01
Related
Where to read next on the same subject.
02
Short answer
Score every candidate workflow on five criteria before anything is built: how often the decision runs, whether the data path exists, who owns the result, how reversible the change is, and whether the outcome is measurable within the quarter. One workflow wins. The scoring session gives the operating team a shared basis for choosing the engagement.
I sat through a selection meeting once where eight candidates were listed on a whiteboard and the conversation went in circles for an hour. Everyone defended their own candidate. The problem was structural: nothing on the board could be compared, so the loudest stakeholder was winning. We drew five columns, scored each candidate in fifteen minutes, and the argument ended. The winning workflow was not the loudest one. It was invoice matching, because it ran daily, the data arrived already, one person owned the result, the recommendation was reversible, and the hours saved could be counted within six weeks.
03
The five criteria
Decision frequency matters because a system learns from repetitions. A decision that runs daily produces feedback every day, and the team gets dozens of improvement cycles in a quarter. A decision that runs quarterly gives four data points a year, and the model barely moves before the sponsor loses interest. When two candidates score similarly, pick the one that runs more often.
Data readiness decides the schedule. If the input arrives in a system with history and a stable definition, the first working slice can land in weeks. If the data arrives by hand, in spreadsheets, or in a source nobody has documented, the project begins with building the path, and that changes the budget conversation. Neither case kills a candidate, but the score must say which one you are buying.
Ownership decides survival. A workflow needs a named person who accepts the result, answers questions during the build, and operates the system after transfer. Committees do not survive contact with a production calendar. If no one will put their name next to the candidate, treat that as a rejection, whatever the other columns say.
Blast radius decides courage. A recommendation that a person reviews before acting can be wrong safely. An automated action that moves money or shuts a line cannot. Early systems earn their autonomy; they do not start with it. Candidates that require full automation on day one should be re-scoped to start with a review step, or deferred until the team has run something smaller.
Measurability decides the story you will tell at day 90. The result has to be countable in the operating team's own units: hours returned, exceptions cleared, error rate against a measured baseline. If the only honest answer is "the business will be better," the workflow cannot show progress, and unprovable systems get cancelled for reasons unrelated to their quality.
04
What rejection looks like
The reject column deserves as much rigor as the winner. Candidates fail for recognizable reasons. No owner accepts the result, and the system would land on a team already carrying more than it can hold. The input arrives by hand, and the project would spend its energy building a data path while calling itself an AI project. A result that only becomes measurable in a year outlasts the attention span of the sponsor. The process is already automated well, and the gain would come from re-platforming, not from intelligence. The sponsor wants a demo, and a demo is a fine thing to want, but it belongs in a different budget.
Writing the rejection reason next to each candidate matters more than the score. Six months later, when someone asks why the chatbot pilot was not first, the answer should be a sentence with a date, not a shrug.
05
The 90-day shape
Selection ends with a schedule, not a document. Days 0 to 15 record the baseline and confirm the data path or plan it. Days 15 to 45 deliver the first working slice, reviewed with the operating owner, ugly but real. Days 45 to 75 run the system in shadow beside the current process, with every difference explained in writing. Days 75 to 90 complete acceptance, cutover, and the transfer of ownership. The candidate that cannot fill this shape honestly is not ready, and the scoring session should say so.
The shadow period deserves emphasis. It is where the operating team builds trust, because they watch the system act beside their own process and see which differences are real. It is also where the delivery team learns the unwritten rules that no requirements document contained. Skipping the shadow run to save three weeks is the most common false economy in a first deployment.
06
Revisit the scoring, gently
Candidates change. The data path for the runner-up may complete during the first build, which promotes it for the next cycle. A new regulation can create a candidate nobody listed. The scoring table stays on the wall, and the team re-scores quarterly, in one hour, with the same five columns. Selection is a process that keeps running, not a meeting that happened once.
07
The result worth naming
The first workflow is a trust investment. When it lands, the second selection session runs faster, the data teams arrive with better questions, and the operating teams ask to be scored in. When it drifts, every later conversation carries the weight of the miss. The five criteria exist to put the odds on your side, and to make the reasoning visible before the first hour of build is spent.
08
Who does the scoring
The scoring session includes the operating team, and that inclusion is the point. A score produced by the delivery team alone measures convenience; a score produced with the planners, the warehouse lead, and the finance partner in the room measures reality. The five criteria get scored out loud, with the disagreements audible and the reasons captured next to the numbers.
Use a simple scale. Each criterion takes a yes, a no, or a named risk. Precision beyond that is theater at this stage: the purpose of the score is to rank candidates and expose the load-bearing uncertainty, not to simulate precision nobody can defend. When two candidates finish close, the tiebreak is the criteria in order: frequency first, because feedback cadence compounds; ownership second, because a workflow without an owner is a science project; the rest follow.
The session also assigns the shadow roles early. Who runs the baseline measurement, who provisions the data path, who signs the acceptance. Names go into the record before the meeting ends. I have watched selections evaporate because the same two people were assumed to own four workstreams and nobody asked them; the names in the record make that conversation impossible to skip.
One more habit from the selection table survives the engagement: the scoring notes. Months later, when the second workflow comes up for selection, the notes remind everyone which data path was completed, which owner proved responsive, and which "measurable this quarter" turned out to be fiction. The table gets better the second time it is used, which is the whole idea of keeping it.
09
Related
The evaluation framework guide shows how the shadow period is scored. The forward-deployed model describes who does the work across those 90 days.