Logiciel Contact Us
Success Stories Tech News Contact Us
whitepaper

30% Of AI Projects Die After The Demo. Selection Kills Them.

Most AI backlogs are sorted by what is technically possible, which is uncorrelated with what any of it is worth. This framework is the selection method a CTO and a COO can run together in one working session, including the six dimensions and the answers that end a candidate's case on the spot.

In depth

Possibility Ranks The Backlog. Value Never Enters The Room.

01

The trap most technology groups walked into: ask engineering what AI could do, get an excellent list of technically possible things ranked by how interesting they are to build, then discover the backlog has three blank columns: who owns the outcome, what the process costs today, and which number is expected to move.

02

What the operators who reach production do: walk the value chain first, drawing candidates only from processes with high volume, high variance, high manual touch and a failure cost someone can quantify, then score each survivor on six dimensions and stop at the first disqualifying answer rather than averaging it away.

The detail

What Separates A Method From A Queue.

Zone · 01

Value Density

Not Hours

Two candidates can both save ten thousand hours a year and differ by two orders of magnitude in value. What separates them is money per unit of work: the loaded cost of one transaction, the revenue at risk in one decision, the penalty attached to one error. Nobody calculates it, so effort flows to cheap work.

Zone · 02

Change Load Counted As Cost

Count the people who must alter a daily habit for the value to arrive, and treat the count as a cost. Four people in one team is a different proposition from four hundred across three regions. BCG puts 70% of program effort in people and process, and McKinsey puts the failure rate of large change program near 70%.

Zone · 03

A Written Refusal List

A selection method that has never produced a refusal is not a method. The scarce resource is senior engineering attention, and it is being divided by the number of live items rather than concentrated on the ones that pay. Name the refusals as output, with a reason against each, and review the list quarterly.

By the numbers

The figures that make it a board-level conversation.

8
median months from AI prototype to production
21%
of organisations using generative AI have fundamentally redesigned even some workflows, and redesign is the practice most strongly linked to EBIT impact
70%
of programme effort sits in people and process change, against roughly 10% in algorithms and 20% in technology and data
Inside the report

What you'll take away.

01

Arrive with the process inventory complete

Process, volume per month, cycle time, error and rework rate, loaded cost and a named owner for each number. Blank cells are not discussed in the room, they are sent back. Six to eight process owners can usually fill this in a week.

02

Cut to a shortlist without technology talk

Fifteen minutes on volume, variance, manual touch and failure cost, led by the COO with no technology discussion permitted. Expect eight to twelve candidates out of an inventory of forty or more, because ranking by ease of demonstration is the failure this prevents.

03

Score six dimensions and stop at the disqualifier

Value density, tolerance for error, data at the decision point, process stability, decision latency in milliseconds, and change-management load. Do not average a disqualifier away against a strong score elsewhere, because that is exactly how unbounded-failure use cases survive a review.

04

Choose the first build for the platform

Separate the highest-value candidate from the correct first candidate, and pick the one that forces retrieval, evaluation, access control, observability and cost governance into existence. Then assign the operations owner, set baseline capture at two weeks, and re-score quarterly.

Questions

Frequently asked.

Does this method require us to stop the AI work already in flight?

No, but it will re-score it. Run the six dimensions across live initiatives and anything hitting a disqualifier goes to the refusal list with the reason recorded. In most portfolios that frees senior engineering attention rather than reducing output.

Who actually owns the scorecard, engineering or operations?

Operations owns the number, engineering owns feasibility, and both score in the same room. The shortlist stage is led by the COO with no technology talk permitted, because ranking by ease of demonstration is precisely the failure this method exists to prevent.

Why should we not simply build the highest-value candidate first?

The first build also constructs your platform: retrieval, evaluation, access control, observability and cost governance. A mid-value case that forces all five beats a high-value case that forces none, since the next two initiatives run on whatever the first one leaves behind.

Our processes are not documented, so can we still run the session?

Not usefully. The inventory is the method. Six to eight process owners can usually produce volume, cycle time, error rate and loaded cost for their own area in under a week, and that exercise alone tends to surface the strongest candidate.

What if the best answer turns out to involve no AI at all?

That is a normal output. Several disqualifiers point straight at a process fix, a data fix or a system of record change. Saying so in the first fortnight is cheaper than discovering it eight months into a build nobody can measure.

Get the whitepaper

Have it emailed to you.

Drop your details and we'll send 30% Of AI Projects Die After The Demo. Selection Kills Them. straight to your inbox - no spam, unsubscribe anytime.

Download whitepaper
Next step

Run the selection first.

Bring your process inventory to a two-hour session with our engineering leads and leave with a scored candidate list and a written refusal list.

Book a selection session