How to pick your first agentic workflow (a scoring method that prevents regret)

Gartner reckons over 40% of agentic AI projects will be cancelled by the end of 2027: escalating costs, unclear business value, inadequate risk controls. MIT’s NANDA group found 95% of enterprise GenAI pilots deliver no measurable P&L impact. Those are grim numbers, and the temptation is to read them as “agents don’t work.”

They do work. We run a four-market CRM business on a fleet of them: finance close, customer onboarding, support triage, marketing production. The failures in those studies are mostly not engineering failures. They’re selection failures. Someone picked the wrong first workflow, it went sideways in public, and the whole programme got quietly shelved.

So before you build anything, score your candidates. Here’s the method we use with clients, and on ourselves.

The five factors

Take every workflow you’re considering and score it 1–5 on each of these. Be honest: the scorecard only works if you resist the urge to bump the sexy option.

1. Hours consumed weekly. How much human time does this workflow actually eat, every week, today? A 5 means someone loses most of a day to it; a 1 means it’s an occasional annoyance. If you don’t know, that’s your first finding: go measure for a fortnight before scoring anything.

2. Rule-clarity. Could you write the workflow down as instructions a competent new hire could follow without asking questions? Fully codifiable is a 5. “It depends, you develop a feel for it” is a 1 or 2. Agents are excellent at following written judgement and terrible at inventing unwritten judgement.

3. Blast radius if it goes wrong, inverted. What’s the worst plausible outcome of the agent making a mistake before anyone notices? Score it backwards: a 5 means the failure is cheap, internal and reversible; a 1 means a customer, regulator or your bank balance feels it directly. This inversion matters, because the workflows that feel most exciting usually score lowest here.

4. Input/output cleanliness. Does the workflow run on structured data that already lives somewhere queryable: a CRM, a ledger, a ticketing system, a spreadsheet with stable columns? Structured in, structured out is a 5. “It starts with whatever’s in Karen’s inbox” is a 2 at best. Agents inherit the hygiene of their inputs.

5. Team appetite. Is there a named human who wants this automated and would champion it: review the output, report the wins, defend it through the awkward early weeks? A genuine volunteer is a 5. “Management thinks the team should want this” is a 2, and it will fail the way most mandated tools fail: quietly, by being ignored.

Add the five scores. 18 or above out of 25 is a good first candidate. Below 15, park it. It might be a fine third or fourth workflow once you’ve built the supervision habits, but it will hurt you as a first one.

A worked example: reporting vs chat

Here’s the comparison that comes up in almost every scoping conversation we have. The business has two candidate workflows: assembling the weekly performance report, and a customer-facing chat assistant. Chat is the one the board has heard of. Let’s score both.

Weekly reporting assembly. Someone (usually someone senior, which makes it worse) spends three to four hours every Monday pulling numbers from the POS, the booking system and the accounting platform, pasting them into a deck, and writing the same six paragraphs of commentary with different numbers.

• Hours consumed weekly: 3–4 hours of a senior person, every single week. 4
• Rule-clarity: the report has looked the same for two years; the commentary follows a known pattern (“flag anything ±10% on last month, note one-offs”). 5
• Blast radius (inverted): the output goes to internal people who read it critically and would catch a wrong number before it mattered. Worst case is a corrected email. 5
• Input/output cleanliness: every input is already structured data behind an API or an export button. 4
• Team appetite: the person who builds it every Monday would pay money to stop. 4

Total: 22. Build it.

Customer-facing chat. Handles enquiries on the website: opening hours, bookings, complaints, the lot.

• Hours consumed weekly: genuinely high, enquiries do eat staff time. 4
• Rule-clarity: half the conversations are codifiable; the other half are complaints, edge cases and people who are already annoyed. Nobody has written down how to handle those because the answer is “it depends.” 2
• Blast radius (inverted): mistakes happen in front of customers, in writing, in real time, sometimes screenshotted. 1
• Input/output cleanliness: free-text questions in, free-text answers out, grounded in a knowledge base that probably doesn’t exist yet. 2
• Team appetite: mixed. The front-of-house team half suspects it’s there to replace them, which is not a champion’s mindset. 3

Total: 12. Not a first project. Possibly a good fourth one, after the knowledge base exists and the team trusts the supervision model.

Notice what the scorecard did there. Both workflows have real hours behind them. Factor one alone can’t separate them. It’s blast radius and rule-clarity that make the call, and those are exactly the factors enthusiasm skips over.

The anti-pattern: starting with chat because it demos well

We understand the pull. A chatbot on the website is visible. You can show the board. It feels like “doing AI” in a way that a report assembler never will.

But demoing well and deploying well are close to opposites. A demo is a two-minute happy path; deployment is every path, including the customer who types in all caps at 11pm. When the visible project fails visibly, you don’t just lose that project. You lose the political capital for the next three. That’s how organisations end up in the cancelled-project statistics: not because agents couldn’t help them, but because the first impression was a public faceplant.

The boring-but-scored-18 workflow does the opposite. It quietly returns four hours a week to a senior person, it builds your team’s supervision muscles where mistakes are cheap, and it gives you a real, internal proof point (“we’ve run this for eight weeks, here’s the error log, here’s the hours saved”) that makes the second project an easy yes.

We followed our own advice (eventually on purpose)

Our first agent at nollie wasn’t customer-facing, and that was deliberate. It assembled internal research digests: market scans, competitor movements, the reading one of us used to do at 6am. Codifiable, structured-ish inputs, zero customer exposure, and one founder who badly wanted the hours back.

It was also wrong sometimes, early on. It over-summarised, it occasionally missed the one item that mattered, and we caught those failures in private, at the cost of a slightly thin digest rather than an embarrassed customer. Those cheap failures taught us how to review agent output, how to write better instructions, and how to know when a workflow is ready to run with less supervision. Every agent we’ve shipped since (including the ones that now touch customers) runs on habits we built where the stakes were low.

There’s evidence this sequencing matters beyond our anecdote. The same MIT NANDA study with the bleak 95% figure found that organisations buying from specialised external partners succeeded about 67% of the time, versus roughly 33% for internal builds. The common thread in the successes wasn’t smarter models: it was disciplined scoping. The scorecard is that discipline, written down.

Do this next week

Get the three or four people who know your operations best in a room for 45 minutes. List ten workflows that annoy someone weekly. Score each one against the five factors (hours, rule-clarity, inverted blast radius, input cleanliness, appetite) and total them.

Anything at 18 or above goes on the shortlist. Anything below 15 goes on a “later, once we’re good at this” list, no matter how impressive it would look in a demo. Then take the top scorer and measure exactly how many hours it consumed this week, so that in two months you can prove what changed.

If nothing scores 18, that’s a useful result too: it usually means your data is messier than you thought, and cleaning one system is your real first project.