Inspiration and credit: This original commentary is inspired by Harold Sinnott's article on data and autonomous networks, published on 2 October 2026. The business examples and practical recommendations below are my editorial interpretation.
Imagine an AI tool spots a customer who appears ready to buy. It prepares a follow-up, selects an offer and schedules the message. That sounds useful until you discover the customer has already purchased, the offer has expired or an unresolved complaint is sitting in a different system.
This hypothetical example shows why a confident response is only part of the story. Before giving an agent responsibility, a business needs to know whether it has the information required to choose an appropriate action.
The practical takeaway
Make the information dependable before expanding automation
Start with one defined task, give the agent access to relevant and current records, and decide which actions still need a person's approval. Assess the outcome before giving it more responsibility.
What autonomous networks can teach smaller businesses
Sinnott examines this problem through telecom networks. His central concern is the move from AI analysing conditions to agents changing live systems. When information is spread across different parts of a network, an apparently sensible fix may overlook its consequences elsewhere.
My reading is that the same decision problem can appear at a smaller scale. A business may keep enquiries in a CRM, bookings in a calendar, purchases in an ecommerce platform and complaints in an inbox. Connecting a new AI tool to one of those systems does not automatically give it an accurate picture of the customer.
The stakes differ from operating a telecom network, but the useful question remains: does the agent know enough to do this particular job?
Begin with the records the task actually needs
You do not need to reorganise every system before trying a limited workflow. Start by identifying which facts determine a good decision. For enquiry follow-up, those might include the most recent conversation, whether a quote has been sent and whether the customer has asked to pause contact.
Then check where those facts come from. Who updates them? How quickly do changes appear? What happens if two records disagree? A connected system can still contain old information, duplicate customers or an empty field that the agent misinterprets.
- Choose an authoritative record: Identify which system owns each important fact.
- Make freshness visible: Preserve update times so old information can be recognised.
- Handle uncertainty: Route missing or conflicting information for review instead of letting the agent guess.
A correct fact can still lead to the wrong decision
Suppose an agent sees that a homeowner has not replied to a window quotation. A follow-up might be reasonable. But if the latest conversation says they are waiting for a survey, another sales message could feel pushy and unhelpful.
The quote status is a fact. The reason for the delay gives that fact meaning. Context includes relationships between records, the stage of a process and exceptions that change what should happen next.
Write these exceptions down in plain language. For example: do not send a sales follow-up while a complaint is open; request review when a promised appointment is missing; stop when the customer has asked not to receive further messages. Confirm that the workflow can actually enforce those rules.
Separate access to information from permission to act
An agent that can read a record does not necessarily need permission to change it. Likewise, preparing a draft and sending it to a customer are different responsibilities.
I would define permissions around each task. An agent might summarise new enquiries and draft suggested replies, while a person approves external messages, changes to prices or commitments about delivery. More consequential actions deserve tighter controls.
Keep a record of what happened, which information supported the decision and who approved it. Assign someone responsibility for handling exceptions. Also decide how to stop a workflow and correct a mistake. A changed field may be recoverable; a message already sent needs a different response.
Test a narrow workflow before giving it more responsibility
A useful first pilot could be drafting replies to a small group of enquiries. Compare each draft with the underlying records and a human-written response. Note errors, missing context and the time spent checking the result.
If the drafts are consistently useful, consider a limited next step with clear rules and approval requirements. Keep unusual cases outside automatic execution until you understand them. Expanding access and expanding authority should be deliberate decisions.
Evaluate outcomes that matter to the business: fewer duplicate contacts, accurate responses, fewer corrections and time saved after review. Speed alone is a poor measure if someone has to repair the work afterwards.
Five questions to ask before enabling an AI agent
- Which records does this task need, and are they current?
- What missing context could change the right response?
- Exactly which actions can the agent take?
- When must it stop and ask a person to review?
- How will we detect, investigate and correct a mistake?
These questions make a pilot easier to evaluate. They also help you compare tools by the controls and visibility they provide, rather than relying on an impressive demonstration.
My view: build confidence through a useful first task
For a smaller business, the most productive starting point is a recurring job with a clear outcome. Understand the information behind it, define the agent's boundaries and review what it produces.
Sinnott's article prompted me to think beyond what AI can generate and towards what a business can responsibly ask it to do. Reliable automation depends on the whole workflow: the records, the instructions, the permissions and the people who remain accountable.
If you are considering automation for your enquiries or marketing, get in touch with me to discuss where a practical first step might fit your business.
Further reading: Read Harold Sinnott's original LinkedIn article for the telecom perspective behind this discussion.