Most AI roadmaps fail the same way — a list of interesting projects with no way to choose which one matters. Ours is a five-stage loop: fast enough to run in a day, rigorous enough to defend to the CFO. Here's exactly how it works.
The floor knows more than the systems. We sit with the people who do the work, understand the operation well enough to rebuild it for AI — not just point AI at the way it runs today — and route that knowledge to the patterns that fit.
STAGE 1 / 5
What do you repeat? What frustrates you? What runs on a schedule?
The floor knows more than the systems. We start by surfacing the widest possible inventory of work — the tedium, the frustrations, the things that fire on a clock — before we narrow anything. Individual thinking first, group debate second.
A wide inventory of named workflows across daily, weekly, and monthly cadence.
STAGE 2 / 5
How does this workflow actually happen — including the workarounds?
For each nominated workflow we walk it end to end as if onboarding a brand-new hire — sit in the seat, follow the process where it actually happens, and keep asking why; don't skip the workarounds. Where does it stall? Where does it fail? Where do you redo work? The breakpoints are where the value is.
An end-to-end canvas per workflow: trigger, inputs, steps, outputs, systems, breakpoints.
STAGE 3 / 5
What shape is this problem — can you write the rule, or does it need judgment?
Classification before solutioning is the discipline that prevents over-building an agent for a problem a script would solve more cheaply. We pick by problem shape, not sophistication. Sub-workflows often classify differently from the parent — we break them out and read each piece on its own.
Each workflow read for its problem shape — rule-writable, data-driven, integration-bound, judgment-heavy — before anyone proposes a solution.
STAGE 4 / 5
Which patterns does this actually need — and how do they stack?
Most pain we see has been seen before, in a known shape. We match each workflow to the pattern — or, more often, the stack of four to six patterns — it needs: script, machine learning, visual workflow, code-first graph, skills, orchestration, harness, retrieval. A pattern library replaces re-invention, with a first read on effort and value.
A pattern-matched proposal per workflow, drawn from the eight patterns of automation, with rough effort and value estimates.
STAGE 5 / 5
What ships first? Who owns it? How do we know it worked?
Every nominee is scored on ROI, confidence, speed, fit, and risk, then sorted. We pick the Day-1 pilot, scope it, and put names on the plan. A portfolio view — not a single project — is what keeps the second pilot from disappearing.
A ranked portfolio and a 30/60/90 plan with named owners and a metric per item.
Anyone can list AI ideas. The method is what makes the list defensible — and what makes the second pilot happen.
We start with opportunities you can name and feel — not a generic list of industry use cases.
Naming the problem shape first prevents over-building an agent where a script belongs — and over-spending on tokens where logic would do.
Pointing AI at today's inefficient process only makes it costly to run. The bigger win is rebuilding the operation so the system runs lean — an operations call as much as an AI one.
Most pain has been seen before in a known shape. We match it to the eight patterns instead of inventing from scratch.
We work the friction alongside the people who own it. A fix the team co-designs is one that survives contact with Monday.
The 30/60/90 view keeps momentum after the first win, so the roadmap doesn't stall at one.
Five questions open the discovery blocks. They surface the workflows where AI pays — so machines can do that work, and your people can act like people.
Same task, same way, by multiple people, on a predictable cycle.
“We do this every day / every week.” · “It's just part of the job.” · “Everyone on the team does the same thing.”
High-volume records people create as part of the job that feed other processes.
“We generate data but never have time to use it.” · “People spend half their day documenting.” · “The answers are in the data somewhere.”
Humans acting as middleware between raw data and a decision or report.
“We have a team that just checks the data.” · “It takes days to get a clean dataset.” · “Can't trust the system without manual review.”
People logging into multiple systems to copy, reformat, or re-enter information.
“We log into 3 systems to do one task.” · “Someone's full-time job is copy-paste.” · “We built a spreadsheet to bridge two systems.”
High-cost talent spending most of its time finding context, not applying expertise.
“Our best people spend 60% of their time finding info.” · “By the time we have the picture, the window's passed.”
The room tells you where the value is in the language it already uses. These are the phrases that flag an opportunity hiding in plain sight.
“Can't do that until Monday.” · “Always behind on that.” · “The data's stale by the time we get it.”
THE OPPORTUNITY
Batch work that could be near-real-time; reporting latency AI collapses to live dashboards.
“It depends on who does it.” · “We catch ~80%.” · “We redo ~15%.”
THE OPPORTUNITY
Inconsistent judgment → classification; a known miss rate → an AI second pass at the point of creation.
“Audit prep takes us offline for weeks.” · “We just got a finding.” · “Policy changed; process hasn't.”
THE OPPORTUNITY
Auto-generated audit trails; continuous compliance posture; an active finding is the fastest-funded engagement.
“Worked at 10, now we have 10,000.” · “We just throw more people at it.” · “Backlog is months long.”
THE OPPORTUNITY
A process that didn't scale → AI absorbs volume without linear headcount; triage to drain the backlog.
“Only two people know how.” · “When she retires, we're in trouble.” · “Everyone asks [name].”
THE OPPORTUNITY
Single points of failure → capture and distribute knowledge; augment juniors to senior-level routine decisions.
“Meeting about the last meeting.” · “Same update to 5 stakeholders in 5 formats.”
THE OPPORTUNITY
Coordination overhead AI summarizes and tracks; one source of truth, many role-personalized views.
Scoring is a group exercise after the matching block. We score every workflow on the same five dimensions, sort, and flag disagreements explicitly — the argument is where the real prioritization happens. The metric we commit to is always a number, never a sentiment. “Faster” isn't a metric; “under 24-hour resolution” is.
What's the dollar value of the time, errors, or risk this removes?
How sure are we the approach works in your environment?
If we shipped this in 30 days, what changes on Monday morning?
Can your team own this after handoff? If not, what's missing?
What's the worst-case failure mode — and how much does it cost?
The method is only as good as the workflows you put through it — and those are yours. Tell us your industry and who'll be in the room, and we'll tailor the day before we arrive.