TL;DR Today, Ampersand launches Subscribe Actions, bringing sub‑second Salesforce & HubSpot webhooks (and more) straight to your AI agents. No triggers, no polling, just clean JSON streams you can wire up in minutes. Read the docs [here](https://docs.withampersand.com/). It's hard to overstate how game-changing this is. So instead, you should hear it from one of our favorite customers, 11x. # Customer spotlight: 11x.ai > “11x is thrilled to partner with Ampersand to deploy Subscribe Actions. The real-time webhooks allow our phone agent, Julian, to focus solely on the exact Salesforce and HubSpot updates that our customers’ sales teams care about. It’s far more scalable than constant polling and expensive filtering. Julian’s speed-to-lead time dropped from 1 minute to just 5 seconds after a form submission, letting Julian call prospects before they even leave the website!” > > Muizz Matemilola, Engineering @ 11x Ampersand builds **for developers first**. Every design choice, from our declarative SDKs, APIs, embeddable components for auth and field mapping, sync observability, strips away busywork so engineers can ship product value, not plumbing. When milliseconds matter, dev-centric tooling is the only way to unlock an agent’s true speed. But this is where technical limitations start to show up. **Salesforce does not support native webhooks**. HubSpot and other CRMs do, but they need a good amount of engineering work to productionize. That pushed us to rethink how time-sensitive and critical events originating in customer CRM tenants reach AI agents. Our shared obsession with “speed-to-value” alongside partners like 11x kickstarted a movement to put real-time CRM events in front of AI agents as fast as physics allows. Let’s unpack the journey. # AI agents and real-time When people say they need data “in real time,” they usually mean so fast that a human, or an algorithm acting on their behalf, can respond before the window of opportunity closes. AI agents collapse that window even further: once an event reaches the model, it can reason and choose an action in tens of milliseconds. Anything slower is human latency imposed on machine potential. Early SaaS platforms (especially CRMs) were understandably tuned for consistent, end‑of‑day reporting rather than sub-second reactions. Their underlying architectures prioritized transactional integrity and overnight bulk‑ETL jobs that kept data safe across thousands of tenants. Real‑time capabilities like triggers, streaming APIs, Change Data Capture (CDC), etc., have been layered in over the years, but they still run inside the guardrails needed for a shared cloud environment. This architectural difference between DIY SaaS systems tuned for batch accuracy versus autonomous AI agents built for sub-second reactions creates a revenue and UX tax for the MCP-native future that seems inevitable with AI. # Where DIY pipelines crumble Getting around this limitation used to mean cutting UX and DX corners. 1. Each hop in the customer-facing pipeline should require its own credential, and keeping all those tokens rotated and secure becomes a never‑ending chore. 1. Sudden traffic spikes easily blow past Salesforce API rate limits, throwing sync errors and negatively impacting user experience. 1. Trying to use Salesforce Change Data Capture or plug gaps with APEX triggers quickly turns into a maintenance nightmare. 1. When queues clog and retries spin out of control, performance is severely degraded, and the on-call engineer is paged. 1. Customers add or rename custom fields frequently, so mappings break unless engineers constantly update code. 1. Debugging a single record’s journey across five microservices at 3 AM is slow, painful, and often inconclusive. Building this infrastructure in-house drains precious engineering time and leaves your AI agents stuck in a bit of a slow motion. # Knock‑on effects for AI agents AI agents can evaluate thousands of data points and propose next steps in **milliseconds,** orders of magnitude faster than a human carrying out the same task. But when those agents sit idle, waiting minutes (or hours) for the CRM to surface a lead change or ticket update, their super‑human speed gets throttled to human pace. - Contact rates drop eightfold when you slide from minute 5 to minute 6. A 45‑minute CRM lag hamstrings an enrichment agent that could score in 50 ms. - A Customer Success Agent labeling an account “healthy” on yesterday’s data recommends an upsell instead of a rescue. - RevOps Agents reconcile stages in microseconds, but overnight batches rewrite amounts afterward, showing the leadership team two conflicting numbers. - Voice Agents and Conversational AI stalls when entitlement or ticket data lags, forcing awkward silences or outdated answers. - Vertical‑AI blind spots – Custom‑field updates in batch windows yield advice out of sync with today’s reality, risking compliance breaches. In short, every extra second of CRM latency widens the gap between what a swarm of AI agents could orchestrate and what the business lets them deliver. If your multi‑agent orchestration layer thinks at machine speed, why chain it (and the entire workflow) to batch‑era plumbing? # Design criteria for a proper solution We reimagined what the scaled-out architectural solution would look like when your AI Agent interoperates with multiple Systems of Record for thousands of enterprise customers. 1. Every customer org needs isolated tokens and adaptive rate limits so one tenant’s spike never starves another. 1. Events must arrive as plain JSON containing state, delta, and timestamp (removing things like custom SOAP wrappers and unnecessary internal metadata). 1. A translation layer should map provider‑specific objects into a common vocabulary, keeping most application logic portable across CRMs. 1. The stream has to guarantee re‑delivery and support historical backfill without CSV uploads. 1. OAuth consent, field selection, and scope mapping should be self‑service and take minutes, not a 3-month professional‑services engagement. With these five boxes checked, engineers and product builders can chase great AI agent outcomes: speed‑to‑lead, churn defense, case sherpa, autonomous quoting, etc...instead of months and years of integration plumbing. # How Ampersand Subscribe Actions fixes this We solve this by building a declarative SDK and deep connector layer that taps directly into Salesforce CDC, HubSpot webhooks, and other Systems of Record, capturing the right events without custom code; each tenant’s data runs through its isolated pipeline with smart adaptive rate-limits, so you are never blowing past any of your customers' API quotas; events arrive as plain JSON and we retain full replay/backfill history so streams stay intact after outages or during onboarding, all without cron jobs; lastly, embeddable React components let end users finish OAuth, pick objects and fields, map custom objects when needed, and grant the right permissions and scopes (headless mode coming soon!). ### Example manifest for phone agent Julian's Salesforce integration ```yaml specVersion: 1.0.0 integrations: - name: realtimeLeadsToJulian provider: salesforce subscribe: objects: - objectName: lead destination: julianWebhook inheritFieldsAndMapping: true createEvent: enabled: always updateEvent: enabled: always watchFields: - status - phone - email deleteEvent: enabled: never read: objects: - objectName: lead destination: julianWebhook requiredFields: - fieldName: id - fieldName: firstname - fieldName: lastname - fieldName: phone - fieldName: email mapToName: emailAddress - mapToName: status prompt: Which field do you use to track lead status? backfill: defaultPeriod: days: 1 # backfill last day of lead activity on install ``` This declarative manifest tells Ampersand to stream only the lead events Julian cares about, land them on his webhook in under a second. ### Field mapping Your **Read** action’s field list, and any `mapToName` aliases automatically carry over to the **Subscribe** stream. Every webhook includes two objects: - `fields` – raw provider names - `mappedFields` – your canonical names (from the read action) That means agents like Julian get clean JSON they can act on immediately, without another lookup layer. ``` { "action": "subscribe", "objectName": "lead", "provider": "salesforce", "result": [ { "fields": { "id": "1234567890" "firstname": "John", "lastname": "Doe", "phone": "4151111111" }, "mappedFields": { "emailAddress": "john.doe@mail.com", "status": "Active" }, "subscribeEventType": "create", ... } ] } ``` # Real‑time sync with customer data, unlocked Subscribe Actions aren’t just an improvement. They’re a rethink of what SaaS should do in an AI-first world. Forget cobbled-together trigger queues that buckle under scale. With Ampersand, you get clean, fast, trustworthy events delivered wherever your application or AI agent needs them. We will be working on Subscribe Actions for heavy-hitters like Microsoft Dynamics 365, other SMB and vertical CRMs (like Zoho, Attio, Pipedrive), Zendesk, Intercom, ServiceNow, Marketo, Gong, NetSuite, QBO, Stripe, Snowflake, DocuSign, and a fast-growing queue of vertical-AI favorites. The principle is the same: **agents can’t be intelligent if their data is stale, low-fidelity, or non-permission-aware.** Ampersand offers architecturally coherent, infinitely scalable LEGO blocks for bidirectional sync with every customer's System of Record: - **Read Actions** for initial backfills, incremental reads, and on-demand updates - **Subscribe Actions** for listening to changes in customer systems - **Write Actions** to push enriched insights back into your customers' Systems of Record Combine the three, and you have a fully closed-loop, real-time integration that keeps AI agents, customer-facing apps, and downstream analytics perfectly in sync. At Enterprise scale. Ready to give your agents millisecond-level reflexes? Give Ampersand a try! [Start building for free](https://docs.withampersand.com/) or reach out directly on my [calendar](https://calendly.com/ayan/ampersand-introduction) to learn more.