The Install
Most AI implementations in B2B fail — despite everyone using it every day. We model the workflows, pick the best-in-class tool for each, install them with a KPI agreed up front, and bring the client's own team up to run it.
Access to AI is now universal, and worth almost nothing on its own.
Every B2B firm has the licences, a handful of enthusiasts, and a management team quietly wondering why none of it shows up in the numbers. The gap is not the model. It is that nobody mapped the work as it is actually done, rebuilt it, trained the people who run it, and put a number on the outcome.
AI does not create the blockers. It exposes the ones already there.
The failure is documented, and it is not a model problem
Why it keeps failing in the same four ways
- The tool arrived before the map. Licences get bought against a category, not against a named workflow with a named owner and a measurable cycle time. Nobody can say afterwards what was supposed to improve.
- The process was never rebuilt. AI gets bolted onto a workflow designed for people doing it manually, so it saves minutes inside a process that wastes days. Marginal efficiency inside an outdated shape.
- Enthusiasts, not systems. Two or three people get very good at it privately. Nothing is written down, nothing is standardised, and when they leave or get busy the gain evaporates.
- No number was agreed at the start. Without a baseline nobody can prove it worked, so the budget conversation next year is a matter of belief. Which is how a working programme gets cut.
The method — four steps, done by people, inside the business
The workflows as they actually run
Not as the process document describes them. We sit with the people doing the work, map each repeatable workflow end to end, and establish the baseline: how long it takes now, how many times a month, who touches it, where it stalls.
The map itself usually earns the fee. Most firms have never seen their own marketing operation drawn out with times against it.
Best-in-class tool per workflow
Chosen against the mapped job, not against a vendor's category page — and benchmarked, including the option of what the client already pays for. Some workflows want a model, some want automation, some want a template and a rule, and some should just be stopped.
We are not reselling anything. The client owns every licence.
Rebuild it, with AI inside
Engineering, not enablement slides. The workflow is reshaped around what the tool is actually good at, run in parallel against the old one until it holds, then switched over with the exceptions and failure modes documented.
This is the step everyone skips, and the step every piece of the research above points at.
Their team runs it
Capability transfer is the deliverable. The people who do the work are trained on the rebuilt version, given the standards and the prompts, and left with something they own rather than something they rent from us.
Then the same pattern is repeated by them on the next workflow. That is the acceleration.
The KPI is agreed before we start
Baseline first, or there is no engagement. Hours returned, cycle time, output per head, cost per asset, pipeline contribution — whichever fits the workflow, named and signed off in week one and reported against at the end.
It is the one discipline that separates the 5% from the 95%, and the reason the budget survives.
Start in marketing
High volume, high repetition, measurable output, and the domain we already own — so the first proof lands fast and is credible. Content production, campaign build, reporting, proposal and pitch production, research, review and approval cycles.
Prove the number there. Then accelerate outward into sales operations, bid, and service.
Why this fits the model
It sells no software. There is no client login, no platform for us to maintain, nothing that asks us to become an engineering company. It is delivered by embedded specialists inside the client's business, and what the client is left with is their own team working better — which is the same sentence we use about our own AI, pointed at their people instead of ours.
It is also the natural land-and-expand engine. A marketing beachhead with a proven number is the cheapest possible route into the rest of the business.
What has to be true before we sell it
- Named engineers and evidenced hours. "AI engineers with thousands of hours" is the claim; it needs the names and the numbers behind it before it appears in front of a client. Selling on an unbacked capability claim is the fastest way to lose the room.
- Two reference installs. Ideally one of them run on ourselves — our own proposal and reporting workflows are the obvious first patient, and the internal delivery work is already pointed there.
- A baseline discipline we will actually hold. If we cannot get the client to agree a measurement in week one, the engagement will end the same way the 95% did.
Where it sits
- Who buys it: the MD or commercial director who has spent money on AI and cannot answer what it returned. Also the marketing director being asked that question by their board.
- Sequence: follows the Strategy & Planning Model naturally — the plan says what the business needs to produce; The Install rebuilds the capacity to produce it.
- Sells with Ground Truth: Ground Truth tells you what content to make; The Install rebuilds the machine that makes it.
- How it is judged: the agreed KPI, reported against the baseline — and whether the client's team runs the rebuilt workflow six months later without us.