AI assistants and agents

Teams whose knowledge is scattered across documents and specific people, who spend part of the day answering the same queries, or who have multi-step tasks that require looking things up and acting across more than one system. It works when reliable sources and clear limits can be set.

• The same internal questions are answered over and over.
• Finding an answer requires knowing whom to ask.
• Procedures and policies live in documents that almost nobody opens.
• Some tasks consist of checking several systems and preparing a reply.
• A generic AI tool has been tried and its answers are not trusted.

• Selecting and preparing the knowledge sources to answer from.
• Assistants that cite where each answer comes from and say so when they cannot find one.
• Agents with bounded actions: what they can look up, what they can do and what they cannot.
• Human approval for actions with impact.
• Evaluation with real questions and cases before they go into use.

• A working assistant or agent, with its scope and limits in writing.
• A test set of real questions and cases, and its results.
• Permissions, an activity log and a review procedure.
• Instructions for the team: what to do and what not to expect.
• A procedure to update the sources and the instructions.

The sources (document manager, intranet, email or databases) and the actions (the systems the agent can act on) are defined in the diagnosis, with minimum permissions. Nothing is connected before it has been decided what it may do.

• The assistant is designed to cite its sources and, if there is no reliable source, to say so.
• It is designed so that each action is bounded and logged, and those with impact require human approval.
• Amounts, totals, deadlines and statuses are calculated by the system with rules, not by the model.
• It is designed so that you can hand over to a person at any time.
• It is designed so that anyone talking to an assistant knows it is an AI.
• It is evaluated before use and again whenever the sources or the instructions change.

• Decisions that must be made by people, such as hiring, granting credit or diagnosing.
• Answers without a source.
• Agents with free access to your systems.
• A guarantee that an assistant will never be wrong: risk is reduced with sources, limits and review, and it is measured.
• Training your own models from scratch.

We measure with a set of real questions and cases: correct and cited answers, queries resolved without intervention, actions approved or rejected by people, and time spent on queries, against the baseline.

Tell us which queries keep repeating or which tasks get stuck. A first conversation is to understand the process and assess whether we are a good fit; if we are not the right option, we will say so.

Contact us →See how we work

• Process automation
• Document intelligence
• Integration and data