To keep AI costs predictable, you need to know what is being counted. Platforms that meter everything as one undifferentiated credit make forecasting impossible. Why AI costs surprise people The unit of consumption is often unclear until the invoice arrives. * One credit pool covering functions with very different costs * No
To read documents with AI at scale, two things have to be true: the extraction has to be reliable enough to trust, and the cost has to be predictable before you start. Where document volume becomes a constraint The work scales linearly with the number of documents and none of
Being able to use AI to edit an existing app changes what AI building is for. Generating a new app from a prompt is a demo; editing one you depend on is a workflow. Why editing matters more than creating Most of an internal app's life happens after
To describe an app so AI builds it, describe the records, the phases, and the rules. A description in terms of screens produces an interface over a structure nobody designed. What a vague description produces The builder fills the gaps, and the gaps are the important part. * Field types chosen
Handling the moment when an AI agent gets something wrong matters more than preventing every error. Errors are expected; what determines whether the deployment survives is how quickly they are corrected. A response sequence Correct first, diagnose second, prevent third. * Correct the record, and the customer-facing consequence if there
To test an AI agent before it acts, you let it propose while a person decides. Its answers become observable without any consequence attached. What to test for Accuracy is one dimension of several. * Whether answers match your documentation, not general knowledge * Whether classification is consistent across similar cases * How
To choose the first process for an AI agent, look for high volume, clear rules, and cheap mistakes. All three matter; two out of three usually produces a disappointing first attempt. Why the first choice decides the rest A poor first deployment ends the conversation about a second. * A process
To build a knowledge base for an AI agent, write for someone answering a question rather than for someone reading front to back. Documentation written as a narrative answers questions poorly. What makes documentation hard to answer from The problem is usually structure rather than content. * Long documents where one
To write instructions for an AI agent, describe its role, its tone, and its limits. What it should know goes in the knowledge base; what it may change goes in field permissions. What belongs where Instructions carry more than they should when the other two are neglected. * Instructions — role, tone,
Understanding the difference between a table title and display label resolves one of the most common complaints in a new build: selection lists showing numbers instead of names. What each one does They are often assumed to be the same field. * The title field identifies the record within the table
To name and order process phases, use names that describe a state rather than an action. A phase called "Send to finance" tells you what someone should do; a phase called "With finance" tells you where the record is. Why phase names cause arguments Ambiguous names
To use tags without creating chaos, decide what deserves one. Tags are easy to create, which is exactly why a tag system becomes unusable faster than any other part of a build. How tag systems degrade Nobody creates a bad tag deliberately. * Near-duplicate tags meaning the same thing * Tags