What week one actually looks like.

A bid-to-delivery process with sixty thousand annual requests, and no obvious place to start.

The valuations business runs an eleven-step lifecycle: lead generation, pitch and RFP, job setup, information collection, calculation, inspection, assessment, deliverable creation, review, finalization and post-close. Roughly sixty thousand requests come in a year and around fifteen thousand are won and go all the way through. A parallel stream of regular established business skips the bid entirely and joins at execution.

The instinct in a process that large is to spend three months mapping it. We started with two engineers, one in the US time zone, one in India, and a commitment that the first sixty days would produce working proofs of concept built live in the workshops, using the platform and Claude Code as the harness, rather than a paper plan with a delivery date on it.

The architecture underneath is deliberately conventional where it should be: React, Node and Postgres, with agents and a durable workflow engine handling the long-running, multi-party orchestration that a process like this actually is. The interesting part is not the stack. It is that a valuation job can sit in-flight for weeks, involve several people and an external inspection, and still need to survive a system restart without losing its place.

The client's critical design review process, a serious architecture gate, was brought in from around week five, running in parallel with the build instead of preceding it. That is a slightly uncomfortable arrangement for everyone and it is the right one: the reviewers get to review something real, and the builders get findings while changing course is still cheap.

Two days of workshops, then a shortlist nobody could argue with.

We ran a two-day discovery series across Operations, Customer Service, HR and Accounting. The teams brought their own use-case inventories, Claims, Traffic and Billing, AR and Credit, which is always a better starting point than a vendor's idea of what should be automated.

The volumes made the priorities obvious in retrospect but not in advance. Open Scan Review alone runs at fifteen to twenty thousand scan reviews a day with a twenty-person team. Claims work was spread across a claims chat need, a website self-service gap and an investigation workflow. AR had unassigned credits and refund handling nobody enjoyed.

Everything was scored on value, feasibility, data readiness, risk and reuse, and cut to three per business area. Nine use cases, with two of them, Claims Chat and Open Scan Review, carrying a hard commitment to complete pilot by day 60, and all nine committed to show measured results within 180 days.

The source systems were exactly as advertised: AS400, an imaging system, Salesforce, Billtrust, a routing system, EDI feeds, driver handhelds, D&B and Power BI. Which is why we spent real effort building a formal LTL freight ontology in OWL 2, NMFC classification, tariff and rating mechanics, terminal operations, exceptions, and mapping it onto systems that predate most of the people using them. Every solution in the portfolio now resolves against it, which is the reason the ninth use case will cost a fraction of the first.

This customer also works with a major ontology platform in other parts of the business. That was useful pressure. It meant the comparison was concrete rather than theoretical, and it made accuracy, measured, on a golden set, published, the thing we chose to be judged on.

“Accuracy was the thing we chose to be judged on, because it was the thing we could prove and everybody else was asserting.”

Sixty plants, and sixty slightly different definitions of a defect.

This relationship started with supplier intelligence and grew into something closer to an operating system: Copilot agents deployed across sixty plants, neural network quality inspection on packaging lines, an HR skills intelligence agent, and eventually the center of excellence that governs all of it.

The hardest problem was never the vision model. It was that a defect class, an asset and a downtime event meant subtly different things at different sites, each difference defensible and locally correct. Rolling out one agent to sixty plants without an ontology layer produces sixty incompatible answers and a credibility problem you cannot recover from.

So the work went into Layer 1 canon for packaging and converting, Layer 2 profiling of each site's actual data, and Layer 3 rules authored by the plants themselves for their genuine local variations. The rule was simple: local differences are allowed, undocumented local differences are not.

Two sites ran as reference programs and became the template, including a formal effectiveness model for the COE itself, so the organization could tell whether the center of excellence was actually working rather than merely existing.

Six years, and the part that mattered was never the model.

The HR service platform we built with this client serves over seventy thousand employees and has been in continuous production for years. It contributed to a CIO 100 award and two filed patents. It has survived at least three generations of underlying model technology.

It survived them because the valuable asset was never the model. It was the accumulated understanding of how policy actually resolves across jurisdictions, entities, job families and collective agreements, encoded, versioned, and owned by the client. Swapping the reasoning engine underneath became a configuration change rather than a rebuild.

That is the argument for putting meaning in a layer of its own, made by six years of evidence rather than by an architecture diagram. It is also why the same relationship has extended into order-to-cash orchestration, legal contract work, financial crime intelligence and the valuations lifecycle: each new domain inherits the discipline, and increasingly the ontology, of the last.

Seven brands, one ontology, and a weekly meeting that finally had an answer in it.

ConsumerEdge went live across seven brands in October 2024 and has run continuously since. The measurable change was not in the dashboards. It was that the weekly category review stopped opening with forty minutes of arguing about whose numbers were right.

Getting there required resolving brand, sub-brand, SKU, pack format, claim, occasion and channel into one ontology that every source mapped onto, because a twelve-pack in club appears under about forty different names across retailer feeds, review sites, service transcripts and internal systems, and no amount of model quality fixes that.

The part the team values most is unglamorous: when a complaint cluster appears, it can be traced back to a plant, a shift and a lot, because the consumer ontology and the operational data resolve against the same entities. That turns a brand problem into a quality work order instead of a theory.