Next-Level AI: From Conversations to Agents That Act

B4 AI Ecosystem Roundtable | 24 September 2026

Marnie Wills | Business With AI Strategist

Lead smarter. Stay human.

Check-in: Intro + where are you today?

Round the table, 90 seconds each

1

Which AI tool(s) do you & your team actually use day to day?

2

What is one thing AI already does well for you, that it didnt go 6-12months ago

3

What is your biggest bottleneck right now, that you think AI should be able to solve.

From prompting to delegating

1

Copy & Paste → Execute

Give a goal, no need for step by step guidance

2

Static Data→ Live Data

Where is your data? Is it in the cloud?

3

Attach → Retrieve

Connect your softwares so AI fetches what it needs.

What is an agent & What makes an agent useful

A traditional chatbot is mainly designed around conversation. You give it a prompt, and it generates a response.

An AI agent can go a step further by reasoning about a goal, planning what needs to happen, using external tools, taking actions, observing the results, and continuing the task based on what it finds.

Think of it this way: a chatbot is mostly question → answer, while an AI agent can work more like goal → plan → action → result → adjustment → completion.


Therefore: AI agents are software agents that use artificial intelligence to make decisions and take actions to complete specific goals. They interact with digital environments like emails, CRMs, and calendars, and analyse your inputs and respond accordingly.

Memory

Knows your business, voice and clients.

Context

Knows the task, the rules and what good looks like.

Tools

Can reach your email, calendar, files and apps.

Workflow

Follows steps to a goal, not just one answer.


Step 1 of going agentic: Customise

*LLM's are my first recommendation ….followed by other Agent Building platforms

Security: your agent safety rules

Least Privilege

Give access to only what this task needs, read-only where possible.

Human Approval

Before anything is sent, posted, paid or deleted.

Data Hygiene

Keep client and confidential data out of tools you haven't vetted. Check where data is stored and whether it trains models.

Business Accounts Only

Not personal ones. Know which agents are running.

Keep a Log

Know what the agent did, and know how to switch it off.

Treat Outside Content as Untrusted

Emails, web pages and documents can all carry hidden instructions.

Why connectors are safe, and where they aren't

No password handed over

A connector isn't handing over your password. You sign in on the app's own login page and approve specific permissions. You get a permission token, not your password shared. Revoke it any time by disconnecting in the app or in the third-party app's security settings.

It only sees what you can see

On Team and Enterprise plans, users connect individually — the AI can only access tools and data that person already has access to. If someone can't see payroll in Xero, their AI can't either.

Admins control what's allowed

Enabling a connector for the organisation doesn't grant everyone access. Each person still authenticates individually. Owners can also disable specific tools that take write actions — allow "read" and block "send, edit, delete".

There's an audit trail

Every call the connector makes is logged — showing time, user, action and resource. The connector needs a business account; personal accounts can't be used. (Microsoft 365 example.)

Step 2 could be other AI Agent platforms

  • Lindy.ai: plain-English agents for multi-step workflows, Slack-native assistants and meeting intelligence, with 1,000+ integrations. It suits inbox triage, scheduling and lead follow-up.
  • AgentKit.ai: your no-code chatbot trained on your own content, embedded links.
  • Zapier Agents: the easiest no-code option for non-developers, using plain-language triggers and actions. It's the natural next step for anyone already running Zaps.
  • Gumloop: the AI inside a workflow can make a judgement call about what happens next, rather than following a fixed chain.

Don't build an agent. Solve a problem


Your AI Agent should be doing all of these Productivity, business efficiency and money generating

Ensure meets all three criteria:

  • Repetitive — at least weekly
  • Predictable output — you know what good looks like
  • Low risk — mistakes get caught before they matter

Action & Or Discussion

How can you add more memory quickly if needed?

What software's can you connect & why?

'What repeatable prompts or processes have you been using in your conversational AI that you could transfer to a skill? (top tip ask your AI)


Live demo:

Memory

Connectors/ plug ins

Skills


Scheduled tasks: The best little hack

Set a task once, and it runs on a schedule - a weekly briefing, a Monday priorities summary, a daily inbox triage.

Best For

Work that is repetitive, predictable and low risk.

Design the Checkpoint

Always design the approval checkpoint: what gets drafted for you, and what is allowed to send itself?

Discussion: What should your AI budget actually buy?a

Most leaders measure AI on speed. The real return is capacity, quality, re-skilling, assets, monetisation

Capacity

What is your teams capacity if they have access to agents?

Quality

Does the output waulity of the human improve

Re-skilling

Whose role is shifting, and how are you supporting that?

Assets / Monetisation

What IP is being created, repurposed and what is the value of this?

Security: the five risks to know

1

Over-permissioned Access

Connectors given far more than the task needs.

2

Prompt Injection

Hidden instructions in an email, webpage or document that hijack your agent.

3

Data Leakage

Confidential client or company data going to the wrong tool, account or model.

4

Irreversible Actions

Sending, paying, publishing or deleting without a human check.

5

Shadow AI

Staff using unapproved tools and personal accounts, so nobody sees the risk.

Security in your business

01

Who owns AI risk?

Who in your organisation owns AI risk today?

02

Near misses

Have you seen, or nearly had, an AI mistake or data slip? What happened?

03

What would make you feel safe?

What would you need in place — policy, training, approved tools — to feel safe giving AI more access?

04

Legal obligations

What does your client contract or UK GDPR obligation mean for how you use AI?

Your clients are more AI-informed than ever

Clients now walk into meetings with AI-generated research, questions and expectations.

How is this changing conversations?

How is this changing the conversations you have with clients and customers?

Help or hindrance?

Where does it help you, and where does it challenge your expertise or pricing?

Your response

What is your response: how do you add the value AI can't — judgement, trust, relationships?

You came in prompting. You're leaving delegating.

Your commitments

Complete these three sentences and share one aloud:

1

Running

By [date], my agent or workflow [name] will be running [how often].

2

Off Your Plate

The task it takes off my plate is ______.

3

Safety Rule

The safety rule I'll put in place is ______.

4

Success Signal

I'll know it's working when ______.


Lead smarter. Stay human.


Next Roundtable

AI Ecosystem Roundtable: 3 December 2026 at B4.