01 / Knowledge
Answers from the
firm’s own knowledge.
Proposals, project files and live business data. Tested with questions about the firm.
100%
Every question passed.
10 questions · 3 runs each
Benchmark
Cuesta 
benchmarked Agnes, then took it firm-wide
Before bringing Agnes to its clients as Strata, Cuesta Partners tested it inside its own firm against the AI tools its people already used. Then it rolled it out to everyone in 17 days.
Cuesta’s benchmark
Share of graded answers that passed
Same 10 questions, 3 runs each,
graded by Cuesta.
- SharePoint documents indexed
- Unstructured270,000+
- business tables, refreshed daily
- Structured28
- defined once in the semantic layer
- Semantic26 metrics · 36 terms
- from kickoff to firm-wide rollout
- Time to value17 days
The story
Beyond AI experimentation, to measurable impact.
That is the promise Cuesta Partners makes to the mid-market companies and private equity sponsors it serves. As a partner bringing Agnes to clients, it decided to hold itself to the same standard first: put a company brain on its own documents, numbers and people, measure it first, and only then take it to clients.
“Instead of looking at slides, we’re looking at an inventory of agents and Agnes. It’s going to totally change the game for us.”
Riley ScottCo-Founder and Partner, Cuesta PartnersThe benchmark scope
Three things a company brain had to prove.
Answers from the
02 / Consistency
The same answer.
Every time.
Each question asked three times. Because a client needs an answer they can rely on.
10/10
Correct answers.
Every single run.
Agnes10/10
Best alternative6/9
03 / Permissions
Access that follows
the role.
Agents use the permissions of the person who starts them. Checked on every request.
0 leaks
Restricted data
stayed protected.
Firm numbers
PartnerAllowed
AssociateRestricted
What Agnes sits on
One brain over the firm’s documents, data and definitions.
Agnes connects what Cuesta knows in three forms: what it wrote, what it measured, and how it defines things.
Inside Agnes
Semantic layer
Your agreed definitions: what each metric means, which terms you use and how tables join. Agnes reads them where you already keep them, in Keboola, your warehouse or Git, or hosts them until they have a home. Before an agent writes a query, it reads the definition, and the answer can say which one it used.
Knowledge graph
A map of what your documents say about clients, projects, people and topics, built by reading each document once. Every relationship carries the exact quote and the file it came from, so an answer can name its sources. When documents disagree, the newer one wins and the older one stays visible as history.
Memory
What your teams learn while working with Agnes becomes shared knowledge. Lessons from sessions and uploaded documents are proposed as memory, and an Admin approves each item before Agnes relies on it. Anyone can keep a session out of memory.
Skills
Written know-how that teaches Agnes how your team works, such as how to build a proposal or which definitions to read before querying. Anyone can write a skill. It becomes visible only after a security and quality review, reaches only the groups it is granted to, and always runs under the access of the person using it.
Agents
Named, focused agents built on Agnes, each with its own role, the data and tools it may use, a budget and its own memory. Ask one a question, put it on a schedule or hand it a task, from chat, the command line or your AI tool. An agent can never reach beyond its owner’s access.
Access & audit
Access follows the person, not the tool: everyone sees only what their groups have been granted, whether they work in Agnes, Slack or Claude. Every change and every sensitive action is recorded with who, what and when, so an Admin can see who did what and check what any person can reach.
Your knowledge
The firm’s knowledge
270k+ documentsTwo years of client work
Roadmap readouts
Statements of work
Diligence reports
Value backlogs
28 business tablesRefreshed daily by Keboola
- time_entry
- opportunity
- account
- engagement
- billing
Agnes
Company context
One connected context
26 metrics · 36 termsA shared semantic layer
- Revenue = billings
- Margin includes contractors
- Utilization · Win rate
10 entity types12 relationships
- worked on
- for client
- in industry
- part of
- evidenced by
Benchmark results
Same questions, five setups.
Cuesta fixed the questions and the scoring rubric before any system saw its data. Every setup received the questions word for word. The four setups with data ran them three times each and every answer was graded; plain Claude, with no data, is the baseline.
- 10questions, each testing one capability
- 5setups: 4 graded, plus a no-data baseline
- 3runs per graded setup
- 7weighted scoring dimensions
10
questions, each testing one capability
5
setups: 4 graded, plus a no-data baseline
3
runs per graded setup
7
weighted scoring dimensions
- Claude, no data (baseline)
- Claude + SharePoint
- ChatGPT + SharePoint
- Claude + SharePoint + context pack
- Agnes
Example questions
Finding the closest past project
“A prospect is a building-products manufacturer, PE-owned, and wants to know if we've done this kind of rapid AI roadmap work before. What's our closest precedent and why is it close?”
Hard because: The right precedent is filed under a different service name.
Counting across many folders
“Which PE sponsor has driven the most of our AI assessment engagements this year, and across all of those engagements, what industries have we touched?”
Hard because: The answer spans dozens of folders and a CRM field that is often blank.
Spotting a question with two meanings
“Show me our manufacturing work.”
Hard because: It can fairly mean two different things. A good answer says so.
Other questions asked for live business metrics, where an answer came from, how two companies are related, and one that has no right answer.
| 30graded answers for Agnes10 questions × 3 runsPassPartialFailNo accessOne dot per run | Connected company data and knowledgeAgnes30 / 3010 of 10 passed all three runs | Document connectorChatGPT21 / 276 of 9 passed all three runs | Document connectorClaude19 / 274 of 9 passed all three runs | Documents and curated definitionsClaude + context16 / 275 of 9 passed all three runs |
|---|---|---|---|---|
| Structured metrics (live data) | Agnes: pass, pass, pass | ChatGPT + SharePoint: no access, no access, no access | Claude + SharePoint: no access, no access, no access | Claude + SharePoint + Cuesta's context pack: no access, no access, no access |
| Precedent recall | Agnes: pass, pass, pass | ChatGPT + SharePoint: pass, partial, pass | Claude + SharePoint: pass, partial, partial | Claude + SharePoint + Cuesta's context pack: partial, partial, partial |
| Closest match | Agnes: pass, pass, pass | ChatGPT + SharePoint: pass, pass, pass | Claude + SharePoint: pass, pass, partial | Claude + SharePoint + Cuesta's context pack: fail, fail, pass |
| Ambiguity | Agnes: pass, pass, pass | ChatGPT + SharePoint: partial, partial, partial | Claude + SharePoint: fail, pass, partial | Claude + SharePoint + Cuesta's context pack: partial, partial, partial |
| Counting | Agnes: pass, pass, pass | ChatGPT + SharePoint: pass, pass, pass | Claude + SharePoint: pass, fail, pass | Claude + SharePoint + Cuesta's context pack: partial, partial, partial |
| Entity resolution | Agnes: pass, pass, pass | ChatGPT + SharePoint: fail, fail, pass | Claude + SharePoint: fail, fail, pass | Claude + SharePoint + Cuesta's context pack: pass, pass, pass |
| Every setup passed every run | ||||
| Synthesis | Agnes: pass, pass, pass | ChatGPT + SharePoint: pass, pass, pass | Claude + SharePoint: pass, pass, pass | Claude + SharePoint + Cuesta's context pack: pass, pass, pass |
| Specific facts | Agnes: pass, pass, pass | ChatGPT + SharePoint: pass, pass, pass | Claude + SharePoint: pass, pass, pass | Claude + SharePoint + Cuesta's context pack: pass, pass, pass |
| Provenance | Agnes: pass, pass, pass | ChatGPT + SharePoint: pass, pass, pass | Claude + SharePoint: pass, pass, pass | Claude + SharePoint + Cuesta's context pack: pass, pass, pass |
| Negative case | Agnes: pass, pass, pass | ChatGPT + SharePoint: pass, pass, pass | Claude + SharePoint: pass, pass, pass | Claude + SharePoint + Cuesta's context pack: pass, pass, pass |
“We were able to get everything for the last two years in our customer folders. And we measure that in days, not months.”
Zach PapinAI Transformation Lead, Cuesta PartnersReal impact
What changed in the work.
Proposals
Proposals start
from precedent.
A brief and discovery notes become a draft deck and SOW, each claim tied to a past engagement.
~20minutes
A full cross-sell list
for a client segment.
Two Cuesta leaders asked for every opportunity across a segment.
“We never would’ve been able to do that just with Claude and SharePoint alone.”Zach PapinAI Transformation Lead, Cuesta Partners
Definitions
One number for
revenue and margin.
Two definitions the firm had never pinned down now live in the semantic layer.
18
18 reusable skills.
Shared across the firm instead of rebuilt by each person.