Practical guidance for established businesses deciding what to retain, connect, automate or build.
The Met Office names a storm when it expects damage significant enough for an amber warning — so the call surge behind it is announced days ahead. What the naming criterion actually commits to, why a compressed evening breaks an office that copes all year, and what is worth deciding inside the warning window.
The clever agents broke and the boring ones are still running. A worked example — a two-stylist salon where the agent reads the diary before deciding whether to transfer — plus what breaks, and the one rule we will not bend: it never denies it is AI.
Discovery, diagnosis, implementation and support — what each stage involves, what you should have at the end of it, and the six signs that separate a supplier selling a diagnosis from one selling a product. Includes what it costs and when you do not need one.
The government spent twelve weeks on business systems integration and explicitly ruled out mandating it. What the call for evidence covered, the question it never asks — which system is right when two disagree — and why the measurable gap is rarely the expensive one.
A menu asks the caller to navigate your business; an agent navigates it for them. What IVR still does better — including the payment case where a keypad genuinely wins — and the conditions where a conversation is worth more than a menu.
ATS, CRM, LinkedIn, job boards, email, calendars, phone, video interviewing — and a spreadsheet doing something more important than anyone admits. Why UK agencies still carry so much admin, and why the answer is usually connecting what is already there rather than buying another AI tool.
Valuation response, buyer and tenant qualification, out-of-hours calls, viewing bookings, feedback, maintenance triage and dormant CRM records. Where AI agents genuinely earn their keep in an agency — and the parts that should stay human.
Almost every citation of the five-minute rule gets it wrong. The 100× figure is MIT, not Harvard, and compares five minutes with thirty, not ten. What the studies measured, what they cannot tell you, and the finding nobody quotes — 23% of companies never replied at all.
Enquiry response, appointment booking, invoice processing, quote follow-up, onboarding, approvals. Twelve processes set out as the manual version, the automated version, and the condition that has to hold before automating pays. Plus what does not automate well.
An agency is paid to build what you asked for. A consultancy should sometimes be paid to say do not. Why the distinction stopped meaning much after the 2024–26 course boom, and seven questions that separate a supplier who delivers from one who will not.
Ruby bills per minute, Smith.ai per call, a build is setup plus monthly — and that decides your cost far more than any headline price. Which model suits which business, including the cases where buying from them beats building with us.
Employed, human service, message-taking or AI: four options priced on four different units, converted here to a common basis. Where the cost hides in each, and the arithmetic for working out which is cheapest at your call volume.
Capture, validation, two- and three-way matching, approval routing, posting. The stages of an automated accounts payable workflow — and why the exception path, not the happy path, decides whether it survives real invoices.
Spreadsheets multiplying, the same information keyed twice, job status that lives in somebody’s head. The operational warnings that a business has outgrown its systems, and which of them are actually worth acting on.
Buy, configure, integrate, automate or build — in that order, because most problems are solved before the last one. How to tell which your situation needs, and when custom software is the expensive wrong answer.
Enquiries, bookings, reporting, finance and memberships. Where automation genuinely saves time in hospitality and service businesses, where it introduces risk, and which decisions should keep a person in the loop.
Frequency, the cost of getting it wrong, whether the rules exist, whether the data is trustworthy. A prioritisation order for choosing the first workflow to automate — and the conditions that mean you should wait.
Which workflows suit an AI agent, how much autonomy is appropriate, and what has to be in place before one touches production: confidence thresholds, human review, logging, and a defined answer for when the model is wrong.
APIs, webhooks, data ownership, security, failure handling and support. What to confirm before connecting two business systems — including the questions that decide who actually owns your data afterwards.
These are written to be useful before you speak to anyone, including us. Each answers a decision rather than describing a service.
Everything here is written by Thomas Barrie, who builds the systems rather than commissioning the writing. Where a piece cites research, the sources are linked and the limitations stated — including where the research does not support what it is usually quoted as saying.
Bring the workflow, the systems and the operational cost. We’ll help establish the sensible next step.