Quite a lot, as long as the job is well defined.An agent can now work through a multi-step task by itself: read a request, check it against policy, draft a reply and hand the awkward ones to a person. It works when the process is written down and the rules are clear. Point one at a process nobody has ever documented and you just get the mess faster. We start with high-volume, fairly dull work, and keep someone on the exceptions until the agent has earned more room.
Telus employees regularly using generative AI (with time saved per use), Google Cloud report
blog.googleFirms (planning agentic AI) with mature governance models, Deloitte global survey 2026
deloitte.comCustomer calls handled via Vapi’s AI voice agent platform (cumulative), company data
techcrunch.comShare of organisations actively seeking to implement AI agent capabilities, 451 Research/S&P Global
press.spglobal.com3 briefings this month, with 4 new figures that passed our source checks.
subscribers read every briefing in full, and get each one on WhatsApp. subscribe free · how we research and check sources
| copilot | AI agent | |
|---|---|---|
| Who is in control | A person, at every step | The agent, within set permissions |
| What it does | Drafts, suggests and summarises | Carries out multi-step tasks across systems |
| Best for | Knowledge work that needs judgement | High-volume, rules-based processes |
| Main risk | Over-reliance on its output | Wrong actions in live systems |
| The human's role | Accepts or edits each output | Reviews outcomes and exceptions |
Agents are a workflow decision before they are a technology decision. The first question is which process should change and who owns it, not which agent framework to buy.
We sequence from copilot to autopilot: people work alongside the agent first, the process is redesigned around what it does well, and autonomy grows as the controls and the evidence build.
Tools make individuals faster; orchestration makes the business faster. Most AI use today stays in personal productivity, with each tool in its own silo and ad hoc hand-offs between people, agents and systems, which is why it rarely shows up in the P&L.
how we help: enable the organisation. →our frameworks: the Lumo method · the value framework
A copilot assists a person who stays in control of each step. An agent is given a goal and carries out the steps itself, calling tools and systems along the way, with a person reviewing the outcome or the exceptions.
Start with a high-volume, rules-based process that is already documented and has a clear owner, such as request triage or document checks. Measure the baseline first so the value of the agent can be proven.
Agents act, so errors compound: wrong actions in live systems, data exposure and runaway costs. Scoped permissions, logging, spending limits and human review of exceptions keep the risk proportionate.
They replace tasks before they replace roles. In most organisations the near-term effect is redesigned roles, with people moving to review, exception handling and work that needs judgement.
Rebuilding a workflow so people, agents, automations and systems of record pass work between them cleanly, with shared state and context. An orchestrated workflow has been redesigned rather than just augmented, is instrumented end to end so you can see where people step in and errors leak, and has feedback loops so it improves the more it runs.
Because the gains stay with individuals. People use AI tools every day, but the workflows they work inside have not changed, so process throughput does not move. Value appears when one high-volume, repeatable, expensive workflow at a time is rebuilt, starting from the work rather than the tool.