An agent that knows when to stop.
Every AI vendor sells autonomy. The useful part is the opposite — an assistant that does the first three messages properly and hands back the moment a person should be doing it.
Thirty minutes · no predetermined package
Three ways it stops, and it cannot override any of them.
These are not prompt instructions. They are conditions in the code, defaultChecked before every outbound message.
A human replies
The rep answers from their own phone and the assistant stands down mid-thread. It does not resume, and it does not ask permission first.
The person asks for a person
Any request to speak to someone real ends the automated thread and creates a task for the rep, with everything captured so far attached.
The message ceiling
A fixed number of outbound messages per conversation. When it is reached the assistant stops, whether or not it finished collecting anything.
What it will never do, and why that is a feature.
Most of the design work on an assistant is deciding what it is not allowed to attempt.
What we have actually built.
An SMS assistant inside a lending platform
Texts as the rep who owns the record, captures the qualifying floor, and hands the thread back the moment a human should have it.
A content agent on a schedule
Runs unattended on a cadence, produces work that a person reviews before anything is published. Autonomy in the execution, not in the decision.
Where an assistant earns its place.
Six situations where automated lead follow-up is doing work a person should not be doing.
An enquiry arrives at nine at night and the first human contact is the following afternoon.
Follow-up stops after one attempt, because nothing in the system enforces the second.
The same four qualifying questions get asked by a person, on every single enquiry.
Reps spend the first ten minutes of every call collecting facts that could have arrived with the record.
Leads that went quiet are never revisited, because nobody owns the list of them.
An appointment setter costs more per booked meeting than the meeting is worth at your volume.
Six questions before anyone builds an AI agent.
The answers usually decide the shape of the thing, and sometimes decide against building it.
Five stages, and the boundary comes first.
Write the boundary
the allowed listExactly what the assistant may say, may ask, and may never attempt. This is the deliverable of week one and everything else follows from it.
Wire the guardrails
conditions in codeConsent, suppression, quiet hours, message ceiling and stand-down, implemented as checks rather than instructions.
Fields, not transcripts
structured captureEach answer maps to a field on the record. If there is nowhere to write it, the assistant does not ask for it.
Run alongside a human
supervisedLive on a small share of volume with a person reading every thread, until the transcripts stop producing surprises.
Instrument and hand over
cost visibleTokens, spend and outcome per conversation on a dashboard the operator owns, plus the runbook for changing the allowed list.
Custom AI agent development is not always the answer.
Three options, and the last one is the most common. We will tell you when it applies.
A custom agent on your own data
Reads and writes your records, obeys your compliance rules, and stops on your conditions. Built on the Anthropic or OpenAI API against your system.
A packaged AI SDR or setter
Fast to start and genuinely fine for a standard outbound motion. The limit arrives when it needs to know something only your system holds.
Nothing — fix the routing first
Most “we need AI” conversations turn out to be a speed-to-lead problem. An assistant on top of broken routing automates the delay.
The sector changes the script. It does not change the guardrails.
Lending and financial services
High enquiry volume, strict consent rules, and a qualifying floor that is the same on every file.
Home services
Demand arrives in weather-driven bursts, and the call that goes to voicemail goes to a competitor.
Legal intake
After-hours enquiries where the first responder usually wins, and a conflict check that must precede anything substantive.
Professional services
Low volume, high value, and a follow-up cadence that everyone agrees on and nobody maintains.
What people ask about AI agents.
What is an AI agent, in a business context?
An AI agent is software that carries out a defined task across several steps without a person driving each one — reading a record, sending a message, writing an answer back. The useful ones have narrow permissions and a stop condition.
How does an AI SMS assistant qualify a lead?
It asks a fixed set of qualifying questions in conversation, maps each answer to a field on the lead record, and stops when the floor is met or a human takes over. The capture is structured data, not a transcript somebody has to read.
What are AI guardrails and why do they matter?
Guardrails are conditions defaultChecked in code before an agent acts — consent, suppression, quiet hours, message ceilings, stand-down triggers. Prompt instructions are not guardrails, because a model can be talked out of an instruction.
Should we build a custom AI agent or buy one?
Buy when the motion is standard and integration depth is not the constraint. Build when the agent has to read and write your own records, obey rules specific to your business, and stop on conditions only your system knows.
What does custom AI agent development cost to run?
The build is the larger number; the running cost is usually smaller than people expect. A qualifying conversation of a few messages typically costs cents in tokens, which is why per-conversation cost should be visible from day one.
Can an AI assistant text customers legally?
Only with recorded consent, defaultChecked before every message, and with identification in the first message. TCPA damages are per message, so consent and suppression have to be code paths rather than policies.
What happens when the assistant cannot answer?
It stops and creates a task for the person who owns the record, with everything captured so far attached. The failure mode to avoid is an assistant that improvises rather than escalating.
How long does it take to put an agent into production?
The allowed list and the guardrails are the work; the model call is not. Expect to run supervised alongside a human for a period before anything operates unattended.