top of page

Should We Build This With AI? Worksheet

S. West
1 hour ago
5 min read

Should We Build This With AI? Worksheet

The assistant launched on schedule. A quarter later, the month-end reconciliation still took two days, and now it had a model bill.

An operations group was asked to put AI into its monthly supplier reconciliation, a two-day job for one analyst. The build took a quarter. Afterwards the team found that most of the two days went on waiting for finance to export the ledger. The model did its part well; its part was never the slow one.

(Composite example, assembled from patterns common in organisations under pressure to show AI adoption. The figures are illustrative.)

The organisations getting returns from AI aren't using it everywhere. BCG's 2024 survey found that AI leaders pursue about half as many opportunities as their peers and invest in a few high-priority ones. This worksheet starts from the refusal end: it looks for reasons to stop before asking whether the idea could work, and ends with a written no that is hard to dismiss as resistance. Take one candidate workflow and work from top to bottom.

Step 1: screen against the five wrong-fit categories

Tick any statement that is true of the candidate. One tick is a hard stop.

  • ☐ The answer is computable. A known procedure gets it right every time: tax, card-number validation, an order lookup, routing, a string match. Write the code instead: it is quicker, cheaper and auditable.

  • ☐ The decision is high-stakes and hard to reverse. A wrong call is costly, hard to undo and often legally exposed: hiring, loan approval, large trades, legal advice, sentencing, or medical diagnosis without clinician sign-off. Let AI support the accountable person, who stays the decider. Removing the human is a common way these projects fail, and the usual repair is putting the human back in the loop. In the EU, the AI Act (2024) classes uses such as recruitment and credit scoring as high-risk and requires human oversight of them.

  • ☐ The volume is low. The manual workflow takes under roughly 100 hours per quarter across the organisation (ShiftQuality's rule of thumb; adjust for your costs). Costs are fixed; savings arrive one task at a time, and at low volume they never catch up. In our experience, this category catches more would-be projects than any other.

  • ☐ The bottleneck isn't intelligence. The time goes on data access, approvals, unclear requirements or waiting on other teams. Measure where the hours go first. If, say, 80% of the hours go on waiting, approvals or coordination, AI can improve at most the remaining 20%.

  • ☐ AI would be the actor in a reliability-critical system. Even 95% reliability is not acceptable: surgical control, air-traffic control, grid management, transaction settlement, dispensing medicine. A model can analyse and recommend here, but should rarely act. This category is narrower than it looks: most business workflows absorb some errors.

No ticks? Go on to Step 2.

Step 2: run the three diagnostics

You need three yeses before AI counts as a candidate.

#

Diagnostic

The question

Answer

1

Pattern fit

Does the work mean recognising patterns in messy input such as language, images or behaviour? Exact calculation, lookups and rule application fail, even if a model could do them.

☐ Yes ☐ No

2

Payback

Will lifetime savings clearly exceed the cost to build and run it?

☐ Yes ☐ No

3

Error tolerance

Can the workflow absorb an occasional wrong output? Model output is probabilistic.

☐ Yes ☐ No

Size diagnostic 2:

  • Lifetime savings = hours saved on each task × task count per period × number of periods: ______

  • Integration cost = engineering hours + ongoing operations + model spend: ______

  • Ratio (savings ÷ cost): ______. Our rule of thumb: aim for 5× or more, because hidden costs such as monitoring, evaluation and incident response eat thinner margins.

Place diagnostic 3:

Tolerates errors (candidate)

Doesn't tolerate errors (stop or narrow)

A draft someone reviews before it goes out

An autonomous action that changes real-world state unchecked

A summary the reader can check against the source

A decision that is hard to undo

A classification with a verification step after it

Output that reaches users with no review

A recommendation where a person makes the final call

______

Tally: ☐ Three yeses: go to Step 3. ☐ Any no: choose an alternative (better software, a process fix, clearer requirements) or narrow the scope and rerun Step 2.

Step 3: apply scope discipline before you commit

Confirm all three before funding it.

  • ☐ Portfolio level: this is one of a short, deliberate list of high-impact workflows. A few well-supported deployments usually return more than many thin ones.

  • ☐ Deployment level: the AI has the smallest scope that still delivers the value.

  • ☐ No creep: we've resisted "just one more capability", a common cause of reliability problems in production AI. Expansion means a new deployment.

Signs the portfolio already says yes too often:

  • ☐ Several teams run overlapping tools on duplicated infrastructure.

  • ☐ No deployment gets enough operational attention to be good.

  • ☐ Each project looked cheap, but the combined spend surprises finance.

  • ☐ Governance reviews are skipped, or are the queue everyone waits in.

  • ☐ Asked what the programme returned, the team offers usage numbers, not outcomes.

Step 4: send a reasoned no

A specific refusal with an alternative is harder to overrule than a vague objection.

  • Workflow: ______

  • Decision: not a fit for AI at this time.

  • Reason: ☐ not enough volume to pay back ☐ the bottleneck is data access or approvals, not intelligence ☐ the reliability needed is beyond what current models deliver ☐ other: ______

  • What we recommend instead: ______

  • Revisit date: ______ (a no this year might be a yes next year, so keep the list)

Refusals need a senior sponsor. Name one before the first goes out.

How to use it

  • Who: the lead who owns the proposal, with the person who runs the workflow today. In our experience, fifteen to thirty minutes per candidate.

  • Output: a screened candidate with diagnostic answers, a payback ratio and a scope statement, or a reasoned no with a revisit date.

  • Done when: the candidate has cleared Steps 1 to 3 with a named sponsor and a single-purpose scope, or has a written no, an alternative and a revisit date.

Where this changes

A small team weighing one tool can run Step 1 and the payback sum in one conversation. A large or regulated organisation needs more: a register of approved and refused candidates, formal risk classification, and documented human oversight where the AI Act or sector rules apply.

The argument behind this worksheet is in choosing what not to build with AI. Before a passing candidate launches, use the pilot-to-production readiness checklist for AI features. If the question becomes whether to make it yourselves, work through build, buy or AI.

Final takeaway

Many AI proposals should stop at Step 1 or Step 2, and that is the worksheet working. Refuse the wrong fits in writing, fund the few that pass, and keep each to a single job.

Sources

bottom of page