What we build

Claude Assistant tasks and coach views, installed inside IBM BAW

Most enterprise AI sits beside the work. Someone leaves the system of record, pastes context into a chat window, gets an answer, and carries it back by hand. The pilot demonstrates capability and changes nothing about cycle time.

We build components that live inside the process instead — assistant tasks that run as steps in a flow, and coach views that appear in the screens your people already use. They inherit the case, the assignment, the escalation and the audit trail, because they are part of the process rather than a conversation next to it.

Receive Classify assistant task Extract assistant task Validate rules, not model Route deterministic Review human task End COACH VIEW — INSIDE THE BAW SCREEN Claude — produces data Deterministic — decides routing Human — approves Claude appears twice and routes nothing.
Assistant tasks run as ordinary automated steps; a coach view rides inside the human task’s own screen. Routing between them is always deterministic.

Two integration surfaces

There are exactly two places an assistant needs to live inside BAW, and they solve different problems.

Surface one

The assistant task

A step in the process with no human in it. The token arrives, the task runs, it writes its output into the run data, and the token moves on. Modelled as an ordinary automated activity, so it can be routed around, retried, or skipped by a decision like anything else.

Each one is configured with a system prompt defining its narrow job, a tool set it may call, and an output contract — the schema of what it must return and the run-data keys those values land in.

Surface two

The coach view

The assistant appearing inside a screen a person is already using. The case worker does not leave BAW; a panel in their task interface offers a reading of the case, drafts a response, or explains what the next step needs.

Coach views keep the human in the loop by construction. The model proposes, the person disposes, and the process records what was accepted — which is the right shape for any judgement that carries consequence.

What the assistant does

Configured per process, not generic.

Classify documents with vision

Attachments of unknown type and quality, labelled against a closed taxonomy. Anything it cannot place comes back as unknown — a valid answer rather than a guess.

Extract structured data

Schema-defined extraction per document type, with a confidence signal and the verbatim source span for every field — so review is fast and provenance is checkable automatically.

Suggest the next action

Guided instructions for the person doing the work: what this case needs, what is missing, what the policy says. Advice in a panel — never an action taken on their behalf.

Generate safe configuration

Validated JSON that drives application behaviour — including report pages a business user composes themselves. Configuration, never code.

See it run, step by step

A document intake process, one step at a time. Click a step, or walk it with the buttons — watch where Claude produces data and where deterministic rules decide what happens next.

Walk a case through the process
1 / 7
Receive intake Classify Claude Extract Claude Validate rules Route deterministic Review human Complete audit written

The one that matters most

Self-service reporting, without showing Claude your data

A business-unit manager wants a report the standard screens do not show — instance progression, ageing by step, volumes by some dimension nobody anticipated. Today that is a ticket, a developer and a release.

The naive fix — give a model database access and let it write queries — is the pattern the industry spent 2026 learning not to deploy. We avoid it by never granting the capability at all:

Schema only

Claude receives a description of the data structures. It never receives rows — not real, not sampled, not anonymised.

Restricted surface

Only pre-approved sanitised views are described. Restricted columns are absent, so they cannot be named by a query that does not know they exist.

Read-only, validated

Execution uses a read-only principal scoped to those views, and every statement is parsed and checked before it runs.

YOUR SYSTEMS Sanitised views restricted columns absent Data rows never leave this box ✕ Read-only principal granted on those views only BOUNDARY schema description only no rows, ever Claude sees table and column names, types, semantics — and nothing else JSON config Validate schema + parse the SQL Human publishes recorded approval executed read-only, against the approved views
The safety property is structural: rows are never on the outbound path, so there is nothing to leak and no prompt-injection surface.

The safety property does not depend on the prompt, or on the model behaving well. It follows from what was never made reachable. There is also no prompt-injection surface, because the model never reads data in the first place.

The rules we build to

Structural, not advisory — principles erode and structures do not.

The model never routes

It classifies, extracts, drafts and suggests. Where the case goes next is decided by a deterministic condition language over a published, immutable version of the flow — so a routing decision is still explainable a year later.

It only acts through a tool surface

Every capability is a named tool with a schema'd payload. No raw query interface, no arbitrary execution. The blast radius is a fixed, reviewable list rather than a matter of how well the prompt was written.

It sees the least it can

Context is assembled per task from the specific fields that task needs. Passing the whole case object is easier, and it is where most real data-exposure incidents start.

Every output is validated before it is trusted

Schema conformance is necessary and insufficient. Business rules are checked separately and deterministically: totals reconcile, dates fall in range, references resolve, enumerations are members of the set.

Read the architecture in full

Where are you starting from?

Tell us which processes are eating your delivery capacity, and we will show you what they look like with an assistant inside them — and what it would take to get there.

Start a conversation

Claude is a product of Anthropic. IBM and IBM Business Automation Workflow are trademarks of International Business Machines Corporation. EFTEDRA is independent and not endorsed by either company. Components described on this page are in active development; working demonstrations are available on request.