Your software already has AI. When is that enough?

Much of the software small businesses pay for now has an AI layer. Three kinds come up most with us, described from the vendors’ own pages:

• Hotel systems. On 7 October 2026 Mews, a hotel property management system, announced Guest Messaging and Automations: WhatsApp, SMS, Booking.com and its guest portal in one thread linked to the guest profile and reservation, where “AI can handle common questions automatically, while conversations that need human attention are surfaced to your team” (Mews). It suggests replies from live reservation data and the property’s knowledge base (Mews, guest messaging).
• Helpdesks and shared inboxes. Zendesk says its AI agents can read and write data in external systems; beyond the automated resolutions each plan includes, its Spanish pricing page lists €1.50 per resolution when you add resolutions to your plan, or €2.00 pay as you go, counted only when the AI agent resolves a request without passing it to a person (Zendesk). Intercom states that “Fin is priced at $0.99 per outcome” (Intercom). respond.io includes AI Agents from its Growth plan, with AI usage inside a credit allowance, and bills WhatsApp fees separately (respond.io).
• WhatsApp itself. On 3 June 2026 Meta opened Meta Business Agent to businesses of all sizes. Meta says it answers questions specific to your business, recommends products from your catalogue, books appointments and lets you decide when a team member steps in (Meta). On the WhatsApp Business Platform, Meta charges it per token from 1 August 2026: one global rate of $2.00 per million tokens, with a typical message using 20,000 to 25,000 tokens, which Meta puts at about 4 to 5 cents (Meta, pricing updates). We compare it with a custom agent in Meta Business Agent vs a custom WhatsApp assistant.

They share one trait: each is good at what it can see. A hotel system sees reservations and its knowledge base; a helpdesk sees tickets, help articles and connected systems; Meta’s agent sees your business information and catalogue. So the real question is not whether the AI is good, but whether your customers’ questions can be answered from what that tool can see.

Take 50 real messages from last month and ask seven questions:

1. Which questions does it answer? Sort them: answered from data inside the tool, needs something outside it, needs a person.
2. What can it read? Only what lives inside the tool, plus its listed connectors. A price list, timetable or documents kept elsewhere are invisible to it.
3. Who writes and maintains the rules? Can you add your own exceptions (the client who always pays by transfer, the room that cannot take a cot) and export them if you leave?
4. What happens when the answer needs another system? Does it say so, hand over, or guess? Try a real case.
5. How does it hand over? To whom, with the history, and what is the customer told?
6. Which languages and channels? Your customers’ languages, and every channel they write on: WhatsApp, email, your website, the phone.
7. What does it cost on top, and how is it billed? Per resolution, per outcome, by credits or by token, plus WhatsApp’s message fees. Check what counts as a resolution or an outcome.

The built-in AI is enough when:

• most questions are answered from data inside that one tool: availability, booking details, order status, help articles;
• your customers write on the channels the tool already covers;
• the vendor’s way of setting rules fits how you work, and exceptions are rare;
• a person can comfortably take whatever is left.

Your own agent earns its cost when:

• answers need data from several systems at once: the booking system plus a spreadsheet of extras, the ERP plus a price list, the customer record plus documents;
• your business runs on rules and exceptions that a settings screen cannot hold;
• customers reach you on several channels and you want one set of rules behind all of them;
• the tool has no API, or leaving it would mean leaving your rules and history behind;
• you want the rules and the test set to be yours: written down, testable and portable.

Your own agent rarely replaces the system you have. It sits beside it: the booking system, the till or the ERP stays where the data lives, and the agent reads from it, and from everything else, to answer.

If you are unsure, do not decide on a demo. Switch on the built-in AI you already pay for and log, for four weeks, every question it could not answer or passed to a person, and why: missing information, another system, one of your own rules, or a person.

Then read the log. If most misses were missing information, fill in the knowledge base; the built-in tool will do. If most needed another system or your own rules, that list is the specification for your own agent, and the business case is the hours your team spends on those messages today. Buying in stages works the same way: one channel and the most repeated questions first.

The other worry we heard often: customers feeling they talk to a machine. We think the split matters more than the tool. The agent answers the repetitive questions at any hour: opening times, documents, availability, published prices. People keep what builds the relationship: the complaint, the special request, the regular customer. Write that split down before switching anything on, and test it with real cases. We explain how in how to test an AI agent.

The choice is also less final than it was, because the way agents receive instructions is becoming a shared standard. On 22 October 2026 Anthropic donated Agent Skills to the Agentic AI Foundation, under the Linux Foundation, where it had already donated the Model Context Protocol (MCP), the standard for connecting agents to other systems. In Agent Skills, a skill is “a folder containing a SKILL.md file” that teaches an agent a task, and Anthropic says more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, have integrated it (Anthropic).

The practical lesson: whatever you choose, keep your rules and test cases as documents you own, not only inside a vendor’s settings screen. Built-in tools may not read these formats yet, so ask; your own documents move with you either way. More in how to write instructions for AI agents.

When the built-in AI is enough, we say so. When it is not, we build the agent on your own WhatsApp number and email, next to the systems you already use, with the rules and test set in your name. See our WhatsApp customer service AI agent.