Custom AI agent or a template chatbot, and how to tell which your questions need

The dividing line, stated first

Use an off-the-shelf or template chatbot when the questions are few, stable, and answerable from material you have already written. Build a custom agent when the assistant has to read live data from your own systems, take an action such as booking or updating a record, or handle the way your customers genuinely write rather than the way a demo script writes. Answering from a page is a template problem. Looking something up or changing something is a custom problem.

That line matters because the cost difference is not marginal. A template product is configured in days by somebody in the business. A custom agent is a build with integrations, testing and ongoing maintenance behind it, and it should only be bought when the template genuinely cannot reach.

The usual mistake is buying the custom version for a set of questions that a well-written page and a template would have covered completely.

When a template chatbot is genuinely enough

This is the honest half, and it applies to a large share of the businesses asking the question. If your enquiries cluster into a short list of repeated questions with answers that do not change weekly, a template product will handle them, and it will handle them next month without anyone maintaining an integration.

The clue is in what your team is already doing. If the same five or six questions arrive every week and staff answer them by pasting from a document, the document is the product and the chatbot is just a faster way to deliver it. Nothing about that requires custom engineering.

There is a second advantage that gets undersold. A template you configure yourself can be corrected in minutes by whoever spotted the wrong answer, without a ticket, a developer or a deployment.

• The questions are a short, repeated list: opening hours, delivery areas, return policy, what you sell and where you are.
• The answers change rarely, and when they do, somebody in the business can edit them directly.
• No action is required at the end of the conversation beyond passing the enquiry to a person.
• Volume is modest, so the cost of a wrong answer is a follow-up message rather than a lost order.
• You want to learn what people actually ask before designing anything. A template running for a month is the cheapest research available.

When a custom agent is the only thing that works

The template stops being adequate at a specific point: when a correct answer depends on something only your systems know. Nobody can configure a stock level, an order status or a customer's outstanding balance into a list of canned replies, because those values change while the conversation is happening.

The second trigger is action. An assistant that books an appointment into a real calendar, updates a record, checks eligibility against your rules or opens a ticket with the right details attached is doing operational work, and operational work needs connections into the systems that hold the truth.

The third is how your customers write. In markets where a single message mixes two languages, arrives as a voice note, or is a photograph of a payment confirmation, a template built for tidy typed English will mishandle a large share of real conversations. Designing for that is not a refinement, it is the difference between something that survives contact with customers and something that only works in a demonstration.

• The answer depends on live data: stock, order status, availability, an account balance, a delivery slot.
• The conversation should end in an action rather than a sentence: a booking made, a record updated, a ticket raised with the right fields filled.
• Your material is large or contradictory enough that answers must be grounded in specific documents and shown with a source.
• Rules vary by customer, region, product line or contract, so one answer is not correct for everybody.
• Customers write in mixed languages, send voice notes and attach screenshots, and the assistant has to cope with all of it.
• The conversation has to be logged against a customer record that other people in the business can see.

Comparing the two on what actually differs

Ignore the feature lists and compare on these. The last row decides more implementations than any of the others.

Criterion	Template chatbot	Custom agent
What it can answer	Anything you have written down in advance	Anything your systems know at the moment of asking
Can it do things	Collect details and hand over	Book, update, check eligibility, raise a ticket
Time to live	Days, configured by somebody in the business	Weeks, because the integrations are the work
Who fixes a wrong answer	Whoever spotted it, immediately	Whoever maintains it, through a change process
Ongoing cost	A subscription, and occasional editing	A subscription plus maintenance, because connected systems change
What it demands of you	Written answers that are actually correct	Working access to the systems, and rules stated precisely enough to encode

The prerequisite most buyers skip

Neither option can answer a question the business has not decided the answer to. Plenty of assistant projects stall at the point where somebody asks what the return policy actually is for a discounted item, and three people give three answers. Writing the answers down is the project. Choosing the software is the smaller half.

A custom agent has a harder prerequisite still. It needs working access to the systems that hold the truth, and that access is often the longest item on the critical path. If nobody can grant an integration user to your order system this quarter, a custom build is not the right thing to buy this quarter, whatever its merits.

There is also a case for buying neither. If your enquiries are low in volume and high in value, where the customer is deciding on trust and the conversation is the sale, putting an assistant in front of them removes the thing that was working. Use automation to route those enquiries to a person faster instead.

The rule to decide by

Take a week of real enquiries, not a sample somebody remembers, and sort them into two piles. Pile one is answerable from something you have already written. Pile two needs a lookup, an action, or a rule that varies by customer.

If pile one is the large majority, buy a template, fix the underlying documents, and revisit in three months. If pile two is large or growing, a template will keep handing conversations to a person at exactly the moment they became useful, and a custom agent is the right purchase. If both piles are small, the honest answer is that you do not have a chatbot problem yet, and the money is better spent on answering faster.

Sorting the two piles yourself, and what you keep either way

Sort the two piles with your own team before anybody quotes, and expect the result to be uncomfortable. Most businesses find pile one is larger than they assumed, which means a template chatbot over material you have already written is the correct purchase and the custom build is premature. The prerequisite named above is in-house work too: deciding what the return policy actually is belongs to the business, and no supplier of either option can decide it for you.

Whichever pile wins, insist on the same thing. The documents, the prompts, the tool definitions where there are any, and the conversation history sit in accounts under your own logins, so you own the system. On a template that mostly means being able to take out the content you wrote. On a custom agent it means the whole apparatus including the run traces, which are the only honest record of what it did on the day somebody complains about it.

The published version of the custom half is Big Texas Land Buyers, where voice agents run more than 500 calls a day and categorise into the CRM without being asked, and Center for Sight in New York, booking appointments largely hands-free. Both are pile two: a correct answer depends on something only those systems know. Ibrahim owns build and workflows at Wobble, and the rule he applies is that whichever you buy, somebody in your building should be able to change an answer without ringing anybody.

Common questions

Is a template chatbot good enough for a small business?

Often yes. If your enquiries are a short list of repeated questions with answers that rarely change, a configured template handles them and can be corrected by anyone in the business in minutes. The point where it stops being enough is when a correct answer depends on live data or requires an action in another system.

What can a custom AI agent do that a template cannot?

Read from your own systems at the moment of the question, and change something as a result. Checking availability, quoting from current stock, updating a record, booking into a real calendar and applying rules that vary by customer are all outside what a pre-written answer set can do.

Should the assistant hand over to a person?

Always, and the handover should be designed rather than left to the model. Money in dispute, complaints, cancellations and any promise the business would be held to belong with a person. So does any customer who asks for one, immediately and without being talked out of it.

Can I start with a template and move to a custom agent later?

Yes, and it is a sensible sequence. Run a template for a month and keep the transcripts. They will tell you which questions actually arrive, which ones needed a lookup and which ones the template got wrong, which is far better input to a custom build than a workshop full of guesses.

What makes a custom agent give wrong answers?

Usually the material it was given rather than the model. Superseded documents left alongside current ones, policies that never name themselves, and tables separated from their headers all produce confident answers drawn from the wrong source. Fixing the documents removes more errors than changing the model does.

How do I keep the cost of a custom agent from growing?

Limit the number of systems it touches, agree in advance who maintains it when a connected application changes, and keep the answerable content in one place rather than four near-copies. Most of the ongoing cost of a custom agent is keeping its connections and its knowledge current, not running the model.