Henry Stewart DAM  ·  Session handout

Handout  ·  Henry Stewart DAM

The four-tier agent ladder

Your DAM's most active user is no longer a person. This is the model from the talk, plus the questions to run against every agent you have already connected. Every agent sits on exactly one rung, and promotion is a decision, not an accident.

See where you stand in 2 minutes

01 The model

Four tiers of agent access

Pick the lowest rung that makes the agent useful. Moving an agent up a rung should require somebody to decide to move it, and should be as visible as granting a person admin rights.

Tier 0Readsuggests, never writes
What it does
Finds assets, summarizes a collection, proposes tags into the conversation rather than into the record.
Gate
Read-only, scoped to named collections, no bulk export, no download of originals.
Watch for
Read-only is not risk-free when the scope is the whole library. An agent that can read everything and summarize it is an export path. Scope matters more than mode.
Tier 1Proposewrites into a queue
What it does
Enriches metadata, classifies against a taxonomy, fills gaps at volume.
Gate
Writes land in a review queue, batched and attributed to the agent. One action accepts a batch, one rejects it.
Watch for
Graduate an agent off the queue per field, once you have a hit rate you trust for that field. Never wholesale, and never because the queue got tedious.
Tier 2Derivemakes new assets
What it does
Crops, renditions, format conversions, cutdowns, generated variants from an approved master.
Gate
Derivatives inherit rights, territory and expiry from the master. The agent cannot write to any rights field. Lineage is recorded.
Watch for
This is where restrictions get dropped. A transform pipeline copies pixels and loses constraints, so a locked-down master quietly produces an unrestricted child. Test it by asking who can name every child of a revoked master.
Tier 3Publishdistributes outward
What it does
Pushes to the CDN, the storefront, the portal, the social endpoint, the partner feed.
Gate
A named human approves each action, and the approval shows what is being published and to whom. No trusted-agent exemption.
Watch for
An agent that has been correct for six months is the one nobody is watching. Reversibility decides whether a human stays in the loop, not accuracy. If an agent publishes hundreds of times a day, the workflow is wrong rather than the control.

Where does your team stand today? Take the 5-question check →

02 Before you connect anything

Five questions before you connect an agent

Run these against every agent you have already connected, not just the next one. Most organizations discover on question two that the answer is "everything." Answer for your setup today, and you will see where your gaps are.

Whose identity is it acting as?A named person it is acting on behalf of, not a shared integration account. If every agent action logs as "API user," you have no attribution and no way to reconstruct an incident.
Whose identity is it acting as?
What is the smallest scope that still makes it useful?Collections rather than the tenant. Specific fields rather than the whole record. Start from nothing and add what the job needs, rather than starting from the connector's default.
What is the smallest scope that still makes it useful?
Where does untrusted text reach it?Filenames, captions, descriptions, embedded XMP, delivery manifests. Anything written outside your organization is an instruction the moment an agent reads it to decide what to do.
Where does untrusted text reach it?
Is each action reversible, and if not, who approves it?Sort the agent's actions by whether you can undo them. Everything irreversible or externally visible needs a person, every time, however reliable the agent has been.
Is each action reversible, and if not, who approves it?
Can you reconstruct exactly what it did, per agent?Not "an API call happened." Which agent, on whose behalf, which assets, which fields, in what order, and can you reverse that batch in one operation.
Can you reconstruct exactly what it did, per agent?

0 of 5 answered

03 Take these to your DAM vendor, mine included

Four questions for your DAM vendor

  1. Can you issue credentials per agent, acting on behalf of a named person, rather than one shared integration account?

  2. Are usage rights and restrictions enforced at the API, or only in the interface?

  3. Can I see every action a given agent took, and reverse a batch?

  4. Where is the kill switch, and can my team reach it without opening a support ticket?

04 The argument in sixty words

Why DAM is unusually exposed

Your metadata is written by strangers

Agency deliveries, contributor portals, self-service tenants and freelance shoots all put free text into your fields. That was harmless while the only readers were a search index and a person. An agent that reads a caption to decide what to do is taking instructions from outside your organization.

Your rights controls live in a dialog box

The warning that stops someone using an uncleared asset is a modal, an approval step, a colleague who knows the campaign. None of it exists at the API layer, and the API layer is all an agent ever sees. Good rights metadata does nothing if nothing enforces it.

Let’s Talk About Your Content.

See how Tenovos helps your team manage and activate content at scale, with the visibility and intelligence to prove its business value.

Get a Demo
Neil Grant, VP of Content Intelligence neil@tenovos.com Questions about your own setup are welcome, including the awkward ones.