Follow-up that runs itself
No answer becomes a retry on a schedule that respects local hours and attempt caps. A soft yes becomes a booked slot on the right calendar. The sequence keeps moving without a human deciding to move it.
Codiste Portfolio · Voice AI for GTM
From Sequence
to Conversation
"What if every lead got a real conversation before the tab closed?"ReachOut's voice agents call in seconds, then write the CRM themselves.
Discuss your requirementA voice layer inside the platform
teams already run.
ReachOut had the sequences, the inbox, the lists and the CRM sync. What it did not have was a way to actually talk to anyone. Every conversation still depended on a rep having a free hour and the will to dial.
Codiste built that missing layer directly into the product. Not a separate dialer bolted to the side, and not a chatbot wearing a phone number. A native voice capability that reads the same triggers, respects the same field config, and writes back to the same records the rest of ReachOut already uses.
Trigger-driven dialingInstant or scheduled
Parallel callingMulti-line dial engine
CRM write-backSalesforce and HubSpot
Automated follow-upRetry and reschedule
Multi-language11 markets at launch
The Challenge
Where the sequence
goes quiet.
Four points where an outreach motion stalls, and what a voice layer does about each of them.
01
Speed-to-lead decay
A prospect fills the form and moves on with their afternoon. The rep opens the task four hours later, and the moment that made the lead worth calling has already passed. The follow-up is now an interruption instead of a reply.
02
The dialer tax
Most of a calling block is spent on ring tone, voicemail greetings and gatekeepers. The talking, which is the only part that books meetings, is a thin slice of the hour that the rep pays for in full.
03
CRM drift
Notes get written from memory at the end of the day, or they do not get written. Forecasts and pipeline reviews then run on a record that nobody in the room fully trusts.
04
A team-shaped ceiling
Coverage stops at the languages the team speaks and the hours they are awake. Every new region needs a new hire before it can be given a strategy, so expansion is gated on recruiting rather than demand.
Codiste's Solution
A voice layer, not a bolt-on.
Three systems, built into ReachOut's existing product surface and shipped entirely behind their brand.
01
Trigger-driven dialing
A call fires the moment a trigger does. Form submit, pricing page visit, reply sentiment, or a window the rep sets by hand. The agent checks local calling hours and consent state before it ever picks up the line, so speed never costs compliance.
02
The parallel dial engine
The agent opens several lines at once, drops the ones that land in voicemail, and stays with the one that answers. Ring time stops being something a human has to sit through, and the rep only ever joins a live conversation.
03
Write-back by default
Every call returns a transcript, a summary, a disposition and the field updates, straight into Salesforce or HubSpot against the right record. Nobody types a note, and the next step in the sequence fires off what actually happened on the call.
What we shipped.
11Languages live at launch
Each market gets its own voice, its own script, its own calling window and its own consent language. Adding the twelfth is a config change, not a release.
Voices that survive a real call
Barge-in, backchannel and recovery from cross-talk. The agent stops talking the moment the prospect starts, which is the difference between a conversation and a recording.
Qualification the rep defined
Scoring criteria come from ReachOut's own field configuration, so a sales lead can move the bar on a Tuesday without filing an engineering ticket.
Compliance in the path, not around it
Consent capture, do-not-call enforcement, per-region recording rules and a full audit trail on every call. The checks run before the dial, not in a report afterwards.
Shipped inside their stack,
not beside it.
Their repo, their conventions
We worked inside ReachOut's codebase, branching model and review process. There is no parallel service for their team to maintain and no second dashboard for their customers to learn.
Their brand, not ours
The feature ships as ReachOut. Codiste does not appear in the product, the docs, or the call audio. The credibility of the release belongs to the client.
No migration asked of anyone
Existing sequences, lists and CRM mappings kept working throughout the rollout. Voice arrived as another step type in a builder people already knew, so adoption needed no retraining.
Handover, not dependency
Runbooks, load-test results, prompt and voice tuning notes, and two weeks of paired work with their platform team before we stepped back. The client can change the thing we built without calling us.
The tools under the hood.
Real-time media and speech on one side, orchestration and system-of-record sync on the other.
TwilioTelephony
LiveKitReal-time media
DeepgramSpeech to text
ElevenLabsSpeech synthesis
OpenAIReasoning
TemporalOrchestration
Node.jsServices
PostgreSQLSystem of record
RedisSession state
SalesforceCRM sync
HubSpotCRM sync
AWSInfrastructure
Value & ROI
The numbers behind it.
< 30 sec
Time to first call
Median gap between a trigger firing and the prospect's phone ringing. The previous flow put that same lead in a task queue where it waited on a human opening a tab.
8x
Dial throughput
Connected conversations per rep hour, measured against the platform's previous single-line dialer. The gain comes from removing wait, not from talking faster.
100%
Calls logged
Every call returns a transcript, a summary and a disposition to the CRM. Manual note-taking left the workflow entirely, and pipeline reviews stopped running on memory.
11
Languages live
Each with its own voice, script and local calling window, all live on the day the feature shipped. Coverage stopped being a function of who the client had hired.
Need a voice layer inside your own product?
We build voice AI into platforms that already have customers, on the client's stack and under the client's brand.
Discuss your requirementHow Codiste shipped it in 45 days.
Five sprints from kickoff to a production rollout behind a feature flag, run alongside the client's own release train.
Day 01 - 05
Telephony and architecture spike
Carrier routing, the concurrency model, and a latency budget agreed in writing before any product code was committed. Getting the budget wrong here is expensive to discover in week five.
Day 06 - 15
The real-time voice pipeline
Speech in, reasoning, speech out, tuned until a turn came back inside a second on the median call. Anything slower and the prospect starts talking over the agent, which no amount of good scripting recovers from.
Day 16 - 25
Parallel dialing and CRM write-back
The dial engine, the disposition model, and bidirectional sync with Salesforce and HubSpot mapped against the client's existing field configuration rather than a schema of our own.
Day 26 - 35
Languages and voice tuning
Eleven markets, each with a voice chosen with the client's regional leads, a calling window that matches local norms, and consent language reviewed per jurisdiction.
Day 36 - 45
Load, compliance and rollout
Peak-volume testing against the client's busiest historical hour, a compliance pass per region, then a flagged release to design partners before general availability.
The sequence didn't get smarter.
It got a voice.
Outreach tools have spent a decade getting better at deciding who to contact. The harder problem was always the contact itself, and that is the part we built.