Skip to content

Roles in an AI Agent Team: Who Does What

Most AI agent projects in mid-sized companies do not stall on technology. They stall on ownership: nobody is sure who decides what the agent may do, who checks its work, or who keeps it running once the project team moves on. This guide sets out a role model that works for small teams and explains which roles can be combined and which can be covered by outside help.

What is an AI agent team, and why does it matter?

An AI agent team is the set of people who own an agent's purpose, build it, run it safely and decide where it fits among the company's priorities. It is an allocation of responsibility rather than a new department. Without it, pilots work in a demo and then quietly fade.

Agents differ from conventional software because they make decisions inside a process. Anthropic describes agents as systems where language models "dynamically direct their own processes and tool usage", in contrast to workflows that follow predefined code paths (Building effective agents). That autonomy changes the work around the system: someone has to keep judging whether those decisions are correct for the business.

When roles are missing, the symptoms are predictable:

  • The agent performs well in testing, but nobody in the business feels it is theirs to use.
  • Bad outputs are noticed but not reported, because there is no obvious person to tell.
  • API keys and credentials sit with the one developer who built the prototype.
  • New agent ideas compete for the same few people with no way to rank them.

Which four roles does a small agent team need?

A small team needs four core roles: the business owner, the AI engineer, operations and the AI lead. Together they answer four questions: what should the agent do, how is it built, how does it run safely, and why this agent rather than another. Each is a responsibility, not necessarily a full-time job.

Business owner

The business owner comes from the process the agent supports: the head of inside sales for a quotation agent, for example, or the service manager for an agent that triages fault reports. This person:

  • describes how the process really works, including the exceptions nobody wrote down,
  • supplies real examples with expected outcomes, such as 30 past enquiries and the quotes that followed,
  • reviews the agent's outputs during development and once it is live,
  • decides which steps the agent may complete on its own and where a human must approve.

This is the most important role and the one most often underestimated. Without it, engineers build against guesses.

AI engineer

The AI engineer turns the process into a working agent: model choice, prompts, tool and API integration, test cases and evaluation. Anthropic's guidance stresses that tool definitions deserve as much prompt engineering attention as the prompts themselves, and that success depends on measuring performance and iterating (Building effective agents). Making the agent testable is as much the job as building it.

Typical skills are solid programming (usually Python or TypeScript), comfort with APIs and data formats, and the ability to talk to a business owner about processes without jargon. Most agents build on existing language models, so a machine learning research background is rarely needed.

Operations

Operations keeps the agent reliable and secure after launch. The role usually sits in IT or with a managed service provider and covers:

  • access rights, permissions and the handling of API keys,
  • the runtime environment, whether a cloud service, a company server or an on-premises GPU machine,
  • logging and monitoring: uptime, error rates, cost and response times,
  • model and library updates, plus backups.

The guide to AgentOps: running AI agents reliably covers this in detail. Agents connect to ERP, CRM and document systems. Every one of those connections is an access path that someone has to manage.

AI lead

The AI lead connects individual agents to business goals. This person decides which use cases come first, sets rules for data, models and vendors, and reports to senior management. In smaller firms a managing director often takes this on, or a part-time specialist does. The options, from a full-time head of AI to an external arrangement, are compared in a separate guide.

The US NIST AI Risk Management Framework asks that roles, responsibilities and lines of communication for AI risk be documented and clear across the organisation, and that executive leadership take responsibility for decisions about AI risk (NIST AI RMF Core). That is a fair description of what the AI lead must make happen. For how the wider agenda comes together, see the guide to AI strategy for mid-sized companies.

Who else needs a seat at the table?

Beyond the core four, several groups need defined checkpoints rather than day-to-day involvement: data protection, information security, employee representatives and the people who will use the agent. Bringing them in at the end is the most common cause of launch delays.

  • Data protection: checks what personal data the agent sees, which vendors process it and on what legal basis. The aim is an agent that can be operated in line with the GDPR.
  • Information security: reviews permissions, data flows and what happens if the agent receives manipulated input, such as a prompt injection hidden in an email.
  • Employee representatives: in Germany, works councils have co-determination rights under Section 87(1) no. 6 of the Works Constitution Act when technical systems designed to monitor employee behaviour or performance are introduced. Other countries may have their own consultation rules. Whether an agent is in scope depends on its design; early briefing builds trust. This is not legal advice.
  • End users: they work with the agent every day. Their feedback is a valuable source of improvements, provided there is a clear channel for it.

How can a small company split the roles?

Few mid-sized companies staff four roles with four full-time people. Two or three people with agreed time allocations is more realistic. What matters is that every role has a named person and that the time is formally planned, not squeezed in around other work.

Three patterns are common:

PatternWho does whatWorks when
Business-ledOwner in the business unit, external engineering, IT runs operations, a director acts as AI leadthe first agent supports one well-defined process
IT-ledIT manager acts as AI lead and runs operations, an in-house developer becomes the AI engineer, owner in the businessthere is a strong internal IT team with development capacity
Mixed core teamPart-time AI lead, one in-house AI engineer, an owner per agent, operations in IT or with a providerseveral agents are planned in parallel

One combination to avoid: business owner and AI engineer in the same person. Whoever builds the agent tends to judge its outputs too kindly. Keeping the two apart is a cheap form of quality control.

Which roles can be covered externally?

The roles that need specialist agent experience but are not yet a full workload are the natural candidates: AI engineer and AI lead. The business owner stays internal, and so does control over credentials. The key is a handover plan, so the company does not stay dependent on outsiders for every change.

RoleOutsource?What to watch
Business ownernoprocess knowledge and accountability stay in-house
AI engineeryes, especially for the first agentcode, prompts and test cases belong to the company and live in its own repository
Operationspartlyhosting can sit with a provider; the company keeps control of keys and permissions
AI leadyes, part-timeplan a handover to an internal person once several agents are running

A part-time AI lead is most useful where nobody in-house has run agents before but decisions on vendors, data and priorities are already on the table. One way this can be set up is described on the page about a fractional head of AI.

What does the team need to get started?

The team needs time, access and competence. Time means protected hours for the business owner. Access means read access to the data and systems the agent will use. Competence means everyone involved understands what a language model does well and where it goes wrong.

Time. The owner's workload peaks during the build, when examples must be gathered and outputs reviewed. If that time is not planned, review falls to the engineer, who does not know the process.

Access. Projects often lose weeks waiting for test access to ERP, CRM or document stores. Operations should arrange it before kick-off, with the narrowest permissions that work.

Competence. Since 2 February 2025, Article 4 of the EU AI Act has required providers and deployers of AI systems to take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff. It does not demand a guaranteed level, but it applies regardless of risk class. In practice, owners, operations staff and users each need a briefing suited to their role, and a hands-on workshop is one way to deliver it. What the duty means in practice is covered in the guide to AI literacy under Article 4.

If an agent falls into the high-risk category, more applies. Article 26 then requires deployers, among other things, to assign human oversight to people with the necessary competence, training and authority, to monitor operation, and to keep automatically generated logs under their control for at least six months. Whether an agent counts as high-risk depends on its intended purpose; either way, the classification should be recorded.

How do you build the team step by step?

Grow the team around the first agent rather than ahead of it. Designing an organisation first and then looking for a use case produces roles with nothing to do. The sequence below is the one that has worked in our own operations.

  1. Pick the use case and the owner together. A process nobody wants to own is a poor first candidate.
  2. Name an AI lead. Internal or external, but a named person whose job includes recording priorities and rules.
  3. Staff the engineering. An in-house developer with allocated time or an external engineer, with code and tests in the company's repository from day one.
  4. Involve operations early. Settle access, runtime and logging before the first test.
  5. Brief privacy, security and employee representatives. Before the pilot, with a short note on data, vendors and what the agent is allowed to decide.
  6. Write the roles down. One page is enough: role, person, time share, deputy, escalation path.
  7. Review after the pilot. Did each role work? Where was time short? Only then decide on dedicated hires.

Good first use cases have many similar cases and clear outcomes, such as quotation or technical service. An overview of typical agents is on the AI agents page.

What are the limits and risks of a small team?

Small teams move quickly but are fragile. The biggest risk is dependence on the one person who built the agent and understands it. Others include an owner with no time, roles that exist only on paper, and an AI lead without real authority.

  • Single point of knowledge. If only one person knows the prompts, keys and tests, the agent stops when they are away. Counter it with documentation in the repository, a named deputy and a shared secrets store.
  • Owner without time. The most frequent reason agents stagnate. Agree the time allocation with the owner's line manager, not just the owner.
  • Lead without a mandate. An AI lead who cannot enforce priorities becomes an advice desk. Senior management has to back the decisions.
  • Too much structure too soon. Steering committees and role descriptions for agents that do not exist yet consume time without results.
  • Permanent outsourcing. Outside help makes sense at the start. A company that still has no internal know-how after several agents will depend on others for every change.

When does this model not fit? Companies that only use AI features built into standard software do not need four roles; one accountable person and normal IT processes will do. At the other end, companies building AI into their own products need a much larger team with specialist roles.

How do I get started?

Start with one question: which process has someone who wants to improve it and will be given time to do so? That person becomes your first business owner. Staff engineering, operations and leadership around them, internally or externally, and record the arrangement on a single page.

A simple template for that page:

  • Agent: name and the process it supports
  • Business owner: person, time share, deputy
  • AI engineer: person or provider, repository location
  • Operations: person or provider, runtime environment, where keys are kept
  • AI lead: person, reporting line to senior management
  • Consulted: data protection, information security, employee representatives, with dates
  • Escalation: who is told when the agent produces wrong results, and who is allowed to switch it off?

The last line matters most. An agent that nobody may or will switch off is a liability; an agent with clear ownership is a tool that can keep improving. If no one in-house can take the lead yet, a part-time AI lead can bridge the gap until internal experience has grown.

Frequently asked questions

Do I need to hire anyone for the first agent?

Usually not. Existing staff can take on the roles with a fixed share of their time. Dedicated hires make sense once several agents are in production and need ongoing development.

Which role should never be outsourced?

The business owner. Only someone from the process itself can judge whether the agent's output is right, and that person keeps responsibility for the process after the project ends.

Where do legal, privacy and works councils fit in?

They are not part of the core team but need defined checkpoints before a pilot goes live. In Germany, for example, works councils have formal co-determination rights when monitoring technology is introduced.

Does the EU AI Act require specific roles?

Article 4 requires organisations using AI to take measures to ensure sufficient AI literacy among their staff. Additional oversight duties apply only to high-risk systems; this article is guidance, not legal advice.

Sources

  1. Anthropic: Building effective agents (December 2024) anthropic.com
  2. NIST AI Risk Management Framework, Core functions airc.nist.gov
  3. EU AI Act, Article 4: AI literacy artificialintelligenceact.eu
  4. EU AI Act, Article 26: Obligations of deployers of high-risk AI systems artificialintelligenceact.eu
  5. German Works Constitution Act, Section 87 gesetze-im-internet.de