ChatGPT
Perplexity
An AI backlog usually starts as a list of requests, often ordered by whoever pressed hardest in the last planning cycle. The business case behind each item then prices the work in wages saved, which filters out anything whose real cost is a customer waiting.
Gartner’s AI maturity survey shows how much rides on selection. Among leaders at low-maturity organizations, 37% named finding the right use case as a top barrier.
High-maturity organizations choose based on business value and feasibility, and 45% keep AI projects in production for three years or longer, compared with 20% at low maturity.
The better approach is to rank candidates by where elapsed time, rework, and customer exposure build up in the processes you already log.
From there, four questions decide whether a step needs a flow, a copilot, an agent, or a redesigned process: how much the work varies, whether its actions can be undone, whether a person must own the judgment, and how many systems it crosses. So how do you read that data and act on it?
Sponsorship and wage math undervalue the work where customers wait
A nominated use case comes with a sponsor who can see the effort involved. The delay a customer sits through has no such sponsor, because it builds up between queues that nobody owns.
Case in point
Brandon Micci, Head of AI Strategy and Business Transformation at a global financial institution, described that gap on our ECAF podcast.
A client asking about a $50 million wire can trigger a ticket that someone picks up four days later, after the transaction has already processed. Proposals to fix that path kept meeting the same objection, which was why anyone would spend $5 million automating work when the company could hire more people at $9 an hour.
Getting past that objection took leadership they could trust. His leadership’s first instinct was to survey call-center staff on their own time savings, and the answers were too wild to defend, so Brandon’s team moved to measured process data.
When should a request be funded before the data is in?
A regulatory deadline or a customer-facing risk can justify funding a request before the data is in, though it still needs a baseline. Every other request should compete on the same terms, which means adding a wait-time line to the business case template so your CFO sees what delay costs next to the labor it absorbs.
Your event logs already show where time and rework build up
What process mining reveals about delay and rework
Process mining reads the timestamps your systems of record already keep and rebuilds the path each case took, loops included. Task mining covers the desktop steps between systems, and in the Microsoft stack both sit inside Power Automate.
Forrester predicts that process intelligence will rescue 30% of failed AI projects in 2026.
Microsoft’s customer and partner organization shows what that analysis surfaces. When it mined its demand-generation process, it found an average of about 60 days across five variants and 33 activities, with 20% of cases needing rework.
Intake took about a tenth of that time, and the review loop was the largest bottleneck, so a team that automated intake first would have addressed only a small share of the delay.
Where process data stops being enough
Judgment exercised in email and meetings leaves no event log. Microsoft’s telemetry shows how much of the day runs through those channels, with employees interrupted by a message or meeting every 2 minutes.
The map is also only as reliable as the logs behind it. Where records are scattered across legacy systems, fix that data plumbing before trusting the ranking, as Brandon argues.
Task mining adds its own condition, since it records employee desktops and needs consent settled before it runs. Cost narrows the field further. Beyond the trial, more process mining capacity means Premium licensing plus a separate add-on, which low-volume processes rarely justify.
Start with the processes that run through your ERP or CRM at volume, since those carry both the logs and the customer wait.
A step’s variance and risk decide whether it needs a flow, a copilot, an agent, or deletion
When a flow is enough
Few variants and structured inputs call for deterministic automation, and so do irreversible actions. Microsoft’s Copilot Studio guidance keeps payments and record deletions in strictly authored logic with no AI interpretation.
Forrester expects fewer than 15% of firms to turn on agentic features in their automation suites in 2026, as ROI and governance keep most organizations on rule-based automation. A flow-first choice puts you in the mainstream.
When a copilot fits
When work lives in documents and conversations, and judgment must stay with a person, a copilot shortens their time in the step without taking the decision away. It can pull the relevant records, summarize a long case thread, or draft the next response, while the employee decides what the information means and what happens next.
The work needs interpretation, but the actions that follow are few and owned by the person, so an agent’s autonomy would add risk without adding speed.
When an agent earns its cost
An agent is justified when cases arrive unstructured in many variations and spend most of their time moving between systems, so the wait between handoffs is the number it should shrink.
Microsoft’s guidance puts that kind of agent in a hybrid layer for medium-risk work, where it operates within set boundaries and an approval step or value limit forces escalation.
Simform built this shape for a US HVAC group. Azure AI Foundry routes technicians’ photos and dictated notes to field operations or billing agents that act directly on the Procore system of record.
When the step should go
Brandon warns that AI on a broken process only speeds up the breakage. If rework comes from a handoff nobody needs, remove the handoff first.
The same AI request can lead to an agent, a flow, or process redesign
Only the invoice case below is illustrative.
The wire-status queue.
In Brandon’s example, the request on the table was more staff at $9 an hour. Framed around the client’s wait, the answer he describes is an agent that identifies each request, pulls the needed records, and drafts the reply, with a person checking it before it reaches the client.
The metric is request-to-response time, and he expects it to drop from days to minutes.
An invoice process that asked for an agent.
The data shows one dominant path and few exceptions. A flow with document extraction handles it; exceptions go to a person, and the touchless rate shows whether it worked.
A logistics firm that fixed the handoff first.
Freight forwarder ALPI used process mining to find the root causes of its fulfillment errors, which traced to how departments communicated.
Better cross-department communication cut fulfillment errors by 75%, and fewer sales steps lifted sales by 15%. Both gains came from redesigning how work moved between people.
The baseline expires the day the automation ships
What to track after automation goes live
Once a step is automated, its old baseline describes a process that no longer exists. Gartner now argues for continuous visibility into how work gets done. Re-mining after launch is how you get it, both to set a new baseline and to find where the bottleneck moved. Brandon likens that loop to a CI/CD pipeline for agents.
Track cycle time, rework, variant drift, and human corrections, and for agents, add the success and fallback rates Microsoft recommends. If agent success falls, or fallback and correction rates climb past the threshold agreed at launch, tighten approvals or return the step to human handling while the team investigates the cause.
Agree on a review window before the pilot starts, and move the budget if the baseline hasn’t shifted by the end of it.
When the process baseline needs to be rebuilt
System changes reset the map faster than drift does. Brandon’s team was about eight months into its first use case when a merger restarted governance, because the bank it was combining with had planned the same use case on different systems. Name an owner for the process map and re-run discovery after any consolidation.
Process data changes what your AI intake team is for
Use elapsed time, rework, and customer exposure to decide which workflows deserve attention first. Then let variability, reversibility, and the need for human judgment or cross-system action settle whether the answer is a flow, a copilot, an agent, or a redesigned process.
Applying that model falls to whoever runs AI intake, and Brandon’s read on the centers of excellence he meets is that most exist to keep the business happy. With process data in hand, the same group can tell a sponsor their idea is sound and show why a different process will return more.
The question for your next planning cycle is whether your AI intake can decline a request with evidence, and who owns the data that lets it.
If you’re deciding which workflows need a flow, a copilot, or an agent, Simform’s Low-code AI & Agents team can assess your current processes and prioritize work on Power Platform and Copilot Studio by feasibility, effort, and impact.
Brandon has led data and AI work across banking and aviation, and on ECAF Voices he explains how his team moved a regulated institution from survey guesses to measured, funded AI use cases.
