// AI Agents · Flagship

AI Agents that don't just answer. They act.

An assistant that can only answer questions saves a search. An agent that can read your systems, apply your policy and complete the task saves the task. That is the difference we build for.

// How an Agent Works

From a question to a completed action

Every step here is a place the agent can be constrained, logged and evaluated. An agent you cannot inspect at each of these points is a demo, not a system.

  1. Employee asks

    In plain language, in the tool they already have open — Teams, email, a portal.

  2. Agent interprets

    Works out what is being asked, what it is allowed to see, and which tools it needs.

  3. Enterprise knowledge

    Retrieves from policies, documents and records — scoped to that person's permissions.

  4. Systems of record

    Reads and writes through SAP, CRM, ERP and internal APIs rather than guessing.

  5. Decision

    Proposes an action with its reasoning and the sources behind it.

  6. Action

    Executes, or routes to a human when confidence or policy says it should not.

// Where Agents Earn Their Keep

Agents we build

Each of these sits on a real system boundary. The value is not the conversation — it is that the agent can reach the system that holds the answer and, where permitted, change it.

SAP landscape

SAP Assistant

Answers questions against SAP data and executes routine transactions, without the user learning a transaction code.

  • Order and delivery status
  • Master data lookups
  • Guided transaction entry
  • Report retrieval

People operations

HR Agent

Handles the questions that consume an HR team's week — policy, leave, payslips — and files the routine requests.

  • Policy questions
  • Leave balances and requests
  • Onboarding checklists
  • Payslip retrieval

Finance & controlling

Finance Agent

Chases the reconciliations and variance questions that otherwise fill month-end.

  • Invoice status
  • Variance explanations
  • Approval routing
  • Month-end checklists

Source to pay

Procurement Agent

Turns a request into a compliant purchase, applying the policy rather than asking someone to remember it.

  • Requisition creation
  • Supplier lookup
  • Policy compliance checks
  • PO status

Plan to deliver

Supply Chain Agent

Surfaces the exception before it becomes a shortage, and proposes what to do about it.

  • Stock and shortage alerts
  • Supplier delay impact
  • Reallocation options
  • Demand queries

Data & reporting

Analytics Agent

Answers questions of the governed semantic layer, so the numbers match the dashboards.

  • Ad-hoc questions
  • Metric definitions
  • Trend explanations
  • Report generation

Service & support

Customer Service Agent

Resolves what it can from real account data and hands over cleanly when it cannot.

  • Order and account queries
  • Troubleshooting
  • Ticket creation
  • Escalation with context

// Before You Build One

What decides whether an agent survives production

Agent pilots rarely fail on model quality. They fail on the four things below, so these are the conversations we have first.

01

Most tasks do not need an agent

If the steps are known in advance, a deterministic workflow is cheaper, faster and easier to audit. We reach for an agent when the path genuinely varies with the request — and say so when it does not.

02

Permissions decide whether it is safe

An agent that can read everything is a data breach waiting for a prompt. Access is enforced in the retrieval and tool layers, under the identity of the person asking, never in the system prompt.

03

Actions need a blast radius

Reading is reversible; writing to SAP is not. Every tool an agent can call has an explicit scope, an approval threshold and an audit record of what it did and why.

04

Evaluation is the whole game

Without a scored evaluation set, nobody can tell whether a change to the prompt improved anything. We build that set before the agent, and it is what makes ongoing changes safe.

If a process is fully deterministic, we will tell you to automate it without an agent — it will be cheaper to run and easier to audit.

// How We Build Them

The stack underneath

Chosen per engagement. Retrieval quality and tool design decide the outcome far more than the choice of model does.

LLMsRAGVector searchTool callingAgent orchestrationEvaluation harnessesSAP integrationREST & OData APIsIdentity-scoped retrievalAudit loggingPythonAWS

Which process would you hand to an agent?

Bring us the one that consumes the most of someone's week. We will tell you honestly whether an agent is the right answer for it, or whether a plain workflow would serve you better.