Marketing agency, paid ads, websites and SEO for local and online businesses
GoHighLevel Agency Setup: Empty Account to AI Agents in 5 Layers
A representative full-account solution build, documented end to end. The source labels it as showing how a setup of this kind is designed rather than a single named client deployment, and that framing is kept here.
- Our role
- GoHighLevel CRM architect
- Timeline
- Full-account build, designed as a reusable snapshot
- Sector
- Marketing agency, paid ads, websites and SEO for local and online businesses
At a glance
- Build areas
- 10: across 5 dependent layers
- Pipelines
- 3: sales, onboarding, retention (15 stages)
- Workflows
- 8: with shared exit rules
- AI agents
- 2: Conversation AI and Voice AI
- Calendars
- 3: including round-robin
- Tools consolidated
- 5 → 1 account
- Reusable output
- 1 snapshot for redeployment
- External dependencies
- A2P 10DLC registration, DNS records, social messaging
What was breaking
Five tools is not a tooling problem. It is a data problem. When forms live in one system, scheduling in another, email in a third, texting on personal phones and deal tracking in a spreadsheet, there is no single record of a lead. Nobody can answer what this person has already been told, who owns the conversation, or where the deal actually stands, because the answer is distributed across five systems that do not know about each other.
The consequences were predictable and all of the same shape. Enquiries waited hours because nobody was watching every inbox. Proposals went out and were never chased. No-shows were never rebooked. Nights and weekends were simply unanswered, which in a market where prospects contact three agencies means the other two got the conversation. And with no pipeline, there was no close rate, no pipeline value and no channel attribution, so the agency could not tell which of its own marketing worked.
Two of the problems were risks rather than inefficiencies. Business texts were being sent from personal phones with no consent records, which is a compliance exposure, not a workflow gap. And the agency's email went out from a domain with no authentication configured, quietly degrading deliverability for every message, including proposals. Those two needed fixing before any volume of automated sending was switched on, not after.
No single record of a lead
Five disconnected tools meant nobody could see a prospect's full history, so conversations repeated themselves and ownership was ambiguous.
Enquiries waited hours
Response time depended on who was watching which inbox, and prospects contacting several agencies rewarded whoever replied first.
Proposals and no-shows were never chased
The two highest-intent follow-ups in the funnel (someone who received a proposal, someone who booked and missed) relied entirely on memory.
Nights and weekends unanswered
A large share of enquiries arrive outside working hours, and all of them waited until morning.
No pipeline, no numbers
Without opportunities there was no close rate, no pipeline value and no attribution, so marketing spend could not be judged.
Texting from personal phones without consent records
A compliance exposure rather than an inconvenience, and one that becomes materially worse at automated volume.
Unauthenticated sending domain
Every email the agency sent was less likely to be delivered, with no error to reveal it.
What we built
The build is organised as five layers, deliberately ordered so each depends on the one below it. This is the part most agency setups get wrong: they start with the exciting layer (workflows, AI) on top of a foundation that has not been laid, and then spend months fighting symptoms. Field conventions, consent and deliverability are unglamorous and they determine whether everything above them works.
Layer one is foundation, set up once: account settings, users and roles, the custom fields every later workflow will read, custom values for the links and details shared across templates, and a tag naming convention (source, status and service prefixes) agreed before anyone starts creating tags ad hoc. Layer two is compliance, and it comes before any sending: A2P 10DLC brand and campaign registration, email domain authentication, an unticked consent checkbox on every form, and published privacy and messaging policies.
Layer three is capture, covering every way a lead can arrive: website and funnels, forms, a chat widget serving the site, SMS and social messages, and a business phone number with missed-call text-back. Layer four is the sales engine, three pipelines, calendars, workflows and opportunities. Layer five is AI and reporting: Conversation AI, Voice AI, dashboards, and the snapshot that makes the whole structure redeployable.
The three pipelines separate stages that genuinely behave differently. Sales runs new lead through contacted, discovery booked, proposal sent, negotiation to won or lost. Onboarding picks up at contract signed and moves through kickoff booked, assets received, in build, to launched. Retention tracks active, review due, renewal and upsell. Keeping them separate means each has stages that reflect reality and the agency can see where work actually sits, rather than one long pipeline where half the stages never apply.
Eight workflows cover the journey: speed to lead, which texts and emails within a minute, notifies the owner and creates the opportunity; missed-call text-back; booking confirmations with day-before and hour-before reminders; no-show recovery; proposal follow-up on days two, five and ten before creating a task for a human; long-term nurture every fortnight until the prospect replies or books; onboarding kickoff; and review requests. All eight share the same exit rules (quiet hours, stop-on-response, defined goals and consistent tagging) which is what prevents the classic failure where a prospect who has already replied keeps receiving the sequence.
The two AI agents cover different gaps. Conversation AI handles web chat, SMS and social messages: it qualifies on service, budget and timeline, then books into the round-robin calendar. It was started in suggestive mode, where it drafts and a human approves, before being moved to autopilot, which is the right sequence for anything that speaks to prospects in your name. Voice AI answers the main phone number after hours and when the line is busy, books appointments, sends a follow-up text, transfers genuinely urgent calls, and writes summaries and recordings back to the contact record. Both are constrained to the agency's knowledge base with an explicit handover path, because an AI agent inventing a service you do not offer is worse than an unanswered call.
Compliance and deliverability are the foundation, not the finish
A2P registration, consent capture and email authentication go in before anything sends at volume. Retrofitting consent onto a system already sending is far harder than building it in, and the failure mode is regulatory rather than technical.
One trigger, one owner
Two workflows reacting to the same form submission is a race condition that produces duplicate messages. One workflow owns each trigger; the others listen for the tags it sets. This single rule removes a whole class of bug.
Layer the build so dependencies run downwards
Fields and tag conventions before workflows; compliance before sending; pipelines before AI booking into them. Building top-down means rework at every level.
Start AI in suggestive mode
Conversation AI drafted replies for human approval before going autopilot. You learn where it strays on real conversations at zero risk, and the agency gains confidence in it rather than being asked for faith.
Shared exit rules across every workflow
Quiet hours, stop-on-response and explicit goals applied uniformly. Without them, sequences keep firing at people who already replied, which is the fastest way to make automation feel like spam.
Build it as a snapshot
Structuring the account for redeployment turns a one-off build into an asset. The same structure can be stood up again without rebuilding it from memory.
Non-negotiables
- Compliance before capability. Nothing that sends at volume gets switched on until A2P registration and email authentication are in place.
- One contact record ties everything together: every channel must write to the same record rather than creating its own.
- The founder sells and two account managers deliver; the build has to suit that shape rather than assume a large sales team.
- The whole account must be reproducible as a snapshot, so the structure can be deployed again rather than rebuilt by hand.
- No external automation platform and no custom code, the agency wanted one system to operate and one bill.
The pipeline
Drawn as text and shapes rather than shipped as an image, so the step labels stay readable and downloadable.
The five-layer build map
Each layer depends on the one before it. Compliance sits beneath every sending capability, and the AI layer books into pipelines that already exist.
- Data store or CRM
- Branch / filter
- Trigger
- Processing step
- AI / model call
- Delivered output
- Primary path
- Flow
Lead to client, across three pipelines
Capture routes into the sales pipeline; a won deal moves to onboarding; a launched client moves to retention. Separate pipelines mean each has stages that actually apply.
- Trigger
- Processing step
- AI / model call
- Data store or CRM
- Delivered output
- Primary path
How the work runs
Each stage, what it does, and the specific failure it is built to prevent.
Foundation: fields, values and a tag convention first
Account settings, users and roles, then the custom fields every later workflow will read and the custom values holding links and details shared across templates. The tag naming convention (source, status and service prefixes) is agreed here, before anyone can start inventing tags. Retrofitting a convention onto hundreds of existing tags is miserable; agreeing one takes an hour.
Prevents tag sprawl, which is the single most common reason a mature GoHighLevel account becomes unmaintainable.
Compliance: before anything sends
A2P 10DLC brand and campaign registration, email domain authentication, an unticked consent checkbox on every form, and published privacy and messaging policies. Sequencing matters: carriers reject campaigns whose consent language is not already live on the site, so the website wording goes up before the registration is submitted.
Avoids the common A2P rejection loop, and avoids building volume on a sending reputation that is already damaged.
Email deliverability: one SPF record, warmed gradually
A sending subdomain is configured with SPF, DKIM, a tracking record and DMARC. The domain already carried an SPF record for another service, and only one is permitted, so the two were merged rather than added alongside, a second SPF record silently invalidates both. Sending then warms up gradually, starting with engaged contacts.
A duplicated SPF record is a classic misconfiguration that breaks authentication while appearing to be correctly set up.
Capture: every channel into one record
Website and funnels, forms, a chat widget covering the site plus SMS and social messages, and a business phone number with missed-call text-back. Duplicate rules on email and phone keep the same person from becoming three contacts, which is what makes 'one record' true rather than aspirational.
Duplicate detection on email and phone prevents the fragmented history the whole build exists to eliminate.
Three pipelines, fifteen stages
Sales, onboarding and retention are separate because their stages behave differently. One combined pipeline would mean most stages never apply to most records, which makes pipeline value and close rate meaningless. Separate pipelines give each phase of the relationship stages that reflect what actually happens.
Eight workflows with shared exit rules
Speed to lead, missed-call text-back, booking confirmations, no-show recovery, proposal follow-up, long-term nurture, onboarding kickoff and review requests. Every one of them honours quiet hours, stops on response, and has an explicit goal. Applying those rules uniformly is what keeps eight workflows from feeling like eight separate senders.
Stop-on-response prevents the worst automation experience there is: continuing to chase someone who already replied.
One trigger, one owner
Two workflows initially reacted to the same form submission, producing duplicate messages. The rule that resolved it is structural: exactly one workflow owns each trigger, and anything else that needs to react listens for the tags that first workflow sets. It is a convention rather than a feature, and it eliminates an entire category of race condition.
Removes duplicate-send races, which are hard to reproduce and easy to prevent.
Round-robin booking with even distribution
Three calendars including a round-robin for discovery calls. The first configuration overloaded one person, so distribution was set to even and per-user calendar sync enabled, otherwise the system books people who are not actually free, which is worse than not booking at all.
Conversation AI: suggestive first, then autopilot
Handles web chat, SMS and social messages, qualifying on service, budget and timeline before booking into the round-robin. It ran in suggestive mode first, drafting replies a human approved, so its weak spots surfaced on real conversations before it spoke unsupervised. Its answers are limited to the agency's knowledge base, with an explicit handover path when a question falls outside it.
Knowledge-base scoping stops the agent describing services the agency does not sell, which is the failure that actually costs money.
Voice AI: the nights and weekends gap
Answers the main number after hours and when the line is busy. It books appointments, sends a follow-up text, transfers genuinely urgent calls to a human, and writes a call summary and recording back to the contact record. This directly closes the gap where every out-of-hours enquiry previously waited until morning.
Dashboards and a reusable snapshot
Dashboards surface pipeline value, close rate and channel attribution, the numbers that did not exist before, because there was no pipeline to derive them from. The whole account structure is then captured as a snapshot, turning the build into something redeployable rather than a one-off.
What changed
| Aspect | Before | After |
|---|---|---|
| Systems of record | Forms, scheduling, email, texting and deal tracking across five separate tools. | One account, one contact record that every channel writes to. |
| Response time | Enquiries waited hours, depending on who was watching which inbox. | Acknowledged within a minute on any channel, day or night. |
| Out of hours | Nights and weekends went unanswered until the next working day. | Voice AI answers, books, texts a follow-up and transfers urgent calls. |
| Proposal follow-up | Sent and then never chased. | Automated touches on days 2, 5 and 10, then a task for a human. |
| No-shows | Not rebooked. | A recovery sequence with a rebooking link, then nurture if unused. |
| Business texting | Sent from personal phones with no consent records. | Registered A2P 10DLC campaign with consent capture and STOP/HELP handling. |
| Email deliverability | Sending from a domain with no authentication configured. | Dedicated sending subdomain with SPF, DKIM and DMARC, warmed gradually. |
| Reporting | No pipeline value, close rate or channel attribution. | Dashboards over three pipelines with attribution by source tag. |
| Repeatability | Any new build would start from scratch. | The structure is captured as a reusable snapshot. |
What it runs on
Every tool listed with the job it actually does in this build.
Platform
GoHighLevel
The whole stack: CRM, pipelines and opportunities, calendars, forms, websites and funnels, workflows, chat widget, custom fields and values, tags, duplicate rules, smart lists, dashboards and snapshots.
AI
GoHighLevel Conversation AI
Qualifies inbound chat, SMS and social messages on service, budget and timeline, then books into the round-robin calendar. Scoped to the agency's knowledge base with a handover path.
GoHighLevel Voice AI
Answers the main phone line after hours and when busy, books, texts a follow-up, transfers urgent calls, and writes summaries and recordings to the contact.
Delivery
LC Phone and LC Email
Native calling, texting and sending, so conversations land on the contact record instead of on someone's personal phone.
A2P 10DLC registration
Carrier-level brand and campaign registration with consent handling and STOP/HELP keywords, the prerequisite for compliant business texting at volume.
DNS (SPF, DKIM, DMARC, tracking)
Email authentication on a dedicated sending subdomain, with the pre-existing SPF record merged rather than duplicated.
Integration
Facebook and Instagram messaging
Social DMs routed into the same inbox and the same contact record as every other channel.
What stopped being manual
Snippets are illustrative and generic. No client data appears in any of them.
Speed to lead across every channel
Replaces someone noticing a form submission in an inbox. A lead arriving on any channel is acknowledged within a minute, the owner is notified, and the opportunity is created, regardless of the hour or who is at a desk.
text
trigger: form | chat | social DM | inbound call
-> send SMS + email within 1 minute
-> notify owner
-> create Sales opportunity (stage: New Lead)
-> apply tags: source-*, status-contacted
// honours quiet hours; stops immediately on responseThe two follow-ups that were never happening
Replaces memory for the two highest-intent moments in the funnel. A proposal that goes unanswered and an appointment that was booked and missed are both strong signals, and both were previously dropped.
text
proposal sent -> day 2 follow-up
-> day 5 follow-up
-> day 10 follow-up
-> create task for a human // automation stops here
appointment no-show -> recovery sequence with rebooking link
-> if no rebooking: long-term nurtureOne trigger, one owner
Replaces a duplicate-send race condition. Rather than two workflows both reacting to a submission, one owns the trigger and the other reacts to the tag it sets, making ordering deterministic.
text
// BEFORE: both fire on form submission -> duplicate messages
// workflow A: on form submitted -> send welcome
// workflow B: on form submitted -> notify owner
// AFTER: one owner, others listen for its tags
workflow A: on form submitted -> send welcome
-> apply tag status-contacted
workflow B: on tag status-contacted -> notify ownerAI constrained to what the agency actually sells
Replaces an unanswered phone, without introducing the worse failure of an agent inventing services. The AI answers from the knowledge base and hands over when a question falls outside it.
text
agent scope:
answer ONLY from the agency knowledge base
qualify on: service, budget, timeline
book into: round-robin discovery calendar
on out-of-scope question -> hand over to a human
on urgent call -> transfer immediately
always -> write summary + recording to the contact recordWhat it was worth
The structural change is that a lead became a single record. Everything else in the build (attribution, close rate, ownership, knowing what someone has already been told) follows from that one fact, and none of it was possible while five tools each held a fragment of the truth. This is why the foundation layer comes first: it is the layer that makes the others mean anything.
The gaps that closed are the ones that were pure loss. Proposals now get chased, no-shows get rebooked, and out-of-hours enquiries get answered instead of waiting for morning. These were not inefficiencies to optimise; they were revenue leaving silently, and each of them is now covered by something that runs whether or not anyone remembers.
Two of the fixes are risk reduction rather than growth. Registered A2P messaging with real consent records replaces texting clients from personal phones, and an authenticated sending domain replaces email that was quietly less likely to arrive. Neither shows up as a win in a dashboard. Both would have shown up eventually as a problem.
On honesty: the source material labels this a representative solution build, a documented design showing how a full agency setup of this kind is assembled, rather than a single named client deployment. The only before-and-after pair it states is 'waited hours' becoming 'within a minute', and the 'before' there is qualitative. I would rather present an accurate architecture with honest framing than attach invented percentages to it.
Tools consolidated into one account
Forms, scheduling, email, texting and deal tracking.
Build layers, each depending on the last
Foundation, compliance, capture, sales engine, AI and reporting.
Pipelines, 15 stages
Sales, onboarding and retention kept separate so stages actually apply.
Workflows with shared exit rules
Quiet hours, stop-on-response and explicit goals applied uniformly.
AI agents
Conversation AI across chat/SMS/social, Voice AI on the phone line.
Reusable snapshot
The structure can be redeployed rather than rebuilt.
Handed over
- A five-layer account build covering ten build areas, ordered so each layer depends only on those beneath it.
- Three pipelines with fifteen stages: sales, onboarding and retention.
- Eight workflows (speed to lead, missed-call text-back, booking confirmations, no-show recovery, proposal follow-up, long-term nurture, onboarding kickoff and review requests) sharing quiet hours, stop-on-response and explicit goals.
- A2P 10DLC brand and campaign registration with consent capture on every form and STOP/HELP keyword handling.
- A dedicated sending subdomain with SPF, DKIM, tracking and DMARC records, with the pre-existing SPF record merged rather than duplicated.
- Multi-channel capture: website and funnels, forms, chat widget across site/SMS/social, and a business number with missed-call text-back.
- Three calendars including an evenly-distributed round-robin with per-user calendar sync.
- Conversation AI qualifying and booking across chat, SMS and social, introduced in suggestive mode before autopilot.
- Voice AI covering after-hours and busy-line calls, with transfers, follow-up texts and summaries written back to the contact.
- Duplicate rules on email and phone, a documented tag naming convention, smart lists and dashboards.
- A reusable snapshot of the entire account structure.
Questions about this build
What teams usually ask when they recognise their own process here.
Bottom-up, in five layers. Foundation first, settings, users, custom fields, custom values and a tag naming convention. Then compliance: A2P registration, email authentication and consent capture, before anything sends at volume. Then capture across every channel, then the sales engine of pipelines, calendars and workflows, and only then AI and reporting. Most struggling accounts were built in the opposite order, with workflows and AI sitting on a foundation that was never laid, and the resulting problems look like workflow bugs when they are actually missing fields and conventions.
Because carriers check that your consent language is already live on your website and that your sample messages match what you actually send. If you build and launch the workflows first, you submit sample messages for a campaign that is already sending, and a rejection means you are non-compliant in production rather than merely waiting. Registering first also means the inevitable back-and-forth happens while you are still building, not while leads are arriving.
No. Start in suggestive mode, where the agent drafts and a human approves before anything is sent. You find out where it strays on real conversations at no risk, and the team gains justified confidence instead of being asked for faith. Move to autopilot once the drafts are consistently what you would have sent. Scope it to your knowledge base with an explicit handover path either way, an agent that confidently describes a service you do not offer costs more than an unanswered message.
Because sales, delivery and renewal behave differently, and one combined pipeline means most stages never apply to most records. That makes pipeline value and close rate meaningless, since you are averaging across stages that are not comparable. Three pipelines give each phase stages that reflect what actually happens, and the handoffs between them (won to onboarding, launched to retention) become explicit events you can automate and measure.
It is the rule that exactly one workflow fires on any given trigger, and anything else that needs to react listens for the tags that workflow sets. Without it, two workflows both reacting to a form submission race each other and the prospect gets duplicate messages, a bug that is intermittent, hard to reproduce, and visible to your leads. It costs nothing to adopt and removes the entire category.
It is a representative solution build: a full account design, documented end to end, showing how a setup of this kind is assembled, not a single named client deployment. The source material labels it that way and I have kept that framing deliberately. The architecture, the obstacles and the fixes are all real and transferable; what I am not claiming is measured performance in a specific named account.
Most of it. The foundation, compliance, capture and workflow layers are generic to any business that generates leads and books appointments, the vocabulary changes and the pipeline stages change, but speed to lead, missed-call text-back, no-show recovery and stop-on-response are universal. What is agency-specific is the three-pipeline split reflecting sell, deliver, retain, and the snapshot being built for redeployment across client accounts.
Capabilities this used
Read next
All gohighlevel services workHave a process that works like this one did?
Book a strategy call. We will walk the workflow with you and say plainly which parts are worth automating and which are not.