Many companies still treat enterprise AI as a model-selection exercise.
Which model is smartest? Which model is fastest? Which one should every employee use?
Those questions matter, but they are not an operating strategy. A capable model can reason about a task. It cannot decide, by itself, which company data is trusted, which tools it may use, what a good result looks like, when it must stop, or who approves a consequential action.
Teradata's latest announcement is useful because it points beyond the model. The company describes its new Tera offering as a combination of a governed context engine, an execution harness that routes work across skills, tools, data and models, and task-specific Agent Skills. Teradata also says the harness loads memory and guardrails before execution, supports human approvals, preserves state and recovers from failures.[1]
Those are vendor-reported capabilities, not independently tested findings. I am not recommending the product here.
The architecture is the signal.
The enterprise AI stack is becoming an execution layer. For an SME, that layer needs five parts: context, orchestration, skills, controls and recovery.
1. Context: decide what the agent is allowed to know
A model can produce a fluent answer from poor information. That is one reason a chatbot demo can look impressive while a live business workflow fails.
Enterprise context is more than uploading a folder of documents. It should identify the source of truth for each decision.
For a sales workflow, that might mean:
- customer and opportunity records from the CRM;
- the current price list and discount policy;
- signed agreements rather than draft contracts;
- approved product information;
- recent communication linked to the correct account.
The agent should also know which sources are advisory and which ones are authoritative. A salesperson's note may be useful context. It should not silently override a signed order form.
This is where data readiness becomes operational. Clean data helps, but provenance matters just as much. The workflow must be able to show where an input came from, when it was updated and whether the agent was permitted to use it.
2. Orchestration: route work by capability, cost and risk
One powerful model does not need to handle every step.
A recurring workflow may include document extraction, classification, calculation, retrieval, drafting, validation and a final action in an ERP or CRM. These steps have different requirements.
Use deterministic software for calculations and rule checks when possible. Use a lower-cost model for routine classification or first-pass drafting. Use a stronger model when the task needs judgment across messy evidence. Require a person to approve actions that affect money, customers, legal commitments or sensitive records.
That is orchestration.
It replaces the vague instruction to "use AI" with an execution plan. Each step has an owner, an input, a tool, a success condition and a fallback.
This also keeps cost discussions honest. Token price is only one part of the cost. Repeated retries, unnecessary model calls, human rework and failed downstream actions can cost more than the initial inference.
The right metric is not cost per prompt. It is cost per completed, verified outcome.
3. Skills: turn expert methods into reusable execution
Teradata describes Agent Skills as reusable, task-specific capabilities that package methods for data engineering, analysis and data science.[1] The same principle applies outside data teams.
A skill is a reusable set of instructions that teaches AI how to complete a task, use the right tools and check its work.
For example, an accounts-receivable skill might define:
- which ERP fields to read;
- how to identify overdue invoices;
- which customer promises count as payment commitments;
- how to draft a follow-up;
- when to escalate a disputed amount;
- what evidence to attach before an employee approves the message.
The value is not the prompt. The value is the operating method.
This changes who can build enterprise AI. The domain expert who understands exceptions, evidence and quality standards can become the architect of the skill. A developer may still connect systems and enforce permissions, but the expert defines how good work is done.
Do it once. Skill it up. Do it again.
4. Controls: place approval before the consequential action
Many workflows add a human review box at the end and call that governance. That is too vague.
A control should be attached to a specific action.
An inventory agent may recommend a reorder on its own. It may create a draft purchase request within an approved range. It should not commit a large purchase, switch a supplier or accept a changed payment term without the right approval.
The boundary should be explicit:
- what the agent can read;
- what it can draft;
- what it can update automatically;
- which thresholds trigger approval;
- who can approve;
- what gets logged;
- how the action can be reversed.
This is how policy becomes executable. "Use AI responsibly" is not a control. "Require finance approval before any supplier commitment above S$10,000" is.
Controls should run before the action, not after the damage appears in an audit report.
5. Recovery: design for long-running work and failure
Business workflows do not always finish in one session.
An agent may need to wait for a manager's approval, a supplier reply, a document upload or a nightly system update. If it loses its state, someone has to reconstruct the work. If it retries carelessly, it may create duplicates.
Recovery needs a few basic elements:
- a durable record of the current step;
- the evidence collected so far;
- a unique idempotency key for the transaction;
- checkpoints before external or destructive actions;
- a clear retry and rollback rule;
- an exception queue with a human owner.
Teradata says its harness is designed to preserve execution state, recover from failures and resume after pauses.[1] Whether an SME uses that platform or builds a smaller system, the requirement is the same: the workflow must survive interruptions without guessing what already happened.
Build the operating layer around one workflow
An SME does not need to buy an enterprise platform or launch twenty agents to use this architecture.
Start with one recurring workflow where the current process is visible and the outcome matters. Lead qualification, invoice follow-up, purchase-order preparation, customer onboarding and monthly reporting are all reasonable candidates.
Then map the five parts:
- Context: Which records and documents are trusted?
- Orchestration: Which steps use rules, tools, models or human judgment?
- Skills: What method and checks should run every time?
- Controls: Which actions need limits or approval?
- Recovery: How does the work pause, resume, retry and roll back safely?
This is the move from Adoption to Agency in the Agent Orchestration maturity path. At Adoption, a team automates useful steps. At Agency, it manages a dependable system of digital coworkers with clear boundaries, evidence and accountable human owners.
The model will change. The tools will change. Your operating method should survive both.
Before buying another AI tool, diagram the execution layer for one recurring workflow. If you cannot show the context, routing, skill, control and recovery path on one page, the workflow is not ready for more autonomy.