6 steps from a chatbot to a Business Agent: what Amazon’s AI playbook means for small businesses

Why a brochure for enterprises matters to a plumber in Bolton

The guide, 6 steps to fast-track generative AI from prototype to production, is written for “technical leaders” at large companies and, unsurprisingly, every step ends with a recommendation to use Amazon’s own platform. But the problem it describes is exactly the one we see with small businesses every week.

Three numbers from it are worth knowing:

• 92% of companies plan to invest more in generative AI over the next three years (McKinsey, January 2025).
• McKinsey puts the productivity potential of generative AI at $4.4 trillion a year.
• Gartner predicted that at least 30% of generative AI projects would be abandoned after the proof-of-concept stage by the end of 2025, because of poor data, weak controls, rising costs or unclear business value.

That last one is the important one. Big companies with big budgets are giving up on a third of their AI projects. If you have ever switched on a “free AI chatbot” for your business, watched it give a customer a wrong price, and quietly turned it off again, you have lived the small-business version of that statistic.

The guide’s answer is a six-step process. Here is each step translated from enterprise-speak into what it means for a UK small business.

First, the ladder: traditional AI, generative AI, agentic AI

The guide draws a useful line between three kinds of AI:

A chatbot lives on the middle rung. It can talk. A Business Agent lives on the top rung. It can do. Amazon’s point, and ours, is that the value is on the top rung; the six steps are how you get there without it falling over.

Step 1: Choose the right model (translation: don’t obsess over it)

Amazon says different AI models suit different tasks and offers over a hundred to choose from. For a large bank, that matters. For a small business, it does not. The models behind a good WhatsApp assistant are more than capable of understanding “can you do Thursday after 4?” and answering from your price list. What matters is everything that sits around the model, which is steps 3 to 6.

What to do: nothing. Pick a provider who has already made this decision and will change the model quietly when a better or cheaper one appears.

Step 2: Optimise cost, speed and accuracy (translation: fixed monthly price, instant replies)

The guide talks about “prompt caching”, “model distillation” and cutting costs by up to 75%. The small-business version is simpler: you want a fixed monthly price so you are not watching a meter, replies in seconds so customers do not wander off to a competitor, and accurate answers on the things that matter (prices, hours, areas you cover).

What to do: ask any provider two questions. “What is the total monthly cost, including WhatsApp’s own fees?” and “How fast does it reply at 9pm on a Sunday?” Our WhatsApp Business API pricing guide explains what Meta charges and what is free.

Step 3: Customise the model with your data (translation: this is the whole game)

This is the step that separates a toy from a tool. Amazon describes feeding the AI your own documents so it answers from your information rather than from the internet. In enterprise language it is called “retrieval-augmented generation”. In small-business language it means the assistant knows:

• your services and your actual prices;
• your opening hours, areas covered and parking;
• your policies (deposits, cancellations, what to bring);
• how you talk, so it sounds like your business and not a call centre.

Every “the chatbot told a customer we do X and we don’t” story we have ever heard comes from skipping this step.

What to do: before switching anything on, write down the twenty questions customers ask most and the answers you want given. That document is worth more than any technology choice. When we set up a Business Agent we build it from exactly this, then keep it updated as your prices and services change.

Step 4: Guardrails (translation: it must know when to stop and hand over)

Amazon’s guide says production AI needs guardrails to protect customers and the brand: blocking topics it should not discuss, filtering harmful content, and keeping it on task. For a small business the guardrails that matter are:

1. Stay in lane. A plumber’s assistant should not give medical advice or opinions on the news, however politely it is asked.
2. Never invent. If it does not know a price, it says so and takes a message rather than guessing.
3. Hand over to a human. Complaints, anything sensitive, anything high-value, and anyone who asks for a person should reach you, with a summary of the conversation.
4. Respect “stop”. If a customer says stop, it stops, and no follow-ups go out.

What to do: test it yourself as an awkward customer before it goes live. Ask it something it should not answer and see what happens. Our agents are built with a human-handover path and a stop rule from day one; see how WhatsApp automated messages should behave.

Step 5: Security (translation: official platform, your data stays yours)

Enterprises worry about encryption, access control and compliance certifications. The small-business equivalent is two questions: is this running on Meta’s official WhatsApp Business Platform, or on an unofficial workaround that could get your number banned? And who can see your customers’ messages?

What to do: insist on the official API. Unofficial “WhatsApp bots” that run through a phone or a browser session are cheaper for a reason. Ask where conversation data is stored and for how long, and make sure you can export or delete it.

Step 6: Agents that transform workflows (translation: let it do the jobs, not just the talking)

The guide’s final step is where the money is: agents that “autonomously complete complex workflows”. For a large company that means processing claims or reconciling invoices. For a small business it means the assistant does the repetitive jobs you currently do on your phone at 10pm:

• answers the enquiry and qualifies it (what, where, when, budget, a photo);
• offers real free slots and books the appointment;
• sends the reminder the day before;
• chases the quote that went quiet;
• asks the happy customer for a Google review;
• tells you what happened, in one message, when it matters.

We listed the 12 jobs a Business Agent does on WhatsApp, with a week in the life of a Manchester plumber, if you want the full picture.

The real lesson: don’t build it, buy it working

Read between the lines of Amazon’s guide and the message is that most organisations fail at AI not because the technology is weak but because they skip the unglamorous steps: their own data, guardrails, handover, upkeep. Gartner’s 30% abandonment figure is the cost of skipping them.

A small business does not have an IT department to work through six steps. The sensible route is the one Amazon is quietly selling to enterprises too: use a platform where the six steps are already done, and put your effort into the one thing only you can supply, which is knowing your business. That is what a done-for-you Meta Business Agent setup is. You supply the twenty questions; we supply steps 1 to 6 and keep them working.

Frequently asked questions