On June 25, 2025, Gartner announced that more than 40% of agentic AI projects launched inside companies would be abandoned by the end of 2027, for lack of a clear strategy beyond the prototype. That figure says nothing against artificial intelligence: it says that a lot of organizations pick their automation tool out of enthusiasm rather than diagnosis. For an SME leader looking at their stack of automations (invoicing, client onboarding, follow-ups), the useful question isn’t “should I use AI?” but “does this process need to stay predictable?”
Two families of automation, one fundamental difference
An automation platform like Zapier or Make runs on a simple principle, spelled out in their official help docs: a trigger (an email received, a row added to a spreadsheet) launches one or more actions written in advance. If a case doesn’t match what was planned for, the automation stops or fails visibly, which is actually a strength: anyone on the team can reopen the chain, spot the step that’s stuck, and fix it without having written the original recipe.
An AI agent runs on a different logic. Anthropic, whose reference framework on this topic is widely echoed across the industry, tells the two categories apart by the degree of autonomy left to the system: a workflow orchestrates tools via predefined code paths, while an agent dynamically directs its own process and how it uses its tools (Building Effective Agents, Anthropic). Put plainly, in the first case you wrote the itinerary; in the second, you described the destination and let the system pick the route. The unplanned path of the second can impress on a prototype tested a dozen times, and turn into an unpleasant surprise the day real-world volume outpaces what was ever tried.
Take a follow-up on an unpaid invoice. A deterministic scenario watches the due date and sends an email at day 7, 15, and 30: if the client responds outside that script, the chain stops and someone steps in, with no ambiguity about where it got stuck. An agent tasked with “following up intelligently” can adapt tone and channel to each client’s history, which looks great on a ten-profile demo. The day the portfolio holds hundreds of clients, nobody can reliably say anymore why a given follow-up went out twice or not at all, short of reopening the model’s decisions one by one.
Four criteria that sort the workflow from the agent, process by process
The choice isn’t made at the level of the whole company but process by process: invoicing and client onboarding don’t necessarily call for the same answer. Four criteria come up, in the order they should be asked.
- Someone other than the person who built it will eventually need to maintain this automation: a deterministic workflow can be read and fixed without having written the original line, an agent is rarely debugged without the skill of whoever designed it.
- The process touches payment, contracts or compliance: a duplicated invoice reminder or a badly filled contract field is costly precisely where the company has the least room for the unexpected.
- The connections already exist as standard between the CRM, email and accounting: a platform like Zapier or Make already covers this ground, which makes the agent redundant, not more capable.
- The logic is genuinely specific to the business, tolerates iteration, and the technical skill to evolve it exists in-house: this is the only case where the agent becomes worthwhile, precisely because someone will be able to step in when it derails.
Three criteria out of four point toward the deterministic option. That’s no accident: most of an SME’s recurring processes, invoicing, onboarding, follow-ups, are exactly the ones where predictability is worth more than flexibility.
The trap few leaders see coming: the gap widens at debugging time
An AI agent looks approachable to anyone on the team: you talk to it, describe an outcome, and it executes. That ease of use says nothing about ease of maintenance, and the gap between the two is exactly where SME automation projects get stuck. Anthropic says as much plainly in its own guide: the autonomous nature of agents drives higher costs and a risk of errors compounding from one step to the next, which is why it recommends sandbox testing and proper guardrails before any real production deployment.
Gartner documents the same gap from the market side, under a different name: analyst Anushree Verma calls it “agent washing”, older tools (chatbots, classic automations) rebranded as agentic AI with no real change in autonomy. A Gartner survey from January 2025 found that only 19% of companies surveyed had invested significantly in agentic AI, with 42% staying cautious: a sign that the ground, once you step past the demo, remains uncertain for most.
None of this is a case against AI: an agent chosen well, on the right process, with the skill in place to catch it when it drifts, delivers on its promise. What derails an automation project is often the order of the questions, more than the technology itself: picking the agent because it’s new, then hunting for a process to fit it into, reverses the sensible order of the exercise.
The two families coexist in automation that holds up
An SME that automates well doesn’t pick a side once and for all. It keeps the deterministic approach for invoicing and follow-ups, precisely because technology alone isn’t enough to secure a critical process: it also needs anyone on the team to be able to pick it back up. It reserves the agent for the pockets where the logic is too fluid for a path fixed in advance, and only there if the skill to maintain it already exists in-house or is deliberately built before deployment, not after the first incident.
On client onboarding, the line is directly testable. If the journey is identical for the large majority of new clients (sending the contract, creating the account, welcome email), a workflow is enough and fixes in one click when a step changes. If every client arrives with a configuration need different enough that no fixed recipe survives more than three cases in a row, the logic becomes specific enough to justify an agent, provided someone in-house can read what it decided to do.
The real work, then, isn’t choosing between workflow and agent, but mapping your processes by their tolerance for the unexpected, then assigning the tool that fits each one. One or two mismatched automations (an agent on an invoice reminder, a rigid chain on a negotiation that changes with every client) cost less to fix now than after the incident. If one of your processes seems misclassified, we can talk it through.
Photo credits: StartupStockPhotos, cloudhoreca, RonaldCandonga (Pixabay).
