Artificial Intelligence
Assistants, automation and AI features, built to run inside your systems.
Some of it answers questions from a company’s own material. Some of it classifies and extracts without anyone opening the file. Some of it starts something on its own. The further a system goes without a person in the middle, the more it matters what it was allowed to see, and what had to be true before it acted.
What we build
Six shapes this work usually takes. Most projects are one of them, and a few are two.
Internal assistants
Answers drawn from a company’s own documents, records and policies, with a way back to the source.
Document pipelines
Classification, extraction and routing for material that arrives as PDFs, scans and email.
Search over private data
Finding by meaning across a body of material too large to browse or tag by hand.
Conversational interfaces
Chat and voice that stay inside a defined set of actions, in the language the user already speaks.
AI inside an existing product
Features added to software that already has users, without a rebuild around them.
Agents that finish a task
Workflows a system completes end to end, with a person approving rather than operating.
What the system is allowed to do
Whatever shape it takes, the work sits on one of three levels. The model is rarely what decides whether it holds up; the level is. The higher it sits, the less there is between the system and the consequence.
It answers
Retrieval over a private corpus. The system finds what is relevant and shows where it came from; a person reads it and decides what to do with it.
When it’s wrong A wrong answer is visible to the person who asked, and costs them time.
ADA matches companies across the IDB’s ConnectAmericas, a network of hundreds of thousands.
It decides
The system commits to a judgment: a classification, a route, a match. The work continues on top of it. Keystone’s investment contracts are sorted by type and turned into structured data every day, without anyone opening them.
When it’s wrong A wrong decision is visible to nobody. It flows into a record that other decisions are built on.
Processing went from hours to minutes per document, at hundreds of contracts a month.
It acts
The system starts something. EmpowerHealth reaches patients by voice and SMS across several care programs running at once, and handles what comes back.
When it’s wrong A wrong action reaches a person who acts on it, and there is no step where someone could have caught it first.
The market is selling the third level. The question worth asking first is which one the problem actually needs.
What gets settled first
Two things shape everything built after them, and both get decided before a model is involved: where the material stays, and who is allowed to see which part of it.
Where the material stays
Keystone’s contracts arrive as PDFs and scans, uploaded into Egnyte, the document system they already had. We didn’t move them. A scheduled process reads them where they live, sorts them by contract type, and extracts structured data with AI-assisted OCR.
Moving a corpus to make it readable creates a second copy to keep in sync, and a second place to get access wrong. Read the case
Who is allowed to see what
A system that answers everything to everyone is a compliance problem wearing a chat interface. EmpowerHealth runs several care programs in parallel under HIPAA, where what may be said, stored and sent is a design input, settled before the first prompt is written and not patched afterwards.
The same question decides the shape of the retrieval index, not just the interface on top of it. Read the case
We ran into this often enough to build our own
A Onetree product
An AI agent that answers across a company’s own data and documents: the knowledge itself, not a search box.
It exists because the same problem kept arriving from different directions. Datialog is that problem solved once, in production, and sold on its own.
Visit datialog.com →
We’re a Microsoft Solutions Partner, and this is what we build with.
Microsoft has checked our work, we build with their tooling, and when the thing that’s broken is their service and not our build, there’s a direct line.
Models
Foundation models accessed through Azure OpenAI, chosen per task rather than by default. We don’t train models from scratch, and we’ll say early when a project doesn’t need one.
Retrieval
Azure AI Search for semantic and hybrid retrieval over a private corpus, with the index built around who is allowed to see what.
Documents
Azure Document Intelligence for classification and extraction from PDFs and scans, including material that was never born digital.
Conversation
Azure CLU for intent and language understanding where the exchange has to stay inside a defined set of actions.
Around it
.NET and Python for the services, React for the interfaces, and Twilio where the conversation leaves the screen for voice or SMS.
ADA, the matchmaking platform for the IDB’s ConnectAmericas, was built in close collaboration with the Microsoft AI Lab.
Questions we get before the first call
Do you build agents?
Yes, and the useful question is which level you need. An agent that acts on its own is the right answer when the workflow is well defined and the cost of a wrong action is bounded and reversible. When it isn’t, the same problem is better solved one level down, where a person stays between the system and the consequence. Datialog, our own product, is an agent; EmpowerHealth’s outreach acts without being asked each time.When is AI not the answer?
More often than the market suggests. A good part of what gets asked for as an assistant is a search index with a decent interface, and a good part of what gets asked for as a model is a rule somebody never wrote down. We’ll say so before the contract rather than during it, and it costs us the larger engagement every time.Do you build your own models?
No. We work with foundation models and put the effort where the difference actually is: the corpus, the retrieval, the access rules and what the system does when it doesn’t know. Training from scratch is the right call for very few organizations, and none of them need to ask us.Can you work with our data where it already lives?
Usually yes, and it’s the better option. Keystone’s contracts stay in Egnyte, the document system they already had; a scheduled process reads them there. Moving a corpus to make it readable creates a second copy to keep in sync, and a second place to get access control wrong.
Bring us the one you’re not sure is an AI problem.
In a 45-minute working session we’ll tell you whether it’s an AI problem, which of the three levels it needs, and what would have to be true for it to work. Bring the material; we’ll bring the questions.