Much of the coverage of AI presents it as a future development: conference keynotes, magazine covers, robots in stock photos. Google I/O last month focused on what the Assistant will do next.
Meanwhile, the small businesses we work with already use machine learning every day, generally without identifying it as such.
This article sets out what works now, in mid-2019, for a business with no data scientist and no budget for one, along with one technology we recommend postponing.
How useful are smart replies and suggested text in email?
Gmail has been suggesting short replies ("Sounds good," "Thanks, see you then") for a few years, and Smart Compose, the grey text that completes sentences, arrived last year. For a business that answers the same five questions forty times a week, as a dive shop does, these features save real time.
The key is to use them for routine messages only. A suggested "Thanks!" is appropriate; a suggested reply to a complaint is not. It reads as machine-generated because it is, and a dissatisfied customer notices immediately.
We therefore recommend using these features for routine correspondence and writing sensitive messages personally.
What can photo tagging and search do for a business?
For businesses with large photo collections, this feature is highly valuable. Google Photos can find "boat," "beach," "dog" or "cake" in a library that was never tagged. For a dive business with thousands of underwater images, or a caterer with years of event photos, it turns an unorganised archive into a searchable resource.
A photo of the boat for a flyer can now be found by typing "boat" and scrolling. Five years ago, the same task could take an afternoon of searching through folders.
Since 2016, Facebook has also generated automatic photo descriptions for people who use screen readers. The descriptions are basic, such as "Image may contain: one or more people, ocean, outdoor," but this is one of the few cases where machine learning has delivered a clear accessibility benefit without anyone having to request it.
One recommendation applies: if a photo matters on your website, write the alt text yourself. The automatic version is a starting point, but it cannot identify that the image shows your nudibranch rather than "a sea creature."
When is machine translation appropriate?
Google Translate moved to a neural system at the end of 2016, and the improvement in quality was immediately apparent. For reading a supplier's email in another language, it is now more than adequate. In our context, between English, Indonesian and whatever language a guest writes in, it is used extensively.
For any content a customer will read, it is not yet adequate. A machine-translated page is recognizable as such, and a native speaker identifies it within one sentence. It signals that the business did not invest the effort to get it right, which is the opposite of the intended effect.
The rule is therefore to use the machine for the draft and a fluent person for the final version. For businesses serving customers in two languages, that person must have a thorough command of both.
How should a business manage ad platforms that optimise themselves?
This is where most small businesses use machine learning without realizing it. Google Ads offers automated bidding strategies, called smart bidding, which adjust the cost per click based on how likely each person is to convert. Facebook decides who sees an ad based on who it predicts will respond.
These systems often outperform manual settings. They will also spend a budget very readily if allowed to.
The standard recommendation is to "let the system learn." That is reasonable, but the advertiser pays for the learning period. For a large brand, a few weeks of costly learning is negligible. For a dive shop on a small island, it can consume an entire month's marketing budget before the algorithm identifies the right customers.
What do appropriate limits look like in practice?
- A daily budget you could afford to lose entirely while the system learns.
- Conversion tracking on a real outcome, such as a booking or a call. Clicks alone are misleading.
- A geographic boundary that matches where your customers actually come from.
- A set date to review the results and decide, rather than letting the campaign run indefinitely.
We expect the platforms to continue in this direction, with more settings disappearing behind automated management. This makes the one element you still control, the goal, more important. The machine optimises for whatever it is told to pursue; with the wrong target, it will reach the wrong target very efficiently. Effective Google Ads management for a small business depends less on sophisticated settings than on ensuring the system pursues the right objective.
Why do we recommend postponing chatbots?
A few years ago, Messenger bots were presented as the next major development, and nearly every conference featured a session on them. Most of the bots we have tested are decision trees presented as conversation: press 1 for hours, press 2 for prices, and questions that fall outside those options go unanswered.
Customers dislike them, and we expect that many of the bots launched with such enthusiasm a few years ago have since been discreetly deactivated.
Language technology may eventually close this gap. OpenAI's GPT-2 produces remarkably fluent text, so a genuine advance may emerge from it. However, the tools a small business can purchase today do not understand anything. If a bot makes it harder for someone to reach a person, it costs the business bookings. Answering messages personally and promptly is more effective than any bot we have seen.
Where should a business owner start learning?
For an accessible introduction, Google has gathered free material on its Learn with Google AI page, including a machine learning crash course. Much of it targets developers, and those sections can be skipped. The beginner material explains what a machine actually does when it "learns," which in itself makes for a better-informed buyer.
Our main recommendation is to use the AI already built into the tools you pay for before purchasing any product labelled AI. Very few businesses, ourselves included, are drawing the full value from the tools they already have.