The short answer: If the inputs are tidy and the steps are the same every time, you need automation. If a person wants a quick answer from information you hold, you need a chatbot. If the job means reading messy material and working through several steps to produce something a person then checks, you need an agent. Many processes need a mix, and the simplest tool that does the job is the right one.
The first article in this series set the three side by side in a table. This one is the decision guide: three questions to ask about any job, then four everyday jobs worked through.
As a reminder, automation follows fixed rules, a chatbot answers the message in front of it, and an agent takes a goal and works through the steps within limits you set.
Three questions that sort most jobs
1. Is the input tidy or messy?
Tidy input is the same thing in the same place every time: form fields, spreadsheet rows, files in a fixed format. Messy input is everything else: emails, PDFs, call notes, photographs of paperwork, other people's websites.
Rules handle tidy input well and messy input badly. If the input is messy, something has to read it first, and that points towards an agent or a chatbot.
2. Are the steps the same every time?
Try drawing the job as a flowchart. If you can do it without a box marked "use your judgement", automation will do. If the next step depends on what turned up in the last one, such as which document to open, what to look up or whether there's enough to go on, you need an agent.
3. Who starts it, and who uses the result?
A person with a question who wants an answer now is a job for a chatbot. An event, such as a form arriving or a date passing, is a trigger for automation. A piece of work to be done and handed to someone for checking is a job for an agent.
| If the work looks like this | Start with |
|---|---|
| Tidy input, same steps every time, started by an event | Automation |
| A person asking questions of information you hold | A chatbot |
| Messy input, several steps, a result somebody reviews | An agent |
| Fixed steps with one messy step in the middle | Automation, with an agent for that one step |
| A judgement call with real consequences | A person, with software preparing the ground |
Start with the simplest tool that works
Automation gives the same output for the same input, which makes it the easiest of the three to test and the cheapest to run. An agent reads material and decides how to work through it, so its output needs checking and it costs more to run.
So don't ask which of the three is best. Split the process into steps and ask the three questions of each step. You'll usually find most steps want plain automation and one or two want something that can read. Buy the agent for those steps only.
Four jobs, worked through
Sales follow-up
The job: make sure every enquiry and every promised callback gets followed up.
Start with the reminder. "No contact for a set number of days, so put it on someone's list" is a fixed rule working on tidy records: automation. CallHub, our own call system, has gap and callback reports for this reason, so follow-ups don't slip.
Writing the follow-up is different. A useful message refers to what the customer said, and that's in a call note or an email thread, which is messy input. An agent can read the history and draft a message, and the rep reads it, changes it and sends it.
Whether to chase at all, and how hard, is the rep's call.
Verdict: automation for the reminder, an agent for the draft, a person for the send.
Quoting
The job: turn an enquiry into a price.
If your enquiries arrive through a form and the price is quantities multiplied by a price list, you need a calculator, and a spreadsheet or your existing software probably already does it.
If they arrive as emails with a description, some measurements and three photographs, somebody has to work out what's being asked for before any sum can be done. An agent can read the enquiry, pull out the items, match them to your price list, note what's missing and lay out a draft quote. A person who knows the work checks it and decides the price.
Verdict: automation for the arithmetic, an agent for reading the enquiry, a person for the price. Never let software send a quote on its own: a quote is a commitment.
Checking documents
The job: confirm that a document says what it needs to say. A subcontractor's insurance certificate, for example, has to cover the right dates and the right kind of work.
Checking that a form is complete, with every box filled and a date in the future, is automation. Checking what a document means is reading, and documents from outside your business arrive in every format going. An agent can read each one against requirements you've written down and report which are met, which aren't and which it can't tell, with the passage it relied on beside each finding.
We use the same pattern in our own planning work. PlanningContent, the system our editors use for Planning Handbook guidance, splits a draft into individual claims and checks each one against the policy: verified, outdated, unsupported, uncertain or not covered. An editor sees every verdict and citation, and nothing is published until they approve it.
Verdict: automation for completeness, an agent for meaning, a person for the pass or fail that matters.
Customer questions
The job: answer the questions customers ask again and again.
If the answers are already in your own documents, this is chatbot work, with two conditions. It should answer only from those documents and show which one it used, and it should say so when it can't find an answer. PlanningAssist, our planning question tool, which is in beta, works this way: it answers from the local plan and national policy with numbered sources, and if the policy it can find doesn't cover the question, it says so rather than guessing.
The job changes when a customer wants something done: a booking moved, an order amended, money returned. Answering is one thing and acting on an account is another. Acting needs a connection to your systems, limits on what can be changed, and for anything involving money, a person.
Verdict: a chatbot for answers from your documents, and something more carefully built, with a person in it, for requests that change anything.
One system, all three
Real systems mix the three. ProspectHub, which our sales teams use, is an example.
- A trigger. Every new prospect is classified when it's imported. Nobody has to ask.
- An agent's work. When asked, it researches a company: what it does, who runs it and what its customers say, with a link to each source.
- A chat. A built-in assistant knows the prospect's research and notes, so a rep can ask questions before they dial.
- A person. Companies that don't fit the campaign are flagged, and the team decides whether to reassign or exclude them.
No single one of these would do the job alone.
Signs you've picked the wrong tool
- A chatbot doing an agent's job. Staff ask it the same sequence of questions for every case and paste the answers into another system by hand.
- Automation doing an agent's job. The exceptions pile is growing faster than the rules, and someone spends each morning clearing what the rules rejected.
- An agent doing automation's job. It's being asked to do something a formula could do, and you're paying for it to think about a sum.
- Software doing a person's job. Nobody can say who approved a decision that mattered. Where a person should stay in the loop deals with this one.