Which AI pattern does your firm need?

Firms usually ask which AI tool to buy. The more useful question is which shape the system should take, because the same tool can be wired five different ways and the differences matter enormously. A classification task and a payment approval do not deserve the same architecture. Here are the five patterns in plain language, and the one question that tells you which you need.

The question that picks the pattern

Ask what happens when the AI is wrong. Not if. When. If a wrong answer means someone rereads an email, you need almost nothing. If it means a client receives incorrect advice, a filing goes out wrong, or money moves, you need a system designed around catching the error before it escapes.

Everything below is ordered by how much checking is built in. Cost, latency, and complexity rise as you go down the list, so the goal is the simplest pattern that survives being wrong in your particular context.

One: the single answer

The model is asked once and responds once. No tools, no loop, no second look. This is the right shape for classification, extraction, summarizing, and reformatting. Turning an inbound message into structured fields is a textbook case.

It is fast and cheap, and it is the correct answer far more often than vendors suggest. Many firms overbuild here, wiring an elaborate agent for work a single well-specified request handles perfectly.

Two: reason, act, repeat

The model decides a step, takes it using a tool, looks at the result, and decides again until it is finished. This is the shape behind most research assistants: look something up, read what came back, decide whether it is enough, look again.

It handles open-ended work that a single pass cannot, which is its appeal. It is also slower, more expensive, and has more places to go wrong, so it needs limits: how many loops before it stops, what it may call, what happens on failure. An unbounded loop is a budget problem waiting to happen.

Three: plan first, then execute

One component writes a plan as ordered steps. Others carry out each step, sometimes in parallel. Results are merged at the end. This suits long multi-step work: assemble a client report, run an analysis across several sources, process an intake end to end.

The advantage is visibility. Because the plan exists as a separate artifact, a person can read it before anything executes, which is a natural approval point in a regulated setting. You can review intent before you review output.

Four: draft, critique, improve

The system produces a draft, critiques its own work against quality criteria, and revises. Useful where the first attempt is reliably rough: long-form writing, code, involved reasoning.

Here is the limit worth knowing. Self-critique improves polish and catches obvious errors, but it is not independent. A system confidently wrong in its draft is frequently confidently wrong in its critique, because the same reasoning produced both. It raises quality; it does not establish correctness.

Five: nothing moves until a verifier says so

An independent layer checks the proposed output against rules that do not come from the model: a schema, a policy, a calculation, an external system of record. Pass and it proceeds. Fail and it is rejected, retried, or escalated to a person.

This is the pattern that matters most for professional firms, and it is the one your approval workflows already describe without using the name. A filing checked against deadline rules before submission. A fee quote checked against the engagement letter. A client communication checked against advertising rules. The verifier is deterministic where it can be, which is precisely why it catches what a model checking itself does not.

It costs latency and requires you to write the rules down. That second cost is the real one, and it is also the reason the pattern works: writing the rule down forces the firm to decide what the rule actually is.

Real systems combine them

Production systems are rarely one pattern. A common shape: a planner breaks down the work, executors use tools to carry out steps, a reflexive pass improves the draft, and a verifier gates the final action. Memory and logging run underneath, guardrails wrap the whole thing.

The design discipline is choosing the simplest arrangement that handles being wrong, then adding a verifier specifically at the steps where wrong is expensive. Not everywhere. At those steps.

How Aday Interactive, Inc. picks for you

In a Custom AI engagement we start with your tasks, not with an architecture. We sort them by what a wrong answer costs, then match each to the lightest pattern that holds, and put the verifier only where the cost justifies it.

The outcome is usually less system than firms expect. Most work is single-shot or a short loop. The verifier goes around the handful of steps that touch filings, money, or advice, and those are the steps worth the extra engineering.