Choosing Your First AI Workflow: A Scored Decision Tree

The first workflow you automate teaches the second one how to fail, or how to hold. Pick the path that is closest to a deal, that already breaks often enough to see, and that runs often enough for a fix to show up this month. Pick the easiest tool to connect and you will automate a path that does not move pipeline, then wonder why the spend looks like overhead.

BCG's January 2025 AI Radar (n=1,803 C-suite executives) found that firms concentrating on 3.5 use cases generated 2.1 times the return of firms spread across 6.1. The mechanism is compounding. A focused deployment produces cleaner data and a tighter feedback loop for the next decision. A scattered one produces noise. For a four-person marketing team, 3.5 use cases means one workflow finished and half of a second started. The first pick sets the trajectory. The wider frame is the business process automation guide.

Why does the first workflow matter more than the fifteenth?

By the fifteenth, the team knows which fields are reliable and which are blank 30% of the time. The CRM admin knows the webhook format. The first workflow has none of that. You are building the pattern, not following it. That is why a low-friction tool connection is a weak reason to go first, and a high-impact path with a named owner is a strong one.

A weak first pick does more than fail on its own. It teaches the wrong lesson about which fields can be trusted, it writes a handoff the next workflow copies, and it gives the team a reason to decide that automation as a category was the mistake. Teams that pick a path with no owner and no visible failure often do not pick a second path. The fifteenth workflow is easier because fourteen earlier ones paid for the pattern. The first one is where that pattern is either worth copying or worth abandoning.

Focus compounds, which is what 3.5 versus 6.1 is describing.

BCG's January 2025 AI Radar (n=1,803) found that firms spread across more use cases, 6.1 on average, generated less return than firms that concentrated on 3.5, and the concentrated firms generated 2.1 times the return. The mechanism is compounding. A focused deployment produces cleaner data, a tighter feedback loop, and more confidence in the next decision. A scattered deployment produces noisy signals across too many fronts, no pattern worth generalizing, and a team that reports the tools did not work.

For a four-person marketing team, 3.5 use cases is not a portfolio. It is one workflow finished, and half of a second one started, with the same people. Focus means you name the path where a breakdown costs a deal you can point at, or a count of booked meetings you can see each week. You build that path correctly. You write down what the fields actually did. Then you apply that to the next path before you start it. Starting a second build in the same week spends the focus the finding is about.

What four criteria should you score?

Each criterion scores 1 to 3. The maximum is 12. The highest total is the first build, provided it clears the readiness gate below. The four are revenue proximity, error rate under the manual process, volume, and whether the data each step needs is actually stable.

Revenue proximity asks how fast a deal stalls if the path stops.

Revenue proximity is how directly the workflow opens, advances, or closes a deal. Score 3 if it produces or advances a booked meeting, a proposal, or a contract. Score 2 if it produces a qualified lead that a person then books. Score 1 if it produces a marketing metric, such as views or downloads, that reaches pipeline only after several steps. Teams underweight this because volume is easier to count. A high-volume email sequence feels important. An inbound form-to-text handoff that runs 50 times a week and books meetings scores 3. The sequence scores 2 at best. The question is: if this stops today, how fast does a deal stall?

The stall question is the one to ask out loud in the scoring meeting, because proximity is easier to inflate than the other three. A dashboard that leadership looks at every Monday feels close to revenue and usually scores 1, because a stopped dashboard does not stop a deal this week. A form that books a meeting scores 3 even when it runs less often than the newsletter, because a stopped form stops the meeting. Write the score next to the sentence that justifies it, so a 3 has to name the meeting, the proposal, or the contract it moves.

Error rate is how often the manual path fails in a way you can count.

Error rate is how often the manual version fails or needs rework. Score 3 above 20% (missed handoffs, blank CRM fields, wrong order, contacts that fall out). Score 2 for 10 to 20%. Score 1 under 10%. If you have not measured it, pull 30 days and count how many contacts entered and did not finish as designed. An exit at step one is a silent failure. The detection method is in the four workflow error types.

Do not score the error rate from memory. A path everyone describes as "pretty reliable" is often the one with a silent exit at the first step, because the failures never reach the person who would complain. Count entries against finishes for 30 days. A contact who is created and never assigned is a failure even if no one filed a complaint. If you cannot produce the count, the score is not a 1. The score is unknown, and an unknown error rate does not get to win the matrix on the strength of a guess.

How do you score volume and data readiness?

Volume and data readiness finish the matrix. With proximity and error rate, they put every candidate on a number from 4 to 12. The higher total is the stronger case for going first, once the readiness gate is passed.

Volume is how many times the path runs in a week.

Volume is how many times the workflow runs a week. Score 3 at 50 or more. Score 2 from 10 to 49. Score 1 under 10. A low-volume path can still matter. Orchestration return compounds with repetition. A workflow that runs 200 times a week with a 15% manual error rate produces 30 failures a week. Removing those failures is visible. Removing three failures a month is not. That 200-times example is illustrative, not a client result.

The reason to score volume is the speed of the lesson, not the importance of the work. A path that runs three times a week and fails 15% of the time produces about one failure every two weeks. You will not know whether the fix worked until the season changes. A path that runs 200 times a week shows the same failure rate inside a few days, which is soon enough to correct the build while the team still remembers why they started. A workflow that runs three times a week needs many weeks to accumulate the failure count a daily path accumulates before Friday. Score the cadence you actually have, not the cadence you hope the campaign will reach.

Data stability is whether the fields exist and stay put.

Data stability is whether each step's inputs exist and stay put. Score 3 if every field the workflow reads is required on the form or the enrichment job, present on 95% or more of contacts, and unchanged for at least 90 days. Score 2 if fields exist but are blank on 10 to 30% of contacts, or a field was added recently and older contacts lack it. Score 1 if you are guessing. Before you score it, check the last 30 days: is first name on 95% or more, is email required on the form, do lifecycle stage and lead source fill themselves, and were any custom fields added after older contacts were created? A no on any of those pulls the score down.

A no on that checklist is a specific score, not a vague warning. First name missing on more than a small share of contacts means the greeting will break as soon as you automate it. Email that is not required on the form means the path's key can be absent. Lifecycle stage and lead source that a person fills by hand will be blank on the weeks that person is out. A custom field added last month is missing on every contact created before that, so a branch that reads it will mis-sort the older half of the list. Any one of those nos moves the score toward 1 or 2. What each handoff has to pass to the next step is in the workflow data contract.

How do you read the matrix?

List every marketing workflow you run or plan to run. For a growing team that is often 8 to 20 items: form routing, enrichment, follow-up sequences, demo follow-up, trial triggers, churn alerts, content attribution, and alerts to sales. Mapping one path in 60 minutes is how you get the inventory before you score it.

Score every path on all four criteria and total them. The highest total is the first candidate. If two totals match, use the tiebreak below. A table with the workflow name and the four scores, ranked high to low, is the decision. Do not build the top two at once unless a dedicated person owns each.

When two paths tie, pick the one you can turn off.

When two workflows tie, pick the one you can undo. Reversibility is the tiebreak. Ask which path, if it misbehaves in the first month, you can switch off and hand back to the manual steps without rebuilding those steps from scratch. A path with a clean off switch in the automation tool ranks ahead of a path whose rollback means recreating a person's whole afternoon. The tiebreak is not "which one is more interesting." It is which failure you can survive in week two.

A 12 of 12 is a meeting path that already fails in public.

A hypothetical form-to-text handoff that scores 12 has proximity 3 because it books appointments, error rate 3 because phone-only submissions silently miss an email-only path, volume 3, and stable fields for months. That shape is the obvious first build. It is an illustration of the scoring, not a client result. It sits close to a deal, it fails often enough that you will see the fix, it runs often enough that the fix shows up this month, and the inputs have been the same fields for long enough that you are not automating a moving target.

Three categories usually sort themselves once you score them.

Across B2B and small-business teams, inbound routing (form to CRM to sequence) typically scores 9 to 12. Proximity is 3 because someone needs a meeting. The error rate is often 3 because a person copies the form, assigns it, and starts the sequence, and each of those slips when the week is busy. Volume scales with inbound traffic. Data stability stays high when the form requires the fields and the mapping is written down. The usual weak point is the form-to-CRM map: when the CRM field names differ from the form field names, the stability score drops to 2 until a mapping layer exists. Those bands are an operating frame, not a client result.

Content attribution typically scores 6 to 9. Wiring source parameters onto every published link scores high on volume, because every page generates tracked visits, and lower on proximity, because the data informs a decision two or three steps away from the deal. Build it second, not first. Clean source data is how you later tell whether the first path improved pipeline. Without it, a gain in meetings can be a gain in traffic, and you will not know which.

Long follow-up sequences typically score 4 to 7: proximity is 1 or 2, volume is low if the active list is under 2,000, and the revenue is far enough away that a fix does not show up soon. The error rate can still be high. The distance to a deal is what keeps the total down. Run those after routing and attribution are clean. Long follow-up is a nurture sequence, and it scores lower because the revenue sits far from the step you would automate.

What is the readiness gate, and why does it come first?

A perfect 12 cannot go first if the gate fails. PwC's May 2025 CMO Pulse found that 63% of CMOs miss opportunities because they cannot decide fast enough, and the top barrier is unclear ownership plus limited access to data and tools. The gate is three questions. It sits in front of the matrix so a high score cannot talk you into building a path nobody owns.

Three questions have to be yes before you wire a trigger.

Does a named person own the workflow, with a written output metric? Automation without an owner produces undetected failures for months, because nobody is watching contacts in versus contacts finished. A path without that person should not be automated yet. The failures will be real, and no one will be assigned to see the gap between entries and finishes.

Does the source system have a documented field map of what it sends and in what format? Building against an undocumented source means you are betting on field names. The form, the CRM, or the ad platform has to say which fields leave it and what shape they take. If that page does not exist, the data-stability score is a guess, and a guess should not be a 3.

Is there a written rollback if the orchestration fails in the first 30 days? Most platforms make rollback technically easy. It still has to be written before the build, not invented under pressure. The page names who turns the path off, what the manual steps are in the meantime, and how you can tell the failure from a quiet week.

A no on a gate question is the start of the build, not a reason to drop the path.

A no on ownership is 3 to 5 days of alignment. A no on the field map is 1 to 2 days with the CRM admin. A no on rollback is 2 to 4 hours of writing. None of those are reasons to drop a high-scoring candidate. They are the first steps of the build. The gate does not remove a 12 from the list. It names the work that has to finish before you connect a trigger. Do that work, then build. Skipping it is how a high score becomes an unowned path that fails in silence.

What does one scored week look like?

The bands above are an operating frame, not a client result. Here is how a team with four paths uses them in one sitting. Inbound form routing is the path where a person copies the submission into the CRM, assigns an owner, and starts a follow-up. It books meetings, so proximity is 3. In this illustration a quarter of submissions sat unassigned overnight, so the error rate is 3. It runs well over 50 times a week, so volume is 3. First name, work email, and company size are required on the form and have been for a year, so data stability is 3. The total is 12. That is the first build if the gate holds.

Demo follow-up is the contrast. A salesperson is supposed to send a recap the same day. Proximity is 3 because the recap advances a meeting that already exists. The error rate is 2, because the recap goes out and the next-step date is often blank. Volume is 2, because demos land in the teens most weeks rather than the fifties. The fields exist, and a few of them were added last quarter, so older contacts are missing them and stability is 2. The total is 9. It is a real candidate, and it waits until routing is finished. The first build is where the team learns the webhook, the field names, and who watches the failures. Starting with the second-best path skips that lesson.

A monthly content report scores lower on purpose. It produces a marketing metric, so proximity is 1. Someone rebuilds it by hand and the numbers usually match, so the error rate is 1. It runs once, so volume is 1. The source fields have been stable, so data stability can still be 3, and the total is 6. Automating it first would teach the team how to connect a tool. It would not teach them how a deal moves. Write the four numbers in a sheet, sort them from high to low, and the argument about which tool to buy gets shorter, because the sheet already answered it.

What does the first workflow teach you about the second?

The first workflow is a calibration, not only a time saving. Run it for four weeks after it goes live and write down four things: how many runs finished end to end, where the failures happened, whether the fields held up on real contacts, and whether the proximity score was right. If you scored it a 3 because it books meetings, the four-week note should say whether meetings were actually booked, or whether the path stopped one step earlier than you thought.

Four observations from the first build rescore everything you have not built.

Four observations from the first build update the scores for everything left. If the first one worked because the data was unusually stable, the next candidate needs the same stability or a cleanup before you build. If the CRM does not reliably fill lead source, every candidate that branches on lead source drops. A candidate that was a 3 on data stability becomes a 1 until that field is fixed at the source, and it should not stay at the top of the sheet out of politeness.

PwC's AI Agent Survey (22 to 28 April 2025, n=308 US executives) found that 55% of adopters report faster decisions and 66% report higher productivity. That speed accumulates across sequential deployments, not inside a single one. Each finished workflow teaches you how to build the next with fewer false assumptions. The second path should be shorter to build than the first, because the webhook, the field names, and the person who watches failures are no longer new. If the second path is just as surprising as the first, the four-week note was not written, and you are about to repeat the same false assumption on a different form.

Do not route to a nurture sequence while a meeting-booking path is still manual. If you want the scores run against the paths you actually have, get your free Systems Plan. We will start with the readiness gate, then the matrix, and we will not recommend a fifteenth workflow while the first one is still unnamed.

How we know this.

BCG, January 2025, n=1,803, is the source for 3.5 versus 6.1 use cases and the 2.1 times return. Firms spread across more use cases, at 6.1, are the comparison group in that finding. PwC's May 2025 CMO Pulse is the source for 63% of CMOs missing opportunities because decisions are slow, and for unclear ownership plus limited data and tools as the top barrier. PwC's AI Agent Survey, 22 to 28 April 2025, n=308, is the source for 55% faster decisions and 66% productivity. The 1-to-3 scores, the 9 to 12, 6 to 9, and 4 to 7 category bands, the gate-timing ranges, and the 200-runs-at-15% example are an operating frame, not a client result.