Common questions about AI agents and knowledge graphs

An AI agent is software that finishes a piece of work, start to end. It reads what comes in, pulls what it needs from your systems, makes the routine calls the way your team would, and hands a person anything that needs approval. Think of it as a new hire who already knows your process.

Agentic AI is AI that plans and carries out multi-step tasks across your tools instead of replying once and waiting. In a business that looks like reading a referral, checking eligibility and drafting the authorisation, with a person approving the last step. How well it works depends mostly on how well it knows your data and your rules.

A chatbot answers questions. A copilot helps one person work faster inside a tool, and that person still does every step. An agent does the steps itself, across several systems, and brings a person in to approve. If the job is to answer something, a chatbot will do. If the job is to get something done, you need an agent.

They run on models trained on public data, so they know a great deal about the world and nothing about your business. Out of the box they can’t see your systems, can’t tell which record is the right one and don’t know how your team decides. Give a model the right record and your rules and it becomes far more reliable. Our platform supplies both.

It is our name for what the agents run on: one place where everything your business knows is connected, from the records in every system to the know-how in your people’s heads. Agents use it to do the work, and your team can ask it questions in plain English and get the answer with its source.

The best first jobs are repetitive, rule-heavy and spread across several systems. Examples: underwriting a commercial real estate deal from the rent roll and T-12, turning an emailed referral into a booked visit, triaging insurance submissions, or turning an emailed drawing into a costed quote. Anything that leaves the building still waits for a person.

Agents take over tasks, mostly the retyping, checking and chasing that fill a working day. Your people keep the judgement calls, the relationships and the approvals. In practice, teams spend the time they get back on the work they were hired for.

Every answer and draft is checked against its source and your rules before anyone sees it, and the agent stops and asks wherever a person has to approve. Each step is logged, so you can see what it did and why. Nothing is sent, filed or paid without sign-off.

Commercial real estate, healthcare operations, insurance, manufacturing and residential real estate brokerage. Each industry has prebuilt agents: deal origination, underwriting and investor reporting in commercial real estate, for example, or referral intake and claims recovery in healthcare.

Pricing has two parts: a one-time configuration fee to set the agent up on your systems and rules, and a monthly fee to run it. The whole price is outcome-based. We agree up front what the agent should deliver, such as hours saved or money recovered, and set the price against it, so the return on investment is built into the deal.

Most agents are live in 4 to 6 weeks. In that time we connect the systems the process touches, write down how your team makes the calls, test the agent on real work and release it. Each agent after the first goes faster, because the groundwork is already there.

We agree the outcome before anything is built: the hours, turnaround time, error rate or money recovered that the agent should move. Then we measure the process before and after. Because the price is tied to that outcome, the return is part of the agreement from day one.

No. Each agent comes prebuilt for its job, with the reading, checking and drafting already worked out. We configure it to your tools, your systems and your parameters, then release it into the work. That is why it takes weeks rather than months.

The ones you already run. That includes property and portfolio systems such as Yardi, MRI Software, ARGUS and CoStar, EHRs such as Epic, athenahealth and eClinicalWorks, ERPs such as NetSuite, SAP and Microsoft Dynamics 365, and tools like Salesforce, Microsoft 365 and Snowflake.

No. Our platform reads from the systems you already run, where they sit, and writes back only where you have asked it to. Your EHR stays the medical record, your ERP stays the ERP, and your team keeps working in the tools it knows.

An agent sees only what the person it works for can already see in the source system, and every read and write is recorded. Our platform can run in your own infrastructure or in a governed environment we operate, with a BAA where healthcare data requires one.

No. We handle the setup, the connections and the configuration. Your team’s part is explaining how the work is done today and approving what the agent drafts. Your IT team joins for access and the security review.

It keeps learning. Corrections your team makes become rules it follows, definitions stay current, and the next agent starts from everything the first one learned. You can see every step it takes, and we keep tuning it as your process changes.

A knowledge graph stores the things a business runs on, such as customers, patients, policies, parts, properties and contracts, as connected records with typed relationships between them. Because it keeps the connections, you can ask how things relate, for example which leases roll next year at properties held by one fund, and get an answer that follows the links.

A semantic layer defines what your business terms mean and how each metric is calculated, so a term like active patient or net operating income means the same thing in every system, report and answer. It sits between raw data and the people or agents asking questions of it.

The semantic layer answers what a term or a number means. The knowledge graph answers how things are connected. Agents need both: definitions so the numbers are right, and relationships so a question can be followed across systems.

A context layer assembles what an agent needs for one specific task: the relevant records from the graph, the right definitions, the matching documents, and the rules and exceptions your team works by. It also holds the know-how that never made it into a system, captured from documents, email, calls and conversations with your people.

A harness is the instruction manual wrapped around an AI agent. It tells the agent how your business decides, which system and record each question should go to, how to check its work against the source, and when to stop and ask a person. We build a domain-specific harness for each industry, then fit it to each company.

Usually both. Vector search is good at finding passages that say something similar to the question. A knowledge graph is good at questions about relationships that span several records or systems. We run keyword search, semantic search and the graph together, then narrow results by entity and date before the agent answers.

GraphRAG is retrieval-augmented generation that uses a knowledge graph alongside plain vector search. Comparisons of the two tend to find vector retrieval better at single-fact lookups and graph retrieval better at multi-hop questions that join several documents. Most real business questions are multi-hop, which is why the graph matters.

A model hallucinates when it has to fill a gap, so it writes the most likely answer instead of the true one. The gap is usually missing or ambiguous data. We reduce it by sending each question to the right record first, answering only from what was retrieved, citing the source, and flagging low-confidence results for a person.

Every figure and claim carries a line back to its source: the document, page, table and line, or the record in the system it came from. When the system is unsure how it read a document, the result is marked as unsure, and you can route those to a person.

Through each system’s API or export where it has one, and through file shares, inboxes and document stores where it doesn’t. Connections run on a schedule, pick up what changed since the last run and leave the source where it is. Nothing is migrated.

Yes. Our platform can run inside your infrastructure or in a governed environment we operate. Access is role-based: an agent sees only what the person it works for can see in the source system, and every read and write is logged.