Agentic AI is the Future of Business Administration

Since the launch of ChatGPT way back in November 2022, the world has changed a lot.

AI is everywhere; every app you use has a (mostly useless) chatbot. The news cycle is dominated by it, the hype is out of control, and the economy is almost certainly in a bubble of some sort, driven entirely by tech stocks.

Certain industries, for example, software development, have been turned inside out by AI with huge leaps forward in productivity, but the majority are fundamentally still operating as they were, with perhaps some modest improvements in productivity across certain functions, such as marketing.

Why is software development so far ahead? It's because agentic AI tools like Claude Code or Codex have been mature for well over a year.

The difference between modest productivity improvements and a complete sea-change in how your business is run comes down primarily to whether you are using agentic AI or not.

Agentic AI?

With traditional tools like ChatGPT, you ask the LLM a question, and it responds with an answer. It may think for a bit or use a couple of tools in the cloud to produce that answer, but fundamentally it's a single cycle: ask -> respond.

With agentic AI, you run a local "harness" - a small piece of software on your machine. This harness orchestrates a much longer conversation with an LLM on your behalf, with the ability to run tools locally.

Claude Code is a harness, as is Claude CoWork, OpenAI Codex, and Microsoft CoPilot CoWork (which uses Claude under the hood).

With this set-up, your harness executes what is called an "agentic loop", where the LLM analyses your prompt and returns a request to execute a tool locally, which could be reading a document, saving a text file, or running a script of some sort. The harness then runs the tool request and sends the response back to the LLM, which analyses it and comes back with either another tool request or a response.

In some instances, this loop can run for hours without interaction from the user, with many thousands of requests going to the LLM.

This process is heavy on compute, but the results can be astounding.

In the context of software engineering, it allows the LLM to effectively build a fully working, fully tested piece of software on your machine, with minimal input from the user aside from guiding the feature set.

The tools

Part of what makes this process so powerful are the tools that you give your harness access to. At a basic level, the harness can read and write files, write and execute simple scripts, which makes it useful for software development, but it might not be able to access your CRM for example, unless you specifically give it a tool to do that.

Tools like this, which allow your harness to communicate with a third-party platform or piece of software, are called Model Context Protocol (MCP) adapters.

If you want your harness to be able to communicate with your CRM, you install the MCP for that CRM, authenticate and off you go.

You might also want to install an MCP for your ticketing system, your billing system, and SharePoint. The more tools you connect, the more powerful your set-up becomes.

The possibilities are endless

In the context of business administration, the potential applications are striking, for example, you could:

• Produce a detailed client activity summary of all work undertaken over the last month, including tickets completed, time logged, bills sent / outstanding, and sentiment

• Produce both an agency side and client side version of a project update summarising progress, decisions made, outstanding tasks and blockers

• Produce a board pack taking into account previous decisions made, transcripts and performance updates

• Carry out a deep compliance audit, pulling data from a number of different systems and compiling it together into a daily exceptions report

• Open a support ticket, investigate the cause, propose and apply a fix, and reply with a simple approval to proceed. Then, draft a response to the support ticket

• Work through all open tickets in your ticketing system, categorising and escalating if required

For each of these tasks, you would typically spend a bit of time going back and forth within your harness until you get the result that you are looking for. You would then save the process as a "skill".

A skill is like a set of instructions for your harness, which would typically include a prompt, a script or two and maybe some other files like a template.

A skill takes whatever it is you have produced and makes it repeatable. All of the examples I provided above could very easily be crystallised into repeatable skills. You could have simpler skills too, such as a document template skill, which takes a chunk of copy and puts it into a branded document, or a proposal skill, which hoovers up a load of information relating to a project and builds a branded proposal.

Once you have a repeatable skill, you can then share this skill with other members of your team so that they can use it as well.

Think about this kind of setup:

• Each member of your team has an approved agentic harness installed on their machine

• Along with this harness, they have a set of approved tools installed

• Finally, they have access to a carefully curated library of shared skills, which they can install and use

If you have a skill that can produce a high-quality, branded proposal, then all of a sudden, every single member of your team will have the capability of doing this in a matter of minutes.

If you have a skill that can run a detailed compliance audit on a potential new client, your sales team would be able to run that check before going to the initial meeting, without bothering the compliance team.

In theory, all of the monotonous, repetitive work involved in the day-to-day running of a business could be condensed into a skill and delivered in a fraction of the time, allowing team members to get significantly more work done each day.

This isn't theoretical either, not out of reach - the technology is here, it's accessible, it's cheap, and it's easy to set up.

There is a caveat

In order for your agentic harness to work well, your data needs to be in good shape. Your client communications should be stored in a structured platform such as Basecamp or Zendesk, your billing should be handled out of a tool like Xero, your client folders should be organised in a consistent manner, and your meetings should be recorded, transcribed and saved down in an organised manner.

If your data is unstructured, disorganised and spread across SharePoint and email, then you will struggle to find a tool that can easily query that data and pass reliable information through your harness and back to the LLM.

Keeping your data in good shape has always been a prerequisite to running a disciplined business, especially in the finance sector, where good KYC and AML practices are essential. With the advent of agentic AI, it is no longer optional.

Those firms whose data is in good shape have a real head start on this journey because they can start their agentic rollout today.

Firms whose data is a mess can try, but as soon as they point their harness at an unstructured data source, they'll start to see gaps and inaccuracies in the LLM's responses, probably to the extent that they won't be able to rely on them.

Thankfully, sorting out and organising data is something that LLMs are genuinely good at, so if you are concerned about the quality of the data in your organisation, an agentic AI tool will almost certainly make the clean-up process much less painful.

The journey

People often ask me how they can roll out AI within their organisation and achieve the productivity gains that it has promised for so long. As discussed in this article, agentic AI is the key, and the journey to rolling it out into your organisation doesn't have to be complicated or expensive.

1. Sort out your data - ensure that your client data is in a CRM, capture your clients' comms in a structured way, keep your client folders organised, and stay disciplined with new data coming into your organisation.

2. Choose a tool - some agentic harnesses are better than others, but for day-to-day business administration, they are all good enough. Your choice should be governed by your data protection and IT security policies. Are you happy with client data being sent to Anthropic's server in the US, even with a bulletproof data protection agreement? If not, then you should probably look at another provider.

3. Consider your toolset - what platforms do you want your harness to connect to? You don't have to connect everything all at once, but remember, the more you connect, the more useful the harness will be. The impact of the five tools together will be greater than the sum of those parts. For each tool, make sure you are across the permissions - you don't want a user having more access than they should because they are connecting to your CRM via the tool rather than directly.

4. Start experimenting with what you can do with your initial toolset, and flesh out a few simple skills to get you going

5. Spin up your skills library. This could be a SharePoint folder, in Notion, Basecamp, or wherever you store your shared files. Develop a skill that allows your staff to harness the skills library so that they can install skills and update skills. Make sure that within each skill, there is a check that it is running the latest version.

6. Bring a small team of early adopter staff into your agentic ecosystem, training them on the tools and the skills library. Run it in a controlled, beta phase for a period, assessing how the staff are getting on, and rolling out improvements along the way

7. From there, you can start to consider a wider rollout

Once you have completed this journey, you will find that people start using the skills repository more and more. The quality of your team's output will improve, and they will turn work around much more quickly. As it grows, you can start introducing categorisation or team-specific skills, depending on how you are using it.

In conclusion

Few businesses are using this approach to rolling out AI currently, but it's a powerful way to start achieving real gains across general business administration.

Here at Indulge, we've been using agentic AI for software development pretty much since Claude Code was released in Q1 2025. More recently, however, we have been applying agentic AI to our wider business using the shared skills approach, and our library of skills is growing by the day. We are seeing significant gains across the whole business as a result.

Right now, this approach to rolling out agentic AI is at the frontier, but in a couple of years, every business will be doing it, from board level down to customer support.

My advice to any business would be to start now, while your competitors are still hesitating.