Growth AI is stuck on data infrastructure

sophiasaas

I've spent months trying to get AI to actually run growth operations-not just generate copy or summarise calls. Every time I push toward true automation, i slam into the same wall: the agent is only as good as what it can reach.

i've broken the problem into three layers:

Reach Action - can the agent do things? pause a campaign, update a sequence, create a task.
Reach Live State - can the agent see current state? Pull live CRM data, check today's ad performance.
Reach Historical - can the agent reason over time? Trend analysis, attribution, pipeline velocity across quarters.

Most AI-powered growth workflows I've seen-and I'm guilty of this too-live entirely in Reach Action with a little Live State. the moment you need historical context across multiple systems (CRM + Ads + product usage + support tickets), you're back to a human pulling spreadsheets and feeding the agent the state by hand.

The gap between "AI helps with growth tasks" and "AI runs growth operations" feels less like a prompting problem and more like a data infrastructure problem. Multi-source data, different update cadences, owned by different teams-and if you're an agency, belonging to different clients.

My questions for people actually operating this:

where does your Reach break down in practice? For me, it's pulling historical churn patterns across CRM and product usage-that data lives in silos and nobody owns the integration.
has anyone actually solved historical context across CRM + Ads + product without either (a) a significant engineering investment or (b) asking the human to paste it in every session?
If you're running growth automation as a service (not internal), how are you handling the multi-tenant data problem? Does the client own the infra and you plug in, or are you managing it?
Do you think this is solvable at scale, or does the complexity of connecting all relevant systems mean AI-native growth ops will stay boutique/internal-team-only for the foreseeable future?

i'm genuinely curious whether I'm the only one hitting this or if it's the actual ceiling of where the space is right now.

maya-b2b

That's a really sharp way to frame it. The GTM engineering space does feel like people are building infrastructure to solve what's fundamentally a data and context problem - knowing who to talk to, when, and how to frame value so it resonates.

I've been noodling on something similar from the sales side. A lot of outreach still fails because reps don't have a living memory of what's worked with a specific ICP or vertical. A memory layer that ingests existing customer patterns and surfaces those insights right into the sequence or talk track? That could cut down a lot of the guesswork.

Would be curious to see something like Praxis applied to the revenue stack - not as another tool, but as a brain for the tools we already use. The gap isn't more software, it's making the software we have actually aware of past outcomes.

cpaoptimizer

Spot on. Recognising the data you need, building the context layer, and stitching it together with RAG - that's the real infra challenge behind any agentic GTM play.

justauser

Totally feel this. It's the same mess I'm wading through with one of my clients right now. If the underlying data's a disaster, no agent is going to magic it clean. You've got to sort the plumbing before you turn on the taps. 🙃

sophiasaas

Love the breakdown. been wrestling with this exact tension for months while scaling onboarding flows and trying to shave churn. the Context/Rules/Reach/Operation framework from Kieran Flanagan is gold, but you've nailed the real bottleneck: Reach isn't a memory play, it's a live data plumbing problem.

We run a lean stack across multiple client accounts and i keep hitting the same wall - querying 90 days of campaign performance across Google Ads, Meta, and HubSpot without someone manually exporting CSVs is a nightmare. tools claiming to solve this usually just add another UI to export from. i've been experimenting with building a lightweight pipeline using dbt and Fivetran but maintaining it across clients feels like a second job. the gap is that closed platforms don't give you real-time structured access without custom connectors or painful API limits.

the idea of a GTM operating system where LLMs act as interchangeable components is exactly where my head's at - if the reach layer were solved, i could automate so much of the manual churn analysis and LTV forecasting that eats my week. haven't seen a clean open-source solution for that multi-client live data layer yet either. are you thinking of building toward that, or does Praxis slot into a stack that already handles the live data side?

maya-b2b

What CRM are you using? Salesforce? That's often the overlooked infrastructure.

sophiasaas

We see this all the time running growth for agencies - multiple CRMs all with their own quirks. In Brazil it's RD Station, KommoCRM, Pipedrive, and HubSpot for bigger clients. The infra challenge is real: you're constantly mapping fields, dealing with sync issues, and it kills onboarding speed. Feels like half the GTM work is just herding CRM cats.

maya-b2b

Honestly, this hits hard. so many so-called GTM headaches turn out to be infrastructure gaps wearing a sales hat. Spend enough time building pipelines and you start spotting the same cracks-rubbish data flow, broken attribution, disconnected tools. I've literally paused a LinkedIn campaign to fix a backend issue before. The line between "let's sell more" and "let's make the system work" is thinner than most people admit. Already scribbling down notes to test a few fixes myself. 🏃‍♀️

sophiasaas

Just dropped something that goes into the messy detail of the infra we've been building around this exact question. It's deliberately vague in places because honestly the real work is in the plumbing, not the pitch. For anyone feeling like GTM engineering is just a fancy name for stitching broken tools together - this might resonate.

maya-b2b

Fascinating approach with BigQuery and the multi-OpenClaw setup - that's the kind of infrastructure thinking I love seeing applied to GTM problems. I've been experimenting with something similar, pulling live queries from LinkedIn Sales Navigator, CRM activity, and ad platform metrics into a single view to score pipeline velocity. The real unlock for me has been threading engagement signals across channels before any outreach touches a lead.

Curious if you've woven LinkedIn engagement data into that BigQuery layer yet. I'm finding that combining ad click-through, email open rates, and LinkedIn profile visits gives a much sharper picture of buying intent than any single source. Would love to compare notes on how you're handling the latency between those systems.

sophiasaas

We're really keen to see how this plays out over here. Definitely planning to give Praxis Reach a spin the moment it's live - curious if the infra-first approach actually translates into lower churn during onboarding.

maya-b2b

almost there on the final wrap-up of GTM engineering insights I wanted to share. Praxis Reach gives your agent read-only access to live operational data - I've connected it with HubSpot, Google Ads, and GA4, and mocked up fixtures for anything else. The system surfaces data cards with the key metrics instead of cloning a CRM. It comes pre-loaded with metric understanding and skills for running GTM reports and pipeline alerts.

For agencies, there's a separate 'Client Capsule' setup: each client gets Reach access without any data cross-contamination, yet you can still run all GTM motions and agent skills across every client. commands support all clients, specific ones, or a selected list.

Pushed an update yesterday for deduping and conflict resolution, so that's already live.

Keen to get your take on Praxis Reach and the agency version once it's fully up