Most AI agents only talk. They sound helpful and leave your team with exactly the same amount of work, because nothing was written anywhere. I build agents that book the appointment, create the lead, issue the payment link, update the order, and hand off to a person when they should. Whatever the job, on the channels your customers already use, writing into the CRM you already run.
customer
"can you move my Thursday appointment to 4pm?"
policy engine
allowedreschedule · within_business_hours · slot_free
write
201 · 340 msPATCH /appointments/8812 → your calendar
agent
"Moved to Thursday at 4pm. You'll get a reminder an hour before."
if the write had failed
the agent says so and hands off. It never returns a confirmation for something that did not happen.
Where this fits
The channel, the industry and the process are yours to choose. Three ways this usually starts:
The agent reads and writes in tools you already run, and those tools have API docs you can link to. Pick a tier below and we start this week.
Perfectly normal, and it is what the Audit is for. In five days I map what your stack can actually do, and the fee comes off the build.
Hours, address, prices and nothing written anywhere? Meta's own business agent does that for almost nothing. I will show you how to set it up and charge you nothing for it.
Three shapes
An agent is worth paying for when it makes something happen: finds the right answer inside your own material, writes the line into your system, or closes the customer's request without going through anyone. Who it works for is your call.
An agent over the company's own material, in its own panel or inside the app you already run. It searches your documents, answers with the source cited, and each person sees only what they are allowed to see. Contracts, manuals, procedures, precedent, account history.
It executes and records: books, confirms, charges, updates the order, writes into the CRM, the ERP or the spreadsheet. If the system refuses the operation, it escalates instead of reporting that it worked.
WhatsApp, website or phone, wired to the rest of the operation. On WhatsApp it runs on Meta's official Cloud API, on your own number, because an unofficial connection gets the number banned and the ban has no appeal.
What I build
These are the shapes that come up most. If yours is not here, tell me which systems it would write into and we will build that one.
Checks real availability in your calendar, writes the appointment, sends reminders, cuts no-shows.
Try it liveAsks the right questions, scores the lead, creates it in your CRM and routes the hot ones to the right person.
Proposes only what your pricing policy allows. It never invents a discount, because it is not the one deciding.
Status, tracking, exchanges and the second purchase, reading and writing in your commerce backend.
Issues the charge, follows up on the agreed date, records the promise to pay.
Resolves what it can, opens the ticket for what it cannot, and escalates with the full context attached.
Live demo, no signup
A fictional clinic whose ERP is published alongside it. You talk, the row changes in the schedule live, and the decision log shows which rule decided each step. Ask for a last-minute cancellation and watch it refuse. Hit the button that takes the ERP down and watch it escalate instead of inventing a confirmation. That last one is the part nobody demos.
Channels and systems
The channel is a setting, not a product. The same agent answers wherever your customer already is and writes into whatever you already run. Multiple channels share one memory, so nobody has to repeat themselves after switching.
The brand does not decide whether I can build it. One thing does: the system exposes a documented API you can link me to. If it does not, the Audit is the honest starting point, and I say so before you pay for a build.
In every build
These are not tier features. They ship in the cheapest build for the same reason they ship in the most expensive one: without them, an agent that acts is a liability rather than an asset.
A deterministic engine built from your written rules decides what is allowed. The language model only phrases the answer it was authorized to give.
Every decision is stored with the id of the rule that produced it and the version of the policy in force. Months later you can still prove what decided what.
An operation that changes data runs only after the person confirmed that specific operation. No silent writes.
With an alert before it is reached and a kill switch that belongs to you, not to me.
Your keys, your invoice, direct from Meta and the AI providers. I never resell usage with a margin hidden inside it.
If your system is down, the agent says it could not do it and calls a person, instead of confirming something that never happened.
Pricing
Not by how many messages it sends. The number that moves the price is how many of your systems it writes into, because that is where the work actually is.
One process, one system, one channel.
Delivered in 10 days · 2 revisions
Several processes, a decision engine, two channels.
Delivered in 21 days · 4 revisions
Custom policy engine, multichannel, self-hosted option.
Delivered in 35 days · unlimited revisions for 30 days
Five days. I map your processes, test whether your systems can actually be written to, and send back a scope with a fixed price. If the honest answer is that you do not need to hire anyone, that is what the report will say. The fee comes off any build you buy afterwards.
Running costs are yours. Meta, AI providers and telephony bill you directly on your own accounts, with your own keys. I never resell usage. A typical small business spends $15 to $50 a month.
Live phone calls are a different project. Understanding a recorded voice message is included from Standard up. An agent that answers a ringing phone is my AI Phone Receptionist, priced separately.
A revision is a change inside the delivered scope. Tone, wording, behaviour fixes. A new flow, integration or channel is a scope change and gets its own quote.
Multi-agent
Push everything into a single agent with twenty tools and a giant prompt and three things get worse together: it picks the wrong tool more often, cost per task climbs as context grows, and you lose the ability to tell where an answer went wrong. Splitting it fixes that and creates a new problem, coordination. So here the coordination is deterministic code, not one more model deciding.
An orchestrator decides which specialist handles which part, in what order, and merges the result. Each specialist holds few tools and one narrow goal, which is what makes the right choice predictable.
Several agents examine the same thing from different angles and the result only passes with enough agreement. Worth the extra cost when being wrong is expensive.
One agent produces, another is instructed to distrust it and try to knock the claim down. What survives becomes the answer. This is what kills the plausible and wrong reply.
Where the decision is sensitive, the flow stops and waits for a person to approve, with everything done so far on the screen.
Multi-agent is slower and more expensive than a single agent. If your problem fits one agent with four tools, that is what I will recommend, even though it is the smaller contract.
The question that decides is not how many agents. It is how many independent decisions the process contains. Two decisions do not justify four agents.
Multi-agent routing ships in the Advanced tier.
Process
I read your answers and your sample conversations, then send back a flow map and the decision rules in writing. Nothing gets built before you approve that.
The channel goes live. On WhatsApp that means your number registered on your own Meta Business account, with templates submitted for approval. The account is yours from that moment on.
I prove read and write against your sandbox before a single flow is built. This is the step where projects like this fail, so it goes first, while there is still time to change course.
Flows, decision rules and guardrails. You get a test channel early. Then a recorded conversation suite plus abuse tests, with the results shared, not summarised.
Gradual volume ramp-up while your messaging limits climb, monitoring in place, and a kill switch you control. Handover call plus written documentation.
Boundaries
Saying this up front costs me some enquiries and saves both of us the wrong project.
Baileys, Evolution in QR mode, WPPConnect, Venom, Z-API. Meta has been removing these connections since January 2026 and the ban is permanent with no appeal. The number is your asset, not mine, so I will not put it at that risk to save you a hosting fee.
A deterministic engine built from your written rules decides what is allowed. The language model only phrases the answer it was authorised to give. Anything outside policy is blocked before it is sent. That is what keeps this defensible when a customer disputes what your agent said.
Bans for unauthorised API use are permanent and there is no effective appeal. Anyone promising recovery is not being honest with you. What I can do is get you operational on a new number, properly, fast.
If the agent only answers questions and writes nothing anywhere, Meta's own business agent already does that for almost nothing. I will show you how to set it up and charge you nothing for the advice.
FAQ
Not through me. Everything on WhatsApp runs on Meta's official Cloud API. Bans happen to businesses using unofficial libraries. I never use those for client work.
In most cases yes, through Meta's Coexistence flow. Tell me the number when you start and I confirm eligibility before you are committed to anything.
You do, from day one. Your Meta Business account, your number, your provider keys. If you stop working with me, nothing switches off and nothing has to be migrated.
No. WhatsApp, a website widget, Instagram or voice. The agent is the same and the channel is a setting. Multiple channels share one memory, so your customer never repeats themselves.
It executes decisions, it does not invent them. Those come from a deterministic engine built from your written policy, and the agent only phrases the answer it was authorised to give.
Platform and AI providers bill you directly, on your own accounts. A typical small business spends $15 to $50 a month. Note that since October 1, 2026 WhatsApp service messages carry a per-message fee, and I design your flows with that in mind.
Then we start with the Audit instead of a fixed-price build. I map what is possible, including the workarounds worth doing, and the fee comes off whatever we build next.
Most clients move to a monthly retainer for tuning, new flows and monitoring. That is a separate agreement and is never bundled into the build price without you asking for it.
Pick a tier and answer six questions about your systems. The form reads your answers and tells you which tier they actually support before you pay for anything.