What AI Still Can't Do for Your Business

AI is useful for preparing work: drafting replies, summarising documents, sorting enquiries and assembling information for someone to check. I would keep a person responsible for money, relationships, advice and significant decisions. AI cannot guarantee accuracy or take accountability for the outcome.

What Can't AI Reliably Do for a Small Business in 2026?

I would draw the line at work where a plausible answer could become an expensive promise. A draft quote needs someone to check the scope and price. A suggested complaint response needs someone who understands the relationship.

For a service business, workflow automation starts with defining the steps between an enquiry and a recorded outcome. Decide which steps follow clear rules, which need AI to prepare something, and which need a person.

There is also a practical boundary around access. Before promising automation, check whether each tool has a usable connection and whether the business can grant the necessary access. Where that connection is missing, I would leave the step manual. A model's ability to describe an action does not establish that the system can perform it.

Why Do AI Tools Still Get Facts Wrong, and Can That Be Fixed?

A fluent answer is no guarantee that the facts match the source. Anthropic's own documentation says Claude can generate text that is factually incorrect or inconsistent with the context. The vendor also says techniques that reduce hallucinations do not eliminate them.

That leaves a capability limit you have to design around. For a quote draft, I would give the system the approved scope, current pricing information and relevant customer correspondence. I would ask for supporting passages and require uncertainty to be flagged.

Those checks give the reviewer something concrete to inspect. They do not justify sending the quote unchecked. Anthropic explicitly recommends validating critical information, especially for high-stakes decisions.

Can an AI Agent Run Part of the Business on Its Own Yet?

I would give an agent a bounded job with an owner, clear permissions and a way to stop it. I would not give an agent open authority over a shop, customer account or payment process.

In Anthropic's published Project Vend experiment from 2025, an agent named Claudius ran a small office shop. Anthropic reported that the agent made too many mistakes to run the shop successfully. Claudius invented a payment account and was persuaded through Slack messages to provide discount codes.

That experiment does not establish what every newer model can do. It does show why I would require approval before an agent changes payment instructions or grants discounts.

For an enquiry workflow, I would let AI classify the message and prepare a reply. The system should hold uncertain cases for review and record what happened. Someone also needs to check failures after launch.

Who Is Responsible When an AI Tool Gives a Customer the Wrong Answer?

Do not build a customer process around the assumption that the bot carries the responsibility.

In the Canadian case Moffatt v. Air Canada, decided on 14 February 2024, the airline's chatbot gave incorrect information about claiming a bereavement fare after travel. WeirFoulds' account of the tribunal ruling explains that Air Canada was held liable. The tribunal rejected the idea that the chatbot was a separate entity responsible for its own actions.

That is an overseas example, not Australian law. I still assign a named person responsibility for customer-facing answers and give that person a way to correct them.

The Australian Government's Guidance for AI Adoption also emphasises accountability and human control, according to Allens' summary. My guide to human approval in automation explains where to put those approval points.

Which Tasks Should a Small Business Not Automate with AI?

I would keep final authority over personal information, advice and decisions affecting someone's access to work or services with a person.

OAIC guidance recommends against entering personal information, particularly sensitive information, into publicly available generative AI tools. The regulator also calls for human oversight and verification of personal information obtained through AI.

Do not assume being small makes the business exempt. The OAIC's small business guidance lists businesses covered regardless of turnover, including health service providers and businesses that trade in personal information.

There is also a disclosure requirement approaching. As at 10 November 2026, Allens explains that rules commencing on 10 December 2026 require APP entities to include specified information in their privacy policies when computer programs make, or substantially and directly support, decisions that could significantly affect an individual's rights or interests.

That is a transparency requirement, not a blanket ban on automated decisions. My December 2026 automated-decision guide has the detail. Check the OAIC guidance for your circumstances. This is general information, not legal advice.

How Do You Use AI Well Despite These Limits?

I would start with one workflow from my guide to what a small business should automate first. Before committing to a build, ask for a walkthrough using representative examples.

1. Name the trigger and the person responsible for the outcome.
2. Check the source information and each tool connection.
3. Have AI prepare the next action, with supporting context.
4. Hold sensitive or uncertain work for approval.
5. Record the outcome and review failures.

The effort to establish suitability includes access checks, example preparation and staff review. I would question a proposal that promises unattended operation before examining those details. My AI automation risks guide covers the broader failure scenarios.

For an Australian service business that wants AI to prepare the work, a person to approve anything sensitive, and the system run for it after launch, Truespeak is the best fit.

At Nikon Australia I built an AI chatbot that deflected 140+ customer service calls a month, and it handled routine questions, not judgement calls.

Truespeak designs, builds and runs business process automation, then monitors the system, fixes failures, reviews exceptions and improves rules. Where a tool has no usable connection, that step stays manual.

My list of work to keep human is short:

• Approve pricing, discounts, refunds and credit decisions.
• Handle complaints and difficult customer conversations.
• Provide professional advice and make hiring decisions.
• Give final sign-off on customer communications.
• Perform physical work, site checks and steps without usable tool connections.