but operations have no time to experiment and every mistake is punished.

Shokrollahi Yancheshmeh (2026) and earlier work on digital transformation in emerging economies help explain how these constraints reinforce one another. A poor connection makes the tool less reliable; the lack of talent drags out every incident; fear of getting it wrong keeps the team from flagging problems early. What looked like a software purchase ends up colliding with an entire configuration.

Not every constraint can be solved within the same project, and that is exactly why they need to be identified. If connectivity rules out a permanently online operation, the architecture has to tolerate outages. If a single person knows every integration, the priority may be training a backup and documenting. If the culture punishes trial, a small, reversible pilot needs an explicit agreement on what will be learned from mistakes. These constraints are design conditions: they should be shown to the vendor, not hidden away as something shameful.

The opportunity inside scarcity

Many SMEs come to this conversation feeling they have almost nothing. They work with spreadsheets, with applications that do not talk to each other and with decisions that come out of a chat. That scarcity is a risk, but it also holds an opportunity: there is less inherited architecture to dismantle.

A large organization may have data teams and, at the same time, carry twenty systems that define the customer in incompatible ways. The SME, by contrast, can start with modest tools and build a common definition from the beginning. Its advantage is being able to avoid part of the accumulated fragmentation.

Nor should the situation be romanticized. Starting from little digitalization concentrates the work, because it forces the technical base and the human habit to be built at the same time: ordering a source, agreeing on who maintains it, integrating the flow and learning to decide with the new data. The absence of systems does not remove complexity; it only makes it less visible.

That is why the road to AI usually starts far from AI. First, the islands that already exist are connected. Then definitions and owners are stabilized. Only when the data is reliable enough does anyone assess which prediction or automation adds value. Skipping those steps speeds nothing up: it piles up a debt that surfaces when the model reaches production. → from silo to architecture in Concepts of our own.

The sequence, however, is not a rigid ladder. SMEs move forward recursively and episodically: they order a process, slip back when a key person leaves and pick up again from another point. Amanollahnejad and colleagues observed that dynamic in 27 UK SMEs and described it through a sociotechnical lens.⁵ The context is different, but the mechanism is recognizable: adoption depends on sponsorship, capability, data and continuity, and it can come undone.

Measuring maturity is still useful if it is understood as a snapshot that orients, with no guarantee of irreversible progress. IMIA shows where there is an imbalance; sociotechnical hygiene makes it possible to see whether the capability acquired holds over time.

Listening to the request without obeying it to the letter

The SME often arrives with the solution already built into the sentence: “I need a chatbot”, “I want to predict sales”, “we have to automate the back office”. Taking that request seriously does not mean carrying it out exactly as phrased. It means finding out what experience produced it.

“I need a dashboard” may mean the owner receives three versions of the sales figure and no longer trusts any of them. “I want AI to answer customers” may hide growing demand and a catalog that changes without the information reaching the people who serve customers. “I need to automate invoices” may come from someone who spends two afternoons a week copying data from one system to another.

Each sentence opens a different problem. In the first case, the priority is semantics and data governance. In the second, institutional memory may be needed before a conversational model. In the third, a conventional integration may solve more than an AI solution, at lower cost and lower risk.

The diagnosis allows three answers. There are places where the technology pays off now; others where the answer is not yet, because a precondition has to be put in order first; and cases where it destroys value. “Not yet” does not reject innovation: it keeps the available budget from being spent on a project whose premise does not hold.

Choosing a first move

The scale of the first project matters. It has to be small enough to limit the damage and real enough to show how the organization actually works. An isolated demo with prepared data proves little. An everyday, bounded and reversible workflow makes it possible to observe adoption, exceptions and maintenance.

Freeing up repetitive time is usually a good way in, as long as the full cycle is measured. If an automation saves effort but adds review, reconciliation or correction, the benefit may be smaller than expected. The baseline has to include both the old effort and the new one.

Another fertile entry point is improving a recurring decision. Unifying the information a supervisor uses to prioritize complaints can deliver value before any prediction. Once the team trusts that descriptive view and learns to record its reasons, the context for more advanced help is in place.

The first result has to fit the company’s own language: hours recovered, errors avoided, faster responses, fewer arguments about which number is the real one. “Model deployed” does not pay an installment or improve anyone’s workday. The artifact is the means; the result is the change that can be observed.

From the start, it also has to be agreed what would force a stop. If rework goes up, if the team goes back to the spreadsheet or if sensitive data shows up in an unauthorized tool, the pilot needs review. A small project is not permission to govern it later.

Honesty criterion. The value of a project in an SME is not measured in the model left running or the dashboard switched on, but in the person: an hour recovered, a decision that comes out better, a piece of data that can finally be trusted. When the metric shifts from the artifact to the person, it becomes impossible to confuse activity with result —and almost all the projects that do not pay off live in that confusion—.

The bricolage that keeps everything running

SMEs survive on ingenuity. They export data, fix a column, reload it, call someone who remembers an exception. This sociotechnical bricolage is often described as disorder, but it is frequently an intelligent response to tools that do not fit.

Before replacing those moves, the diagnosis takes stock of them. A manual export may be pure waste, or it may be the only moment when someone catches an inconsistency. Automating without looking removes that control as well. The redesign has to separate the repetitive task from the judgment it was quietly carrying.

The “coffee economy” belongs to the same reality. Some important decisions are made through relationships, in a brief conversation between people who know each other. That informality can exclude and leave no trace, but it also transmits context with great efficiency. Standardizing it completely can impoverish the decision. The goal is to keep the reasoning and make it accessible, without forcing every human interaction into the shape of a form.

Duplication is another sign. During a transition, people fill in the new system and keep using the old one, because they do not trust it or because no one decided when to drop it. If that stage has no date, no owners and no exit criterion, the project adds work in the name of efficiency. Tuning the model does not fix it. Clear roles do, along with the decision to shut down the old flow once the new one proves it can carry the operation.

The champion and the day after

Many transformations get going thanks to one persistent person. It may be the owner, a curious administrative employee or someone in IT who connects areas and finds time. That champion is valuable and, at the same time, a sign of fragility.

If every decision, permission and memory runs through that person, one absence is enough to stall the project. Amanollahnejad and colleagues speak of episodic leadership: adoption advances while there is momentum and slips back when attention shifts or the person leaves.

That is why the method asks early on who can take over. The point is to keep the organization’s capability from being confused with one individual, without diluting responsibility. Core decisions have to leave a record, at least one other person needs to understand the flow, and review routines must become part of the everyday agenda.

Transfer also has to be adjusted to turnover. A long training course delivered once may fail in a company with seasonal employment. What works better there are short walkthroughs, help built into the workflow and practices people can teach each other without distorting the judgment.

Distribuidora Norte tries again

Synthetic case. Rebuilt from patterns observed in the field; it does not correspond to an identifiable client. It continues the story begun in The four links.

A year after the at-risk customer alerts, Distribuidora Norte came back with another problem. Deliveries were running late, complaints did not match the routing dashboard and drivers kept spreadsheets the official system did not reflect. The owner saw an opportunity to optimize delivery windows with AI.

The first technical conversation could have started with routing algorithms, but the diagnosis found a scar first. The company had already tried a similar solution; that pilot had been presented as the users’ failure, and no one mentioned it in the first meeting. The drivers remembered a different story: the system did not account for access points, informal schedules or arrangements with customers that they sorted out during the route.

The parallel spreadsheets held part of that knowledge. There was some resistance in them, but they were also an attempt to keep the operation going while the official dashboard showed an incomplete picture. Automating on top of that data would have repeated the failure with a more powerful tool.

The intervention began by unifying complaints and deliveries. Supervisors and drivers took part in defining “on-time delivery”, which varied by zone and type of customer. A view was co-developed that let people record exceptions without forcing anyone to write a full story at every stop. Only then was prediction added, at the points where the team could judge whether it helped.

What mattered was not an isolated precision percentage but the sequence: understand the scar, recover the knowledge behind the workarounds, agree on the data, build trust and only then automate. The same person, or the same team with memory, stayed with the whole journey so the hypothesis would not be lost in a handover.

IMIA can record the change in configuration: better data quality, more integration, more governance. Value metrics, by contrast, have to follow what happens to the supervisor and the driver: whether they decide sooner, whether rework goes down and whether the tool helps them or watches them. Both readings are needed, because an SME can improve its architecture without having recovered trust yet, or get excited about a pilot without having built the capability to sustain it.

Growing without copying

An SME is not a large company in miniature. Its relationships, decision times and dependencies are different. A practice designed for a corporation can load it with committees, licenses and roles it cannot sustain. Conversely, the closeness between those who decide and those who do the work allows quick adjustments and conversations that a larger structure would slow down.

Design has to take advantage of that closeness without turning it into permanent informality. Minimal governance can consist of a few clear rules, identified owners and a monthly review. A modular architecture can grow without demanding a complete platform from day one. A vendor can contribute capability without keeping the only copy of the data or the knowledge needed to operate.

Every SME that builds that capability closes part of the region’s productive gap. The value goes beyond individual gain: more companies able to use data and technology with judgment make up a productive fabric less dependent on imported solutions and on promises that ignore its scale. → closing the gap in Concepts of our own.

Still, responsibility starts with the concrete case. Someone risking their savings does not need to be talked into riding the latest wave. They need to know whether this intervention gives back something their organization can sustain. The bridge applied to the SME does not promise to put AI into every process. It promises something harder: recognizing where it helps, building from what already exists and having the discipline to say “not yet” when the model would multiply the wrong side.

See also: The bridge applied to the public sector · IMIA, the maturity instrument · The four links · Authors and currents · Bibliography