Skip to content
About How It Thinks

We build agents. Then prove how they think.

How It Thinks is a small team that builds AI agents for companies, with evidence of how they make decisions.

Voice is a practical starting point. A conversation puts an agent’s judgement in front of a customer in real time. We focus on what the agent may do, what it needs to verify and when it must hand off to a person.

How we operate

Built by engineers. Supported by AI-assisted workflows.

Building an agent is easier than knowing how it will behave. The same input can lead to different answers, and the stakes rise when an agent books, routes or acts on its own. Engineers own the architecture, tool boundaries, edge cases and final review. AI-assisted workflows support the groundwork, while people remain accountable for the decisions that matter. The diagram illustrates that division of responsibility.

ANATOMY OF A BUILD PODILLUSTRATIVE
HUMANS OWN THE JUDGEMENT
System architecture and tool boundaries
Expert alignment on what "right" looks like
Edge-case judgement and final sign-off
DIRECTED BY BUILDERS, EXECUTED BY AGENT CREWS
AGENT CREWS DO THE VOLUME
How we prove it

Four things we won't trade.

"Stochastic heuristic latency"Paused: waiting on policy ID
01 · PLAIN ENGLISHPlain language, zero black boxes.We explain the job, the decision trace and the limits in words everyone in the room can challenge.
ThinkprintAGREED SCENARIOS → REVIEWABLE EVIDENCE
02 · OBSERVABLE PROOFObservable proof over demos.The Layered Mind examines grounding, reasoning, judgement, consistency and resilience. A Thinkprint records the decisions examined, the boundaries tested and the gaps found. It is evidence from defined tests, not a promise that an agent is perfect. Changes to models, prompts, tools or policies call for the relevant tests to be repeated.
Model AModel BModel C
03 · ZERO VENDOR LOCK-INModel-agnostic by default.We use the best model for each step and orchestrate across providers, tied to none. Tap the task to reroute it.
AGREED DEPLOYMENT ENVIRONMENTData use and retention agreed before work
04 · YOUR ENVIRONMENTYour deployment, agreed together.Before work starts, we agree the deployment environment, data use and retention with you.
VOICE CASE STUDY

Patient outreach, by voice.

The problem

A patient outreach platform had sleep test referrals piling up. Staff called every patient by hand, most calls went unanswered, and appointment follow-ups slipped.

What we built

Voice agents that call each patient, explain the sleep test, answer common questions, book a slot, and follow up before and after the appointment. Anything clinical or unclear goes straight to a person.

How we proved it

Before go-live we mapped how the agent decides when to book, when to reschedule and when to hand off. We ran it through 500 simulated calls, including interruptions, confused callers and wrong numbers, and shared a Thinkprint showing where its decisions held and where we tightened them.

The result

3x

more patients reached each week

42%

of referrals enrolled with no staff time

28%

fewer missed appointments

6 weeks

Live in 6 weeks

Start a build. Explore our thinking.

EXPLORE THE IDEASWe also make explainers about how AI systems decide, sharing the ideas behind our work in plain language.Browse explainers
START A BUILD
Workflow intakemapbuildproveThinkprint
Tell us what you want built, and we will show you how we would evaluate its decisions.Describe your workflow