12× fewer tokensTokens read to answer the question
A standard MCP stack161k
The Doris Ontology14k
The question we asked.
“Across our open deals, which three has the buyer gone quietest on? For each, tell me how many days since the buyer last made contact, what is still outstanding on the deal and who owes it, and whether the close date is at risk.”
A CRM records the last activity on a deal, not the last time the buyer made contact. And it cannot say who owes what, because that was said out loud on a call.
What each side was given.
Both sides pay for every tool they expose, on every round.
A
A standard MCP stack
The three a team already has: a CRM, a notetaker and an email client.
CRM18 tools
Notetaker10 tools
Email and calendar18 tools
B
The Doris Ontology
One server. The same read tools anyone can connect to the public sandbox today.
Doris Ontology8 tools
The measured run.
MCP stackDoris Ontology
Tools the model is given468
Tool calls to answer311
Tokens read161,14513,512
Cost per question$0.565$0.081
A month, at 100 questions a day$1,695$244
Prices are Anthropic's published Claude Opus 5 rates on 15 September 2026. The method, the tool definitions counted and both call traces are recorded with the run.
Where the difference comes from.
01
More servers is not the problem
Connect two more and the bill barely moves. The tools are not what costs you, which is why trimming your stack will not fix this.
02
It is what they hand back
CRM rows, call lists, mail headers, transcripts. All of it dragged into the model so it can find the one fact inside. The ontology hands back the fact.
03
Every round pays for the last
Whatever the model read in round one is sent again in round two, and three, and four. Running calls in parallel helps. It cannot go below the rounds your data forces.
The answer changes too.
A stack rebuilds the meaning of a deal out of raw speech, every time you ask. The ontology reads meaning that was captured once, typed, and linked to what surrounds it.
01
A promise is not a state
A transcript tells you somebody said they would do something. It cannot tell you whether it happened, when it was due, or which side owed it. On the record those are fields, kept current.
02
It says when it cannot answer
Ask for comparable deals where there are none and you get told there are none, and why. A model reading call recordings cannot tell the difference between nothing to find and not having looked.
03
Some questions do not fit
Which past deals hit this same objection, and what did the rep who won them say? One call. The other way is reading every closed deal you have, and there is more of that than a context window holds.
Ask your own question.
Bring a pipeline question. We will answer it through the ontology, live.
