From one assistant to many department agents
Many enterprise teams first met AI through a general-purpose assistant rolled out to everyone at once. It helped with drafting and summaries, but it seldom touched the systems where work actually gets done. The shift now under way in many organizations is toward vertical AI agents built for one department's job: matching invoices, answering policy questions or sorting incoming contracts.
The change is less about smarter models and more about ownership. A department agent is agentic AI with a business owner, a defined job, approved sources and access only to the systems that job needs. That makes it measurable in a way a company-wide assistant is not. It also multiplies the AI systems an organization must govern, buy, secure and support.
Governance starts with an inventory of agents
Once agents multiply, the first risk is not knowing what is running. Departments buy tools, switch on features inside platforms they already use and build their own. An AI governance program starts with a register that lists every agent, what it does, who owns it and what it can touch. If an agent is not on the register, it should not be live.
Governance also sets the shared rules every department agent inherits: which data classes are off limits, which actions always need a person, how logs are kept and how an agent is retired. Privacy law, including PIPEDA where it applies, and sector regulation shape those rules. Legal, privacy and risk teams should help write them, not only approve them at the end.
- Owner and department
- The single job and its finish line
- Systems and data it can read or change
- Actions that need human approval
- Where logs live and how long they are kept
- Next review date and retirement plan
Procurement and data access
Procurement teams are used to buying horizontal SaaS by the seat. Agents rarely fit that model, since they do work rather than host users. Contracts need to cover usage or outcome pricing, caps, data residency, whether your data trains outside models, export rights and what happens to logs at the end of the term. One standard question set saves each department inventing its own.
Data access is where department agents differ most from a general assistant. Each should get the narrowest access its job needs, through a service account security can monitor and revoke. A finance agent doing document processing on supplier invoices needs the payables system, not the HR file share. Least privilege keeps one confused or compromised agent from reaching everything.
Build or buy is a decision made job by job. Where industry software already includes a capable agent, buying is usually simpler. Where the process is distinctive or crosses several internal systems, custom software built around your own rules may fit better. Either way, the agent meets the same governance standard before it goes live.
Change management for the people around the agent
Agents change jobs before they change headcount plans, and staff notice. Using general AI for drafting asked little of anyone's role; a department agent asks more. People who handled the routine part of a process become reviewers, exception handlers and keepers of the agent's rules. Explain that early, involve them in writing the rules, and be honest about what is still undecided.
Start each department with a single-workflow agent and a supervised period in which staff approve every action. Track the measures the department already uses, such as cycle time, backlog and rework, so results are judged in familiar terms. Choosing where to begin is easier with a short review across departments that weighs volume, rule clarity and risk.
When agents start working together
Once several department agents exist, they begin passing work to each other. An intake agent sorts a supplier email, a document agent extracts the invoice, and a finance agent matches it to a purchase order. Multi-agent systems like this need an orchestration layer that assigns each step, tracks every hand-off and stops the chain when something looks wrong.
Risk compounds with each link, because an early error can travel quietly through several agents. Keep a human checkpoint wherever money moves or records change, log the whole chain rather than each step alone, and use plain workflow automation for the fixed parts. Save the agents for steps that genuinely need judgment about language or documents.
What to plan for in 2027
Expect the number of agents inside large organizations to grow faster than the teams governing them. Those that cope will treat agents like any other business system: registered, owned, access-controlled, logged and reviewed on a schedule. Budget for review and support work, not only the build, and revisit the inventory at least once a quarter.
- 01Agent registers becoming a standard audit item
- 02Procurement templates written for agents, not seats
- 03Orchestration and logging that span departments
Questions & answers
They are AI agents built for one department's job, such as invoice matching, policy questions or contract intake. Each has a business owner, approved sources, limited access to the systems it needs and rules for when a person takes over. They differ from a general assistant that everyone uses for drafting and research.
Each agent needs a business owner in the department it serves, who answers for what it does. A central group drawing on IT, security, privacy, legal and risk sets shared rules, keeps the inventory and approves new agents. Ownership without central rules tends to end in sprawl.
Give each agent its own service account with the narrowest access its job needs, and review that access on a schedule. Keep sensitive data classes off limits by default, log every read and write, and make sure security can revoke an agent's access quickly.
Both, depending on the job. Buy where industry software already handles the task well and meets your governance standard. Build where the process is distinctive or spans several internal systems. Apply the same review, logging and access rules to bought and built agents alike.
