Agentic AI: From Tools to Teammates

- Emails from Friday you never got to, and probably more that have come in over the weekend

- A presentation to prepare

- A new client or new project to onboard and scope

- A reconciliation that isn't quite over the line

- Research into global trends and the latest news

You're already working through the list in your head. Keep that picture in mind. We'll come back to it.

Chat AI vs agentic AI

First, a pause to be clear about what agentic AI is and how it differs from chat or generative AI.

Most of us are familiar with chat AI: tools like ChatGPT, Copilot, Gemini. You type something in, it gives you an answer. It's genuinely useful for drafting emails, summarising documents, explaining things, brainstorming, but fundamentally it's reactive. It's always waiting for your next prompt.

Agentic AI works differently. You give it a goal, not a question, and it works out how to achieve that goal. It breaks the goal into steps, uses tools, takes actions, checks its own work, and delivers something back to you.

The shift is from answering questions to actually completing tasks: doing the thing itself, not just returning an arrangement of words. It can complete tasks like building a presentation, making a reservation, or paying something on Amazon.

Four things make that possible:

1. Planning: it breaks a goal into logical steps without being told what they are.
2. Tool use: it can search the web, use APIs, write files, work in Excel or PowerPoint.
3. Memory: it remembers the context of where it is and adjusts based on what it finds along the way.
4. Feedback loops: it checks its own work, iterates, and only delivers when it's satisfied the goal has been met.

The moment everything changed

The rate of change in AI is fast, but there was a particular moment last year that really brought the agentic era into being.

An independent developer, working over a weekend, built an AI agent called OpenClaw. The concept was simple: it was built so it could control a computer. It could click buttons, type into fields, browse websites, and interact with software the same way a human does. (Except passing an "are you a robot" check. For that, the agent would still need to hire a human.)

Behind the scenes, what's happening is that it's calling a large language model in a loop, along with other tools. Take booking a restaurant as an example: "book Hook for 2 people, 7pm next Saturday." It will find the website, send a snippet of the page to an LLM, wait for instructions on where to click and what to fill in, and repeat that loop until the task is done.

OpenClaw itself isn't really what this piece is about. If you're curious to try it, a virtual environment is strongly advisable, rather than giving it access to your whole computer. What made it interesting is that the developer released it as open source: totally free, immediately available to anyone. That, in turn, pushed every major platform to release their own agents and build agentic capability into their products.

Where the major platforms are now

Microsoft Copilot, Google Gemini, Anthropic's Claude, OpenAI: they've all moved quickly to build agent capabilities. These are embedded into tools most people are already using. This isn't something you need to go and adopt from scratch. In many cases, it's already there, waiting.

The question is whether you're ready to use it effectively.

A simple AI capability ladder

Worth asking: where does your organisation sit on this ladder, and where do you think it will sit in one, two, or five years?

- Level 1, personal productivity: most people in the organisation have access to a tool like Copilot for document drafting and summarising.

- Level 2, tools with AI built in: for example, Copilot in Excel to analyse data and create charts, but not yet making the most of it.

- Level 3, AI as part of a process: an automated process categorises and archives old files or tickets, and where the category isn't obvious from the name alone, an AI step reads the document, understands the context, and categorises it correctly.

- Level 4, AI owning multi-step workflows: agents complete whole workflows autonomously and deliver the outcome for review, rather than AI being one step inside a programmatic process.

The hidden cost

A quote from a recent client conversation: "A lot of time is spent by everyone moving around File Explorer to find relevant data in folders."

That's worth sitting with, because it highlights something easy to miss when thinking about where time actually goes. The technical expertise, the knowledge work, understanding a client's risk profile, reconciling accounts, building financial statements: that's the valuable part. But a huge amount of effort happens in between: finding the template, locating the files, re-reading an email, reconstructing the context again and again. That in-between work is exactly what agents are built to remove.

Worked example: meeting minutes and follow-up

Today, a human takes notes during a meeting or tries to reconstruct them afterwards from memory. Later, they reformat them into the firm's template, draft the follow-up email with action points, send it, and then at the next meeting find out what's been completed and what hasn't.

With an agentic co-worker, that process changes significantly. An agent can transcribe from an AI note-taker, format the output into a predefined template, and draft the follow-up email with actions, all before anyone's made it back to their desk from the meeting room.

A human then reviews the minutes for accuracy and tone, decides what's privileged or sensitive, chooses what goes in the email and what doesn't, and presses send when satisfied. You can give the agent authority to send it directly, though that usually takes time to build trust in. From there, the agent can track outstanding actions, nudge people, and surface what hasn't moved before the next meeting.

Worked example: a board pack

This one shows the difference between using a chat tool today and using a true agentic workflow.

With a chat tool today, you might say: "help me write commentary for slide three." It does it. You say: "I'm not happy with that, make it more concise." It revises. That iteration loop is manual; every step needs a prompt.

With an agentic workflow, you give the goal: produce this month's board pack. From there, the agent takes over. It pulls the previous board pack as a template, searches emails and documents for the information, flags rolled-forward elements that need review, perhaps researches relevant news, compiles everything into the format, adds commentary, and delivers it for review.

That feedback loop, checking its own work, iterating, flagging anomalies, is built in. You're not managing it step by step. You're receiving a first draft that's already been through several rounds of internal review. You give the goal; the agent figures out the steps.

Where this fits across industries

IT and operations, research synthesis, reporting, and everyday admin work all have strong potential for agentic AI, and it's already happening. This isn't specific to any one sector.

- IT and operations: an agent can triage helpdesk tickets, monitor alerts, and close routine requests.

- Research synthesis: an agent can monitor news and trends and produce structured briefings.

- Reporting: pulling information together into a template, as in the board pack example above.

Areas like compliance monitoring, regulatory due diligence, and KYC-style checks are places where agentic AI can speed things up significantly, but shouldn't yet be given full autonomous authority. These require human oversight. Take due diligence as an example: there may be hundreds of documents to read, cross-reference, and review. The heavy lifting can be done by an agent, with a human then reviewing and validating, cutting the time required from days to hours.

Anywhere there are structured processes, large volumes of information, or repeated workflows, there is real potential here.

Worked example: regulatory change

Today, staying on top of regulatory change is largely a manual process. Someone monitors news sources and inboxes for updates from regulators. When a change is detected, a compliance officer reads it, maps it to relevant internal policies, assesses the impact, and drafts a response.

With an agentic workflow, human effort moves to a different place. The agent monitors news sources, inboxes, and regulatory feeds continuously. When it detects a relevant change, it compares it to internal policies and procedures, reasons through the likely size of the impact and the actions required, and produces both an impact assessment and a gap analysis, ready for a compliance officer to review. The agent does the gathering, the mapping, and the drafting. The compliance officer does the deciding.

What actually changes

Some work shrinks. Manual data collation, file searching and retrieval, template population, first-draft reporting: agents are very good at these. They won't disappear overnight, but the volume reduces significantly over time.

Some work grows. Reviewing AI-prepared outputs, exercising professional judgement, client-facing conversations, strategic thinking, quality control: these become a larger share of the working day. They should, because that's where real expertise sits.

And some work is entirely new. Designing AI workflows, prompt and process architecture, output validation frameworks, governance and audit trails, AI risk assessments: someone has to own those, and they're new, important roles.

The composition of work changes. What stays, and becomes more valuable, is human judgement.

The common fears

The most common concerns when discussing AI with businesses are data security, and accuracy, specifically bias and hallucinations. These are legitimate, and need addressing in any implementation.

On data security: you don't have to send anything to a global cloud. Private models can run within your own infrastructure or a controlled, permissioned environment, so nothing leaves your network. This is increasingly standard practice for regulated industries, and the tooling to support it is mature.

On bias: this already exists in human processes. AI needs the same controls applied to human decisions: review, validation, oversight. Good prompt engineering and well-designed agent architecture help significantly here.

On hallucinations: agents need to work from trusted, scoped data sources. They should show their reasoning and cite their sources, and a human should always review outputs before anything consequential is acted on. Build those checkpoints in, and hallucination risk becomes manageable.

These aren't reasons not to proceed. They're design requirements.

When an agent goes rogue

With this technology, nobody fully knows what will happen, and without the right controls, a poorly designed system can do the unexpected.

One widely reported example: in February 2026, a safety researcher at a major AI lab was experimenting with an OpenClaw-style agent to clear out an overstuffed inbox. She'd explicitly instructed it not to delete anything without confirmation. It began mass-deleting hundreds of emails anyway, and she had to physically get to the machine and unplug it. The apparent cause was that the model became overwhelmed by the volume of data, ran out of context, and lost track of its own safety instructions. The system hadn't been built with proper governance, and it had been given too much unsupervised control. It's the kind of incident that's pushed the industry to take "kill switches" and stricter human-on-the-loop requirements more seriously.

Source: Business Insider, "Meta AI alignment director's OpenClaw agent deleted her emails," February 2026.

Regulatory highlights

The EU AI Act's core framework isn't enforceable everywhere, but its design principles are worth following regardless. Four things need addressing in any AI tool or framework:

- Explainability: can you describe how an output was reached?

- Accountability: who owns it?

- Human oversight: where does a human sit in the process?

- Audit trails: can you show what the system did, and why?

These aren't obstacles. They're a useful framework for designing AI systems well. Build to these standards, and you're building something trustworthy. Local regulators, including Guernsey's GFSC for regulated financial businesses, have generally taken a supportive view: the expectation is that businesses start adopting this technology, implemented thoughtfully.

Where should the human sit?

On a scale from fully manual to fully automated, there are two useful positions.

Human in the loop means the agent pauses and waits for approval before executing each significant action. This is right for anything regulated, client-facing, or irreversible. If the consequences of getting it wrong are serious, a human should approve every step.

Human on the loop means the agent acts, and a human monitors and can intervene. This suits high-volume, well-defined tasks where mistakes are recoverable: the agent sends a draft email, and you can catch it before it causes harm; the agent files a document, and you can review the log and correct it.

The right model depends on the task, the risk, and the organisation's comfort level.

It helps to think of it as a matrix of stakes against reversibility. Stakes: what's the consequence of an error? Summarising meeting notes is low stakes. Flagging a compliance breach is high stakes. Advising on a client's finances is very high stakes. Reversibility: can the action be undone? Drafting an email is reversible. Sending it isn't. Filing a regulatory return isn't. The more irreversible the action, the more human oversight is non-negotiable.

- Low stakes and reversible: the agent can take on the workflow with full autonomy, for example summarising a weekly team meeting.

- High stakes and irreversible: human in the loop, checking each part of the process.

- High stakes but reversible, or low stakes but irreversible: human on the loop, reviewing the output afterwards.

Anything requiring physicality, for now, still needs a human.

What breaks first

Here's a scenario worth thinking through.

You deploy your first agent. It pulls data from multiple sources and produces a clean, confident output that looks right.

Then another team, solving a similar problem, builds their own agent, using slightly different data and slightly different logic. Now there are two outputs. Both look right. But they're different, and there's no clear audit trail showing how either was created. Which one do you trust?

This is the trap organisations are about to fall into. Nobody made a bad decision; each team just made a local one. It's exactly what happened with the rise of SaaS software: every team adopting its own tools, none of them working together, until eventually there's inconsistency, duplication, silos, and a technology stack nobody can fully see or govern.

The fragmentation trap

This is what fragmented agent adoption looks like in practice: multiple agents, multiple data sources, no shared foundation, no common logic. Over time, outputs diverge, not dramatically, not obviously, but enough to matter when someone needs to rely on them.

The problem isn't capability. The technology works. The problem is coordination: a human and organisational challenge as much as a technical one.

Build the foundation first

The answer is to build the foundation before deploying the agents. Five things need to work together for any agent to be genuinely useful:

- Data: the agent needs to be connected to the systems where your data actually lives, in a form it can reliably use.

- Knowledge: the firm's context, policies, templates, and procedures need to be structured and accessible. The agent needs to understand your world, not just the general world.

- Identity: each agent needs scoped permissions, only accessing what it needs for the task at hand. This is how security and auditability are maintained.

- Governance: audit logs, classification frameworks, controls. You need to be able to see what each agent did, when, and why.

- Reasoning: choosing the right frontier models, deliberately, for the tasks being deployed.

Build these five things as a coherent foundation, and every agent deployed from that point is faster, safer, and smarter than it would have been otherwise. The foundation builds trust, and trust is what lets you scale.

This is where Concise comes in: helping organisations build the foundation for agentic AI.

Back to Sunday evening

Let's come back to that Sunday evening picture. There are emails you didn't get to on Friday. A board pack to create. A client to onboard. Reconciliation to review. Compliance news to research.

Now imagine an agentic co-worker in place. By the time you walk in on Monday morning: your emails are summarised and prioritised by the action required, so you know exactly what needs attention. The board pack is created and ready for review. The supporting documents for the new client review have been gathered, with a one-page summary waiting. The reconciliation is complete, with anomalies flagged for a closer look. And the regulatory changes: monitored, flagged, compared to internal documents, gap analysis completed, recommended actions summarised, ready for a decision.

Every step, every action, every decision those agents made is logged, transparent, and reviewable.

The Sunday scaries don't go away because someone else did the job. They go away because by Monday morning the groundwork is done, and there's space to focus on the part that actually needs a person.

The question worth sitting with isn't whether AI will replace people. It's how to use it to replace the parts of work that don't need people, so the parts that do need human intelligence, judgement, and relationships get the attention they deserve.

None of us knows exactly where this technology takes us. But two things are certain: things will change, and waiting is not a neutral choice. Guernsey has built a genuinely world-class business environment, and that same opportunity exists right now with this technology.