Most AI projects stall. And it’s not the model’s fault.

“Automate everything. Save 80%. Ship next week.”

The deck looks the same in every meeting. An LLM, a workflow tool, a Zapier diagram with arrows, a pilot inside three weeks. The build gets quoted — three weeks of integration, a workshop, a first agent in production, a line on the invoice. What doesn’t get quoted is everything that decides whether the thing survives.

• i.

  The prompt that breaks on the next model upgrade.

  Nobody owns the regression. It worked in the demo; it drifts in production.

• ii.

  The schema change upstream that silently corrupts the output.

  No eval catches it because no one was assigned to run evals.

• iii.

  The vendor who churns and takes their fork of LangChain with them.

  The automation works exactly as long as the vendor relationship does.

• iv.

  The exceptions the demo never showed.

  The 12% of cases that need a human, the API that changes every quarter, the compliance review that didn’t exist when the pilot started.

Three patterns we see every quarter.

The same three failures, across every sector. None of them are technical — they’re shortcuts taken before the work began.

• i. Automating the wrong process

  The team mapped the formal flow, not the real one. The agent learned the slide-deck version — not the seven informal handoffs that actually move the work. Three months in, the operators are doing more work, not less: they correct the agent and run the old process in parallel.

• ii. Tool-first, problem-second

  Someone chose the LLM, the orchestrator, and the vector DB before anyone defined the measurable outcome. The pilot ships on time. Nobody can answer what it’s supposed to move — revenue? cycle time? error rate? — or by how much. After 90 days the project is “running” but can’t survive a single hard question in the budget review.

• iii. No operator inside the loop

  The build was outsourced. Nobody on the client team can change a prompt, retrain an eval, or debug a broken integration without re-hiring the vendor. The automation works exactly as long as the vendor relationship does — and dies the day it ends.

Seven moves. Same loop. No shortcuts.

Six of these moves are not technical. The model only gets useful at move seven — and only because the first six were done right.

• i. Listen

  Discover the real process in the operator’s own words. Watch a Friday afternoon, not just a Monday standup. Find where the work is silently re-done.

• ii. Quantify

  Make the cost of the current process tangible — hours lost, errors caught downstream, deals delayed. If you can’t name a measurable number, you haven’t finished this move.

• iii. Vision

  Co-create the future state with the team that lives it. Describe success in their language. What gets said on a leadership call when this works?

• iv. Targets

  Agree the measurable outcomes that prove the future state happened. Time-bound. Their KPIs, not ours. The number AI is supposed to move — before a single agent is written.

• v. Map

  Identify what actually needs to be built — in the buyer’s shopping list, not a vendor feature matrix. Half of what gets mapped is process change. Some is software. Almost none is a brand-new model.

• vi. Build

  Execute alongside the team. Shape Up cycles — 6-week appetites, bets, ship dates. Senior operators in the room, no juniors, no subcontracting. The team that ships is the team that learns.

• vii. Automate + Prove

  Leave automation behind, not slides. Workflows, agents, dashboards, integrations live on after we leave — measured against the Targets from Move iv. Maintenance is owned, evals are running, the on-call rotation knows what to do.

Want the full mechanics — phases, the pod, the engagement? The Advisory page is the deeper read.

Twelve weeks. First process live by week six.

Each phase ships a concrete artifact that gets used the next day — not a deliverable that gets filed. The work runs against measured Targets agreed by leadership before any agent is written.

• Wk 02

  First Playback. We recite your real process — the one operators actually run — back to you, in your words, and you nod. No automation yet. We’ve listened.

• Wk 04

  Quantified Targets. The cost of waiting is a number you can defend to a CFO. The measurable Targets the framework is supposed to move are signed off by leadership.

• Wk 06

  First process live. One process automated end-to-end and running in production, measured against the week-four Targets. Not a demo — real work, real traffic. This is the six-week promise.

• Wk 09

  Second + third process. Two more shipped on the same loop. Evals running. The team is driving the next one without us in the room.

• Wk 12

  Handoff. Dashboards bound to the Targets, an on-call playbook your team owns, a maintenance plan that doesn’t require us. We leave. Things keep running.

“First process live by week six” is the line we hold ourselves to. If we’re not on track by week four, we say it then — not at the handoff.

Working infrastructure. Not a binder.

Most engagements leave a 60-slide deck, a trained team whose champion will quit in 18 months, a vendor dependency for the next prompt change, and a renewal conversation. The deliverable here is different — it’s the actual work that runs the next day, and keeps running after we’re gone.

• i. Workflows in production

  Running against real traffic, not a sandbox.

• ii. Agents wired to your real signals

  Pointed at the steps we could measure — and only those.

• iii. Dashboards bound to your Targets

  The week-four numbers, tracked where leadership can see them.

• iv. Evals + on-call your team owns

  A maintenance plan you don’t have to renew us to keep alive.

Six companies. The same loop. A different sector each time.

Not vendor case studies. Our founder ran the framework inside these companies as an operator — Head of Engineering, CTO, Head of Innovation, turnaround lead — before it had a name.

• Picap

  LATAM mobility, five countries. Move 07 done right: rebuilt deploy, observability and on-call, then layered ML on routing. 99.9% uptime, 10× volume on the same headcount.
  Read the full case →

• Fluyenta

  B2B enterprise revenue. Moves 02 + 05: rebuilt discovery around procurement reality. 3× deal size, −40% cycle. The motion outlived the consultants — still in production.
  Read the full case →

• RunMyProcess

  Enterprise process automation, EMEA + LATAM. Moves 04 + 06: Targets signed before any workshop, then Shape Up to ship the motion. 2× quota, ramp time halved, two continents on one playbook.
  Read the full case →

• ROEPA

  France + Spain, small ops team. Moves 01 + 07: mapped the real flow, automated only what we could measure. Stable in maintenance at ~8 hrs/month — first Nexus AI-seller target.
  Read the full case →

• LexPro

  Colombian legal tech (portfolio). Move 02 done right: firms named their own cost of waiting. Pricing validated at 10× the original number — set by the buyer, not the seller.

• Amplifica

  Marketing tech (portfolio). Moves 04 + 07: Targets defined per brand before the model scored anything. Live at amplifica.studio — output stays inside the brand’s rails because the framework drew them first.

Each engagement above is a single headline. The full ledger — eighteen distinct broken processes, each tagged by the move that fixed it — lives at The framework, in action →

The full operating portfolio — Lumina, BrainzLab, Sondea, Bite — runs on the same loop. See the portfolio →