When we talk clients out of AI

Roughly one prospect in five leaves our first call with a recommendation not to use AI at all. That is not a sales tactic gone wrong; it is the cheapest advice we give and, over time, the most profitable. Below are three real engagements — anonymised, with the clients’ permission — where the right answer was a rules engine, a spreadsheet, and a firm no. Then the checklist we run before we ever propose a model.

The returns desk that wanted a model

A homeware retailer asked for “AI to decide return approvals”. Before proposing anything, we mapped six months of their actual decisions — about 2,300 of them. Fourteen deterministic rules (days since purchase, item condition category, customer history, receipt present) reproduced 96.4% of the human outcomes. The remaining 3.6% were genuinely judgment calls that would have needed human review under any system, model included.

So we shipped a rules engine. Nine days of work, run cost near zero, every decision explainable in one sentence — which matters enormously when a customer appeals. A model would have cost more to build, more to run, and produced decisions the retailer could not defend at the counter. The rule set is printed on two pages and the operations manager owns it, edits it, and does not need us to change it.

The commission calculation that wanted an agent

An insurance brokerage asked for an “AI agent” to run their monthly commission close, which took three days. One look at the workbook explained the three days: eleven years old, forty-one tabs, broken links, and three cells nobody dared touch. There was no ambiguity anywhere in the task — every commission rule was written down and exact. Ambiguity is what models are for; this was decay.

We rebuilt the workbook in a week: typed inputs, validation on entry, one protected calculation sheet, version history. Monthly close now takes half a day. Run cost: zero. It is the least impressive project on our books and one of the highest-return things we have ever shipped. Nobody writes conference talks about spreadsheet remediation, which is convenient for us, because the demand is endless.

The hiring screen we refused

A growing firm asked us to rank 300 job applicants with a model. We declined the model outright, for three reasons we put in writing. The training signal would have been their past hiring decisions, which bakes yesterday’s bias into tomorrow’s shortlist with a straight face. The errors are invisible — you never meet the good candidate the system filtered out, so the system never looks wrong. And the regulatory exposure around automated employment decisions is real and growing. We built them a structured scoring rubric instead — same screening time saved, every score traceable to a human’s judgment on a stated criterion. Some decisions should stay expensive.

The checklist we run before proposing a model

1. Can a correct answer be verified cheaply? If checking the output costs as much as producing it, automation gains you little.
2. What does one error cost, and who catches it? Misrouted ticket: minutes. Wrong commission payment: a resignation.
3. Do written rules cover 90%+ of cases? Write them and count. If yes, ship the rules — you can always add a model for the remainder later.
4. Is there enough real data to evaluate — not train, evaluate? No eval set, no deployment. This kills more proposals than any other line.
5. What is the monthly run cost at real volume? Including the human review the vendor forgot to mention.
6. Will the people using it be able to override it, and will they trust it? A technically correct system nobody trusts is an expensive decoration.

What honesty does for retention

The 98% of clients who come back for a second project — the number on our home page — is mostly built in these first calls. The returns-desk retailer came back twice; the two later projects together were worth about nine times the model we declined to build them. Telling a prospect “you do not need what you came to buy” converts badly this quarter and compounds for years. We can afford that trade; agencies optimising for this quarter cannot, and their clients eventually notice.

What we'd tell you to do

• Before any AI conversation, spend a day mapping the decisions you actually make. Count how many a written rule could handle.
• Run the checklist above on your own project before a vendor does — especially the eval-set question.
• Ask every vendor: “what would you build here if you were not allowed to use a model?” If they have no answer, you have learned what they are selling — and it is not a solution to your problem.