Customers buy the outcome, then expand the layer.
What our lead enterprise account saw, and how a pilot grows system by system.
by month two, from acting on Optialy's recommendations, as reported by the customer. +20% in month one.
What customers saw in the first months.
From our lead enterprise account's paid pilot, now in beta, and our beta customers.
Campaign efficiency lift
Acting on Optialy's recommendations, as reported by the customer
Analyst hours saved per question
Simple, multi-system and complex questions, across beta customers
Systems connected, one by one
From 1 to 4 system groups, each added at the customer's request
Two more enterprise accounts, in Quebec and British Columbia, just started. Both took part in our accuracy testing.
Across four industries, shaping what we build next.
Across ads, analytics, commerce, CRM, finance and files.
“Which closed-won clients have RFPs due this month?”
And what does their performance across our data say about their trajectory? It needed four systems. Salesforce's own AI couldn't find the RFP dates, because they live in a custom field.
- Optialy found all four sources
- Read the custom field directly
- Returned one combined, cited answer in two to three minutes
Generic assistants reach only what each vendor's integration exposes. Optialy engineers retrieval per platform.
One question. Then system by system.
Our lead account, sponsored by its VP of Strategy.
One high-value question, with an executive sponsor.
Advertising and marketing systems.
Analytics, e-commerce and CRM added on request.
Four system groups, one governed layer. Annual contract under discussion.
Computed beats generated: 97.5% vs 65.8% of figures correct.
4,500 real marketing and commerce questions on live connected data, each graded blind by three AI judges against figures computed from the source systems, then checked manually by our team.
Figures correct. Each square is 0.1% of graded figures.
Answer quality, out of 5
Answers with a chart
Why a RAG baseline
Plain ChatGPT or Claude can't connect to most of these systems, so many of the questions are impossible for them. The fair comparison is a pipeline that can.
Internal benchmark, September 2026. RAG pipeline: retrieval, then the language model reads the data and writes the figures itself (our earlier architecture).
One question. Every relevant source. Trusted wherever work happens.
Start with one question that costs you money today. We'll answer it on your live data.