Done by hand, that’s days of exports, requests and meetings. Most teams stop after the first question and decide from there.
Sage — product intelligence with integrity
Ask the hard product question. Get an answer you can stand behind.
Sage reads what your team already knows, asks before it reaches into anything new, and answers with the evidence attached. Watch it work on three real questions.
Grounded in your evidence · Auditable ·
Problem one: too much for people to connect
Every planning cycle starts with the same question: what gets us to the number?
Answering it properly is a chain. Each answer changes the next question, and each one lives in a different tool, owned by a different team.
01Salesforce · sales ops
Is churn or new business the bigger lever?
Export churned ARR and pipeline, then build the comparison in a spreadsheet.
02Amplitude · data team
Where do accounts drop off?
Ask an analyst for a funnel of churned accounts, then wait for it.
03Zendesk and Intercom · support
Why do they leave at that step?
Read through hundreds of tickets looking for the pattern.
04Salesforce again · sales ops
Were the people complaining from our biggest accounts?
Match ticket authors to accounts and ARR by hand.
05Roadmap · product
Is a fix already planned, and when?
Find the items, then argue about moving them in a planning meeting.
Problem two: AI fills the gaps
So teams hand the question to AI, and it answers with confidence.
A general AI reads fast enough to do the chain. It also produces a polished plan whether or not your data supports one. Ask it the same question and watch what happens.
- It answers from what you pasted in. It never saw the churn export you didn’t think to include.
- It fills gaps with numbers that sound right.A revenue estimate arrives with no source behind it.
- It mixes up what you’re measuring.New-logo deals get counted toward a retention target.
- It never says “I don’t know”.Every question gets a confident plan, including the ones your data can’t answer.
- It agrees with you.Ask it to cut a feature and it writes the rollout plan, even when your biggest accounts depend on it.
Look at it again
You need the reach of AI and the discipline of a good analyst.
The first problem needs something that can read everything your team knows. The second needs something that knows what it read, and what it didn’t.
Most tools solve one at the cost of the other. People are careful but slow, so they stop early. General AI is fast but unaccountable, so its plans can’t be trusted in front of a board.
The useful assistant does both: it follows the whole chain, and it shows what each step rests on.
This anonymized workspace is messy on purpose: missing tags, unmatched accounts, a dirty CRM. Like yours.
Interactive demo · pick a question
How can we increase our revenue by 20% over the next 6 months?
Reading the goal
+20% revenue in six months, so by the end of Q2. Revenue moves through new business, expansion and churn
Planning what to read
Salesforce for churn and pipeline, Amplitude for usage, feedback for reasons
Asking for access to Salesforce and Amplitude
Salesforce: allowed once · Amplitude: allowed once
Reading Salesforce: 3 queries in parallel
Churned ARR, last 6 months$0.72M across 41 accounts
Expansion, last 6 months$0.60M across 26 accounts
Open pipeline and sales cycle$1.9M open · 21% win rate · 5-month cycle
Comparing the two levers
On current trends: about +5% by the end of Q2. Churn is the line that can move most before then
Finding where churned accounts stop using the product
Amplitude: 80% of new accounts start the auto-send setup; more than half of those stop at step C
Reading feedback about step C
Zendesk and Intercom: 23 items. Reviewing answer rules one by one takes days
Checking who complained
17 of 23 from admins at accounts above $40k ARR. 9 of those accounts have already churned ($0.46M)
Searching the roadmap for fixes
2 features address step C. Both are scheduled for Q4
Checking what fits in six months
Accounts stalled at step C today hold $0.38M renewing in Q2, so the fix has to ship in Q1
Sage wants to read revenue data from Salesforce
Churned ARR and open pipeline, to compare the two ways revenue can grow in six months.
salesforce · query_opportunitiesAllowed once
Sage wants to read product analytics from Amplitude
Journey completion by account, to find where churned accounts stopped using the product.
amplitude · get_funnelAllowed once
Twenty percent in six months means adding $1.2M to your $6.0M ARR by the end of Q2. On current trends you’ll add about $0.28M, and churn is the line you can move most before then:
| Next six months, on current trends | Forecast | Can it move before Q2 ends? |
|---|---|---|
| New business | +$0.40M ($1.9M open × 21% win rate) | Little; the average deal takes 5 months |
| Expansion | +$0.60M at the last six months’ pace | Some, mostly at renewal |
| Churn | −$0.72M at the last six months’ pace | Most; $0.38M at risk renews in Q2 |
| Net | +$0.28M, about 5% |
Most of that churn traces to one place. 80% of new accounts start the auto-send setup. Of those, more than half stop at step C, Review answer rules:
| Setup step | Share of accounts that started setup |
|---|---|
| A · Connect inbox | 100% |
| B · Import help center | 93% |
| C · Review answer rules | 88% |
| D · Set confidence threshold | 41% |
| E · Test on sample tickets | 38% |
| F · Go live | 36% |
The feedback says why. 23 items describe reviewing rules one by one as slow and tedious. 17 came from admins at accounts above $40k ARR, and 9 of those accounts have since churned, worth $0.46M. Exit interviews name the rules review in 6 of the 9. It’s a strong pattern, though it doesn’t prove the screen caused every loss.
That revenue is already gone. What you can still keep: accounts stalled at step C today hold $0.38M of ARR renewing in Q2. Two features on your roadmap fix the step, but both are scheduled for Q4, so they need to ship in Q1:
| Change | From | To | Why |
|---|---|---|---|
| Bulk-approve answer rules | Q4 | Q1 | Removes the one-by-one review |
| Group rules by category | Q4 | Q1 | Cuts step C to the rules that matter |
| Custom report builder | Q1 | Q3 | 2 requests, no link to churn; frees about the same capacity |
Keeping that $0.38M lifts the forecast from about 5% to about 11%. It’s the biggest single move, but 20% by the end of Q2 would need expansion well above its recent pace. I can look at which accounts expand after they finish setup. Figures from Salesforce and Amplitude, last 6 months.
Proposed Candidates
Admins abandon setup at the rules reviewOpportunity candidate
Bulk-approve answer rulesSolution candidate
Group rules by categorySolution candidate
- 1.More than half of accounts stop at step C, Review answer rules(Backed by evidence) — “Amplitude funnel, last 6 months: 88% reach step C, 41% reach step D.”
- 2.Admins find the rules review slow and tedious(Backed by evidence) — “Reviewing answer rules one at a time took our team two weeks.”
- 3.Accounts that stall at step C churn more(Backed by evidence) — “9 of 17 complaining accounts churned; exit interviews name the rules review in 6.”
- 4.Faster rule review will lift setup completion(Evidence inconclusive)
Recommendation
Move bulk approve and category grouping to Q1, push the custom report builder to Q3, and track step C completion and time in step every week.
The fix has to be live before the Q2 renewals. Step C completion moves within weeks and churn shows up later, so watch the early number first.
What it costs
Most ideas don’t work. The expensive part is finding out late.
Companies that measure every change report the same thing: most well-argued ideas fail to move the metric they were built for.
1 out of 3
ideas tested at Microsoft improved the metric they targeted.
Kohavi et al., Online Experimentation at Microsoft
~10%
of experiments at Booking.com produce a positive result.
Thomke, Harvard Business Review, 2020
And it lands on people
- A quarter of engineering spent on a bet nobody checked.
- A PM defending the plan to the board with “customers asked for it.”
- The same decision argued in March, and again in September.
- A feature your biggest accounts rely on, cut because the average user ignored it.
A better way to decide
Answers that show their evidence and know their limits.
Five things an assistant needs to do to solve both problems. The first two give it reach. The last three keep it honest.
1
Read everything relevant
Feedback, interviews, usage, revenue and the plan, together.
2
Back strong bets
When the evidence holds, say so plainly and help the work move.
3
Ask before reaching further
New data sources get your approval first, question by question or permanently.
4
Say when it’s guessing
When your data is silent, offer a general answer and label it as one.
5
Say “don’t” when your data does
When a plan would hurt the product, show why and offer a safer path.
How Sage does it
Built on your evidence, kept current.
Sage turns every artifact into claims and the evidence behind them, then keeps checking as new information arrives.
How does Sage ground an answer?
Each claim is claimed, unsupported, corroborated or contradicted, set strictly from accepted evidence links. The model reasons over those states and never sets them. Every turn is marked grounded, degraded or exploratory.
What does it do when nobody is asking?
A background sweep looks for contradicted assumptions, untested claims, orphaned research, stale commitments and conflicting solutions. It is deterministic and reproducible.
Where does our data go?
Each workspace has its own vector collection and a tenant-aware graph, with scope enforced server-side. Analytics arrive as aggregate insights, never session recordings or user-level data. Your data is never used for model training.
What is it built on?
A router, reasoner, verifier, embedder and reranker, working over a Neo4j product graph and Qdrant retrieval. The graph and rules are deterministic; the models work inside them.
Solveo
The answer was already in their CRM.
Solveo’s sales team was sure larger customers wanted a voice channel. Product wasn’t, because the belief lived in sales calls and the numbers lived in a closed-lost export nobody had read.
Sage linked the claim to the export and the exit interviews. Five of nineteen enterprise deals had been lost on voice, worth $346K. The claim went from a sales hunch to something the team could put in front of the board, and voice moved onto the roadmap for teams above 200 agents.
Larger teams need a voice channel before they expand
Stated by sales. No evidence linked yet.
“No voice channel” lost 5 of 19 enterprise deals
$346K in closed-lost revenue.
Larger teams need a voice channel before they expand
Backed by the CRM export and exit interviews.
Bring your hardest product question.
Explore a live workspace without signing up, or start a trial with your own documents.
✓ Free 14-day trial✓ No credit card required"