Confident, wrong document
12% off list.
Pricing policy, 2019 · page 4
Services
Context and Knowledge Graphs
Turning what your company knows into context your AI systems can use.
How we see the problem
The agent is only as good as what it can see.
An agent that answers confidently from the wrong document is worse than one that says it doesn’t know.
What’s Northwind’s discount on renewal?
Says it doesn’t know
I don’t have a source I can cite.
Nothing in view covers this account.
Where agents fail
Most failures trace to missing, stale, or conflicting context, not model quality.
What we build
The content layer your agents read from: sources, structure, permissions, and freshness rules.
Who it’s for
Teams whose agents already work in a demo and break on real company data.
What we capture
Company brain
The pricing exception lives in two reps’ heads and a Slack thread from 2023.
- CRM
- Call transcripts
- Contracts
- Wikis
- Tickets
One governed layerCompany brain
Capture
Pull knowledge out of CRM, call transcripts, contracts, wikis, tickets, and email into one governed layer.
Tacit knowledge
Interview-driven extraction of the rules nobody wrote down: discount logic, escalation paths, customer-specific terms.
Ownership
Every fact has a source, an owner, and a review date, so the brain doesn’t rot after launch.
Access control
Agents inherit the permissions of the person asking.
A fact, after capture
Northwind’s pricing exception
- Source
- A Slack thread from 2023, and a deal-desk interview
- Owner
- Deal desk
- Review
- Quarterly
Asking as the account rep
The agent inherits that person’s permissions.
- Contract and account notes in view
- Other teams’ private threads stay out
How it’s structured
Knowledge graph engineering
Vector search finds similar text. A graph knows that this customer, on this contract, owns that renewal.
- 01
Entity modeling
Accounts, contacts, products, contracts, deals, and policies, with the relationships between them defined up front.
- 02
Graph plus retrieval
Combine graph traversal with vector and keyword search so agents answer multi-hop questions (“which at-risk accounts share a champion who just left?”).
- 03
Entity resolution
Merge the six spellings of the same customer across systems.
- 04
Provenance
Every edge traces to a source record, so answers can cite their evidence.
Jordan Hale, the champion who left, connects to the Northwind account. Northwind holds contract MSA-441, which renews this quarter and covers the Enterprise plan. The champion edge cites the CRM contact record.
Multi-hop
Which at-risk accounts share a champion who just left?
Entity resolution
- Northwind
- North Wind, Inc.
- NW Logistics
One account
Champion edge · source CRM contact record.
What the agent receives
Context engineering
More context makes agents worse once it’s the wrong context.
Agent step
Draft the renewal noteTools · graph lookup, hybrid search
In the window
- MSA-441, renewal clauseContract420 tokens
- Northwind, three hopsGraph neighborhood610 tokens
- Open ticketTicket180 tokens
Held out
- Pricing wiki, 40 pages
- Email thread from 2021
1,210 tokens in a 4,000 budget
Selection
Decide what each agent step gets: which documents, which graph neighborhood, how much of it.
Assembly
Prompt structure, memory design, and tool definitions built around the task, not dumped into one window.
Budgeting
Token cost and latency per call, tracked against answer quality.
Conflict handling
Rules for when two sources disagree: newest wins, system of record wins, or flag for a human.
Proving the context works
Evals
A team ships a retrieval change on Monday and nobody can say by Friday whether answers improved.
Golden set · a rep’s question
Which at-risk accounts share a champion who just left?
- Verified answer
- Northwind. Jordan Hale, champion on MSA-441, left last month.
- Must cite
- The CRM contact record and contract MSA-441.
Layered metrics
- 01
Retrieval
Did the right content surface?
- 02
Grounding
Did the answer stay inside it?
- 03
Task success
Did the work get done?
Regression gates
Every change to content, graph, or prompts runs against the set before it ships.
Retrieval change · held until the set passes
Production monitoring
Sample live traffic, score it, and route failures back into the golden set.
How the work runs
How we engage
Services first, software where it earns its place.
- 01
Assess
Audit your sources, agents, and failure cases; rank where context is costing you most.
- 02
Build
Stand up the graph, the retrieval layer, and the eval harness against one high-value workflow.
- 03
Operate
Hand over dashboards and runbooks, or keep running it with you.
Case studies
Where we’ve put it to work.
Other services
Seven more ways we put AI to work.
- AI TransformationWe write your AI roadmap, build it, and hand it to your team.
- AI SDLCOne agent-ready flow from idea to merged code, scaled team by team.
- AI for GTM & FinanceA data foundation for go-to-market and finance, then the agents that run on it.
- Fine-tuning and EvalsModels tuned to your data, and the evaluations that prove they work.
- ML and Enterprise ForecastingForecasting at enterprise scale, like forecasting Black Friday demand to plan staffing.
- AI Cost OptimizationSeeing where AI spend goes and bringing it down without losing quality.
- Custom Agent and App DevelopmentAgents and AI applications built for one workflow and run in production.