AI agents in operations: promises and first realities
- 1An AI agent is only worth its cost where the handling path genuinely varies; elsewhere, a conventional workflow is more reliable and cheaper.
- 2Reliability erodes with every chained step: eight steps at 95% each give only about two thirds end-to-end success.
- 3The economics turn on supervision cost and data quality, not on model performance alone.
- 4Start with a bounded scope and human validation, then extend autonomy based on measured results.
One word, several realities
The word "agent" now covers very different things. In its strict sense, an AI agent is a system that receives an objective, decides for itself which steps to take, calls tools (email, ERP, document stores), observes the outcome and adjusts. What sets it apart from conventional automation is that the path is not written in advance.
For an executive this distinction matters, because it determines both the promise and the risk. A workflow or software robot follows a defined route: it is predictable, auditable and cheap to run. An agent tolerates the unexpected, at the price of behaviour that is harder to predict. In most SME operations, the majority of cases still belong to the first world, which is good news: most of the value does not require an agent at all.
Where the promise holds
The promise is real on one specific segment: exception handling. Every operational function has a nominal flow that can be automated, and a long tail of atypical situations that occupy skilled people on repetitive work: a mismatch between purchase order and invoice, a poorly worded complaint, an incomplete file to chase, an intervention slot to reschedule.
This is where an agent can contribute: reading a free-text message, retrieving documents across several systems, proposing a resolution and drafting the reply. We have described elsewhere where the lost hours hide in support functions. These rarely measured tasks of searching, cross-checking and chasing are the most natural terrain.
First realities
Reliability compounds, and not in your favour
An agent chains steps, and each step carries its own error rate. Take a deliberately simple assumption: eight successive steps, each succeeding 95% of the time, independently. End-to-end success is then about two thirds. These figures are illustrative, not measured, but the mechanism is general: the longer the autonomy, the more errors accumulate, and the harder they are to spot because they surface late in the process.
The practical consequence is to shorten chains, insert control points, and reserve the agent for the steps where variability justifies it, leaving the rest to rules.
The hidden cost of supervision
Consider a worked example, with its assumptions stated. A company processes 1,500 supplier invoices a month. Assume 15% raise an exception, or 225 cases, and that an employee spends 12 minutes on each, about 45 hours a month. Suppose an agent proposes a correct resolution for half of them, and human validation takes 3 minutes. The saving is 9 minutes on roughly 112 cases: a little under 20 hours a month, which at a loaded hourly cost in the order of CHF 70 is a modest sum over several years.
The calculation is sober, and that is its value. At this volume the gain exists but is small against integration, monitoring and maintenance costs. The trade-off changes at 15,000 invoices a month, or when exceptions take longer to resolve. In most cases, an agent's profitability depends less on model quality than on exception volume and supervision time. For the nominal flow, automating invoice processing with proven methods remains the first lever.
Data and system access are the real work
An agent is only useful if it can read and write in your systems. Yet many SMEs run on imperfect supplier or customer master data, limited interfaces and rules known to a handful of people. A poorly fed agent produces plausible errors, which are more dangerous than obvious ones. Much of the work therefore consists of cleaning reference data, documenting rules and defining precisely what the agent may do: read only, draft, or act.
The same reasoning applies in field operations. In field technician routing, a conventional optimisation algorithm computes the best plan; an agent can usefully be added to handle disruptions (absent customer, missing part, rescheduling), but it does not replace the planning engine.
The regulatory frame, without drama
Two texts frame the discussion. In Switzerland, the revised Federal Act on Data Protection (FADP), in force since 1 September 2023, requires transparency, a clear purpose and information for individuals subject to automated individual decisions. It also bears on supplier choice: an agent that reads customer emails and sends them to a model hosted abroad raises the question of cross-border data disclosure.
In the European Union, the AI Act entered into force on 1 August 2024 with phased application: prohibitions and the AI literacy requirement from February 2025, obligations for general-purpose AI models from August 2025, and most obligations from August 2026. A Swiss company with EU customers or subsidiaries should take note. For most operational uses (sorting mail, matching documents), the risk level is limited, but uses in human resources or in assessing individuals are treated differently. Our advice is to keep a simple register of uses, data processed and the level of autonomy granted.
Implications for leaders
Start from the process, not the technology. Map where exceptions concentrate and how much time they consume. Without that measurement, no serious business case is possible.
Raise autonomy in steps. First an agent that suggests, then one that acts with human validation, then one that acts with sampled control. Each step is earned through observed error rates, not promises.
Bound the scope. One domain, one document type, limited rights, a reversible action. Early successes come from narrow perimeters.
Name an owner. An agent without a business owner drifts. Someone must track quality, cost and incidents, and be able to switch the system off.
Compare with the simple alternative. A rule, a better form or a redesigned upstream step sometimes removes the exception at source. This is often less spectacular and more profitable. The same logic applies on the commercial side, for instance in handling enquiries and quotes, which we address in our Growth & commercial performance expertise.
To place these choices in a wider approach, our page on Operations & processes expertise explains how we link strategy, deployment and steering of results, with AI as a lever where it helps and organisational common sense elsewhere.
If you want to know whether an agent has a place in your operations, and under what conditions, the starting point is a factual assessment: our diagnostic identifies within a few weeks where the costly exceptions sit and which response, technological or not, is the most proportionate.
A personalised first reading in three minutes.