AI adoption in practice: managing a team of humans and agents
At the DII Gathering: The Workshops event on 3 September, Michelle Wallace and Maryrose Lyons ran a session called Closing the Gap with AI: Building the workforce of the future. It opened with a question that had nothing to do with technology, and everyone in the room had to answer in a single word. AI adoption in practice: managing a team of humans and agents
How do you feel about managing a team of humans and agents?
The one-word answers are the point. Most conversations about AI in digital infrastructure get stuck on tools. Which platform, which licence, which pilot to run next. This session treated it as a management problem instead, and that is where the useful thinking started. Michelle Wallace brings 18 years in digital infrastructure leadership, starting at Digital Realty, and is Programme Director of Emerging Leaders in Digital Infrastructure. Maryrose Lyons is founder and managing director of AI Institute, has a Masters in Cyberpsychology and a background in UX, hosts the
ChattingGPT podcast, and has now delivered AI training to more than 2,800 professionals across the built environment.
Between them, the message was the same. The gap holding the industry back is a management gap, not a technology one.
Where AI adoption actually is, and where most organisations are still buying
To place the industry honestly, the session borrowed OpenAI's five stages of AI capability. Stage one, chatbots that hold a conversation. Stage two, reasoners that solve problems at expert level. Stage three, agents that take action. Stage four, innovators that help invent. Stage five, organisations that run whole functions.
The workshop's read on where we actually are: the technology has reached stage three, agents that take action, while most organisations are still buying at stage one.
That mismatch explains a lot of stalled adoption. Buy a chatbot licence, expect an agent's return, and you get neither. The two do not behave alike at all.
They are different products, and more to the point, different management problems.
The uncomfortable evidence: faster tools, heavier jobs
The hardest slide to sit with was not a product demo but a study. In February 2026, Harvard Business Review published AI Doesn't Reduce Work. It Intensifies It. by Aruna Ranganathan and Xingqi Maggie Ye. They tracked a 200-person technology firm in the United States as it adopted AI tools, and three things happened to the people there.
• They worked faster. The tools did exactly what they promised at the level of the individual task.
• They took on more. The scope of what each person was expected to cover widened.
• They worked longer. Work spread into more hours of the day, not fewer.
The productivity gains landed on the individual. Nobodyredesigned the job.
No licence purchase fixes this part. Redesigning a job means making three things explicit: what the person stops doing, what the tool takes on, and what stays human. Skip it, and a faster tool just raises the ceiling on what one person is expected to carry. A lot of what looks like AI fatigue in engineering and construction teams right now has little to do with resisting the technology. It is what happens when you keep adding capability and never redesign the role underneath it.
What a live project cost dashboard shows about practical AI use
The construction example from the session is worth pausing on, precisely because it is so unglamorous. A client had its project cost picture scattered across several systems that were never built to talk to each other.
This is the kind of AI use that survives contact with a live project. No moonshot. A reconciliation job that had been eating skilled people's weeks, handed to something that will happily do it every hour. The people went from assembling the number to questioning it, which is what they were hired for in the first place. You see this pattern over and over in AI training for the built environment, because AEC and digital infrastructure teams are full of it: senior professionals stitching systems together by hand.
Managing humans and managing agents: the same conditions, different mechanisms
This was the spine of the session, and the part most worth taking back to your own team. A good human team needs four things: psychological safety, trust, role clarity and motivation. Agents need the same four. What changes completely is how you provide them. With people you build it through relationship. With agents you build it through structure.
Psychological safety becomes guardrails
With a person, safety protects them so they will take the initiative. With an agent, the guardrail protects the organisation from the initiative. Spell out what it must never do, and it can get on with everything else.
Trust becomes engineered verification
You trust a colleague because they have earned it over years. An agent has no such history, so the trust has to be built deliberately: testing, explainability, accuracy thresholds, real access to the systems where the work actually happens. You rely on it because you have checked it, not because it has been around a while.
Role clarity becomes goal clarity
A person picks up their role from context and fills the gaps with judgment. An agent only knows what you tell it, so the objective has to be written down and it has to be precise. One job, defined tightly enough that you can check it was done. The cautionary tale from the session lands right here. An airline told its agent to resolve complaints efficiently, and it handed out 140,000 dollars in vouchers nobody needed. Same requirement as role clarity, minus any feel for reading between the lines.
Autonomy becomes compliance and a readable log
You give a person autonomy because you trust their judgment. An agent has none to earn it, so its freedom to act rests entirely on the trail it leaves. Every action logged, bounded, and open to inspection by someone who was not in the room. That trail is the only reason you can safely let it off the leash.
If you carry responsibility for responsible AI in a regulated supply chain, that last point is the entire governance argument in a sentence. It is also why sensible AI governance for construction firms begins with logging and boundaries, not with a policy document nobody reads.
What a team could look like next year
To make it concrete, the session sketched a sales function a year from now. A Head of Sales and a Sales Manager, both human, working alongside four agents.
• Research agent. Builds the account brief before a call: company news, org chart, previous deals, competitor position.
• CRM hygiene agent. Logs calls, updates stages, fills the fields reps skip and chases the ones it cannot infer.
• Proposal and bid agent. Assembles first-draft proposals and tender responses from the content library, pulling in the right case studies and pricing.
• Pipeline reporting agent. Weekly movement, deals that have gone quiet and forecast variance, flagged to the Head of Sales before the Monday meeting.
Two people, four agents. The humans spend their time on judgement, on relationships, and on the things a log can never capture.
Maryrose pointed out that the CRM hygiene agent always gets the biggest reaction in a room, because everyone hates doing that job. Which is a good clue for where to start at home. Pick the task nobody will fight to keep.
Two exercises to run with your own team this month
The workshop did not finish on a slide. It finished on two tasks, and both drop straight into your next leadership meeting.
• Task one, map your own function. Take your department as it is today. Draw the people, put their titles on, then add the agents you would want and give each one a name and a single job. The gaps in that picture are your AI skills gap, spelled out far more precisely than any maturity assessment will manage.
• Task two, write down the mechanisms. Two columns, humans and agents. For each of the four conditions, safety, trust, clarity and autonomy, write down how you actually deliver it to each side. Whichever column is harder to fill in is telling you what to build first.
Then ask the opening question again, the way the session did. How do you feel now about managing a team of humans and agents? What was interesting on the day is that the first answers were already warmer than we expected. This was a room of people getting on with the work, not one waiting to be talked round. By the end the mood had lifted again, with people readier to carry on what they had already started. That shift, from one word to a more confident one, is the thing worth measuring.
Do not let AI happen to you
The session closed on two lines. Do not let AI happen to you. And then, become the orchestrator.
That is a leadership brief, not a technology one. The organisations that close the gap will be the ones whose managers can write a clear brief, set a boundary, check an output and read a log. And who redesigned the job before the tools landed, rather than after the burnout did.
If that is the capability you want in your own teams, the good news is it can be taught. AI Institute builds its work around behaviour change rather than tool demonstrations, which is the approach behind the training, and the reason the 2,800 professionals it has trained sit mostly in construction, engineering and the wider built environment rather than in tech. Its AI training courses run across AEC, bid teams, finance and marketing, and cover the governance and responsible AI groundwork as well as the practical skills.
Whichever route you take, the honest starting point is the one the room started with. One word. How do you feel about managing a team of humans and agents?
