Ten builds, told with the failures left in.

The situation

The data existed. It reached the CEO through five people, in five formats, about two weeks late, and shaped on the way. A utilization figure that arrives with an explanation from the person accountable for it is not the same thing as a utilization figure.

Definitions came before extraction, and they were the hardest part: three sessions on what utilization means, when an engagement counts as booked, and what a client is when three subsidiaries buy separately.

What was built

Where it stopped

Refusing to approximate one metric bought credibility for all the others. It became the decision the CEO cited most often.

The situation

The close ran long every month. Coding queries bounced between shared services and the business units for days, and controllers wrote variance commentary that read like last month’s, because it largely was.

The constraints were written before any design work: no posting without approval, no accrual proposed without evidence, no adjustment to consolidated results, and a full log of every proposal with its reasoning. The external auditors reviewed the design before go-live, and their position was minuted.

What was built

Where it stopped

Posting to the ledger is not available, by construction rather than by policy. The assistant is wrong often enough to matter on transactions with no close precedent, and its rejections are reviewed quarterly.

The situation

Two procurement staff keyed supplier documents by hand: confirmations, shipping notices, and invoices in dozens of layouts, arriving as PDFs and email attachments. Mismatches were discovered at goods receipt, and by then the truck has arrived and the only options left are expensive.

Twelve months of documents from the top forty suppliers were sampled first, and match rules, tolerances and exception criteria were agreed line by line with the people who would live with them.

What was built

Documents now land in an extraction pipeline that turns them into structured records, with a confidence score on every field. A three-way match runs orders against confirmations and receipts inside the agreed tolerances, and anything outside them goes to an exception queue, where each item arrives with a proposed resolution and the evidence behind it.

The same pipeline watches lead times, computed from actual deliveries and not from supplier promises, so drift surfaces before a stockout does.

Where it stopped

Handwritten delivery notes from smaller suppliers were trialed and abandoned: checking the extraction cost more than keying the note. Those suppliers were routed to portal submission instead.

The situation

The forecast was not wrong. It could only hold one version of the future at a time, and testing a second one took a week. Ask what happens if lead times stretch by six weeks, and the honest answer was a week. By then the question had moved on.

Definitions took longer than the model. Pipeline meant different things in different regions, and no one had said so out loud. Five years of orders were reconciled to invoiced revenue, and project business was separated from spare parts and service, because blending them had been hiding the signal in both.

What was built

Where it stopped

Two proposed drivers were dropped because they added nothing the others didn’t already carry.

The situation

Response times slipped every peak, precisely when they mattered most, while agents spent the day pasting order numbers into templates. Three months of tickets showed the same pattern in every season: order status, returns, sizing, invoices, and delivery exceptions, all answerable from systems the company already ran.

Before any automation, the policy contradictions got cleaned up. The website said one thing and the macro library said another, so the first job was a single authoritative source both could point to.

What was built

The situation

The bottleneck was never the arithmetic. It was assembling context: the account, the contract terms, the purchase history, the applicable pricing rules. Hours per quotation, living partly in systems, partly in contracts, and substantially in the heads of three long-tenured people.

The pricing logic had lived in those heads for two decades. Writing it down was the real deliverable: volume breaks, contract terms, freight, currency, lead-time surcharges, all of it. The document outlasts the software built on top of it.

What was built

A quotation now starts wherever the request lands: an email, a portal request, or a call note. The agent pulls the account, its contract terms, and its purchase history together, prices the lines under the documented rules, and hands sales a finished document to check and send.

Cost data refreshes on a fixed cadence, and every deviation from the rules is logged with a reason.

The situation

A four-person strategy team tracked markets, competitors, regulation, and tenders by hand, and their findings arrived after the decisions they should have informed. The gap was not breadth. It was language: the Chinese-language sources were effectively invisible, and that was the expensive part.

Signal was defined before anything was collected. Named competitors, technology areas, regulators, and tender portals went on a watch list, with one rule underneath: anything entering a decision paper traces to a named primary source with a date. Aggregators detect. They are never cited.

What was built

Where it stopped

Social listening ran for six weeks. Only a handful of items met the signal criteria, and none had been missed by the other sources. It was dropped rather than tuned, because the cost was attention, and attention was the scarce thing.

The situation

Every advertising platform counted differently, and clients noticed. Monthly reporting consumed most of a week, and the analysts spent it defending numbers rather than reading them.

The foundation was a metric dictionary with a formal sign-off: which source is authoritative when platforms disagree, how taxonomies map, and how hierarchies, time zones, and currencies resolve. A dated, signed definition ends the conversation. Rebuilding the report from scratch used to be the only answer available.

What was built

Every platform now flows through one normalization layer and comes out in the dictionary’s terms, with history reprocessed so year-on-year comparisons hold. Recurring client reports build themselves with first-pass commentary, and the commercial team gets self-service answers in between.

When sources disagree, the report shows both figures with the explanation. Nobody splits the difference, and three advertisers went on to request that discrepancy view specifically.

The situation

Paid search and professional-network advertising were a side duty for two people, run across two platform interfaces with separate logins and separate taxonomies. Testing cost time nobody had, so it rarely happened.

The data job that mattered was mapping conversions back to CRM stages. Until that existed, the team could optimize cost per click and nothing more useful. Historical exports were cleaned, landing pages audited, and one naming taxonomy imposed across both platforms.

What was built

The build is one campaign workspace. The team briefs once and gets keyword sets with volume and cost estimates, plus ad variants drafted for each platform. Tests run against agreed thresholds, with budget guardrails set during a calm period, not negotiated mid-campaign, and both platforms connect through their APIs, so the ad managers themselves go untouched on a normal day.

Every reallocation arrives with its reasoning and waits for a click. Approval stayed a human act.

The situation

Each campaign concept produced one set of assets. The retailer formats, audience segments and local-market adaptations were cut, campaign after campaign, because agency retainers were priced per asset, and that turned every variant into a budget conversation instead of a marketing one.

The longest phase came before any build. Six workshops on voice and constraints. Three years of campaign archive collected, photography rights cleared. And a claims register that had gone two years without review, rewritten and signed off by legal before the system generated a single word.

What was built

Where it stopped

Video was scoped, prototyped, and stopped. On-pack accuracy could not be held to brand standard, and music and talent rights were unresolved. The kill was documented, not buried.