Chatbot theatre versus actual work
The easiest way to sell AI is to put a chat window on a website. It demos beautifully, it is visible to everyone in the business, and it very rarely changes a number that matters.
The unglamorous version is where the money is: a process that used to consume four hours of someone's week now runs by itself, correctly, every time. Nobody sees it. It just quietly stops costing you.
"Agentic" should mean the work gets done, not that there is a chatbot in the corner.
What a real AI automation agency does
- Maps the process before touching a tool. How it actually happens, not how the org chart says it happens.
- Finds the repetitive, rule-based steps. Data moving between systems. Copy-paste. Chasing. Reconciling.
- Fixes the process before automating it. Non-negotiable, and the step most agencies skip.
- Automates the hand-offs, not the judgment. Machines are excellent at moving things and poor at deciding what matters.
- Measures the hours back. If nobody can say what was saved, nothing was, and that means the underlying data has to be trustworthy before anything acts on it automatically.
Notice how little of that is about AI specifically. Most of the value in "AI automation" is plain automation done well, with AI where it genuinely helps, reading messy inputs, summarizing, classifying. The AI is a component, not the point.
The task that should have been automated years ago
The one we find most often: a team is copy-pasting leads between a form, a spreadsheet and the CRM by hand. It works, until it does not. Records get missed, follow-up slows to days, and eventually nobody trusts the numbers in any of the systems.
An agent handling that hand-off end to end fixes three things at once. Follow-up happens in minutes. The reporting finally lines up, because there is only one path the data can take. And the team gets back the hours they were spending on clerical work they never should have been doing.
This is what we mean by leverage, not layoffs. AI agents free people for the work only people can do: the conversations, the judgment, the relationships. Nobody was hired to be a copy-paste machine.
What you should not automate
- A process that is broken. Automate a broken process and you just get to be wrong faster. Fix it first, then encode it.
- Genuine judgment calls. If a step needs someone to weigh context, leave it with a human and automate everything around it.
- Something you do twice a year. The automation will cost more than the task.
- Anything you have never done manually. Do it by hand ten times first. You will discover what actually matters.
- Customer conversations, by default. Some are worth automating. Most people can tell, and resent it.
And be wary of any agency that never says no to a use case. The value is in choosing the right five processes, not automating fifty. That is how our AI automation work is scoped, process first, tools second, hours saved measured at the end.
How to scope an automation so it actually lands
Most automation projects fail the same way software projects fail: too big, too vague, and measured by nothing in particular. The fix is unglamorous discipline.
- Map the process as it really happens. Not the version in the handbook. Sit with the person doing it and watch. You will find steps nobody documented and workarounds nobody admitted to.
- Count the hours. If nobody can say how long the task currently takes, nobody will be able to prove the automation helped.
- Separate decisions from data movement. The moving should be automated. The deciding should stay with a person, supported by better information. Automating a judgment call is how you end up with a confident machine making your worst decisions at scale.
- Fix the process before encoding it. This is the step everyone skips because it is boring and it delays the interesting part.
- Define the number that says it worked. Hours returned, errors eliminated, time-to-response. Agreed before you start.
The category is broader than "AI" suggests. A great deal of what gets sold as AI automation is closer to classical process automation, what the industry has long called robotic process automation, with AI added where it genuinely helps: reading messy input, classifying, summarizing, drafting.
That is not a criticism. Plain automation done well is enormously valuable. But it does mean you should be suspicious of pricing that assumes you are buying something exotic when you are mostly buying careful plumbing, well-built, well-measured, and worth every dollar when it is scoped like this.
Where the AI part actually earns its place
Worth separating the two things being sold together. Most automation is deterministic: if this, then that. Reliable, testable, boring, valuable. AI is useful precisely where determinism fails.
- Reading messy input. Invoices, emails, PDFs, handwriting, the free-text box where customers write anything. Rules break here; models cope.
- Classifying and routing. Is this enquiry urgent? Is this a complaint? Which team should see it? These are judgments with fuzzy boundaries.
- Summarizing. Turning a long thread into the three things a human needs to know before they reply.
- Drafting. A first draft a person edits is faster than a blank page. A first draft nobody reads before sending is a liability.
Outside those, plain process automation is usually the better tool: cheaper, more predictable, and far easier to debug when it misbehaves at 2am. Using a language model to move a row from one system to another is an expensive way to do a simple job unreliably.
The other half of it is data. Automations make decisions on the information you give them, so if your tracking and data are inconsistent, the automation will make confident, consistent, wrong decisions, which is worse than the manual process it replaced, because at least a human notices when something looks odd.
Frequently asked questions
What do AI automation agencies actually do?
The good ones map how a process really happens, find the repetitive rule-based steps, fix what is broken, then automate the hand-offs (data moving between systems, chasing, reconciling) while leaving judgment calls with humans. The measure of success is hours returned, not tools deployed.
How do I spot a bad AI automation agency?
They lead with a chatbot, they never say no to a use case, and they skip straight to tooling without mapping your process. A chat window on a website demos well and rarely changes a number that matters. Ask what they would refuse to automate.
What should not be automated?
Broken processes (fix them first, or you just get to be wrong faster), genuine judgment calls, tasks you only do twice a year, anything you have never done manually, and most customer conversations. Automating a mess makes it a permanent, faster mess.
Does AI automation mean cutting staff?
That is not how we use it, and it is not where the value sits. The point is leverage: nobody was hired to copy-paste between a form and a CRM. Automating that frees people for the conversations, judgment and relationships that only people can do.