- Is it really worth letting AI agents run unsupervised in your company?
- What did Monks actually do, and with what concrete results?
- How does the team structure change when agents work autonomously?
- How does it work in practice for an Italian company: three scenarios compared?
- How do you unify the AI budget with the staff budget: what does Monks do, and what can you do?
- Which AI tools and models can you use to replicate this approach?
- What risks and rules do you need to manage before letting an agent run on its own?
- Why are people talking about this now and not a year from now?
- How do you get started: the steps to introduce autonomous agents in your company?
- What mistakes do companies make when they approach AI agents for the first time?
- How can you tell if the model is working for your company?
- So, is it worth unifying budget and team on the Monks model for your company?
- How do we see it at SHM?
- Frequently Asked Questions
- Monks (S4 Capital) runs AI agents for 36-48 hours with no supervision on coding, social and performance marketing, with plans to extend this to clients.
- The budget for AI tokens has moved into the "talent budget": it's no longer a separate IT line item, but competes with salaries for the same resources.
- You need an agent register, governance rules and human roles who select and steer the output, not just approve it.
- In Italy only 8% of large companies are already experimenting with Agentic Automation, but 25% plan to do so soon: those who start first build a real advantage.
An AI agent working 36-48 hours straight with nobody watching: it's not science fiction, it has already happened in production. The question for your company isn't "when will it arrive," but "how do I organize now that it's already here."
Is it really worth letting AI agents run unsupervised in your company?
It depends on how measurable the work is and how clear the criteria are. Where there are precise requirements and objective feedback loops, autonomous agents cut costs and time; where human judgment is still irreplaceable, supervision remains necessary.
Monks, the operating division of S4 Capital, ran AI agents for 36-48 consecutive hours on coding tasks with nobody checking on them. The result: no significant errors, and speed far above human work. The news, reported by Digiday, raises a concrete question: which activities in your company have the same characteristics?
The answer depends on three factors: how precise the requirements you can write are, whether there is a feedback loop (an objective signal that tells the agent whether it's on the right track), and your risk tolerance if something goes wrong.
| Your situation | Agent autonomy | why |
|---|---|---|
| Work with precise rules and verifiable output (e.g. code, reports) | Yes | The agent can work for hours or days; the requirements act as guardrails |
| Work with KPIs measurable in real time (e.g. social, ads) | Yes | Performance data is the feedback loop that guides the agent |
| Creative work with qualitative judgment (e.g. copy, brand) | Partial | The agent generates options, but an expert is needed to select |
| Work with high reputational or legal impact | No, for now | The risk of unsupervised error is too high |
What did Monks actually do, and with what concrete results?
Monks moved coding first, then social and performance marketing, onto autonomous agents that coordinate with each other, test their own work and correct themselves without human intervention. The financial numbers confirm the model works.
The starting point was code: precise, verifiable, with requirements that can be written sharply. The agents receive the specs, divide the work, test the code they produce against the original requirements and check each other. Until a few months ago, this coordination required instructions written by hand by developers. Today the models do it on their own.
The next step is social and performance marketing. Here the agents wait for an event — a trend taking off, a wave of comments, a post losing traction — and act. Engagement data replaces written requirements as the control signal.
S4 Capital's financials tell the same story from another angle. The 2025 annual report shows net revenue of £673 million, down 10.8%, but with an operating margin up to 12.1% (+70 basis points versus 2024): less revenue, more efficiency. Labor costs fell because agents now do part of the work.
On the creative side, the Monks.Flow platform produced 460 personalized assets for 20 use cases in a single holiday campaign, with a 13% lower cost per acquisition and a 62% higher conversion rate than the control.
How does the team structure change when agents work autonomously?
The work doesn't disappear: it shifts. People stop carrying out repetitive tasks and become responsible for writing requirements, selecting the best output and deciding the direction. It's a change of role, not of headcount — at least in the early stage.
Monks is facing an unexpected problem: agents produce 50 ideas in 10 minutes, all consistent with the brief. Asking for something bolder generates another 50, but only five are really good. Finding those five takes time and expertise. The creative team's job is no longer producing ideas: it's recognizing the ones worth keeping.
The same goes for work rhythms. Agile sprints, two-week cycles with daily meetings, have shrunk to two days. The risk is that people stop thinking and simply sign off. Monks is considering reintroducing deliberate friction to force the team to truly process the work, instead of delegating everything to the agent.
- 1 · Strategist / PMWrites the requirementsDefines goals, constraints and quality criteria before the agent starts. The quality of the prompt determines the quality of the output.
- 2 · AI agentExecutes and self-checksWorks autonomously for hours or days, tests its own output against the requirements, and coordinates other agents if needed.
- 3 · Subject matter expertSelects and guidesChooses among the options produced, flags directions worth exploring, and spots the errors the requirements didn't anticipate.
- 4 · Team governance / Tech councilMonitors costs and qualityKeeps a register of active agents, tracks token spend, and assigns the right models to the right tasks.
- 5 · ManagementDecides and answers for resultsHolds final responsibility: an agent's autonomy doesn't mean a lack of accountability.
How does it work in practice for an Italian company: three scenarios compared?
An e-commerce business, a communications studio and a manufacturing company have different needs, but all three can apply the "unified talent budget" logic, starting with the most structured activities.
The three scenarios below are illustrative: the numbers show the logic of the calculation and are not data from real clients.
| Company type | Task that can be delegated to the agent | Estimated savings (hours/month) | Main risk |
|---|---|---|---|
| E-commerce (20 people) | Replying to comments, optimizing product pages, ads reports | 40-60 hours | Tone not aligned with the brand if requirements are vague |
| Communications studio (8 people) | Copy drafts, A/B variants, creative briefs from client briefs | 30-50 hours | Selecting the output requires a dedicated senior person |
| Manufacturing company (50 people) | Internal reports, technical documentation updates, data analysis | 60-80 hours | Sensitive data: you need an enterprise contract with the AI provider |
The critical variable in all three cases is the same: the quality of the requirements written before the agent starts. An agent with vague instructions produces vague output. Investing time in the setup phase, what Monks calls "decent prompting," is the highest-impact lever you have.
How do you unify the AI budget with the staff budget: what does Monks do, and what can you do?
Monks stopped treating AI tokens as an IT line item and moved them into the "talent budget," the same bucket as salaries. The logic is simple: if an agent does the work a person used to do, it competes for the same resources.
Here's the paradigm shift: until recently, the cost of AI tools sat in the software or IT budget. Now that agents replace hours of human work, putting them in a separate line distorts how you read your costs. Monks moved them into the talent budget, the budget dedicated to people, because that's where they really compete.
The practical problem is that teams keep asking for more tokens (the processing units AI models consume for every task). Monks is measuring its needs for 2027 and has created an internal "tech council" that assigns the most expensive models only to tasks that truly require them. For everything else, it uses cheaper models. This is exactly the kind of optimization you can replicate even with small teams, as explained in the article on how to cut token costs by choosing the right model for each task.
| Cost item | Where it used to sit | Where it goes now (Monks model) | why |
|---|---|---|---|
| AI tokens (processing) | IT / software budget | Talent budget | Competes with salaries for the same tasks |
| Premium AI models | Centralized IT purchase | Assigned by task by the tech council | Avoids waste: not every task needs the best model |
| AI training | HR training budget | Talent budget or operating budget | It's part of the cost of keeping people productive |
| Agent supervision | Didn't exist | Talent budget (senior hours) | The time of whoever selects the output is a real cost |
For an Italian SME, the practical step is simpler than it looks: just add a line to your operating cost plan that compares the monthly cost of tokens with the hours of work they replace. If the ratio is favorable, the budget justifies itself. If it isn't, you're using the wrong model or the wrong task.
Which AI tools and models can you use to replicate this approach?
There's no single platform: the choice depends on the type of task, your budget and the level of control you want to keep. The key distinction is between "always-on" models and "event-driven" agents that activate only when something happens.
| Agent type | When to use it | Example tasks | Relative cost |
|---|---|---|---|
| Continuous-run agent | Long, structured tasks (coding, reports, documentation) | Software development, data analysis, catalog updates | High (many tokens per run) |
| Event-driven agent | Responds to external triggers (trends, comments, KPIs out of range) | Social monitoring, ads optimization, customer care | Medium (activates only when needed) |
| Review agent | Checks the output of other agents | Code QA, fact-checking, brand consistency | Low (short, targeted tasks) |
| Lightweight (low-cost) model | Simple, repetitive tasks | Classification, tagging, short summaries | Very low |
The logic of Monks' tech council, assigning the right model to the right task, can be replicated even without a dedicated team. All you need is a simple matrix: for each process you want to automate, ask how critical an error would be and how complex the required reasoning is. Only high-complexity, high-risk tasks deserve the most powerful and expensive models.
To understand how models are evolving in this direction, it's worth reading our analysis of what the new 2026 Gemini models mean for SMEs and our piece on how AI agents are evolving beyond interfaces, according to OpenAI.
What risks and rules do you need to manage before letting an agent run on its own?
Monks' clients are already asking which models work for them, where they're hosted and whether an enterprise contract stands behind them. Agent governance isn't a technical detail: it's a contractual requirement.
When a Monks client asked about the agents working for them, the questions were precise: which models, where do they come from, is there an MSA (Master Service Agreement, a framework contract) covering them? Nobody wants unknown models processing company data without a clear contractual basis.
Agent governance is closely tied to the issue of content published without supervision, a risk we covered in detail in our article on AI agents and images published without consent. The basic rule: the more autonomy you give the agent, the more robust the control system you build around it needs to be.
| Risk | Probability | Impact | How to mitigate |
|---|---|---|---|
| Output that is off-brand | High if requirements are vague | Medium | Detailed requirements + sample review |
| Token spend out of control | Medium | Medium-high | Monthly budget per agent + automatic alerts |
| Sensitive data exposed to the model | Low with an enterprise contract | High | Enterprise contract with the provider, no unverified models |
| Errors going undetected in production | Low with review agents | High | QA agent that checks the output before publication |
| Loss of in-house expertise | Medium in the long term | High | Keep deliberate friction in the creative process |
Why are people talking about this now and not a year from now?
The Italian AI market grew 50% in 2025, and Agentic Automation (automation based on agents that act autonomously) is already being tested by 8% of large Italian companies. Wait another year and you start behind.
The Italian artificial intelligence market reached €1.8 billion in 2025, up 50% from the previous year, according to the Osservatori Digital Innovation of Politecnico di Milano. This is no niche market: it has become structural.
On autonomous agents, data from the Osservatorio Intelligent Business Process Automation of Politecnico di Milano show that 8% of large Italian companies launched Agentic Automation pilots in 2025, and 25% plan to do so soon. The gap between large companies and SMEs is huge, though: only 8% of small and medium-sized businesses have launched even a single AI project, versus 71% of large ones, according to data published by AI4Business.
The clearest signal comes from the job market: according to data reported by InnovationCity, job postings requiring AI skills grew 93% in Italy in 2025. Companies are no longer looking for people who can use a tool. They want people who can build and govern processes with agents.
How do you get started: the steps to introduce autonomous agents in your company?
Always start with a process that is already structured, with requirements you can write down and an objective quality signal. Don't start with creativity or high-risk processes.
- 1 · Step 1Map the candidate processesList the repetitive tasks with clear rules and verifiable output. Coding, reports, replies to standard comments, product sheet updates.
- 2 · Step 2Write requirements as if a new colleague were reading themThe agent doesn't guess: it executes. The more precise the requirements, the better the output. Test the prompt on real tasks before automating.
- 3 · Step 3Define the feedback loopIdentify the signal that tells the agent whether it's on track: performance KPIs, automated tests, quality checklists.
- 4 · Step 4Create a registry of active agentsKeep track of which agents are running, on which tasks, with which model and at what cost. It's your minimum 'tech council'.
- 5 · Step 5Put token costs in the operating budgetCompare the monthly token cost with the work hours they replace. Update the estimate every quarter.
- 6 · Step 6Keep deliberate friction on creative tasksDon't delegate everything. Ask your team to think before validating the agent's output. Human judgment is the value that can't be automated.
What mistakes do companies make when they approach AI agents for the first time?
The most common mistake is starting with the wrong processes: creative or high-risk ones, instead of structured, measurable ones. The second is not building governance before the problem shows up.
Monks learned the hard way that asking the agent for something 'bolder' produces a lot of noise and little signal. Fifty interesting ideas on paper, but only five really good ones, and finding those five takes more time than you save. This is what happens when you use the agent for tasks that require qualitative judgment without a dedicated expert to do the selecting.
| Error | Consequence | How to avoid it |
|---|---|---|
| Starting with creative or high-risk tasks | Unreliable output, loss of trust in the tool | Start with coding, reports, structured tasks |
| Vague or incomplete requirements | Generic output that saves no time | Write requirements like a detailed brief |
| No registry of active agents | Token spend out of control, no visibility | Create a minimal tracking sheet from day 1 |
| Token budget separate from the operating budget | Hidden costs that grow unchecked | Put tokens in the talent budget or the operating budget |
| Delegating everything without deliberate friction | The team's judgment erodes over time | Keep moments of independent thinking in the process |
| No enterprise contract with the AI provider | Legal risk on sensitive client data | Check the MSA and data processing agreement before you start |
How can you tell if the model is working for your company?
Three clear signals: the cost per unit of output goes down, senior time shifts from production to selection, and average output quality stays stable or improves.
S4 Capital's EBITDA margin rose to 12.1% in 2025 despite a 10.8% drop in revenue: a sign that operational efficiency offset shrinking revenue. For an Italian SME the equivalent KPI is simpler: how much does it cost to produce one unit of output (a report, a product sheet, a post) before and after introducing agents?
| Indicator | How it is measured | Positive signal |
|---|---|---|
| Cost per unit of output | Total € (tokens + senior hours) / number of outputs | Down 20%+ from baseline |
| Senior time on selection vs. production | % of senior hours spent choosing vs. making | Selection exceeds 50% of total time |
| Average output quality (sample) | Weekly sample review on a 1-5 scale | Stable or rising average |
| Monthly token spend | € tokens / month per agent or process | Stable or below plan |
| Errors going undetected in production | Number of post-publication reports | Trends to zero with review agents active |
| Team satisfaction | Quarterly survey (1-5 scale) | Doesn't drop: the team sees the change as positive |
So, is it worth unifying budget and team on the Monks model for your company?
It's worth it if you have structured processes with requirements you can write down, a measurable feedback loop and the willingness to invest time in governance. The operational savings are real and measurable within the first few months.
Somewhere in between if you're in a transition phase: you have some structured processes, but also a lot of creative or high-risk work. Start with the simplest processes, build governance and expand gradually. Don't wait until everything is ready.
Not yet if your work relies almost entirely on qualitative judgment that can't be measured, or you don't yet have the resources to write decent requirements and maintain an agent registry. In that case, it's better to wait and get trained first.
How do we see it at SHM?
For us, AI is a means, not an end. The Monks case is instructive precisely because it shows where automation stops: in selection, in judgment, in deciding which direction to take. Fifty ideas in ten minutes are worth nothing if you don't have someone who can recognize the five that matter.
We believe that integrating autonomous agents into a company doesn't mean reducing people to passive validators. It means giving them back the time to do the work that truly matters: thinking, choosing, building relationships with clients. The deliberate friction Monks is now reintroducing isn't a step backward. It's an acknowledgment that human thinking has a value that can't be measured in tokens.
In the projects where we've supported companies in integrating AI processes, the most critical moment has never been choosing the tool. It has always been defining the requirements: whoever writes the instructions for the agent determines the quality of everything that follows. It's a skill you build, not one you buy.
If you're evaluating how to structure this journey, you can start from our dedicated page on strategic AI integration, dig deeper into business intelligence and data processing or explore the options for virtual assistant development. You can also follow our hub on AI and process automation to stay up to date on how the market is moving.
And if you want to understand how to optimize your company's visibility in AI search engines, the article on OpenAI Astra and computer use offers a useful perspective on where research is heading.
Frequently Asked Questions
It depends on the model you use and the complexity of the task. The most powerful models consume far more tokens (the units of processing) for every operation. Monks' logic is to assign the cheapest model that fits the task: not every job needs the best model. Track spending per agent from day one.
Ask yourself two questions: can I write the requirements so that a new colleague understands them without ambiguity? And do I have an objective signal to tell whether the output is good (a test, a KPI, a checklist)? If the answer to both is yes, the process is a good candidate. If not, work on these two elements first.
Not necessarily, at least in the initial phase. Monks has reduced its headcount, but the model it describes is one of redistributing work, not outright replacement. People move from production to selection and governance. The risk of losing in-house expertise is real over the long term if you don't maintain deliberate friction in the process.
Yes, and some will ask for it. Monks' most advanced clients want an agent registry, want to know which models are being used and want to be able to "switch them on and off." Prepare this transparency before they ask for it: it's a competitive advantage, not an inconvenient obligation.
Before letting an agent process sensitive data, check that the AI provider has an enterprise contract (MSA) with clear clauses on data processing and data residency. Unverified models, or models without a contract, are a concrete legal risk, as confirmed by the concerns of Monks' clients.
On structured processes (coding, reports, product sheets), the first measurable results arrive 4-8 weeks after go-live. Most of the time goes into the setup phase: writing good requirements and building the agent register. Don't skip this phase to speed things up: it determines the quality of everything else.
Agent autonomy isn't the problem to solve. It's the context that makes visible just how much human judgment is really worth.
Related articles
Discover more articles exploring similar topics, selected to offer you a more complete and stimulating perspective. Each piece of content is carefully chosen to enrich your experience.
