Insights
Field notes, not
thought leadership.
What we have learned building AI systems that had to survive real users - including the parts that did not work. No predictions about the next decade.
Six folders
Start where your question is
- What will this cost, and who should build it? Buying and budgeting AI 15 articles
- Which tool, and why do these things break? Workflow automation 14 articles
- How should the system be put together? Building and running it 16 articles
- How do we know it is any good? Measuring whether it works 11 articles
- What do we have to be able to show? Governance and regulation 4 articles
- Why does ChatGPT name a competitor? AI search visibility 5 articles
Buying and budgeting AI
Prices, contracts and build-or-buy calls, with the numbers written down. What an agent actually costs to run, what a proposal should contain, and what you own when the work is finished.
-
Ecommerce support: what to automate first
Sort contact reasons by volume, automatability and the cost of a wrong answer. Why order status comes first, and four intents that look easy and are not.
-
Inheriting an AI system nobody documented
The vendor left, the developer left, or you acquired it. What to establish in the first week, and why you change nothing until the number exists.
-
What an AI roadmap should actually contain
Most are a list of use cases with quarters attached. A useful one is a sequence of decisions, ordered by what unblocks the rest, and it names what you are not doing.
Workflow automation
n8n, Make and Zapier compared on the things that decide the bill, plus the operational detail nobody writes about: silent failures, retries, replay, version control and scaling.
-
Draft replies in a shared inbox: what to automate
Triage and drafting beat an autonomous agent in most inboxes. The 80 percent draft is the expensive one, and the honest metric is how often it is deleted.
-
Cold email deliverability: what actually decides placement
The 2026 sender rules from Google, Yahoo and Microsoft, the thresholds that get a domain blocked, safe volume per mailbox, and why generated volume makes it worse.
-
Version controlling n8n workflows
How to get n8n workflows into git so changes are reviewable and revertable, what the JSON export does not carry, and the promotion path between environments.
Building and running it
Architecture decisions that are hard to reverse, and what changes once the thing is live. Retrieval permissions, structured output, agent tool access, long documents, and the cost of the year after launch.
-
Why retrieval misses the answer that is there
The passage is in your documents and the assistant did not find it. Six reasons retrieval misses, the test for each, and why hybrid search is the missing piece.
-
Guardrails: what to check going in and coming out
A rule in the prompt is a request. A rule in code is a control. What to check on the way in, what to check on the way out, and the false positives nobody measures.
-
Qualifying inbound enquiries without losing the good ones
An assistant on an enquiry form should route fast and gather context, not decide who is worth talking to. The disqualification you never see is the expensive error.
Measuring whether it works
Evaluation sets, baselines, confidence thresholds and latency budgets. The measurement work that separates a system you can improve from one you can only have opinions about.
-
We shipped without an eval set
What the first year looks like when nobody measured: decisions made on anecdote, changes nobody can justify, and the cost of retrofitting too late.
-
Testing an AI system against the human baseline
How to measure what your people currently achieve on the same cases, why that number is usually lower than everyone assumes, and how to compare honestly.
-
Latency budgets for conversational AI
Where the seconds actually go in a grounded answer, what users tolerate in chat against voice, and how to buy back time without losing accuracy.
Governance and regulation
Acceptable use policies, audit trails and the right to human review, written against the primary sources. Educational, not legal advice, and they point at a lawyer rather than at a contract.
-
Automated decisions and the right to human review
When GDPR Article 22 applies to an AI system, what meaningful human involvement means after two EU court rulings, and how the UK rewrote the rule in 2026.
-
Building an audit trail an AI system can defend
What to log at each stage of an AI system so a complaint, an access request or a regulator can be answered from the record, and the 2026 retention rules.
-
What belongs in an AI acceptable use policy
The eight sections an AI acceptable use policy needs, what the vendor training defaults actually are in 2026, and the rules that survive contact with a busy team.
AI search visibility
How assistants choose what to cite, what llms.txt does and does not do, and where generative engine optimisation genuinely differs from the search work you already fund.
-
Measuring AI visibility: the method, and our own baseline
Nine buying questions, 69 cited sources, 64 distinct domains. What the answer layer in this niche is made of, and what it says about a site three days old.
-
Why ChatGPT recommends your competitor
How assistants decide which companies to name, why invisibility is usually a sourcing problem rather than a ranking one, and what you can influence.
-
llms.txt and structured data: what actually matters
What llms.txt is, who reads it and who does not as of 2026, the exact format, what schema markup demonstrably does, and where to spend an hour if you only have one.
Free calculators
Or skip the reading and run the numbers
Each calculator is the executable version of one of these articles - same method, your figures.
Automation payback
Hours returned, annual value and payback in months - checked at a conservative automation rate as well as your own.
How many test cases do we actually need?Eval set size
Margin of error at 95% confidence, sized from per-category counts, and the point where more cases stop buying precision.
Where should the confidence gate sit?Escalation threshold
Compare two gate settings on your own measured numbers: deflection, errors reaching customers, and total weekly cost.
What will this actually cost per month?LLM cost
Monthly spend broken down by input, output, retries and evaluation runs - with cost per resolved task, not per call.
Which automation platform is cheapest for my workload?Zapier vs Make vs n8n
Same workflow, three billing models. Counts tasks, operations and executions separately, and includes what self-hosting really costs.
What should our llms.txt contain?llms.txt generator
A spec-correct llms.txt from a name, a summary and the pages that matter, with the lines that would break it flagged.
Let's talk
Prefer a conversation to an article?
Most of what is written here started as a question a client asked on a call. If you have one, bring it - we answer questions before contracts.