The Humane Agentic Enterprise: How Dubai Teams Thrive With AI Agents at Work

Most talk about the agentic enterprise is about speed, with AI agents finishing in minutes what once took a team a day. A recent post from Salesforce asks a harder question about what all that speed does to the people working beside them. Its researchers want to know how you get the lift from AI without losing the human skill underneath it. For any Dubai business now handing real work to AI agents, the answer should shape how you set them up.

The agentic enterprise in practice

An agentic enterprise is a business where AI agents carry out multi-step tasks while people set goals and check results. A chatbot answers a question and then waits for the next one. Agents go further: they read data, take actions in your systems and report back once the job is done. That shift from answering to acting is what makes this operating model attractive, and it's also where the risk sits.

Autonomous systems can now update records in your CRM, send follow-up emails and sort support tickets without anyone pressing a button. Each of those is a time-consuming manual job built around moving data between tools, and few people miss it. When agents act independently at scale, though, the people who used to do that work get less practice at the judgement behind it. Salesforce's research looks at that trade, and its first findings are useful for any team planning its next step with AI.

Inside the Salesforce research

Salesforce's Responsible AI and Product Research teams are studying human-AI collaboration to learn how people can thrive with AI at work. They interviewed Salesforce staff across 11 jobs, including engineers, lawyers and designers. They also ran micro-interviews with 2,714 Trailblazers at Dreamforce and reviewed 47 academic papers. On top of that, they analysed more than 3,000 real AI work samples.

The people they focused on weren't casual users chatting with a bot. They were builders who turned their own standards into reusable AI skills, agents and reviewers. That detail matters for any team in Dubai moving past simple prompts. The more expertise you encode into AI agents, the more their design decides whether your people keep their skills.

Salesforce says it will share the full research soon, so treat what follows as early findings rather than final results. One theme came through clearly in the first study. Many experts resist passive AI use by building scepticism into the way they work. They trust AI a great deal, but they earned that trust through deliberate routines that steer its output and verify it.

The paradox at the centre of the agentic enterprise

One senior designer-engineer in the study had leaned on AI heavily for months. When his partner asked for help writing something, he froze and felt he'd forgotten how. An engineering architect described the opposite experience. He told the researchers he could now do things he had wanted to do before but couldn't.

Both stories are true at the same time, and that's the point. AI can stretch what an expert does, while the same tool wears down any skill nobody practises any more. Salesforce frames the real question as how you design your interactions with AI so they make you better at the work. Measuring productivity alone misses half the picture.

For a business leader, this changes the brief for any AI project, from marketing to finance. Faster output and lower costs still count, but they aren't the only results worth tracking. You also want to know whether your people are getting sharper or slowly handing their expertise to a tool they can't question. The second outcome is harder to see, and it costs more to fix later.

Automation bias grows as AI agents improve

Even the heaviest users of coding agents in the study kept a last mile of human judgement for themselves. They let the tools handle the bulk of the task execution and held on to the final call. Without that kind of deliberate check, a different pattern takes over. Researchers call it automation bias, which is when people rely too much on automated outputs and recommendations.

Ask someone whether something looks okay often enough, and saying yes stops being a decision. It turns into a reflex. The smooth, frictionless flow that makes an expert faster can also skip past the judgement that makes the expert worth trusting. As each new model gets better, the temptation to skip the check grows with it.

Several people in the study ran into failures that showed this risk clearly. In one case a model described a file it had never opened. In another, an automation reported a task as complete when it hadn't run at all. Both errors would have slipped through if the person had trusted the report without checking the source.

Productive friction, the design move behind a humane agentic enterprise

Their response to those failures wasn't to trust AI any less. Instead, they built scepticism into the system itself, so staying alert didn't depend on willpower alone. Salesforce calls this design move productive friction. It builds on an earlier Salesforce idea called mindful friction, which adds deliberate pauses to help prevent harm.

Productive friction goes a step further than pauses. The aim is to design AI interactions that strengthen people's judgement, creativity and expertise instead of removing every obstacle. Done well, it keeps engagement meaningful without asking people to stay on high alert all day. The check is part of the process, so it runs whether the person is fresh or tired.

Take the technical writer who caught an AI pretending to call a live store of her product data and vocabulary. She added a mandatory confirmation step to her workflow, so the model has to pause and confirm its own certainty before it continues. None of this is a new concept. It's what experts do when they know where a tool tends to go wrong and put a check right at that spot.

For a business, the useful lesson is that your senior staff already know where your processes tend to break. Productive friction means writing their checks into the agent's workflow, so the check runs every time instead of only on a good day.

Eight practices for keeping people at the centre

Salesforce sets out eight emerging practices for keeping expert judgement at the heart of AI systems. It calls them thought starters, and it plans to go deeper in future posts. Below, each one comes with an example of how it could look inside a Dubai company using AI agents today.

1. Put judgement where the decision matters

The first practice gives people a clear chance to apply their expertise at important decision points. Even the most autonomous agent should recognise when a choice needs human judgement rather than more automation. For example, it might pause at a critical step and ask whether a recommendation feels right. If something feels off, the agent can then show the data behind its suggestion before it carries on.

In a sales team, that moment might come when an agent proposes a discount or drops a lead from an outreach sequence. Your sales reps know account history the CRM data doesn't hold, such as a promise made over coffee or a complaint by phone. Let the agent do the research and the first draft, then pause and hand the decision to the rep. Anyone planning AI sales automation for a busy team should build that pause into the first version and keep it there.

2. Let the person go first

The second practice protects independent thinking before the AI offers an answer of its own. One method is a staged reveal, where the model holds back its response until the person has made a first assessment. Another is to use AI as an editor rather than a generator, so the original idea stays yours. Salesforce says this was one of the most common approaches people described at its Dreamforce research wall.

Marketing teams can try this with no new tools. Ask a strategist to write the campaign angle in a few lines before the agent produces options. Or ask a copywriter for a rough first draft, then let the model tighten it. You keep the thinking that makes your brand sound like you, and the AI handles the polish.

Practitioners can also set out their intent, their specification and their plan before they ask an AI to build anything. Human direction then shapes the work from the start, and the agent fills in the detail. Because your team's input comes first, they're in a better position to judge what the agent returns.

3. Engineer scrutiny into the workflow

The third practice moves verification out of people's heads and into the AI workflow itself. One practitioner built an audit log for her coding agent sessions. Each time she corrected a mistake, questioned a claim or changed something the agent had built, the system recorded it automatically. She reviews those logs regularly to track patterns and decide how autonomous the agent should be.

The same idea works well beyond developers. A customer service team can log every agent reply that a supervisor rewrites before it reaches a customer. After a few weeks, the log becomes data that shows which tasks the agent handles safely and which still need a person. Scrutiny then lives in the system, and nobody has to rely on constant vigilance to catch errors.

4. Get a second opinion

The fourth practice adds independent views that challenge assumptions and expose gaps. Instead of trusting a single model, you can have several agents assess the same task and compare their conclusions. One designer-engineer in the study built an 'AI council' of expert personas covering security, engineering and inclusive design. The personas debate possible solutions and report back on the points where they disagree.

This counters the habit some models have of agreeing with whatever you say, and it widens your view of the problem. For a Dubai property firm, a council might set a brand view against a sales view on one marketing campaign draft. You see the trade-off before the campaign goes out, rather than after a customer complains.

5. Use AI to explore complexity

AI agents are good at breaking complex tasks into neat steps. The model behind them is less good at telling you which step hides a judgement call. Many practitioners told Salesforce that AI makes complex issues look simpler than they are, with confident answers that skip important context. The fifth practice uses AI to bring out the information, assumptions and trade-offs behind a decision rather than doing all the sensemaking for you.

An AI system might lay out several reasonable readings of an unclear question side by side. It would then explain how different priorities change the recommendation. For a leadership team weighing an AI-driven hiring or expansion decision, seeing how the same data supports different readings beats a fast answer. It helps people think through what each option gives up before they commit.

6. Own your context and quality standards

The sixth practice helps people define what good work looks like. If your team takes its standards from whatever the model produces, the bar drifts without anyone choosing to lower it. Salesforce points to shared 'second brains' for teams, where people collaborate to build and maintain the context layer their agents rely on. Teams can also create test sets and scoring criteria to judge AI output against standards they agreed on.

For most UAE companies, this is where the real work of an agentic enterprise starts. Your tone of voice, product data and past decisions need to live somewhere the agents can read and your people can edit. Many teams begin by building an AI knowledge base that gives this context one central home. People stay responsible for the content and the quality bar, and the agents work inside those boundaries.

7. Prevent dependency

The seventh practice protects the user's autonomy and avoids designing for attachment nobody needs. Several practitioners described coding sessions where fast output made it easy to spend far more time on a task than it deserved. One fix is to design AI experiences that let people step away once their goal is met. Instead of ending every reply with another follow-up question, a model can signal that the task is done.

Inside your own company, dependency looks a little different. Employees who always ask the agent first can lose the ability to work without it, and nobody notices until the tool is down. Rotate who does key tasks by hand now and then, so the skill stays alive across the team. Any programme of AI corporate training in Dubai should teach people when to switch the agent off, as well as how to use it.

8. Match oversight to the stakes

The last practice lets the cost of getting something wrong determine the level of checking. Not every task needs the same scrutiny. A rough internal draft might need only a quick read. Customer-facing, executive-level or irreversible decisions may need strict verification or explicit approval from a person.

This is how autonomous AI earns a place in serious work. An agent that scores inbound leads using real-time data is usually safe to run with light checks. An agent that refunds a customer, changes a contract or moves money needs a hard stop and a named approver. Matching human oversight to the stakes keeps your people's attention for the decisions where judgement counts most.

Mapping your processes this way is a core part of AI strategy consulting, because it decides where agents act alone and where they wait for a person. It also gives your board a clear answer when someone asks who is accountable for what an agent does.

Agents versus earlier waves of automation

Previous task automation followed fixed rules that someone wrote down in advance, applied to tidy, structured data. Automating simple, repeatable steps rarely touched anyone's judgement, because the rules held it. AI agents work differently, because they reason about a goal and choose their own steps to reach it. Within the limits you set, they act autonomously to achieve an outcome instead of following a script.

That freedom is what makes autonomous agents useful, and it's also why the old approach to quality control falls short. A rule-based system fails in predictable places. The model behind an intelligent agent can fail in new ones, and it can report success while doing so. Deciding where the checks sit is an architectural decision, made when you design the agent rather than left to each person's discipline.

Rule-based tools also never made anyone worse at their job, because people still did the thinking. Agents take over parts of the thinking too, which is why Salesforce's focus on skill and judgement belongs in every enterprise AI plan. It's a question for the implementation stage, well before deploying an agent.

Signs your team is waving agent work through

You can usually spot the drift before it costs you anything. Approvals that once took a few minutes now take a few seconds, and nobody reads past the first line. A marketing manager signs off a week of agent-written posts over one coffee. When the tool goes down for an afternoon, people struggle because they've lost the habit of doing the work themselves.

Another warning sign is a correction log that drops to zero. Either the agent has improved, or nobody is checking any more. You find out by sampling a handful of outputs each week and checking them against the source data. A model update can also change how an agent behaves, and people who no longer check won't notice.

None of these signs is a reason to switch the tools off. They tell you where to add friction back in, with a pause, a log or a second agent checking the first.

The UAE businesses this affects

Salesforce has published research findings rather than a product release, so nothing changes in your software today. The findings matter to any UAE business that has moved beyond chatbots and now lets AI agents execute tasks in its systems. Marketing teams running campaigns through AI tools are in scope. So are sales and service teams whose agents update CRM data or respond to customers directly.

The post also speaks to two groups by name: leaders whose teams are adopting AI, and people who build AI products. In Dubai, that covers family groups rolling out enterprise AI across departments and start-ups building AI solutions for regional clients. Both face the same choice about where people stay in charge.

Businesses that work in two languages have one more reason to care. When an agent replies in Arabic, the reviewer needs the language and a feel for the brand to catch a wrong tone. If that skill fades because the agent always goes first, nobody catches a mistake on a sensitive topic before a customer does.

Steps a Dubai business should take now

For leaders building an agentic enterprise, Salesforce gives one practical starting point. Find the quality checks your strongest people have already built, often without giving them a name. Ask each of them which check someone else would miss if they tried to scale a slice of their work without the same expertise. Their answers show you exactly where an agent needs a check before it goes ahead.

From there, implement a simple plan for AI adoption one workflow at a time.

1. List the tasks you automate with agents and the data each one touches, then rank them by the cost of a mistake.

2. Add a human approval step before any autonomous workflow acts on anything customer-facing, irreversible or tied to money.

3. Turn your team's quality rules into written standards and test data that every agent is checked against.

4. Log each correction your people make as data, then review it every month to adjust each agent's instructions.

5. Set permissions for which agents may use customer data, and re-run your test sets whenever the underlying model changes.

6. Choose platforms with enterprise-grade security and clear data privacy controls before connecting them to live systems.

Many teams now build AI agents on low-code tools, which makes it easy to launch autonomous workflows fast and skip the checks. Good governance and clear guardrails give that work a solid foundation. They also cost less to add at the start than after a failure. Establish a small cross-functional group to own the review, including the people who do the work every day.

Over time, the competitive advantage comes less from having agents and more from agents that carry your strongest people's judgement. If you want a partner for that work, our AI agency in Dubai helps companies plan where agents act alone and where their people step in.

Keeping a strategist in the loop

We follow the same principle when AI reads client data for marketing work. The Adnika Dashboard: one live view of a client's ads, analytics and CRM data, with AI recommendations reviewed by the strategist. AI makes the first pass on the data, and the strategist makes the final call.

If you build AI products or agents

Salesforce's advice to builders is to make human thriving an explicit goal next to quality and productivity. AI can multiply good work, but only when the workflow protects people's agency and treats skill as core to the job. In a humane agentic enterprise, a little productive friction belongs in the design from day one.

For UAE tech firms building AI solutions for marketing clients, the work starts with defining the person's role in each task. Write stopping points, approval steps and logs into the specification, and test them like any other feature. An agent can be fully autonomous and still be checked. An agent that knows when to pause is easier to trust, and that trust is what moves it from a pilot into daily use.

Keep yourself at the helm

Salesforce says these patterns came up across many different jobs, and it's still pulling them together. The people who thrive with AI, it argues, aren't the ones who trust it most. They're the ones who decided on purpose where to keep themselves in charge. That choice is open to every agentic enterprise in Dubai, whatever its size, and it starts with the next agent you build.