Business Management

AI in the workplace: redesigning processes for human-agent collaboration

The adoption of AI agents is changing workflows, roles and skills, necessitating a structural rethink of the collaboration between humans and machines within organisations

by Lorenzo Ghidotti*

2 July 2026

6' min read

Translated by AI

Versione italiana

6' min read

Translated by AI

Versione italiana

There are certainly effects on redundancies, with significant figures that could rise exponentially even in the short term. The effects on lost recruitment opportunities are already even more profound today, although difficult to quantify. But the most revolutionary effect of artificial intelligence on the labour market is qualitative in nature. This is because it changes the way work is carried out and organised, particularly with the integration of AI agents into business processes, which are redefining workflows, responsibilities and, consequently, organisational structures. For this reason, a systematic rethinking of human-agent collaboration within organisations is necessary. This will enable us to tackle this revolution and, hopefully, emerge from it stronger.

Let’s start with the figures. General Motors has announced the loss of over 500 jobs, citing AI in the technology sector. In April, Snapchat’s parent company announced the redundancy of 16 per cent of its workforce for the same reason. Meta has implemented restructuring and cut around 8,000 jobs (10 per cent of its workforce) and cancelled 6,000 planned hires. Cloudflare has made 1,100 redundancies. We could go on, but to provide a broader picture: according to the Job Cuts Report by Challenger, Gray & Christmas, from January to September, US companies announced redundancy plans affecting one million people – the highest figure since 2020 (the year of the Covid pandemic) – whilst they announced plans to fill 204,000 positions, the lowest figure since 2009, the year following the Lehman Brothers collapse. The real revolution, however, lies elsewhere.

Research published by Anthropic in March 2026 on the impact of LLMs on US occupations reveals a figure that warrants attention: in the occupations most exposed to AI, recruitment of workers aged between 22 and 25 has fallen by 14% compared with 2022, whilst the rate of new hires in less exposed occupations has remained stable. Computer programmers, customer service staff and data entry clerks top the list of roles most ‘covered’ by the automated use of Claude — 75% respectively, with high exposure also for administrative roles, and 67% coverage for data entry clerks. The same study notes that, so far, there has been no systematic increase in overall unemployment within the most exposed professions. The signs of transformation lie elsewhere: in labour market inflows, in the composition of job roles, and in the redesign of organisational structures. It is here that the issue becomes, even before it is a technological one, an organisational one.

Italian case law is also adapting. In judgment no. 9135 of November 2025, the Court of Rome recognised as lawful the dismissal on objective grounds of a graphic designer whose role had become redundant following a reorganisation carried out, in part, using AI tools. The judge treated AI as just another tool for improving efficiency. The rules have not changed; what has changed is the scope of their application. This is because the organisation of work and businesses has changed – and will change even further.

Cutting, extending, redistributing: a comparison of three strategies

Cases involving large corporations are beginning to mount up and are following very different trajectories. Salesforce has reduced its customer support workforce from 9,000 to 5,000 staff thanks to the Agentforce platform, resulting in a net reduction of 4,000 positions. JPMorgan has rolled out its LLM platform to 200,000 employees, with hundreds of active operational use cases. Klarna presents the opposite and most instructive case: between late 2022 and late 2024, it reduced its workforce from 5,527 to 3,422 employees, stating in February 2024 that a single agent was doing the work equivalent to 700 customer service operators. In May 2025, CEO Sebastian Siemiatkowski publicly changed course in an interview with Bloomberg: cost had become the dominant criterion, service quality had declined, and the company had resumed hiring staff.

Gartner estimates that by the end of 2026, 40% of enterprise applications will incorporate task-specialised agents, compared with less than 5% in 2025. The speed of adoption, combined with evidence that those who cut back indiscriminately are backtracking, suggests that what is at stake is not the level of automation achieved, but the way in which agents, roles and processes are integrated.

The quality front: how the work that remains is changing

The qualitative impact is less visible than the figures on the job cuts, but just as significant. Alongside the reduction in customer support, Salesforce has launched a reskilling programme involving 72,000 employees, structured around three key areas: human skills (adaptability, emotional intelligence, exception handling), agent-related skills (structured human-agent collaboration, supervision of automated outputs), and business skills (data analysis, complex problem-solving). Hundreds of customer support staff have been redeployed to professional services, sales and customer success.

The most sought-after skills are shifting towards those that agents struggle to replicate: handling exceptions, exercising judgement in borderline cases, building relationships, and critically supervising outputs. This is where the quality of the transition is truly at stake. Tacit skills — those that experienced staff have built up by handling anomalous cases, regulatory exceptions and situations that fall outside standard procedures — cannot be transferred to an agent. They are lost along with the people who possessed them, and it is precisely this mistake that forced Klarna to backtrack.

Senior management structures are also changing. Moderna has brought HR and IT together under a single senior executive. The reasoning is that, as agents span all business functions, the distinction between ‘business requirements’ and ‘technological implementation’ ceases to function as an organisational boundary. The AI layer becomes cross-functional infrastructure and must be managed as such.

The framework: four stages for process re-engineering

The practical question for those leading a complex organisation is how to translate these signals into a coherent framework of processes and responsibilities. From the documented experiences, a sequence emerges that is worth formalising into four stages.

Phase one: audit of AI capabilities already in use. Almost all organisations of a certain size already make widespread use of artificial intelligence, often without this being formally recorded — so-called ‘Shadow AI’. Mapping these out is the first step, both for risk management purposes (compliance, security, data quality) and to ensure that efforts are based on reality rather than abstract planning. Without this step, subsequent initiatives risk duplicating or contradicting existing practices.

Stage two: classification of activities. Each business task must be categorised into one of four operational configurations. Human-led activities remain entirely the responsibility of the individual, as they require judgement, interpersonal skills and non-delegable responsibility. In AI-assisted activities, the agent supports the operator without making autonomous decisions. AI-led activities with a human-in-the-loop are managed by the agent, with human supervision in critical cases. Fully autonomous activities are delegated to the agent without direct intervention. This classification is not static: a mechanism for periodic review is required as the agents’ capabilities evolve and the organisation gains experience.

Stage three: placing agents within the organisational chart as designated roles. Every agent in production must have a defined area of responsibility, an identified human point of contact, explicit intervention criteria and thresholds beyond which the decision is passed to a human. Agentforce, in Salesforce, autonomously manages low-complexity conversations — the human contact person intervenes in cases exceeding the threshold according to codified rules, not at individual discretion. The logic is that of a structured workflow, not a generic partnership between a person and an agent that leaves it up to individuals to informally renegotiate responsibilities.

Phase Four: governance and dual reporting lines. The human point of contact reports to their functional line management. The agent reports to an AI Governance function, which in many large corporations is overseen by a Chief AI Officer responsible for the agents’ lifecycle, their alignment with business objectives, performance monitoring and risk management. It is within this function that compliance issues — the AI Act in particular — and system security are overseen.

The three bottlenecks to watch out for

When this sequence is not followed, three recurring issues emerge, as documented by early adopters. The first concerns validation: an agent produces output at a rate that human review struggles to keep up with, and comprehensive checking inevitably degenerates into risk-based sampling — with the problem of establishing the sampling criteria. The second is accountability: without a designated human point of contact with genuine authority to intervene, decisions grind to a halt at the first borderline case, and the organisation loses responsiveness precisely where the agent promised to enhance it. The third is psychological and manifests itself later, in staff turnover: remaining staff perceive the agent as a replacement rather than a tool, and motivation declines — a factor not included in transformation plans but which is measured, after six or twelve months, in terms of operational resilience.

The direction of change

The organisations that will capitalise on this transition are those in which the organisational chart will once again become what managerial tradition has always called for: a coherent, value-driven map of capabilities — both human and artificial. Competitive advantage does not lie in the number of agents deployed, but in the quality of the process design that accommodates them. Redesigning workflows before automating them is essential if AI is to generate efficiency rather than exacerbate existing dysfunctions. It is a task of organisational engineering, even before it is one of technology. And like any engineering task, it requires clear criteria, sequences and responsibilities.

*Strategic Management Partners