How this stuff actually works, explained in plain English.
Notes on how AI systems actually behave: what an agent does between the prompt and the answer, what to automate, how to tell whether the output was any good, and where a person still has to say yes. No jargon, no pitch.
·5 min readFind out what the tool is before you plan around it
A good demo shows what a product can look like, not what it is. Before you plan work around a new tool, check what you would actually be using day to day and what the license costs once it is in real use.
Software evaluationAI toolsSoftware licensingVendor pricingAI adoption
·6 min readAn approval nobody is there to give is just a delay
A "check with a person first" step only protects you if the person sees the request in time. Before you add one, decide where the request shows up, who sees it, and what happens if nobody answers.
Human in the loopApproval workflowsAI automationBusiness processAI agents
·6 min readPass along the original, not your summary
When you relay feedback to whoever does the work, person or AI, your paraphrase drops the details that decide whether the revision lands. Send the raw words. Summarise for yourself, never for the worker.
AI promptingFeedbackDesign revisionsDelegationHuman in the loop
·6 min readKeep the working notes out of the answer
An AI system works things out before it replies, and that working is not for the customer. Treat the reasoning and the reply as two different documents, and check the one that goes out.
AI output qualityLLM reasoningCustomer experienceAI guardrailsHuman in the loop
·4 min readGive the automation its own paper trail
When you hand a repeating task to software, give it its own address, its own log, and its own name on the record. Not for tidiness. So that weeks later you can still tell what it did.
Business automationOperationsProcess designAuditabilityAI oversight
·5 min readWhat your AI is not allowed to say
Accuracy is not the only setting worth having. Deciding in advance what a system may never produce, and checking for it before a customer sees the answer, is a business decision rather than a technical one.
AI governanceAI oversightGuardrailsBusiness automationHuman in the loop
·4 min readIt said it worked
Most automation reports on its own success, which is not evidence that anything arrived. Confirmation has to come from the other end, and deciding where that is belongs to you.
Business automationAI oversightProcess designOperationsHuman in the loop
·5 min readHow to evaluate a technical product when you are not technical
You will never be able to inspect the system you are buying. You can still judge it, by evaluating the account someone gives of it rather than the technology itself.
AI adoptionVendor evaluationBuying softwareDue diligenceProfessional services
·4 min readNot every task deserves the expensive tool
Once a tool bills by usage, the real decision is which work earns the premium tier and which gets the cheap one. Most owners pay top rate for everything and call the bill the cost of AI.
AI costsModel selectionOperating costsAI adoptionVendor pricing
·4 min readTwo screens, two answers
When two parts of the same product answer the same question differently, you lose trust faster than by being plainly wrong once. Consistency is a feature, not polish.
Product qualityAI featuresData consistencyUser trustSoftware design
·4 min readA working pilot is not an offer
Proving a system works and knowing what someone can buy are two different projects. The gap between them is where most promising pilots quietly stall.
AI pilotProduct definitionScopingProfessional servicesAI adoption
·4 min readThe bottleneck is rarely the tools
Firms that already run a capable AI stack still lose hours every week. The loss is not inside any tool, it is in the handoffs between them, which is why buying another one does not fix it.
Workflow automationAI adoptionProcess designSystem integrationKnowledge work
·5 min readWhat the step makes besides its output
Some work produces a second thing that never appears on the process map. Automate the step without noticing it, and the output still arrives while the judgment quietly stops forming.
Workflow automationHuman in the loopProcess designAI adoptionKnowledge work
·4 min readThe work you are not allowed to send anywhere
Every firm has a category of work it would automate tomorrow if the files were allowed to leave the building. The hardware constraint that made that a dead end has quietly lifted.
On-premise AILocal LLMData privacyDocument automationHuman in the loop
·6 min readThree signs your intake process needs a second look
A short diagnostic for professional services firms. Each sign is easy to spot, and each one costs more than it appears to.
Client intakeBusiness automationProfessional services
·6 min readFour oversight patterns, running in a real product
Alpine Velo is an AI cycling coach. The interesting part is not the coaching, it is the four mechanisms that keep the AI honest without a person watching it.
AI oversightHuman in the loopAI agents
·6 min readHow much time would AI actually give your team back?
Vendors quote multipliers they cannot possibly know. Here is the arithmetic for sizing the real number inside your own operation, before you spend anything.
Business automationOperationsProcess auditHuman in the loop
·6 min readWhat an automation project actually looks like
Four steps, in order, with the decisions that happen at each one. Including the two places where a project should be allowed to stop.
Business automationProfessional servicesAI strategy
·6 min readThe real cost of typing the same thing twice
Duplicate data entry looks like a small tax on everyone's week. It is actually three separate costs, and only one of them shows up as time.
Business automationClient intakeProfessional services
·6 min readAgent skills vs. agent tools
Two words that get used interchangeably, and shouldn't. The difference decides what your AI agent can do, what it's allowed to touch, and who has to say yes first.
AI agentsHuman in the loop