Read-only by design
Writes are refused at the database, not left to the model.
Give Refix a goal for any KPI. Without being asked, it digs into your data until it finds what is holding that KPI back and what to change.
Free and takes about a minute.
Used by growth teams at
You connect your data and approve the change, while Refix does the rest. You approve the change. Refix does the rest.
33.8% of free trials become paid plans, against a 37% target.
Trial-to-paid
Reaching 37% is worth about $4.4M a year.
The AI coach is a false lead.
Matched on days logged
Members who break a streak pay 18%, less than members who never had one.
The 0 feels like failing.
Trials with the weekly count pay 36.2%, against 33.8% without it.
About $3.3M more a year
From acquisition to retention, every goal comes with its own playbook. For anything else, describe your goal and Refix writes the playbook.
Refix ties campaign spend to paying customers, so cheap clicks cannot hide a high acquisition cost.
CAC playbook
At a language learning app
Refix pinpoints the exact step where new users miss first value and the change that caused it.
Activation playbook
At a meal planning app
You see the block that stops the most trials from paying, then the next one when that block lifts.
Trial conversion playbook
At a team workspace app
Most releases leave checkout untouched, so only the change that moves it reaches you.
Checkout conversion playbook
At a shopping app
Before tracing what changed, Refix separates failed renewals from decisions to leave.
Churn playbook
At a fitness subscription app
Refix follows every first order until the customer should return, then finds why the second purchase never came.
Repeat purchase playbook
At a skincare app
Refix builds a full understanding of your business as it works on your goals. Every goal shares it. You never explain your business twice.
Trials that start a paid plan, out of all trials started.
The goal to increase trial-to-paid conversion starts here.Days in a row with a logged meal. Opening the app does not count.
Members who break a streak pay less often than members who never had one.
“3 of 4 days this week” beat the day streak. Every trial has it now.
Refunds inside 7 days do not count as paid.
Launched to every trial. Trial conversion did not move.
One row for each trial, in BigQuery.
One row for each paid plan, in Stripe.
Product lead. Every paywall change goes through Sam.
Finance. Her dashboard settles any dispute about a figure.
Shown on day 3 of the trial, with the yearly plan first.
A day with at least one meal logged.
Sent each time a member logs a meal, in Amplitude.
Android reminders failed for two days. Logging dropped, then recovered.
Paid members who cancel in a month.
The goal to reduce subscription churn starts here.Members who had a streak of three days and are now at zero.
Twelve months paid up front. Most revenue comes from it.
New members who log a first meal within a day.
The goal to improve onboarding conversion starts here.The figures finance reports. Every definition has to match it.
Refix builds it from every source you connect, with no setup project and no context files.
A definition counts only when it reproduces the figures your team already trusts.
Every investigation, correction, release and result is written back into it.
A new goal starts with it on day one, and questions in Slack use it too.
You can see, correct, approve and export it, and nothing changes silently.
A wrong number sends your team after the wrong fix. That is why Refix works every goal on Refix Convex.
DataAgentBench asks 54 questions across 12 messy, multi-source datasets. Its hints are notes about each dataset.
With its hints, Refix answers 88.9%, first on the leaderboard, ahead of agents built on Claude Opus 5.5 and GPT-5.5.
Without them, Refix answers more than 65%, at least 6.7 points ahead of the best agent that ran without them.
DataAgentBench asks 54 questions across 12 messy, multi-source datasets. Hints are notes about each dataset.
Refix is first on the leaderboard with hints, and 6.7 points ahead of the best agent without them.
Read the full resultsLiveSQLBench tests 180 PostgreSQL read questions across 18 business databases.
After correcting broken reference answers, Refix answers 134, or 74.4%. The best model on the published board answers 47.8%.
Read the full resultsQuestions answered correctly
Refix brings an analyst’s judgement to every KPI, every day.
| Analytics and AI tools | Manual analysis SQL, spreadsheets and dashboards Manual | Refix | |
|---|---|---|---|
| Investigates every KPI, every day | No | No | Yes |
| Starts without a question or a ticket | No | No | Yes |
| Finds the cause nobody asked about | No | Yes | Yes |
| Learns your business on its own | No | Yes | Yes |
| Follows each finding through to the result | No | No | Yes |
| No hiring and no setup project | No | No | Yes |
Choose what Refix can access and set the approvals it needs. Review the data and reasoning behind its recommendations.
Writes are refused at the database, not left to the model.
Require your team’s approval before Refix takes action in connected tools.
Decide who connects data, approves actions, and manages billing.
Daily spend in dollars, with an alert before you reach your limit.
Model providers keep nothing and never train on your data.
Set as many goals as you like. Credits pay for the work Refix does on them.
Includes a forward-deployed PM
$150 /mo
$125 /mo, $1,500 billed yearly
Start with 14 days free and 70,000 credits.
Custom
For companies that need self-hosting, security reviews and yearly terms.
Choose the KPI, its target and any constraints. Refix brings the playbook, writes the brief and asks only for the data it needs.
One read-only source is enough to start: product analytics, payments or your warehouse. Refix asks for more only when a finding needs it.
You connect one source, read-only. Refix opens the goal’s playbook and checks its KPI. It posts in Slack when it has a cause, the evidence and the revenue at stake.
Refix Convex reads the values, not just the column names, and asks your team when a field is unclear. Messy, multi-source data is what DataAgentBench tests, and Refix Convex ranks first on it.
A dashboard waits for you to read it. Refix keeps watching the goal’s KPI, works out why it moved across your tools, sizes the revenue at stake and suggests what to do.
Refix starts from everything it already knows about your business, reads the evidence for the goal’s KPI, and comes back when it finds a change worth your attention.
In Refix, and in Slack if you connect it. Each finding names the goal, shows the evidence and the money at stake, and suggests a next step.
Database and analytics access is read-only. Actions in connected tools follow the approvals you set. Nothing changes without your team’s OK.
Yes. Every goal uses the same ontology, memory and playbooks. Correct a business rule once and every goal uses the correction.
Overage bills automatically, up to a monthly cap your workspace owner sets. Near the cap, Refix checks less often. At the cap, the work pauses until next month, unless the owner raises it. Anyone can still set goals.
Yes. Enterprise runs in your Kubernetes cluster, on your models and access controls, with no Refix usage bill.