Everyone wants an agent, we are all talking about them, and the hype man has us believing they will solve every problem.
At Concise we are big believers in AI, in technology, but are built on a couple core principles: the data is the most important thing and must be built as a good foundation first, and one should never overengineer - you should match the technical solution to the problem.
We see many businesses falling into the trap of finding use cases for agents that are actually just use cases for automation, bog standard data engineering (ETL): extract, transform, load. No fancy AI, just traditional technology and programming.
Most conversations start with "Where can we use an AI agent?" We think the better question is "What is the simplest way to solve this problem well?"
Sometimes that will be an automation. Sometimes it will involve AI in part. And sometimes an agent is appropriate.
So what is the difference?
At the simplest level, an automation follows a process you have already defined. Something happens, a set of steps follows, and you get an expected result. It's where there are well defined rules that are easily understood by a computer:
IF column is TRUE then perform TASK A
ELSE perform TASK B
For example: a form is submitted on your website, the details are added to your CRM, the right person gets a notification, and a confirmation email is sent to the customer. That is automation. There is very little judgement involved. You already know what should happen, and you are using technology to make it happen consistently without someone doing it manually every time.
An AI agent is useful when the next step is not always obvious. It might need to interpret information, decide what to do next, use different tools, ask for more information or change its approach depending on what it finds.
Imagine a small business has one general email inbox. Every morning, someone goes through the messages and forwards them to the right person. Invoices go to finance. New enquiries go to sales. Existing customers go to support. Job applications go to HR.
It is repetitive and it takes time, so someone suggests building an AI agent to manage the inbox. That might work, but it is worth looking at the process first.
If most emails are easy to categorise, you could build something much simpler. An automation could look for known senders, keywords or mailbox rules. AI could be added to classify anything that cannot be handled reliably with those rules. Anything still uncertain could stay in the inbox for a person to review.
You might automate 80 or 90% of the work without giving an AI system free rein to decide what to do with every email, and that is often the better starting point. The aim is to remove the repetitive work that doesn't need a human, while leaving genuine judgement calls to people.
There is also quite a lot in between. You might have a normal automation that uses AI for one particular step, such as summarising a document, extracting information from an email or classifying a customer enquiry. That does not necessarily make the whole thing an agent.
When we talk about agentic systems and AI native, there is a wide range in between. What's true for all these is that you must build everything on top of a clean, well structured and governed knowledge layer - get your data right first!
Start with the process, not the technology
This is where I think a lot of AI projects go wrong. A business starts with "we want to use AI," or "could we build an agent for this?" and then tries to find a problem for the technology.
It is usually much more useful to start by looking at the work itself. What is someone doing manually today? Where are they copying information between systems, or repeatedly checking the same thing? Where are they making a genuine judgement call, and where are they simply following a process that everybody already understands?
When a normal automation is probably enough
Automation is a good fit when the process is predictable. For example, an organisation might automate:
- sending an invoice reminder seven days before payment is due;
- copying new website leads into a CRM;
- creating a task when a deal reaches a certain stage;
- producing and emailing a weekly management report;
- checking a spreadsheet for missing information;
- sending a customer an update when the status of an order changes;
- moving files into the right folder based on their name or type.
None of these are especially exciting. They are often exactly the sorts of small, repetitive jobs that absorb time across a business, and fixing ten of them can have a much bigger impact than building one impressive AI demo.
When an agent starts to make sense
Agents become more useful when there is genuine variation in what needs to happen next.
Imagine instead that you want a system to research potential new clients. It might need to look at your existing customers, work out what a good prospect looks like, research companies online, decide which ones appear relevant, find useful information about them, compare them against your criteria, produce a shortlist, and explain why each company might be worth approaching. There is no single fixed route through that task. The system has to gather information, make decisions and adapt as it goes. That is much closer to an agent.
Another example could be producing a monthly business performance summary. An automation can collect the numbers and refresh a dashboard, but an agent might look across sales, finance and customer-service data, identify unusual movements, investigate possible reasons and draft a summary of what management should pay attention to. It interprets what it finds and adapts accordingly.
How do you know what to use?
Before deciding whether something needs an automation or an agent, we would ask:
- Do we already know the steps? If somebody could write the process down as a fairly simple flowchart, start with automation. If the person doing the task regularly has to stop, investigate something and decide what to do next, an agent may be more relevant.
- How much judgement is actually involved? This is an important one. A process can feel complicated simply because it is badly organised. Lots of spreadsheets, emails and manual checks do not automatically mean the work requires intelligence. Sometimes it just means the process needs tidying up.
- What happens when the system gets it wrong? Sending an internal report to the wrong folder is annoying. Sending the wrong information to a customer, approving a payment incorrectly or making a compliance decision is something else entirely. The consequences of a mistake should determine how much autonomy you are comfortable giving the system.
- Could we automate most of it and leave the difficult bit to a person? This is probably the question we would ask most often. Businesses can get obsessed with reaching 100% automation, but if you can remove 80% of a repetitive task and leave the genuinely unusual cases for a person, that can already be a very good result. The final 20% is often where complexity and cost start increasing quickly.
Your data still matters
Whether you choose an automation or an agent, both depend on the information underneath them. If customer information is spread across three spreadsheets, two systems and somebody's inbox, an AI agent does not magically solve that problem. If there are three different definitions of an "active customer," the agent still has to work with that ambiguity. If nobody trusts the sales data, adding AI on top of it does not make it more trustworthy.
This is why we still think the less glamorous work matters: getting systems connected, agreeing definitions, improving data quality, making sure the right people have access to the right information. Those things make traditional reporting better, but they also make automation and AI much more useful. Good data foundations are not something you finish before doing anything with AI. They are part of making AI work properly.
The simplest solution is often the best one
There is a huge opportunity for organisations to use AI, but at Concise we do not think the opportunity is in turning every process into an autonomous agent. It is in looking properly at how work happens today and asking where technology can remove friction.
The important thing is that you choose the technology because it fits the problem, not because it is the newest thing in the sales pitch. Start with the problem. Understand the process. Use the simplest tool that solves it. Then add complexity when the complexity actually earns its place.
Not sure whether your next project needs an automation, an agent, or a bit of both? Contact us and we'll help you work it out.