The ACT lane
The Power Automate Decision Fabric
Last updated by Errin O'Connor, Founder & Chief AI Architect, EPC Group
Automation that acts on a single AI opinion is automation somebody has to supervise. The Decision Fabric brings several models into Power Automate and Power Apps at the exact moment a judgment call is required, compares what they return, and lets the flow continue only when the decision clears the threshold the organization set. High-stakes actions always retain a human approval step.
Key Facts
- The AI Consensus Gate: two or more models evaluate independently; disagreement escalates to a human queue.
- Financial, HR, customer-facing and compliance actions always keep a human approval step.
- Models are matched to decision types; sensitive payloads stay inside the Microsoft 365 trust boundary.
- Every decision leaves an audit record explaining why the flow acted as it did.
- Models can be substituted later without rewriting the flows around them.
The Knowledge Mesh governs what your models are permitted to know. This governs what your automations are permitted to do about it. That distinction is the whole design: retrieval mistakes produce a bad answer someone can question, whereas automation mistakes produce a posted invoice, a mis-routed case, or an approval nobody intended.
The single-model automation problem
A flow that calls one model and acts on the result has quietly accepted that model’s worst day as its operating floor. One hallucinated classification becomes a payment against the wrong cost centre. One misread contract clause becomes an obligation nobody reviewed.
Teams usually respond by adding a manual review queue over everything — which is a rational reaction that also erases the automation’s entire business case. You end up paying for the flow and reviewing its output anyway.
The D-E-C-I-D-E method
D — Define the decision points
Walk each flow and mark every step where judgment replaces deterministic logic — classification, extraction, routing, drafting, exception handling. Steps that are genuinely rule-based stay rule-based; putting a model where an if-statement belongs adds cost, latency and doubt for nothing.
E — Evaluate model fit per decision
Match decision types to model classes rather than standardizing on one. Sensitive payloads route to models deployed inside the Microsoft 365 trust boundary; high-volume low-stakes classification routes to whatever is fastest and cheapest that clears the bar.
C — Consensus Gate
The signature control. For designated decisions, two or more models evaluate independently and their outputs are compared. Agreement above the configured threshold lets the flow proceed; disagreement escalates to a human queue instead of guessing. Disagreement is treated as information, not as failure.
I — Integrate governance
DLP policies, connector governance, environment strategy, service-principal hygiene, and an audit record for every decision the fabric makes.
D — Deploy with human checkpoints
Approval steps on financial, HR, customer-facing and compliance actions are non-negotiable in the reference architecture. Not a configuration preference — a design constraint.
E — Evolve
Review the decision log, tune thresholds against real disagreement rates, and substitute models as the market moves — without rewriting the flows around them.
What the Consensus Gate actually changes
Ordinarily a flow has two states at a judgment point: proceed, or fail. The gate adds a third — contested — and that is where the value sits. Items where models agree flow through untouched. Items where they disagree are precisely the items a human should have been looking at, and now they arrive in a queue already flagged, with each model’s reasoning attached.
The disagreement rate itself becomes a management signal. A step where models rarely disagree is a candidate for full automation; a step where they constantly disagree is telling you the underlying data or the instructions are ambiguous, which is worth knowing regardless of AI.
Governance and safety architecture
SharePoint and Microsoft 365 permissions determine which information the authenticated user is authorized to retrieve, while EPC Group’s governed integration architecture controls what approved context is subsequently transmitted to a selected AI model.
Around that sit DLP policies, connector governance, a deliberate environment strategy, service-principal hygiene, and per-decision logging. This is a governance project before it is a technology project — the models are the easy part, and the controls are what make the result defensible when someone asks why a flow did what it did.
Where this sits: KNOW, ACT, ANALYZE
Three lanes, deliberately separate. The SharePoint AI Knowledge Mesh governs retrieval (KNOW). Multi-model architecture for Power BI governs interpretation (ANALYZE). The Decision Fabric governs action (ACT). A client can adopt any one without the others.
Getting started
Engagements begin with a decision-point map of flows already in production, since the fabric is retrofitted onto real automation rather than built alongside it. Work is fixed-fee and fixed-scope; pricing is confirmed after a scoping call.
Frequently asked questions
Can Power Automate call Claude or ChatGPT today?
Yes, through connectors or governed HTTP calls to enterprise APIs, inside your DLP and environment policies. That is not the same as those models running natively inside Microsoft 365 — Copilot is the native experience; everything else is integration EPC Group designs and governs.
What is an AI Consensus Gate?
A control that asks two or more models the same question independently, compares their answers, and only allows the flow to proceed when agreement clears a threshold the organization sets. When they disagree, the item escalates to a human queue rather than proceeding on one model’s opinion.
Does consulting several models slow flows down?
Yes, at the judgment points, by design — you are buying a second opinion and second opinions take time. Deterministic steps are untouched. Consensus is applied only where a wrong call is expensive, which is why the D stage maps decision points before anything is routed.
Which decisions should never be fully automated?
Anything that moves money, changes someone’s employment or benefits, communicates externally on the organization’s behalf, or creates a compliance record. Those keep a human approval step in the reference architecture regardless of how well the models agree.
How is a Decision Fabric engagement structured?
It begins by mapping decision points in flows you already run, because the fabric is retrofitted onto real automation rather than built beside it. Engagements are fixed-fee and fixed-scope, and pricing is confirmed after a scoping call.
If the flows already exist and the question is what to do about them, start with Power Automate consulting or Power Platform governance. If the harder question is who owns AI decisions across the organization, that is a vCAIO conversation.
Patterned on 300+ Copilot initiatives and 11,000+ engagements — the approval checkpoints in this architecture come from watching which automated decisions organizations later wished a person had seen.