A European travel and tourism group was assembling a new central digital function out of teams that had never reported to each other. A first phase was due to finish the structural blueprint by the end of August: boxes, reporting lines, a name. What the blueprint could not answer was how the function would actually run, and in particular how it would create the delivery hubs meant to carry the work, including the hubs that would take over work from an outside partner. This proposal set out that missing half, scoped for delivery across the three months that followed, from the start of September to the end of November.
The challenge
A structure chart tells you who reports to whom. It does not tell you how a piece of work becomes a standing team, who approves that, against what criteria, or what has to be true before an external partner hands its work over.
Around 30 processes were identified as needing description before the new function could be run rather than drawn. Some were existing processes being rewritten for a new structure. Others existed nowhere in the group: customer experience, customer service management, ideation, hub management, resource and practice management, and service-management service delivery. The second kind is the harder kind, because there is no incumbent to interview and no as-is state to document.
Two organisational transitions ran underneath all of it. The predecessor digital portfolio had to move into the new function, and a federated, source-market IT estate had to consolidate into a single global IT model. Both ran alongside a group-wide reorganisation, so every overlapping area and touch point had to be reconciled rather than ignored.
There was also a data dependency, named up front rather than discovered later. The global IT model design was supposed to be informed by a fact-based cost and resources review that was not finished, and might not finish in time to be useful.
The approach
The hub creation framework was specified as five components rather than a policy statement: set-up criteria, governance in the form of an approval process, the organisational set-up processes themselves, the transition from work-stream to hub, and the transition from partner to in-house. Each one answers a question that otherwise gets asked in the wrong order. Should this hub exist. Who signs. What has to be stood up before it opens. When does a temporary work-stream become permanent. And what moves in-house, on what trigger.
The framework sat inside a three-stream programme: design of the global IT model, implementation of the organisational structure, and a customer-lifecycle-value model that would rank the idea portfolio against the customer journey using both soft and hard benefits. Supporting the hub creation framework was written as a line item under the first of those, not as a separate initiative.
The data dependency got a dated fallback instead of an escalation path. If the cost and resources review was not complete by mid-September, the organisational structure would be designed from HR headcount and job title data instead. The degraded input was written into the proposal in advance, with a date attached.
Process design was sequenced hardest first: financial processes, capacity management, setting up a hub, gating, and product end-to-end management, each with the named process owner in the room for the design of their own process. Depth was rationed deliberately. Level 2 for 10 main processes, then level 2 across the remainder, then level 3 for only 5.
- 30
- Processes scoped for description
- 10
- Planned for level 2 first
- 5
- Planned for level 3
The delivery plan was cut into three lanes across September, October and November: organisation design, operating model, governance. Its three recurring steps were workshop, refine and validate, written on the plan as work in their own right rather than assumed. Recruitment milestones sat inside the same three months as the design, so the model was staffed while it was still being drawn. And the plan did not stop at "operating model designed". It carried a milestone reading "operating model in use, people trained".
The commercial shape matched the delivery shape. Three people: 5 days at £2,550 per day at the most senior level, 39 days at £1,973 for design and workshop facilitation, and 70 days at £980 for production of the detailed process flows, totalling £158,297 in professional fees. The two known risks were priced separately as a dated option rather than left as an open change request: 25 days at £1,311 to take 5 key processes into detail, and 20 days at £1,030 to analyse HR data, £53,375 in total, to be confirmed or dropped by 25 September.
The outcome
This was a proposal and a design specification, not a results report. Nothing below was delivered at the time of writing; it is what the document put its name to. The outputs were listed per dimension of the operating model: rules of engagement with distribution, marketing and the source markets; an extended functional model covering both the source-market estate and the target single IT model; level 2 process flows with roles for 30 processes plus the training material to walk people through them; report mapping at levels 1 and 2 with role descriptions, gaps and a mapping of individuals to the new design; a described gating cycle with customer, product and executive forums; and a strategic dashboard. The value now is in the shape of that specification, and the shape has aged well.
Read it now, more than a decade on, and the hub creation framework is a specification for something most enterprises are currently failing to build for AI. Substitute "agent" for "hub" and the five components hold. Set-up criteria: which work justifies a standing agent at all, rather than a prompt somebody runs by hand. Approval governance: who signs before it touches a customer. Organisational set-up: who owns it, who reviews its output, who is on call when it is wrong. Work-stream to hub: the step where a pilot becomes a supported capability with a budget line and a rota, which is exactly where most agent pilots quietly stop. Partner to in-house: which parts stay bought from a vendor and which get built and run internally, decided on a trigger rather than on enthusiasm.
The level-3 rationing transfers just as directly. At the time depth cost money, so only five processes got it. An agent cannot execute a process nobody has described at that depth, because the ambiguity a human absorbs silently becomes a wrong action or an invented step. What has changed since is the cost of the first draft: a model can now read tickets, call transcripts, logs and existing documentation and propose the level 3 description in hours. That is loop and harness engineering rather than authorship. The loop drafts a description, runs it against real cases, and marks the steps where it breaks. The harness around the loop holds the fixtures, the evaluation set, the permission boundary and the audit trail that make an unattended run worth trusting. What has not changed is validation. The workshop, refine and validate loop with a named process owner is still the binding constraint, and it is the part that cannot be handed to the machine, because the process owner is the person who has to live with the answer.
The three months in this plan were three months of human throughput: workshop, refine, validate, three lanes, hardest first. Loop and harness engineering compresses the drafting side of that, not the deciding side, and the compression is largest exactly where this group's work is thickest, which is documents. Rules of engagement with distribution, marketing and the source markets are contracts. A federated source-market estate consolidating into one global IT model is years of supplier paperwork, service schedules and system inventories held in different languages and formats. And the outputs the plan promised are drawings before they are prose: level 2 process flows with roles, report mapping at levels 1 and 2, a strategic dashboard, each carrying its meaning in the layout rather than in a paragraph. A model that reads a page as drawn rather than as a stripped transcript of it turns the tier costed here as production of the detailed flows into a pass a person checks, and the same reading feeds the process descriptions instead of sitting in a folder next to them.
The gating cycle is where agents stop being assistants and start carrying a share of how the group runs. The plan described customer, product and executive forums and a strategic dashboard, each of them fed by hand. Wire them as one connected pipeline instead. A hub candidate is scored against the set-up criteria by an agent reading the same operational data the dashboard reports. The approval gate becomes a real gate in that pipeline, so a forum opens an assembled case with its evidence attached rather than a pack somebody built the week before. The organisational set-up steps fire as tasks with owners on them. The partner-to-in-house trigger stops being a conversation somebody remembers to start and becomes a threshold the pipeline watches, with the handover checklist drafted the moment it is crossed. Autonomous here means the run proceeds without a person in every step, not without a person answerable for the outcome. The executive forum still signs.
The net-new problem repeats too. Back then the processes with no as-is were customer experience, ideation and hub management. Today they are agent supervision, evaluation and rollback, model change management, and AI incident response. None has an incumbent to interview, so each needs a workshop rather than an interview, and each should be budgeted as invention rather than documentation.
The dated fallback is the last piece worth copying. Every AI programme carries the same kind of dependency this plan carried: the clean dataset, the labelled examples, the consolidated system of record. Agree now, in the contract, what you will build on if that input is not ready by a named date. It turns a predictable slip into a decision somebody already made.
And hold the programme to the same finish line this plan set for itself. Not deployed. In use, by people who were trained to use it, with an owner whose name is on it. The difference today is how long the run between the criteria and that finish line has to take, and the answer is much shorter than three months of workshops if the loop and the harness are built first.