Every claim below is paired with its limit. The limit is the part worth
reading: it is what tells you whether the capability survives contact
with your data, your auditors and your escalation policy.
C-01
Grounded retrieval
Answers come from your documents, with the passage attached.
We index the sources you nominate — policies, product data, past tickets, contracts — and every answer carries the passages it was built from. A reviewer can open the source without leaving the thread.
Where it stops If the answer is not in the corpus, the agent says so and escalates. It does not fall back on general knowledge to fill the gap.
C-02
Tool calls into your systems
The agent reads and writes where the work already happens.
CRM, ticketing, ERP, data warehouse, internal APIs. Each tool is declared with an explicit schema, so a call that does not match is rejected before it reaches your system rather than after.
Where it stops Every tool is allow-listed by name. There is no general-purpose HTTP call, and no tool is added without a scoping decision.
C-03
Scoped permissions
Read and write are separate decisions, per object.
An agent gets its own service identity, never a shared human account. Read scope and write scope are declared separately, field by field where the system allows it.
Where it stops We refuse deployments that need a blanket admin credential. If the target system cannot express the scope, we say so at scoping, not at delivery.
C-04
Approval gates
You decide which actions never happen unattended.
Thresholds are business rules, not model behaviour: a refund over an amount, a contract clause, a named account, a first contact. Anything crossing a gate is drafted and queued for a person.
Where it stops A gate is a hard rule evaluated outside the model. It cannot be argued around by a cleverly worded input.
C-05
Business test sets
Quality is measured against your cases, not a public benchmark.
During scoping we build a test set from your real history — the ordinary cases, the edge cases and the ones that went wrong. Every change to prompt, model or retrieval is scored against it before it ships.
Where it stops A test set covers what it contains. We report the cases it does not cover rather than quoting a single accuracy figure.
C-06
Full run traces
Every run is reconstructable, months later.
Input, retrieved passages, tool calls, gate decisions, output, and the identity that approved it — all logged with a run id, and queryable from your side.
Where it stops Retention is set by you and enforced by deletion, not by hiding rows from a view.
C-07
Rollback
A bad run is undone without archaeology.
Writes are designed to be reversible: staging tables, drafts, or field history. The run id links every write it produced, so reverting is one query rather than a manual hunt.
Where it stops Some actions cannot be taken back once they leave your systems — a sent email, a posted ledger entry. Those sit behind a gate by default.
C-08
Supervision in production
Drift is caught by monitoring, not by a complaint.
Escalation rate, gate hits, retrieval misses and cost per run are tracked from day one, with alerting on the thresholds agreed at scoping. Ninety days of supervision are included with every build.
Where it stops Monitoring reports what changed. Deciding whether a change is acceptable stays a business call, and we bring it to you.
Mechanism
What one request actually does.
Six steps, in this order, every time. The model appears at step three
and proposes; it never executes. That separation is what makes the
rest auditable.
01
Receive
The request arrives from your channel and is classified against the declared scope.
02
Ground
Relevant passages are retrieved from your corpus. Nothing is retrieved outside the nominated sources.
03
Decide
The model proposes an action. The proposal is a candidate, never an execution.
04
Check
Business gates run outside the model. Anything over a threshold is held.
05
Act
Allowed tool calls execute under the agent identity, inside the declared write scope.
06
Record
Input, sources, calls, decisions and approver are written to the trace under one run id.
Steps four and five swap only when an agent is explicitly configured to
act first and report after. We have never shipped one that way.
Integrations
Where agents plug in.
Named connectors we have shipped and maintain. Anything with a
documented API can be added during scoping — that work is estimated
explicitly rather than assumed.
Customer
Zendesk
Intercom
Freshdesk
Front
Revenue
Salesforce
HubSpot
Pipedrive
Operations
SAP
Microsoft Dynamics
NetSuite
Odoo
Knowledge
SharePoint
Confluence
Notion
Google Drive
S3
Identity
Okta
Entra ID
Google Workspace
SAML / OIDC
Data
Snowflake
BigQuery
Postgres
Elasticsearch
Not on offer
Three things we turn down.
A capabilities page that only adds is a brochure. These are the
requests we decline, and why.
An agent with no scope
A general assistant wired to everything cannot be tested, cannot be approved and cannot be rolled back. We build narrow agents and add scope once the first one is measured.
A number we cannot reproduce
We do not quote accuracy from a vendor deck. Every figure on this site names its period, its population and what it excludes.
A model you cannot change
Models are a dependency, not the product. Prompts, retrieval, gates and test sets belong to you, and the agent is built so a model can be swapped and re-scored.
See it on real deployments.
Three agents in production, with their scope, permissions, approval
path and the figures we can stand behind.
