The AI assistant is becoming the AI colleague
A prompt ends when the answer arrives. A persistent agent may watch a process, follow up and pick up work days later. That is not just a faster assistant; it is continuing responsibility assigned to software.
A new kind of delegation
Most chat assistants wait for someone to ask a question. A persistent agent can be assigned a goal, keep working in the background, follow activity across tools and return to a task later. The shift is from prompt to task, from task to delegation, and potentially from delegation to ongoing responsibility.
Microsoft’s 25 September 2026 announcement describes Autopilot as a cloud-hosted, persistent agent that can watch channels, follow up on threads, run recurring work and resume a project days later. Microsoft says it can have its own identity, memory, computer and workspace, with permissions, audit and governance. Home and Code were to roll out through Frontier, while Autopilot was expanding to private preview at the end of September; this is not evidence of general availability for every Microsoft 365 customer. These are Microsoft’s product and rollout claims, not an independent evaluation. Read Microsoft’s announcement.
Do not start with the product name
A small accountancy practice could spend time asking clients for missing information, checking whether it arrived and preparing a file for review. An agent might monitor a defined inbox, match replies to an approved checklist, remind a client once and assemble a draft status report. That is a candidate workflow—not a reason to give it access to every client email or permission to approve the accounts.
Start with the repetitive responsibility: monitoring, information gathering, coordination, preparation or exception detection. Understand how staff do the job today, where judgement is required and what an acceptable result looks like. Then ask whether AI is appropriate, whether a rule-based automation would be more reliable and what happens when the process changes.
Give the agent a job description
- Role and owner: What recurring job does it own, and who is accountable for the process?
- Objective and measure: What outcome should improve? Set a baseline for quality, time, exceptions and cost.
- Inputs and tools: Which records and systems may it use? Give only the access needed for its role.
- Authority: What can it read, draft, send or change? Which decisions stay with a person?
- Exceptions and review: When must it stop or hand over? Who reviews value and removes the agent when the work ends?
An agent preparing a supplier review may need to read approved vendor records and draft questions; it may not need to change bank details or accept terms. Log what it did and let the owner pause access when the role changes.
Persistent work still needs management
A manager remains responsible for priorities, exceptions, quality, cost and whether the process is worth doing. An agent that continues without a prompt can continue a mistaken assumption too. Set review points for changed instructions, stale records, unusual requests and work without a clear finish condition.
There is also a risk of agent sprawl. Separate teams may create agents for sales follow-up, finance preparation, recruitment coordination and IT requests without a central view of owners, identities, connected systems and spend. Keep an inventory and retire an agent when the process ends or stops creating value.
Microsoft’s announcement introduces FinOps for AI capabilities to help organisations manage usage and spend. That is a signal to watch agent economics, not proof that software activity equals business value. Measure work completed, turnaround time, errors and rework, human supervision and operating cost against a baseline. The important metric is useful work delivered—not how many times an agent ran.
A colleague metaphor should come with colleague-level clarity
Calling an agent a colleague can make delegation easier to understand. It should not obscure that software has no independent accountability. People still need to know who owns the process, what the agent may do and who handles exceptions. For a boundary checklist, see Your AI Agent Needs a Job Description — and Boundaries. To assess the commercial case, read AI agents change the economics of automation.
The question is not whether an agent can keep working while the team is away. It is whether it has a worthwhile job, a named owner and a measurable result.