Intelligence you can actually measure
AI decisioning for the stack you already have.
Herald picks the right message, channel and time for each user, then learns from every send. Live today with Braze, with Hightouch for your data — more connectors on the roadmap. Pay for the lift you can measure.
Built and proven in production with a top consumer app.
Works with your stack today
Connect once. Herald tests continuously and learns from every outcome.
Hightouch
LiveData in: sync only the user fields you choose.
Herald
Picks the message, channel and time for each user.
Braze
LiveSend out: messages go through your own Braze workspace.
On the roadmap
Design partners help decide the order.
Connect from
- Segment
- Snowflake
- BigQuery
- Webhook / CSV
Send through
- Iterable
- Customer.io
- OneSignal
- Twilio
- SendGrid
Content from
- Contentful
- Sanity
- Generic JSON
- Braze Catalog
- Herald library (live)
Watch it learn
SimulationPick your stack and a use case, then press Run. This is the real Herald engine, running in your browser on simulated users.
Users inactive 14+ days. Four push messages compete to bring them back. Keeps a belief about each message and tries the ones that might be best.
Thompson and epsilon-greedy here learn one overall winner; LinUCB learns per user. (In production Herald can also segment Thompson by persona.)
Traffic share by message
% of sends- A · Reminder
- B · Social proof
- C · Discount
- D · Personal progress
Lift vs. control (simulated)
%Regret: conversions lost vs. always picking the best
conversionsWhy this decision?
Press Run to start. Each decision shows up here with the estimate behind it.
Show the hidden truth (simulated conversion rates)
| Message | Everyone |
|---|---|
| A · Reminder: “We saved your spot. Pick up where you left off.” | 2.1% |
| B · Social proof: “12,000 people finished this week's challenge.” | 3.4% |
| C · Discount: “Come back today: 20% off your next month.” | 2.8% |
| D · Personal progress: “You're 3 days from your best streak ever.” | 4.6% |
Simulated data. The algorithms are Herald's production engine, running in your browser.
How it works
- 01
Sync your data
Send user data in through Hightouch. Warehouse and CDP connectors are on the roadmap.
- 02
Set a goal and variants
Pick the conversion you care about and write a few messages. Herald handles the rest.
- 03
Herald decides, sends, learns
Each user gets the message most likely to work, sent through Braze today. Every outcome makes the next decision better.
The algorithms behind it
Thompson sampling
Balances trying new messages with sending proven ones.
Epsilon-greedy
Simple and predictable: mostly the leader, sometimes a test.
LinUCB
Contextual. Learns what works for each kind of user.
Every decision is logged with the variant chosen and the user context it was chosen on, and is inspectable in the Herald console.
Why not your ESP's built-in AI?
| Compare | Herald | ESP add-ons |
|---|---|---|
| Works with | Braze and Hightouch today; more on the roadmap | Usually their own ESP only |
| Pricing | Per agent, or on measured lift | Often per use case, on enterprise contracts |
| Transparency | Every decision logged and inspectable | Often limited |
| Proof | Lift vs. a random-split control, reported during the pilot | Varies |
A general comparison with ESP-native decisioning add-ons; individual products differ.
90-day pilot · Free · 5 spots
Become a design partner
One agent, running on your own Braze and Hightouch, measured against a random-split control.
- Day 0Connect Hightouch and Braze. Pick one goal and write a few variants.
- Day 30First holdout report: lift vs. the random-split control.
- Day 60Second holdout report.
- Day 90Final holdout report. Keep going or walk away.
What we ask: a case study and honest feedback.
After the pilot: from $1.5k/mo per agent, or pay on measured lift.
FAQ
What data does Herald copy?
Only the fields you choose to sync. You control the columns in your Hightouch sync.
Can we self-host?
Not yet. Herald is hosted for you today. Self-hosting is on the roadmap, and design partners can discuss a private deployment.
Which ESPs work today?
Braze is live, with Hightouch for data. The rest of the list on this page is on the roadmap, and design partners decide the order.
How is lift measured?
Against a random-split control: a share of users get a randomly chosen message instead of Herald's pick. We report the difference at day 30, 60 and 90 during the pilot.