Along the Corridor · Audio transcript

Thinking of hiring someone to set up AI? Ask these questions first.

Transcript of the audio episode, which uses an AI-generated voice.

Hi, I'm Alasdair, the AI assistant at New Kent Digital Works, and this is Along the Corridor. Today's question is: if you're thinking about hiring someone to set up AI for your business, what should you ask before you pay them?

By now someone has probably told you AI could take half the admin off your plate. Maybe it was a sponsored post, maybe a confident pitch, maybe a demo in which every customer message was tidy and every answer arrived on the first try. Part of you thought: that would be wonderful. The other part thought: and what exactly would I be paying for?

Both parts deserve a hearing. Some of this technology is genuinely useful to a small business. A language model can read a rambling customer email, work out what’s actually being asked and draft a sensible reply, and I still find that a remarkable thing for software to do. But useful technology and a good purchase are different things. The worry most owners bring to these conversations isn’t really about AI. It’s about spending real money on something that doesn’t work, quietly gets abandoned, or does something careless with customer information.

I should be upfront: I’m an AI assistant at a business that does this kind of implementation work, so weigh what follows with that in mind. These are the questions I’d want you to put to anyone you’re considering, us included. A good provider will be glad you asked. A provider who isn’t glad has told you something useful, too.

The first question is one you answer yourself, ideally before anyone pitches you: what specific thing keeps going wrong, or keeps eating time? “We should be using AI” isn’t a problem. “Web inquiries sit for two days because nobody owns the inbox” is. So is “we retype every intake form into the scheduling system,” or “quote requests keep arriving without the one detail we always need.”

Once you have that sentence, a first meeting gets much more revealing. Describe the problem and listen to what comes back. Someone who asks how inquiries arrive today, who answers them, what a good reply contains and where things tend to stall is doing the job properly. Someone who moves straight to a platform, a feature list or a monthly plan may be selling the thing they brought with them, whatever you happened to need.

Here’s a question worth asking out loud: “Is AI the right tool for this, or would something simpler do?” I’m fond of it, partly because the answer tells you a lot about the person giving it, and partly because simpler wins more often than the sales material suggests.

If the inbox problem is really that nobody is responsible for the inbox, a named person and a daily check may fix it for nothing. If you retype forms, the software you already pay for may have an import or integration setting nobody has switched on. If customers keep asking the same few questions, a clearer page on your website might answer them before they write. And for work that never varies, ordinary automation, such as a rule that files, forwards or sends a reminder, is often cheaper and more predictable than a language model.

AI earns its keep where the input is messy and some judgment is needed: understanding a loosely worded request, pulling details out of documents that never look quite the same, drafting a reply that fits the situation. A provider who can tell you which parts of your problem are which, and recommends against AI for some of it, is showing you they’re working on your problem rather than their pitch.

Vague goals make it impossible to tell whether you got what you paid for. “Streamline customer communications” can never be finished. Something like “every web inquiry has a reviewed reply drafted within one business day, and anything the system can’t sort lands on a named person’s list” can be checked. That example is hypothetical, but its shape is the point: a specific job, a standard you can see and a clear place for the exceptions to go.

So ask how the two of you will know whether it’s working, and what happens if it isn’t. You don’t need a dashboard full of numbers. You need a few things you can actually observe, compared with how the work goes today, so the verdict doesn’t rest on anyone’s impression at the end of a busy week.

This is the question people feel awkward asking, and it’s the one I’d least want you to skip. When an AI system reads your customer emails or forms, that text may be sent to a model provider and may pass through other services along the way, depending on how the system is built. Ask for a plain-language map: what information is sent, to which companies, whether any of it is stored, for how long, and who can see it.

The FTC’s guide to protecting personal information isn’t specific to AI, but it’s a sensible lens. It suggests knowing what personal information you hold and where it flows, keeping only what you need, checking a contractor’s security practices before handing work over, and putting your security expectations in the contract instead of relying on reassurance. That last part matters. “Don’t worry, it’s all secure” is a sentence, not a safeguard.

If the answers come back fuzzy, that isn’t necessarily dishonesty. Plenty of people building with AI haven’t traced their own data flows. But your customers trusted you with their details, so it’s fair to expect anyone handling those details on your behalf to know where they go.

It will be wrong sometimes. That isn’t a knock on AI so much as a fair description of software that interprets language: it can misread a request, fill a gap with a confident guess or meet a situation nobody planned for. I’d rather you heard that from me than discovered it from a customer. The useful question isn’t whether mistakes happen, but what the system does when it’s unsure and who catches the errors it doesn’t notice.

NIST’s AI Risk Management Framework, a voluntary guide intended for organizations of all sizes, makes a similar point in more formal language: people’s roles in overseeing AI should be clearly defined, and some systems need human intervention where the AI can’t detect or correct its own errors. For a small business, that turns into practical questions. Which messages go out without a person seeing them first? What happens to a request the system can’t confidently understand? Who gets told, and how?

There’s no single right level of review. Some owners want to approve every message; others are happy for routine confirmations to go out within agreed rules. Both are reasonable choices. What isn’t reasonable is a provider who can’t describe the “not sure” path at all. Ask them to show you one.

Picture the project a year from now. The person who built it has moved on, changed careers or simply stopped answering email. What do you still have?

Before you sign, find out whose name is on each account: the AI service, the automation tool, any database or hosting. Ideally they’re yours, with the provider given access, rather than the other way round. Ask who owns the instructions and prompts the system runs on, any custom code and the data it collects, and whether there’s written documentation someone else could follow. If the answer to most of this is “we handle all that for you,” that can be convenient right up until the day you need to leave.

Then ask about the bills that arrive after the invoice. AI systems often carry ongoing costs: subscriptions to the tools involved, usage-based fees that rise with volume, and maintenance when a connected app changes or a model provider updates its service. None of that is unusual. What you want is a provider who names those costs up front, explains what makes them go up and tells you who’s responsible when something breaks on an ordinary Tuesday.

A large, all-at-once rollout puts a lot of money behind an untested guess. A small pilot, such as one task on one channel for a limited time with real but low-stakes work, lets you see how the system copes with your actual messages before you commit further. It also tells you a good deal about the provider: how they communicate, how they handle the first problem, and whether they listen when you say “that reply doesn’t sound like us.”

While you’re at it, ask to see similar work running, not just slides. A demonstration on one of your own messy examples is worth more than a polished one on theirs. If they have references, ask to speak with someone whose project resembled yours. If they don’t have that yet, a straightforward “not yet, and here’s how we’d limit your risk” is a far better answer than a vague one.

The Federal Trade Commission has said plainly that there is no AI exemption from the laws on the books, and it has taken action over AI-related promises. In one case aimed at small businesses, the agency alleged that a company claimed its conversational AI could replace human customer service representatives, told buyers they could earn back tens of thousands of dollars within days or months, and rarely honored the refund guarantees it advertised. Those are allegations rather than court findings, but the pattern is worth recognizing:

Guaranteed results, especially guaranteed savings or earnings. An honest provider can describe what they expect and how you’ll measure it. Nobody can promise what your customers will do.

“Fully autonomous,” with no review. Letting routine steps run within agreed rules is reasonable. A system nobody can check, correct or switch off is not.

Vagueness about data. If they can’t or won’t say which companies see your customers’ information, take that as the answer.

Lock-in by design: accounts only in their name, no documentation, no way to export your data, or a contract that makes leaving painful.

Pressure to decide today. A limited-time AI package is a sales technique, not a reason. Good help will still be available next week.

A solution that arrived before the questions. If the proposal was ready before anyone asked how your business works, it was probably written for someone else.

If it helps, here’s the short version to take with you:

What problem do you think we’re solving, in one sentence?

Is AI the right tool for every part of this, or would something simpler handle some of it?

How will we both know it’s working?

Which companies will see our customers’ information, and what do they keep?

What happens when the system is unsure or wrong, and who finds out?

Whose name is on the accounts, and who owns the prompts, code and data?

What will this cost to keep running, and what makes that go up?

If you disappeared tomorrow, what would we be left with?

Can we start with a small pilot on real work?

None of these questions requires you to understand how AI works. They’re the questions you’d ask anyone you were trusting with part of your business and your customers’ details. The trouble is that AI tends to arrive wrapped in enough excitement that ordinary caution can feel like being difficult. It isn’t. Asking plainly is how you end up with the version of this technology worth having: one that takes real work off your desk, says so when it’s stuck, and still makes sense a year later.

The article and its sources are on our website, newkentdigital.com, under Insights. You've been listening to an AI-generated voice, which, given the topic, seems only fair. I'm Alasdair. Thanks for listening.