Two different things wearing one name
It helps to separate two ideas that get bundled together, because they have very different costs and reliability.
Rule-based automation is deterministic. When a form is submitted, create a record, send an acknowledgement, notify the owner, remind after two days. It does exactly the same thing every time. It is cheap, predictable and easy to verify — and it should be used wherever it can be.
AI automation adds a layer that handles language and light judgement. Reading a message written in mixed Hindi and English, working out that it is a bulk order enquiry rather than a complaint, extracting the quantity and delivery location, drafting a suitable reply.
Good systems use both, with rules doing as much as possible and AI reserved for the genuinely messy parts. A provider who reaches for a language model to solve a problem a rule could handle is adding cost and unpredictability for no benefit.
How to tell whether a task is worth automating
Four questions, in order. If a task fails the first two, stop.
- Does it happen often? Several times a week, at minimum. A task that runs twice a month rarely repays a build.
- Is it broadly the same each time? Not identical, but recognisably the same shape. Tasks that are different every time need judgement, and judgement is the expensive kind of automation.
- Is the information written down somewhere? If the answer lives only in one person's head, that has to be documented first — which is worth doing regardless, but it is the real project.
- What does failure cost? A wrong acknowledgement is recoverable. A wrong quotation sent to a customer is not. High-cost failures need human approval steps, which changes the design and the value.
Tasks that pass all four are usually unglamorous: enquiry capture, acknowledgement, follow-up, data transfer between systems, status updates, review requests, report assembly. That is the honest list.
Where to start: the enquiry path
If we can only fix one thing for a business, it is almost always what happens between an enquiry arriving and a person replying.
In most small and mid-sized businesses the picture looks like this. Enquiries arrive on four channels — website form, WhatsApp, phone, social messages. There is no single list. Whoever sees one deals with it. On a quiet day this works. On a busy day, an unpredictable share of enquiries are answered late or not at all, and nobody knows which ones, because there is no record of what was missed.
The fix is not sophisticated:
- Every enquiry, from every channel, into one list, with its source recorded.
- An immediate acknowledgement, at any hour, that tells the person what happens next.
- An owner assigned, with a reminder if nobody has responded within an agreed time.
- A follow-up sequence for enquiries that go quiet, which stops when the customer replies.
- An outcome recorded — won, lost, not relevant — so reporting reflects reality.
It is measurable, it requires nobody to change how they work, and the improvement usually shows within the first month. Everything more ambitious is easier to justify once this exists.
What AI adds on top
Once capture and follow-up are solid, AI extends what is possible in a few specific directions.
- Understanding unstructured messages. Classifying an enquiry by type, urgency or service without the customer filling in a form.
- Answering from your own material. A chatbot or assistant that responds using your price lists, policies and service details rather than general knowledge. See what an AI chatbot is.
- Extracting data from documents. Pulling structured details out of invoices, purchase orders and specifications instead of retyping them.
- Qualifying before a human joins. Asking the questions your sales process needs, so the conversation starts informed. See what an AI agent is.
- Drafting. Preparing a reply, a summary or a quotation for a person to check and approve.
Note how many of these end with a human. That is deliberate, and it is what separates automation that helps from automation that creates work.
What it costs
Three separate numbers, and any proposal that blends them is worth questioning.
Build fee. One-off, driven by how many systems are involved, how complex the logic is, and how much source material needs preparing. A single well-defined workflow is a small project. Multi-system operational automation is not.
Running costs. Ongoing and usage-based: AI model usage, messaging fees, automation platform subscriptions, hosting. These scale with volume, so they rise in a busy month. They should be billed to accounts in your own name, so you can see exactly what is being used and can leave.
Maintenance. Optional but worth budgeting for. Third-party platforms change their APIs and policies without notice, and automations need tuning after launch. A build with no maintenance arrangement will eventually break, and the question is only whether you find out gracefully.
The guardrails that should be non-negotiable
Every automation touching customers should have these, and you should ask for them explicitly.
- Defined scope. The system has a specific job. Requests outside it get redirected, not attempted.
- Grounded answers. Responses come from your approved material, and the system says so when it does not know.
- Escalation rules in writing. Complaints, refunds, negotiation, and anything medical, legal or financial go to a person by rule.
- Logging. Every conversation and action reviewable by your team.
- Disclosure. Customers are told they are dealing with an automated assistant.
- An off switch you control. You can pause it without calling your provider.
Most of the reputational damage from business automation comes from missing one of these, not from the technology failing.
When automation is the wrong answer
We recommend against it more often than people expect.
- Low volume. If the task happens twice a month, a reminder in a calendar is cheaper and more reliable.
- The process is not settled. Automating a workflow that is about to change means building it twice. Fix the process first, then automate it.
- The real problem is a decision nobody has made. Frequently the reason enquiries are handled inconsistently is that nobody has decided who owns them. Automation will not make that decision for you.
- Every case genuinely needs expertise. Some work is judgement all the way down. Automating around the edges may help; automating the core will not.
- The data is a mess. Automation applied to inconsistent data produces inconsistent results faster.
A realistic first project
For a business new to this, the sensible first project is narrow, measurable and boring.
Pick the enquiry path. Define success as a specific number — for example, that every enquiry receives an acknowledgement within two minutes and no enquiry goes more than one working day without a human response. Build the capture, acknowledgement, assignment and reminder. Run it for a month. Measure whether the number held.
That project is typically days to a couple of weeks of work. If it delivers, you have both a result and a foundation. If it does not, you have learned something cheaply, which is the other reason to start small.
Want to know what is worth automating?
PresenticAI maps how work actually happens in your business, identifies the workflow with the best return, and tells you honestly when automation is not the answer.