AI in IT support: what changes, what does not, and how to judge a vendor

What genuinely changed

For decades, first-line IT support was priced by the hour because it was delivered by the hour. That made continuous cover a luxury: a small business could afford a person for part of the week, or an agreement with a response target measured in hours, but not someone available at 11pm on a Sunday.

When the marginal cost of handling a routine question approaches zero, the shape of the service changes. Instant response stops being a premium tier. Out-of-hours stops being a surcharge. And people stop rationing their questions — which surfaces the long tail of small problems that previously went unreported and quietly accumulated into larger ones.

That is the real change, and it is an economic one rather than a technological triumph. It is worth being clear about, because the industry frequently sells the technology when the customer benefit is the pricing model.

Where AI support fails, specifically

Physical reality is absolute. A machine cannot replace a failed drive, reseat a cable, receive a delivery or attend a site. Any AI support offer must exclude physical work, and any that does not is misrepresenting itself.

Confident error is the characteristic failure mode. A language model can produce a fluent, plausible, wrong answer, and the fluency makes it more dangerous than an obvious error, because the user acts on it. The mitigation is not 'a better model' — it is narrow scope, grounding in documented behaviour, and a low threshold for handing back rather than guessing.

Judgement and accountability are the third gap. When a request is unusual — a director asking for a mailbox export on their last day, an urgent permission change during a dispute — a human weighs context and sometimes declines. AI applies its rules. Rules are more consistent and less wise, which is precisely why anything touching money or bulk data should require a human decision.

• Anything physical — hardware, cabling, on-site attendance
• Ambiguous problems where the user has described the wrong symptom
• Situations requiring judgement about an unusual or political request
• Undocumented, bespoke systems the model has no grounding in
• Anything requiring professional accountability for the outcome

The controls that make it safe

Zero standing access is the most important. If the AI holds permanent administrative credentials in your tenant, then any prompt injection, any model failure and any compromise of the vendor becomes a full compromise of your environment. Credentials should be brokered per task, scoped to the specific action, and expire.

Identity verification before account-affecting actions is second. An AI that resets MFA on request, with no verification, is a social engineering machine. The verification step is the security value of the service, and it should be non-negotiable and identical for everyone including the chief executive.

Then: a published scope, so the limits are known before they are discovered; complete audit logging of every action with who authorised it; and a defined hand-back that produces a written diagnosis rather than a dead end. Hand-back quality is an underrated differentiator — an AI that gives up usefully, with a summary you can pass to a specialist, has done real work.

Questions to ask any AI support vendor

The useful questions are about limits and architecture, not capability demonstrations. Ask what the service explicitly does not do, and be sceptical of an answer that struggles to name anything. Ask whether the AI holds standing access to your tenant, and if so what that access is. Ask what happens when it cannot resolve something, and what you receive at that point.

Ask what identity verification is performed before an MFA reset, and whether senior staff can bypass it. Ask where your conversation data is processed and stored, how long it is retained, and whether it is used to train models. Ask for the contract position on liability when the AI does something wrong — a vendor unwilling to discuss this in writing is telling you something.

And ask a question the vendor cannot rehearse: when is your service the wrong choice? Anyone who cannot answer that honestly either does not understand their own product or is not being straight with you.

• What is explicitly out of scope?
• Does the AI hold standing administrative access, or is access brokered per task?
• What identity verification happens before an account-affecting action?
• What exactly do I receive when it cannot resolve something?
• Where is my data processed and retained, and is it used for training?
• What is the contractual liability position when the AI gets it wrong?
• When is your service the wrong choice for a business like mine?

Where this is heading

The likely direction is not AI replacing IT providers but the first line separating economically from everything else. Routine, documented, high-volume work moves to automation because its cost approaches zero; physical work, project delivery, judgement and accountability stay with humans and become the visible part of what providers sell.

For a small business the practical implication is to stop buying first-line support twice. If an AI service handles the password and mailbox traffic, the human provider you keep should be scoped and priced for infrastructure, projects and judgement — not for a helpdesk you have stopped using.